From e933ce06a465f00aae12e4144cf75abc5205dab9 Mon Sep 17 00:00:00 2001 From: Bruno Fargnoli Date: Sun, 28 Dec 2025 23:29:57 +0100 Subject: [PATCH] Added autotune_cache file for Triton + Fixed "Mesh Texturing" + "PreProcess Image" node --- autotune_cache.json | 48757 +++++++++++++++++++ nodes.py | 83 +- trellis2/pipelines/trellis2_image_to_3d.py | 12 +- 3 files changed, 48831 insertions(+), 21 deletions(-) create mode 100644 autotune_cache.json diff --git a/autotune_cache.json b/autotune_cache.json new file mode 100644 index 0000000..a520407 --- /dev/null +++ b/autotune_cache.json @@ -0,0 +1,48757 @@ +{ + "*": { + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm.sparse_submanifold_conv_bwd_input_implicit_gemm_kernel": { + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm.sparse_submanifold_conv_bwd_weight_implicit_gemm_kernel": { + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_input_implicit_gemm_kernel": { + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_weight_implicit_gemm_kernel": { + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_input_implicit_gemm_splitk_kernel": { + "(7, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_weight_implicit_gemm_splitk_kernel": { + "(7, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_input_implicit_gemm_splitk": { + "(2^7, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^9, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^11, 1024, 1024, 27)": { + "SPLITK": 4 + }, + "(2^13, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 256, 27)": { + "SPLITK": 1 + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_weight_implicit_gemm_splitk": { + "(2^7, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^9, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^11, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^13, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 256, 27)": { + "SPLITK": 8 + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm.sparse_submanifold_conv_bwd_input_masked_implicit_gemm_kernel": { + "(18, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm.sparse_submanifold_conv_bwd_weight_masked_implicit_gemm_kernel": { + "(16, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_bwd_input_masked_implicit_gemm_kernel": { + "(18, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_bwd_weight_masked_implicit_gemm_kernel": { + "(16, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_bwd_input_masked_implicit_gemm_splitk_kernel": { + "(16, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_bwd_weight_masked_implicit_gemm_splitk_kernel": { + "(18, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_bwd_input_masked_implicit_gemm_splitk": { + "(2^18, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^16, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^16, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^14, 256, 1024, 27)": { + "SPLITK": 2 + }, + "(2^14, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^12, 512, 2048, 27)": { + "SPLITK": 4 + }, + "(2^12, 512, 512, 27)": { + "SPLITK": 4 + }, + "(2^10, 1024, 4096, 27)": { + "SPLITK": 8 + }, + "(2^10, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^12, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^14, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^16, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^18, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^11, 512, 2048, 27)": { + "SPLITK": 8 + }, + "(2^11, 512, 512, 27)": { + "SPLITK": 4 + }, + "(2^9, 1024, 4096, 27)": { + "SPLITK": 16 + }, + "(2^9, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^11, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^13, 256, 1024, 27)": { + "SPLITK": 4 + }, + "(2^13, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^13, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^17, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^15, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^15, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^15, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^17, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^8, 1024, 4096, 27)": { + "SPLITK": 16 + }, + "(2^8, 1024, 1024, 27)": { + "SPLITK": 16 + }, + "(2^19, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^17, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^17, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^15, 256, 1024, 27)": { + "SPLITK": 2 + }, + "(2^15, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^13, 512, 2048, 27)": { + "SPLITK": 4 + }, + "(2^13, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^13, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^15, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^17, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^19, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^11, 1024, 4096, 27)": { + "SPLITK": 16 + }, + "(2^11, 1024, 1024, 27)": { + "SPLITK": 4 + }, + "(2^10, 512, 2048, 27)": { + "SPLITK": 16 + }, + "(2^10, 512, 512, 27)": { + "SPLITK": 4 + }, + "(2^10, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^20, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^20, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^18, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^18, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^18, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^16, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^16, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^16, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^12, 256, 1024, 27)": { + "SPLITK": 8 + }, + "(2^12, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^12, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^21, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^21, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^19, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^19, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^14, 512, 2048, 27)": { + "SPLITK": 4 + }, + "(2^14, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^14, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^19, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^7, 1024, 4096, 27)": { + "SPLITK": 32 + }, + "(2^7, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^14, 128, 512, 27)": { + "SPLITK": 4 + }, + "(2^14, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^14, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^16, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^16, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^22, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^22, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^13, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 512, 27)": { + "SPLITK": 1 + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_bwd_weight_masked_implicit_gemm_splitk": { + "(2^18, 64, 64, 27)": { + "SPLITK": 128 + }, + "(2^16, 128, 512, 27)": { + "SPLITK": 16 + }, + "(2^16, 128, 128, 27)": { + "SPLITK": 32 + }, + "(2^14, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^14, 256, 256, 27)": { + "SPLITK": 4 + }, + "(2^12, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^12, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^10, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^10, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^12, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^14, 256, 64, 27)": { + "SPLITK": 16 + }, + "(2^16, 128, 32, 27)": { + "SPLITK": 64 + }, + "(2^18, 64, 16, 27)": { + "SPLITK": 128 + }, + "(2^11, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^11, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^9, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^9, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^11, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^13, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^13, 256, 256, 27)": { + "SPLITK": 4 + }, + "(2^13, 256, 64, 27)": { + "SPLITK": 16 + }, + "(2^17, 64, 64, 27)": { + "SPLITK": 64 + }, + "(2^15, 128, 512, 27)": { + "SPLITK": 4 + }, + "(2^15, 128, 128, 27)": { + "SPLITK": 16 + }, + "(2^15, 128, 32, 27)": { + "SPLITK": 32 + }, + "(2^17, 64, 16, 27)": { + "SPLITK": 128 + }, + "(2^8, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^8, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^19, 64, 64, 27)": { + "SPLITK": 128 + }, + "(2^17, 128, 512, 27)": { + "SPLITK": 16 + }, + "(2^17, 128, 128, 27)": { + "SPLITK": 32 + }, + "(2^15, 256, 1024, 27)": { + "SPLITK": 2 + }, + "(2^15, 256, 256, 27)": { + "SPLITK": 8 + }, + "(2^13, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^13, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^13, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^15, 256, 64, 27)": { + "SPLITK": 8 + }, + "(2^17, 128, 32, 27)": { + "SPLITK": 64 + }, + "(2^19, 64, 16, 27)": { + "SPLITK": 128 + }, + "(2^11, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^11, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^10, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^10, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^10, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^20, 64, 64, 27)": { + "SPLITK": 128 + }, + "(2^20, 64, 16, 27)": { + "SPLITK": 128 + }, + "(2^18, 128, 512, 27)": { + "SPLITK": 32 + }, + "(2^18, 128, 128, 27)": { + "SPLITK": 64 + }, + "(2^18, 128, 32, 27)": { + "SPLITK": 64 + }, + "(2^16, 256, 1024, 27)": { + "SPLITK": 4 + }, + "(2^16, 256, 256, 27)": { + "SPLITK": 4 + }, + "(2^16, 256, 64, 27)": { + "SPLITK": 8 + }, + "(2^12, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^12, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^12, 256, 64, 27)": { + "SPLITK": 8 + }, + "(2^21, 64, 64, 27)": { + "SPLITK": 128 + }, + "(2^21, 64, 16, 27)": { + "SPLITK": 128 + }, + "(2^19, 128, 512, 27)": { + "SPLITK": 32 + }, + "(2^19, 128, 128, 27)": { + "SPLITK": 128 + }, + "(2^17, 256, 1024, 27)": { + "SPLITK": 8 + }, + "(2^17, 256, 256, 27)": { + "SPLITK": 8 + }, + "(2^14, 512, 2048, 27)": { + "SPLITK": 2 + }, + "(2^14, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^14, 512, 128, 27)": { + "SPLITK": 4 + }, + "(2^17, 256, 64, 27)": { + "SPLITK": 32 + }, + "(2^19, 128, 32, 27)": { + "SPLITK": 128 + }, + "(2^7, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^7, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^14, 128, 512, 27)": { + "SPLITK": 4 + }, + "(2^14, 128, 128, 27)": { + "SPLITK": 16 + }, + "(2^14, 128, 32, 27)": { + "SPLITK": 32 + }, + "(2^16, 64, 64, 27)": { + "SPLITK": 32 + }, + "(2^16, 64, 16, 27)": { + "SPLITK": 64 + }, + "(2^22, 64, 64, 27)": { + "SPLITK": 8 + }, + "(2^22, 64, 16, 27)": { + "SPLITK": 128 + }, + "(2^13, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 512, 27)": { + "SPLITK": 1 + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_implicit_gemm.sparse_submanifold_conv_fwd_implicit_gemm_kernel": { + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_implicit_gemm_splitk.sparse_submanifold_conv_fwd_implicit_gemm_kernel": { + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_implicit_gemm_splitk.sparse_submanifold_conv_fwd_implicit_gemm_splitk_kernel": { + "(8, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 16, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_implicit_gemm_splitk.sparse_submanifold_conv_fwd_implicit_gemm_splitk": { + "(2^8, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^10, 1024, 1024, 27)": { + "SPLITK": 2 + }, + "(2^12, 1024, 1024, 27)": { + "SPLITK": 2 + }, + "(2^14, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^16, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^18, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^20, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^7, 1024, 1024, 27)": { + "SPLITK": 16 + }, + "(2^9, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^11, 1024, 1024, 27)": { + "SPLITK": 4 + }, + "(2^13, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^19, 128, 128, 27)": { + "SPLITK": 1 + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_masked_implicit_gemm.sparse_submanifold_conv_fwd_masked_implicit_gemm_kernel": { + "(18, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_fwd_masked_implicit_gemm_kernel": { + "(18, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_fwd_masked_implicit_gemm_splitk_kernel": { + "(16, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 256, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 16, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 32, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 64, 27, 256, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 512, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 512, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 512, 27, 256, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 128, 27, 256, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 128, 27, 512, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 128, 27, 1024, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 16, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 512, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 512, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 128, 27, 256, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 128, 27, 512, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 16, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 256, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 256, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 64, 27, 256, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(6, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(6, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(6, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(6, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(6, 1024, 1024, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(6, 1024, 1024, 27, 256, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 4, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_fwd_masked_implicit_gemm_splitk": { + "(2^18, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^18, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^16, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^16, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^14, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^14, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^12, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^12, 512, 128, 27)": { + "SPLITK": 16 + }, + "(2^10, 1024, 1024, 27)": { + "SPLITK": 4 + }, + "(2^10, 1024, 4096, 27)": { + "SPLITK": 4 + }, + "(2^12, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^14, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^16, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^19, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^19, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^17, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^17, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^15, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^15, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^13, 512, 512, 27)": { + "SPLITK": 2 + }, + "(2^13, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^11, 1024, 1024, 27)": { + "SPLITK": 4 + }, + "(2^11, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^13, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^15, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^17, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^11, 512, 512, 27)": { + "SPLITK": 4 + }, + "(2^11, 512, 128, 27)": { + "SPLITK": 16 + }, + "(2^9, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^9, 1024, 4096, 27)": { + "SPLITK": 8 + }, + "(2^11, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^13, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^13, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^13, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^17, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^17, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^15, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^15, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^15, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^8, 1024, 1024, 27)": { + "SPLITK": 16 + }, + "(2^8, 1024, 4096, 27)": { + "SPLITK": 4 + }, + "(2^10, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^10, 512, 128, 27)": { + "SPLITK": 32 + }, + "(2^10, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^20, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^20, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^18, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^18, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^18, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^16, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^16, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^16, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^12, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^12, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^12, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^21, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^21, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^19, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^19, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^14, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^14, 512, 128, 27)": { + "SPLITK": 4 + }, + "(2^14, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^19, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^7, 1024, 1024, 27)": { + "SPLITK": 4 + }, + "(2^7, 1024, 4096, 27)": { + "SPLITK": 16 + }, + "(2^14, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^14, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^14, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^16, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^16, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^22, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^22, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^12, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^10, 256, 256, 27)": { + "SPLITK": 16 + }, + "(2^12, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^13, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^13, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^20, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^20, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^18, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^18, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^18, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^20, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^16, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^16, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^14, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^14, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^16, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^23, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^23, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^21, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^21, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^19, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^19, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^17, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^17, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^15, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^15, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^17, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^19, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^21, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^11, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^11, 256, 64, 27)": { + "SPLITK": 32 + }, + "(2^14, 64, 64, 27)": { + "SPLITK": 2 + }, + "(2^14, 64, 16, 27)": { + "SPLITK": 4 + }, + "(2^12, 128, 128, 27)": { + "SPLITK": 4 + }, + "(2^12, 128, 32, 27)": { + "SPLITK": 32 + }, + "(2^10, 256, 64, 27)": { + "SPLITK": 128 + }, + "(2^8, 512, 512, 27)": { + "SPLITK": 64 + }, + "(2^8, 512, 128, 27)": { + "SPLITK": 256 + }, + "(2^15, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^15, 64, 16, 27)": { + "SPLITK": 4 + }, + "(2^13, 128, 128, 27)": { + "SPLITK": 2 + }, + "(2^13, 128, 32, 27)": { + "SPLITK": 32 + }, + "(2^9, 512, 512, 27)": { + "SPLITK": 8 + }, + "(2^9, 512, 128, 27)": { + "SPLITK": 256 + }, + "(2^13, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^13, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^9, 256, 256, 27)": { + "SPLITK": 8 + }, + "(2^9, 256, 64, 27)": { + "SPLITK": 32 + }, + "(2^6, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^18, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^18, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^20, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^20, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^22, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^22, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^24, 64, 64, 27)": { + "SPLITK": 1 + } + }, + "flex_gemm.triton.gemm.gemm_nn.gemm_nn_kernel": {}, + "flex_gemm.triton.gemm.gemm_nn_splitk.gemm_nn_kernel": {}, + "flex_gemm.triton.gemm.gemm_nn_splitk.gemm_nn_splitk_kernel": {}, + "flex_gemm.triton.gemm.gemm_nn_splitk.gemm_nn_splitk": {}, + "flex_gemm.triton.gemm.gemm_nn_splitk_atomic.gemm_nn_splitk_atomic_kernel": {}, + "flex_gemm.triton.gemm.gemm_nn_splitk_lock.gemm_nn_splitk_lock_kernel": {}, + "flex_gemm.kernels.triton.grid_sample.indice_weighed_sum_bwd.indice_weighed_sum_bwd_input_kernel": {}, + "flex_gemm.kernels.triton.grid_sample.indice_weighed_sum_fwd.indice_weighed_sum_fwd_kernel": { + "(23, 4194304, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 4194304, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 32, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 1941851, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 7877533, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 2226123, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 12748156, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 1911611, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 5017088, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 4194304, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 1112821, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 33122502, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 1284070, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 10949861, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 1571524, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 12857151, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 16 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 1641261, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 15495770, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 1808517, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 9814710, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 32, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 1780790, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 14310670, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 32, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 2113485, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 21145957, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 1782768, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 14828658, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 1577958, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 17012742, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 32, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 2345014, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 12887107, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 2168973, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 12754203, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 1873811, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 12506390, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 2223717, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 14387524, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + } + }, + "NVIDIA GeForce RTX 3080 Laptop GPU": { + "flex_gemm.kernels.triton.grid_sample.indice_weighed_sum_bwd.indice_weighed_sum_bwd_input_kernel": {}, + "flex_gemm.kernels.triton.grid_sample.indice_weighed_sum_fwd.indice_weighed_sum_fwd_kernel": { + "(23, 4194304, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 4194304, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 32, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 1941851, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 7877533, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 2226123, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 12748156, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 1911611, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 5017088, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 4194304, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 1112821, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 33122502, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 1284070, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 10949861, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 1571524, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 12857151, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 16 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 1641261, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 15495770, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 1808517, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 9814710, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 32, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 1780790, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 14310670, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 32, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 2113485, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 21145957, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 1782768, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 14828658, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 1577958, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 17012742, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 32, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 2345014, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 12887107, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 2168973, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 12754203, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 1873811, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 12506390, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 2223717, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 8, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 14387524, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 1957618, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 2256858, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 2241291, 6, 8, 'torch.float32', 'torch.uint32', 'torch.float32', 'torch.float32')": { + "kwargs": { + "BM": 16, + "BK": 8 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 2, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm.sparse_submanifold_conv_bwd_input_implicit_gemm_kernel": { + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm.sparse_submanifold_conv_bwd_weight_implicit_gemm_kernel": { + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_input_implicit_gemm_kernel": { + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_weight_implicit_gemm_kernel": { + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_input_implicit_gemm_splitk_kernel": { + "(7, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_weight_implicit_gemm_splitk_kernel": { + "(7, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_input_implicit_gemm_splitk": { + "(2^7, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^9, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^11, 1024, 1024, 27)": { + "SPLITK": 4 + }, + "(2^13, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 256, 27)": { + "SPLITK": 1 + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_weight_implicit_gemm_splitk": { + "(2^7, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^9, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^11, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^13, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 256, 27)": { + "SPLITK": 8 + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm.sparse_submanifold_conv_bwd_input_masked_implicit_gemm_kernel": { + "(18, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm.sparse_submanifold_conv_bwd_weight_masked_implicit_gemm_kernel": { + "(16, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_bwd_input_masked_implicit_gemm_kernel": { + "(18, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_bwd_weight_masked_implicit_gemm_kernel": { + "(16, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_bwd_input_masked_implicit_gemm_splitk_kernel": { + "(16, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_bwd_weight_masked_implicit_gemm_splitk_kernel": { + "(18, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.uint32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_bwd_input_masked_implicit_gemm_splitk": { + "(2^18, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^16, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^16, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^14, 256, 1024, 27)": { + "SPLITK": 2 + }, + "(2^14, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^12, 512, 2048, 27)": { + "SPLITK": 4 + }, + "(2^12, 512, 512, 27)": { + "SPLITK": 4 + }, + "(2^10, 1024, 4096, 27)": { + "SPLITK": 8 + }, + "(2^10, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^12, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^14, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^16, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^18, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^11, 512, 2048, 27)": { + "SPLITK": 8 + }, + "(2^11, 512, 512, 27)": { + "SPLITK": 4 + }, + "(2^9, 1024, 4096, 27)": { + "SPLITK": 16 + }, + "(2^9, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^11, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^13, 256, 1024, 27)": { + "SPLITK": 4 + }, + "(2^13, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^13, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^17, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^15, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^15, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^15, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^17, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^8, 1024, 4096, 27)": { + "SPLITK": 16 + }, + "(2^8, 1024, 1024, 27)": { + "SPLITK": 16 + }, + "(2^19, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^17, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^17, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^15, 256, 1024, 27)": { + "SPLITK": 2 + }, + "(2^15, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^13, 512, 2048, 27)": { + "SPLITK": 4 + }, + "(2^13, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^13, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^15, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^17, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^19, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^11, 1024, 4096, 27)": { + "SPLITK": 16 + }, + "(2^11, 1024, 1024, 27)": { + "SPLITK": 4 + }, + "(2^10, 512, 2048, 27)": { + "SPLITK": 16 + }, + "(2^10, 512, 512, 27)": { + "SPLITK": 4 + }, + "(2^10, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^20, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^20, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^18, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^18, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^18, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^16, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^16, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^16, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^12, 256, 1024, 27)": { + "SPLITK": 8 + }, + "(2^12, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^12, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^21, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^21, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^19, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^19, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^14, 512, 2048, 27)": { + "SPLITK": 4 + }, + "(2^14, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^14, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^19, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^7, 1024, 4096, 27)": { + "SPLITK": 32 + }, + "(2^7, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^14, 128, 512, 27)": { + "SPLITK": 4 + }, + "(2^14, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^14, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^16, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^16, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^22, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^22, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^13, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 512, 27)": { + "SPLITK": 1 + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_bwd_weight_masked_implicit_gemm_splitk": { + "(2^18, 64, 64, 27)": { + "SPLITK": 128 + }, + "(2^16, 128, 512, 27)": { + "SPLITK": 16 + }, + "(2^16, 128, 128, 27)": { + "SPLITK": 32 + }, + "(2^14, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^14, 256, 256, 27)": { + "SPLITK": 4 + }, + "(2^12, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^12, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^10, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^10, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^12, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^14, 256, 64, 27)": { + "SPLITK": 16 + }, + "(2^16, 128, 32, 27)": { + "SPLITK": 64 + }, + "(2^18, 64, 16, 27)": { + "SPLITK": 128 + }, + "(2^11, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^11, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^9, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^9, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^11, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^13, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^13, 256, 256, 27)": { + "SPLITK": 4 + }, + "(2^13, 256, 64, 27)": { + "SPLITK": 16 + }, + "(2^17, 64, 64, 27)": { + "SPLITK": 64 + }, + "(2^15, 128, 512, 27)": { + "SPLITK": 4 + }, + "(2^15, 128, 128, 27)": { + "SPLITK": 16 + }, + "(2^15, 128, 32, 27)": { + "SPLITK": 32 + }, + "(2^17, 64, 16, 27)": { + "SPLITK": 128 + }, + "(2^8, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^8, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^19, 64, 64, 27)": { + "SPLITK": 128 + }, + "(2^17, 128, 512, 27)": { + "SPLITK": 16 + }, + "(2^17, 128, 128, 27)": { + "SPLITK": 32 + }, + "(2^15, 256, 1024, 27)": { + "SPLITK": 2 + }, + "(2^15, 256, 256, 27)": { + "SPLITK": 8 + }, + "(2^13, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^13, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^13, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^15, 256, 64, 27)": { + "SPLITK": 8 + }, + "(2^17, 128, 32, 27)": { + "SPLITK": 64 + }, + "(2^19, 64, 16, 27)": { + "SPLITK": 128 + }, + "(2^11, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^11, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^10, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^10, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^10, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^20, 64, 64, 27)": { + "SPLITK": 128 + }, + "(2^20, 64, 16, 27)": { + "SPLITK": 128 + }, + "(2^18, 128, 512, 27)": { + "SPLITK": 32 + }, + "(2^18, 128, 128, 27)": { + "SPLITK": 64 + }, + "(2^18, 128, 32, 27)": { + "SPLITK": 64 + }, + "(2^16, 256, 1024, 27)": { + "SPLITK": 4 + }, + "(2^16, 256, 256, 27)": { + "SPLITK": 4 + }, + "(2^16, 256, 64, 27)": { + "SPLITK": 8 + }, + "(2^12, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^12, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^12, 256, 64, 27)": { + "SPLITK": 8 + }, + "(2^21, 64, 64, 27)": { + "SPLITK": 128 + }, + "(2^21, 64, 16, 27)": { + "SPLITK": 128 + }, + "(2^19, 128, 512, 27)": { + "SPLITK": 32 + }, + "(2^19, 128, 128, 27)": { + "SPLITK": 128 + }, + "(2^17, 256, 1024, 27)": { + "SPLITK": 8 + }, + "(2^17, 256, 256, 27)": { + "SPLITK": 8 + }, + "(2^14, 512, 2048, 27)": { + "SPLITK": 2 + }, + "(2^14, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^14, 512, 128, 27)": { + "SPLITK": 4 + }, + "(2^17, 256, 64, 27)": { + "SPLITK": 32 + }, + "(2^19, 128, 32, 27)": { + "SPLITK": 128 + }, + "(2^7, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^7, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^14, 128, 512, 27)": { + "SPLITK": 4 + }, + "(2^14, 128, 128, 27)": { + "SPLITK": 16 + }, + "(2^14, 128, 32, 27)": { + "SPLITK": 32 + }, + "(2^16, 64, 64, 27)": { + "SPLITK": 32 + }, + "(2^16, 64, 16, 27)": { + "SPLITK": 64 + }, + "(2^22, 64, 64, 27)": { + "SPLITK": 8 + }, + "(2^22, 64, 16, 27)": { + "SPLITK": 128 + }, + "(2^13, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 512, 27)": { + "SPLITK": 1 + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_implicit_gemm.sparse_submanifold_conv_fwd_implicit_gemm_kernel": { + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_implicit_gemm_splitk.sparse_submanifold_conv_fwd_implicit_gemm_kernel": { + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 128, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_implicit_gemm_splitk.sparse_submanifold_conv_fwd_implicit_gemm_splitk_kernel": { + "(8, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 16, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_implicit_gemm_splitk.sparse_submanifold_conv_fwd_implicit_gemm_splitk": { + "(2^8, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^10, 1024, 1024, 27)": { + "SPLITK": 2 + }, + "(2^12, 1024, 1024, 27)": { + "SPLITK": 2 + }, + "(2^14, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^16, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^18, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^20, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^7, 1024, 1024, 27)": { + "SPLITK": 16 + }, + "(2^9, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^11, 1024, 1024, 27)": { + "SPLITK": 4 + }, + "(2^13, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^19, 128, 128, 27)": { + "SPLITK": 1 + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_masked_implicit_gemm.sparse_submanifold_conv_fwd_masked_implicit_gemm_kernel": { + "(18, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_fwd_masked_implicit_gemm_kernel": { + "(18, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(23, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 256, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 512, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 1024, 4096, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(21, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 32, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 16, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 32 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(19, 128, 128, 27, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(18, 512, 2048, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(20, 256, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 128, 128, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(22, 128, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(24, 64, 64, 27, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float16')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_fwd_masked_implicit_gemm_splitk_kernel": { + "(16, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 1024, 4096, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 512, 2048, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 256, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 4096, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 128, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 64, 16, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 256, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 256, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 1024, 4096, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 512, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 256, + "BK": 64 + }, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(16, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(17, 512, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 16, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(14, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(12, 128, 32, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(10, 256, 64, 27, 256, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 512, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 512, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 512, 27, 256, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 128, 27, 256, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 128, 27, 512, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(8, 512, 128, 27, 1024, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 16, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(15, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 128, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 128, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 128, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 32, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 32, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 32, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 32, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 128, 32, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 512, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 512, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 512, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 512, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 128, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 128, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 128, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 128, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 128, 27, 256, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 512, 128, 27, 512, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 64, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 64, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 16, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 16, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 16, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(13, 64, 16, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 256, + "B2": 64, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 256, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 128, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 256, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 256, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 256, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 256, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 64, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 64, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 64, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 64, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 64, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 256, 64, 27, 256, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 32 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(6, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(6, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(6, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(6, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(6, 1024, 1024, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(6, 1024, 1024, 27, 256, True, 'torch.float16', 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 64 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(7, 1024, 1024, 27, 4, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 64, + "B2": 64, + "BK": 64 + }, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(9, 1024, 1024, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + }, + "(11, 1024, 1024, 27, 4, True, 'torch.float32', 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.int64', 'torch.float32')": { + "kwargs": { + "B1": 128, + "B2": 128, + "BK": 32 + }, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 4, + "num_buffers_warp_spec": 0, + "num_consumer_groups": 0, + "reg_dec_producer": 0, + "reg_inc_consumer": 0, + "maxnreg": null, + "pre_hook": null + } + }, + "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_fwd_masked_implicit_gemm_splitk.sparse_submanifold_conv_fwd_masked_implicit_gemm_splitk": { + "(2^18, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^18, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^16, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^16, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^14, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^14, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^12, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^12, 512, 128, 27)": { + "SPLITK": 16 + }, + "(2^10, 1024, 1024, 27)": { + "SPLITK": 4 + }, + "(2^10, 1024, 4096, 27)": { + "SPLITK": 4 + }, + "(2^12, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^14, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^16, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^19, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^19, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^17, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^17, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^15, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^15, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^13, 512, 512, 27)": { + "SPLITK": 2 + }, + "(2^13, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^11, 1024, 1024, 27)": { + "SPLITK": 4 + }, + "(2^11, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^13, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^15, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^17, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^11, 512, 512, 27)": { + "SPLITK": 4 + }, + "(2^11, 512, 128, 27)": { + "SPLITK": 16 + }, + "(2^9, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^9, 1024, 4096, 27)": { + "SPLITK": 8 + }, + "(2^11, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^13, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^13, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^13, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^17, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^17, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^15, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^15, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^15, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^8, 1024, 1024, 27)": { + "SPLITK": 16 + }, + "(2^8, 1024, 4096, 27)": { + "SPLITK": 4 + }, + "(2^10, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^10, 512, 128, 27)": { + "SPLITK": 32 + }, + "(2^10, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^20, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^20, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^18, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^18, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^18, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^16, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^16, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^16, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^12, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^12, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^12, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^21, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^21, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^19, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^19, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^14, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^14, 512, 128, 27)": { + "SPLITK": 4 + }, + "(2^14, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^17, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^19, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^7, 1024, 1024, 27)": { + "SPLITK": 4 + }, + "(2^7, 1024, 4096, 27)": { + "SPLITK": 16 + }, + "(2^14, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^14, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^14, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^16, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^16, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^22, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^22, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^12, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^10, 256, 256, 27)": { + "SPLITK": 16 + }, + "(2^12, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^13, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^13, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^15, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^20, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^20, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^18, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^18, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^18, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^20, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^16, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^16, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^14, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^14, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^16, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^23, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^23, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^21, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^21, 128, 32, 27)": { + "SPLITK": 1 + }, + "(2^19, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^19, 256, 64, 27)": { + "SPLITK": 1 + }, + "(2^17, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^17, 512, 128, 27)": { + "SPLITK": 1 + }, + "(2^15, 1024, 1024, 27)": { + "SPLITK": 1 + }, + "(2^15, 1024, 4096, 27)": { + "SPLITK": 1 + }, + "(2^17, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^19, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^21, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^11, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^11, 256, 64, 27)": { + "SPLITK": 32 + }, + "(2^14, 64, 64, 27)": { + "SPLITK": 2 + }, + "(2^14, 64, 16, 27)": { + "SPLITK": 4 + }, + "(2^12, 128, 128, 27)": { + "SPLITK": 4 + }, + "(2^12, 128, 32, 27)": { + "SPLITK": 32 + }, + "(2^10, 256, 64, 27)": { + "SPLITK": 128 + }, + "(2^8, 512, 512, 27)": { + "SPLITK": 64 + }, + "(2^8, 512, 128, 27)": { + "SPLITK": 256 + }, + "(2^15, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^15, 64, 16, 27)": { + "SPLITK": 4 + }, + "(2^13, 128, 128, 27)": { + "SPLITK": 2 + }, + "(2^13, 128, 32, 27)": { + "SPLITK": 32 + }, + "(2^9, 512, 512, 27)": { + "SPLITK": 8 + }, + "(2^9, 512, 128, 27)": { + "SPLITK": 256 + }, + "(2^13, 64, 64, 27)": { + "SPLITK": 1 + }, + "(2^13, 64, 16, 27)": { + "SPLITK": 1 + }, + "(2^9, 256, 256, 27)": { + "SPLITK": 8 + }, + "(2^9, 256, 64, 27)": { + "SPLITK": 32 + }, + "(2^6, 1024, 1024, 27)": { + "SPLITK": 8 + }, + "(2^18, 512, 512, 27)": { + "SPLITK": 1 + }, + "(2^18, 512, 2048, 27)": { + "SPLITK": 1 + }, + "(2^20, 256, 256, 27)": { + "SPLITK": 1 + }, + "(2^20, 256, 1024, 27)": { + "SPLITK": 1 + }, + "(2^22, 128, 128, 27)": { + "SPLITK": 1 + }, + "(2^22, 128, 512, 27)": { + "SPLITK": 1 + }, + "(2^24, 64, 64, 27)": { + "SPLITK": 1 + } + } + } +} \ No newline at end of file diff --git a/nodes.py b/nodes.py index 043e46b..ca7e5e3 100644 --- a/nodes.py +++ b/nodes.py @@ -18,6 +18,7 @@ import copy import pymeshlab import cumesh as CuMesh +import o_voxel import meshlib.mrmeshnumpy as mrmeshnumpy import meshlib.mrmeshpy as mrmeshpy @@ -164,6 +165,8 @@ class Trellis2LoadModel: def process(self, modelname, backend, device, low_vram, keep_models_loaded): os.environ['OPENCV_IO_ENABLE_OPENEXR'] = '1' os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True" # Can save GPU memory + os.environ["FLEX_GEMM_AUTOTUNE_CACHE_PATH"] = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'autotune_cache.json') + os.environ["FLEX_GEMM_AUTOTUNER_VERBOSE"] = '1' os.environ['ATTN_BACKEND'] = backend torch.backends.cudnn.benchmark = False @@ -1042,6 +1045,7 @@ class Trellis2Remesh: "fill_holes": ("BOOLEAN", {"default":True}), "fill_holes_max_perimeter": ("FLOAT",{"default":0.03,"min":0.001,"max":99.999,"step":0.001}), "dual_contouring_resolution": (["Auto","128","256","512","1024","2048"],{"default":"Auto"}), + "remove_floaters": ("BOOLEAN",{"default":True}), }, } @@ -1051,9 +1055,12 @@ class Trellis2Remesh: CATEGORY = "Trellis2Wrapper" OUTPUT_NODE = True - def process(self, mesh, remesh_band, remesh_project, fill_holes, fill_holes_max_perimeter, dual_contouring_resolution): + def process(self, mesh, remesh_band, remesh_project, fill_holes, fill_holes_max_perimeter, dual_contouring_resolution, remove_floaters): mesh_copy = copy.deepcopy(mesh) + if remove_floaters: + mesh_copy = remove_floater(mesh_copy) + aabb = [[-0.5, -0.5, -0.5], [0.5, 0.5, 0.5]] vertices = mesh_copy.vertices @@ -1334,9 +1341,10 @@ class Trellis2PostProcess2: "required": { "mesh": ("MESHWITHVOXEL",), "fill_holes": ("BOOLEAN", {"default":True}), - "fill_holes_max_perimeter": ("FLOAT",{"default":0.03,"min":0.001,"max":99.999,"step":0.001}), + "fix_normals": ("BOOLEAN", {"default":False}), "fix_face_orientation": ("BOOLEAN", {"default":True}), - "remove_duplicate_faces": ("BOOLEAN",{"default":True}), + "remove_duplicate_faces": ("BOOLEAN",{"default":True}), + "remove_infinite_values": ("BOOLEAN",{"default":True}), }, } @@ -1346,7 +1354,7 @@ class Trellis2PostProcess2: CATEGORY = "Trellis2Wrapper" OUTPUT_NODE = True - def process(self, mesh, fill_holes, fill_holes_max_perimeter, fix_face_orientation, remove_duplicate_faces): + def process(self, mesh, fill_holes, fix_normals, fix_face_orientation, remove_duplicate_faces, remove_infinite_values): mesh_copy = copy.deepcopy(mesh) vertices_np = mesh_copy.vertices.cpu().numpy() @@ -1357,26 +1365,24 @@ class Trellis2PostProcess2: print(f"Initial mesh: {len(trimesh.faces)} faces") print(f"Is winding consistent? {trimesh.is_winding_consistent}") - if fix_face_orientation: - # 2. Fix the normals - # This aligns the winding of all faces to be consistent + if fix_normals: print('Fixing normals ...') - trimesh.fix_normals() - - # 3. Handle inverted orientations - # If the mesh is "inside out" (normals pointing inward), - # this attempts to flip them to point outward based on the mesh volume. + trimesh.fix_normals() + + if fix_face_orientation: if trimesh.is_watertight: print('Mesh is watertight, fixing inversion ...') - Trimesh.repair.fix_inversion(trimesh) + Trimesh.repair.fix_inversion(trimesh) + else: + print('Mesh is not watertight, cannot fix inversion') - # 4. Optional: General Cleanup - # AI meshes often have duplicate vertices or degenerate faces if remove_duplicate_faces: print('Removing duplicate faces ...') trimesh.remove_duplicate_faces() - - trimesh.remove_infinite_values() + + if remove_infinite_values: + print('Removing infinite values ...') + trimesh.remove_infinite_values() if fill_holes: print('Filling holes ...') @@ -1391,7 +1397,46 @@ class Trellis2PostProcess2: del trimesh gc.collect() - return (mesh_copy,) + return (mesh_copy,) + +class Trellis2OvoxelExportToGLB: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "mesh": ("MESHWITHVOXEL",), + "resolution": ([512,1024],{"default":1024}), + "texture_size": ([512,1024,2048,4096],{"default":2048}), + "target_face_num": ("INT",{"default":2000000,"min":500,"max":16000000}), + }, + } + + RETURN_TYPES = ("TRIMESH",) + RETURN_NAMES = ("trimesh",) + FUNCTION = "process" + CATEGORY = "Trellis2Wrapper" + OUTPUT_NODE = True + + def process(self, mesh, resolution, texture_size, target_face_num): + mesh_copy = copy.deepcopy(mesh) + + glb = o_voxel.postprocess.to_glb( + vertices=mesh_copy.vertices, + faces=mesh_copy.faces, + attr_volume=mesh_copy.attrs, + coords=mesh_copy.coords, + attr_layout=mesh_copy.layout, + grid_size=resolution, + aabb=[[-0.5, -0.5, -0.5], [0.5, 0.5, 0.5]], + decimation_target=target_face_num, + texture_size=texture_size, + remesh=True, + remesh_band=1, + remesh_project=0, + use_tqdm=True, + ) + + return (glb,) NODE_CLASS_MAPPINGS = { "Trellis2LoadModel": Trellis2LoadModel, @@ -1410,6 +1455,7 @@ NODE_CLASS_MAPPINGS = { "Trellis2PreProcessImage": Trellis2PreProcessImage, "Trellis2MeshRefiner": Trellis2MeshRefiner, "Trellis2PostProcess2": Trellis2PostProcess2, + "Trellis2OvoxelExportToGLB": Trellis2OvoxelExportToGLB, } NODE_DISPLAY_NAME_MAPPINGS = { @@ -1429,4 +1475,5 @@ NODE_DISPLAY_NAME_MAPPINGS = { "Trellis2PreProcessImage": "Trellis2 - PreProcess Image", "Trellis2MeshRefiner": "Trellis2 - Mesh Refiner", "Trellis2PostProcess2": "Trellis2 - PostProcess Mesh 2", + "Trellis2OvoxelExportToGLB": "Trellis2 - Ovoxel Export to GLB", } \ No newline at end of file diff --git a/trellis2/pipelines/trellis2_image_to_3d.py b/trellis2/pipelines/trellis2_image_to_3d.py index c81a7fd..e8c57e5 100644 --- a/trellis2/pipelines/trellis2_image_to_3d.py +++ b/trellis2/pipelines/trellis2_image_to_3d.py @@ -1001,7 +1001,9 @@ class Trellis2ImageTo3DPipeline(Pipeline): """ Preprocess the input mesh. """ - vertices = mesh.vertices + mesh = mesh.copy() + vertices = mesh.vertices.copy() + vertices_min = vertices.min(axis=0) vertices_max = vertices.max(axis=0) center = (vertices_min + vertices_max) / 2 @@ -1011,7 +1013,9 @@ class Trellis2ImageTo3DPipeline(Pipeline): vertices[:, 1] = -vertices[:, 2] vertices[:, 2] = tmp assert np.all(vertices >= -0.5) and np.all(vertices <= 0.5), 'vertices out of range' - return trimesh.Trimesh(vertices=vertices, faces=mesh.faces, process=False) + + mesh.vertices = vertices + return mesh def encode_shape_slat( self, @@ -1069,9 +1073,11 @@ class Trellis2ImageTo3DPipeline(Pipeline): texture_alpha_mode = 'OPAQUE', double_side_material = True ): + vertices = mesh.vertices faces = mesh.faces - normals = mesh.vertex_normals + normals = np.asarray(mesh.vertex_normals).copy() + vertices_torch = torch.from_numpy(vertices).float().cuda() faces_torch = torch.from_numpy(faces).int().cuda() if hasattr(mesh, 'visual') and hasattr(mesh.visual, 'uv') and mesh.visual.uv is not None: