SolitaryThinker 73b56f0a63 [feat]: FlashInfer SageAttention3 backend + attention shape logging ...

Adds FLASHINFER_SAGE_ATTN3 wrapping flashinfer's dense SM120 NVFP4
attention (flashinfer-ai/flashinfer#3640), falling back to torch SDPA
with one warning when flashinfer/SM120 is unavailable, and an env-gated
JSONL attention shape logger (FASTVIDEO_ATTN_SHAPE_LOG) hooked into the
new backend and SAGE_ATTN_THREE to collect real workload shapes for
SM121 (DGX Spark) kernel tuning.

2026-07-09 12:48:18 -07:00

..

__init__.py

[misc]: demote ROCm-unavailable startup message to DEBUG (#1360)

2026-05-15 22:12:16 -07:00

cpu.py

[feat] pre-commit support 120 col num (#1167)

2026-03-11 20:19:30 -07:00

cuda.py

[feat]: FlashInfer SageAttention3 backend + attention shape logging

2026-07-09 12:48:18 -07:00

interface.py

[feat]: FlashInfer SageAttention3 backend + attention shape logging

2026-07-09 12:48:18 -07:00

mps.py

[feat] pre-commit support 120 col num (#1167)

2026-03-11 20:19:30 -07:00

npu.py

[feat] pre-commit support 120 col num (#1167)

2026-03-11 20:19:30 -07:00

rocm.py

[feat] pre-commit support 120 col num (#1167)

2026-03-11 20:19:30 -07:00