Aryan Kumar 72f4928065 [feat]: dense FP4 attention QAD for Pika H3 Flex (FA4 NVFP4 Q/K numerics) ...

fp4_dense_qat routes FlashAttention layers through the FP4 training kernel with every
key block listed (Q/K NVFP4, P and V BF16, straight-through gradients), matching the
FlashAttention-4 NVFP4 deployment; qad.fp4_attention enables it in the Flex profile.
check_fp4_attention_parity.py compares it with the real FA4 NVFP4 kernel on a GPU.

2026-10-31 21:26:55 -07:00

..

minimax_h3_nvfp4_decoder

[feat]: FastH3 OmniRef inference on consumer devices (RTX 5090 / PRO 6000 / DGX Spark / 4090) (#1980)

2026-10-19 14:57:36 -07:00

minimax_h3_omniref_qad

[fix]: make the OmniRef eval manifest safe across shards

2026-10-19 13:10:19 -07:00

minimax_h3_v2_qad

[fix]: align the QAD configs and docs with the released provenance

2026-10-19 12:38:43 -07:00

pika_h3_flex_qad

[feat]: dense FP4 attention QAD for Pika H3 Flex (FA4 NVFP4 Q/K numerics)

2026-10-31 21:26:55 -07:00

v1_distill_dmd_wan_VSA.sh

[misc]: Refactor separate training pipeline runner (#1696)

2026-10-07 03:19:21 -07:00

v1_distill_dmd_wan.sh

[misc]: Refactor separate training pipeline runner (#1696)

2026-10-07 03:19:21 -07:00