Commit History
-
README: fix H1 title + direct-install instructions for new build/ layout
Verified signature6011704 -
Update repo-id in build.toml after migration to SNC org
Verified signature0e62022 -
Update repo-id references after migration to SNC org
Verified signature72ebb10 -
README: add backward validation results, real perf numbers vs vLLM FA2
Verified signatureAdds validated 40/40 backward pass result, fresh-venv standalone install confirmation, and a measured perf table (8 Qwen3 prefill shapes, HF FA4 vs vLLM FA2) replacing the prior speculative perf section.c3c086c -
Remove obsolete torch-ext/ (replaced by build/torch-cuda/)
Verified signaturef467875 -
Restructure: build/torch-cuda/ layout with relative imports for self-contained kernels.get_kernel() use
Verified signaturePreviously the kernel had files at torch-ext/flash_attn_4_sm120_sncbl/*.py with absolute 'from flash_attn.cute.X' imports, which required users to also install Dao-AILab/flash-attention separately. Now the package follows the kernels-community/flash-attn4 layout: build/torch-cuda/*.py with relative imports throughout (160 import lines rewritten across 27 files). Self-contained, loadable via kernels.get_kernel() without external flash-attn install.02fc236 -
Add README.md as the model card; remove CARD.md
Verified signatureHF renders README.md on the model page; CARD.md was not being shown. New README expands setup instructions, usage examples (basic / GQA / varlen / paged KV), API reference, validation matrix, known limitations, performance expectations, and credits.63bd811 -
Apply Dao #2484 SM120 init-time fix: pack_gqa=False on Sm120/Sm120Tma
Verified signatureResolves GQA / MQA crd2idx compile errors. Validation: 64/64 pass on SM121a (MHA + GQA + MQA, dense + varlen, bf16 + fp16). See CARD.md and CONFLICTS_LOG.md.07431ad
-
Initial commit: 5-PR SM120 bundle validated on SM121a
Verified signatureb7d73d5 -
initial commit
Verified signature46a0968