biondizzle a782ac00ce Integrate CSA/SDPA attention into vLLM for Blackwell ...

- Add vllm/patches/layers/csa_attention.py: pure PyTorch replacement
  for FlashMLA + fused CUDA kernels that don't work on SM100
- Patch deepseek_v4_attention.py: detect SM100+ and dispatch to
  _forward_blackwell() which uses:
  1. fused_qnorm_rope_kv_insert_py() instead of C++ kernel
  2. full_sdpa_attention() instead of FlashMLA
  3. BF16 inverse RoPE + BMM for wo_a (same as existing BF16 path)
- Add csa_attention.py to Dockerfile

The Blackwell path:
  GEMM projections (CuTeDSL) → RMS norm → q_b → RoPE (PyTorch) →
  SDPA attention → inverse RoPE + wo_a BMM → wo_b → output

2026-05-19 08:04:07 +00:00

..

kernels/linear/nvfp4

Fix OOM: add --max-model-len=876544 + revert CPU dummy weight

2026-05-19 07:35:43 +00:00

patches

Integrate CSA/SDPA attention into vLLM for Blackwell

2026-05-19 08:04:07 +00:00

cutedsl_quant_method.py

Fix OOM: add --max-model-len=876544 + revert CPU dummy weight

2026-05-19 07:35:43 +00:00

nvfp4_cutedsl.py

Replace autograd.Function with torch.library.custom_op for Dynamo compat

2026-05-19 01:54:48 +00:00