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

3.9 KiB

Raw Blame History

View Raw