biondizzle 450793311c Wire CuTeDSL kernels into vLLM: replace all BF16 dequant with native NVFP4 ...

- CuTeDSLNvfp4Method: custom quant method that creates CuTeDSL runners
  during process_weights_after_loading, then swaps to CuTeDSLNvfp4LinearMethod
  for forward dispatch
- Attention projections (fused_wqa_wkv, wq_b, wo_b) now route through
  CuTeDSLNvfp4Linear (cosine 0.992-0.996 vs BF16 reference)
- Shared expert now uses CuTeDSLSharedExpertRunner (cosine 0.992 vs BF16)
  with monkey-patched forward for fused L1+SiLU+L2 pipeline
- Deleted all BF16 dequant code (_dequant_nvfp4_to_bf16, _post_quant_fix,
  input_scale fixes)
- Deleted _post_quant_fix hook from utils.py
- Fixed SwiGLU clamp: gate clamped BEFORE SiLU (matching SiluAndMulWithClamp)
- Cleaned up all debug prints
- Updated Dockerfile with new kernel files

2026-05-18 20:27:42 +00:00

..

kernel

refactor: copy CuTeDSL kernel into repo with local imports

2026-05-16 02:57:54 +00:00

__init__.py

refactor: copy CuTeDSL kernel into repo with local imports

2026-05-16 02:57:54 +00:00

bridge.py

add kernel compile caching — compile once, invoke on subsequent calls

2026-05-16 20:45:46 +00:00

moe_pipeline.py

Add pipeline test with real model weights, add swiglu_limit to reference moe_pipeline

2026-05-17 18:07:44 +00:00

nvfp4_linear.py

Wire CuTeDSL kernels into vLLM: replace all BF16 dequant with native NVFP4

2026-05-18 20:27:42 +00:00

shared_expert_pipeline.py

Wire CuTeDSL kernels into vLLM: replace all BF16 dequant with native NVFP4

2026-05-18 20:27:42 +00:00