biondizzle db5192fe41 Patch from Docker image's vLLM (0.20.2rc1) instead of newer upstream ...

The nightly Docker image uses an older vLLM that doesn't have
NormGateLinear, breakable_cudagraph, etc. Patching the Docker
image's own files ensures compatibility.

- deepseek_v4.py: Patches from Docker image + NVFP4 mapper + wo_a BF16
- deepseek_v4_attention.py: Patches from Docker image + inv rope BF16
  + weights_proj quant + removed QuantFP8/GroupShape imports

2026-05-19 03:35:15 +00:00

..

kernels/linear/nvfp4

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

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

patches

Patch from Docker image's vLLM (0.20.2rc1) instead of newer upstream

2026-05-19 03:35:15 +00:00

cutedsl_quant_method.py

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

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

nvfp4_cutedsl.py

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

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