vllm/vllm/model_executor at fix/eplb-nvfp4-modelopt - vllm - Gitea: Git with a cup of macha paffe

Files

T

History

Tyler Michael SmithandClaude Opus 4.6 36b60baf60 Fix EPLB + NVFP4 modelopt: use direct Parameter assignment for g1/g2_alphas

PR #34646 uses replace_parameter() to register g1_alphas and g2_alphas,
but these are never pre-registered in create_weights(), so
replace_parameter's getattr(mod, name) raises AttributeError.

Switch to direct nn.Parameter assignment (matching the compressed_tensors
path in the same PR) so the parameters are properly registered on the
layer. This ensures get_expert_weights() includes them in the EPLB
rearrangement set, and the quant config holds a reference to the same
tensor that EPLB modifies in-place.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>

2026-03-07 16:50:51 -05:00

kernels

[Bugfix][CI] fix typos (#34934 )

2026-03-05 17:05:46 +00:00

layers

Fix EPLB + NVFP4 modelopt: use direct Parameter assignment for g1/g2_alphas

2026-03-07 16:50:51 -05:00

model_loader

[Hardware] Replace torch.cuda.empty_cache with torch.accelerator.empty_cache (#30681 )

2026-03-04 09:49:47 +00:00

models

Adding support to Sarvam's MoE models (#33942 )