diff --git a/vllm/model_executor/layers/fused_moe/config.py b/vllm/model_executor/layers/fused_moe/config.py index 905a9bea3c..d7ad59b46c 100644 --- a/vllm/model_executor/layers/fused_moe/config.py +++ b/vllm/model_executor/layers/fused_moe/config.py @@ -861,6 +861,7 @@ def nvfp4_w4a16_moe_quant_config( g2_alphas: torch.Tensor, w1_scale: torch.Tensor, w2_scale: torch.Tensor, + gemm1_clamp_limit: float | None = None, ) -> FusedMoEQuantConfig: """ Construct a quant config for 16-but activations and nvp4 weights. @@ -872,6 +873,7 @@ def nvfp4_w4a16_moe_quant_config( g1_alphas=g1_alphas, g2_alphas=g2_alphas, weight_dtype="nvfp4", + gemm1_clamp_limit=gemm1_clamp_limit, ) diff --git a/vllm/model_executor/layers/fused_moe/oracle/nvfp4.py b/vllm/model_executor/layers/fused_moe/oracle/nvfp4.py index 93bc81c22b..8603ca85b8 100644 --- a/vllm/model_executor/layers/fused_moe/oracle/nvfp4.py +++ b/vllm/model_executor/layers/fused_moe/oracle/nvfp4.py @@ -174,6 +174,7 @@ def select_nvfp4_moe_backend( NVFP4_BACKENDS_WITH_CLAMP = { NvFp4MoeBackend.FLASHINFER_TRTLLM, + NvFp4MoeBackend.MARLIN, } if config.swiglu_limit is not None: @@ -423,6 +424,7 @@ def make_nvfp4_moe_quant_config( g2_alphas=w2_scale_2, w1_scale=w13_scale, w2_scale=w2_scale, + gemm1_clamp_limit=swiglu_limit, ) elif backend == NvFp4MoeBackend.EMULATION: return nvfp4_moe_quant_config(