diff --git a/vllm/model_executor/layers/fused_moe/experts/nvfp4_emulation_moe.py b/vllm/model_executor/layers/fused_moe/experts/nvfp4_emulation_moe.py index f93c67a97d..cd862cac59 100644 --- a/vllm/model_executor/layers/fused_moe/experts/nvfp4_emulation_moe.py +++ b/vllm/model_executor/layers/fused_moe/experts/nvfp4_emulation_moe.py @@ -372,6 +372,11 @@ class Nvfp4QuantizationEmulationTritonExperts(TritonExperts): def expects_unquantized_inputs(self) -> bool: return True + @property + def a1_scale(self) -> torch.Tensor: + # Used in experts/triton_moe.py and passed to moe_kernel_quantize_input. + return self.a1_gscale + @staticmethod def supports_lora() -> bool: return False diff --git a/vllm/model_executor/layers/fused_moe/experts/triton_moe.py b/vllm/model_executor/layers/fused_moe/experts/triton_moe.py index 8c756e2570..3196667b3f 100644 --- a/vllm/model_executor/layers/fused_moe/experts/triton_moe.py +++ b/vllm/model_executor/layers/fused_moe/experts/triton_moe.py @@ -245,7 +245,7 @@ class TritonExperts(LoRAExpertsMixin, mk.FusedMoEExpertsModular): lora_unquantized_hidden_states = hidden_states hidden_states, a1q_scale = moe_kernel_quantize_input( hidden_states, - self.a1_scale or self.a1_gscale, + self.a1_scale, self.quant_dtype, self.per_act_token_quant, self.block_shape,