diff --git a/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16_moe.py b/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16_moe.py index 581972032..0d6a9b896 100644 --- a/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16_moe.py +++ b/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16_moe.py @@ -425,6 +425,7 @@ class CompressedTensorsWNA16MoE(CompressedTensorsMoEScheme): num_bits=self.num_bits, is_k_full=self.is_k_full, routed_scaling_factor=self.moe_runner_config.routed_scaling_factor, + clamp_limit=self.moe_runner_config.swiglu_limit, workspace=layer.workspace, ) return StandardCombineInput(hidden_states=output)