Fix compressed-tensors NVFP4 MoE W13 layout (#32430)

python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py

@@ -1035,6 +1035,7 @@ class CompressedTensorsFusedMoEMethod(FusedMoEMethodBase):
         the necessary parameters for the layer. See LinearMethodBase for param
         details
         """
+        self.load_up_proj_weight_first = layer.scheme.load_up_proj_weight_first
         layer.scheme.create_weights(
             layer=layer,
             num_experts=num_experts,

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_scheme.py

@@ -68,6 +68,8 @@ class CompressedTensorsMoEScheme(BaseMoEScheme):
     of different quantization schemes supported by CompressedTensors.
     """
 
+    load_up_proj_weight_first = False
+
     @classmethod
     def get_min_capability(cls) -> int:
         """

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_nvfp4_moe.py

@@ -42,6 +42,11 @@ class CompressedTensorsW4A4Nvfp4MoE(CompressedTensorsMoEScheme):
         self.group_size = 16
         self.use_flashinfer_trtllm = get_moe_runner_backend().is_flashinfer_trtllm()
 
+    @property
+    def load_up_proj_weight_first(self) -> bool:
+        """Load W13 as ``[up; gate]`` for CUTLASS; TRT-LLM reorders post-load."""
+        return not self.use_flashinfer_trtllm
+
     @classmethod
     def get_min_capability(cls) -> int:
         # Requires sm100(blackwell) architecture