From 3005af09412cedb988a5871a3441ae49424a13bb Mon Sep 17 00:00:00 2001
From: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
Date: Tue, 28 Jul 2026 05:56:42 +0800
Subject: [PATCH] Fix compressed-tensors NVFP4 MoE W13 layout (#32430)

---
 .../quantization/compressed_tensors/compressed_tensors.py    | 1 +
 .../compressed_tensors/schemes/compressed_tensors_scheme.py  | 2 ++
 .../schemes/compressed_tensors_w4a4_nvfp4_moe.py             | 5 +++++
 3 files changed, 8 insertions(+)

diff --git a/python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py b/python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py
index f9d3e2ea0..76e5c139e 100644
--- a/python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py
+++ b/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,
diff --git a/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_scheme.py b/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_scheme.py
index 25c7168e7..6f9ef0498 100644
--- a/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_scheme.py
+++ b/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:
         """
diff --git a/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_nvfp4_moe.py b/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_nvfp4_moe.py
index 001734c96..07c11416e 100644
--- a/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_nvfp4_moe.py
+++ b/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