From 9b59fc5db514e3748a0f6bb4ac90128602d09c87 Mon Sep 17 00:00:00 2001
From: YAMY <74099316+YAMY1234@users.noreply.github.com>
Date: Mon, 21 Sep 2026 23:45:58 -0500
Subject: [PATCH] [ModelOpt][PP] Keep BF16 shared experts out of the NVFP4
 fusion so TP1 pipeline stages can load (#40628)

---
 .../srt/layers/quantization/modelopt_quant.py |  9 ++++++++
 .../quantization/test_modelopt_nvfp4.py       | 22 +++++++++++++++++++
 2 files changed, 31 insertions(+)

diff --git a/python/sglang/srt/layers/quantization/modelopt_quant.py b/python/sglang/srt/layers/quantization/modelopt_quant.py
index 129b17b5a..7a2536933 100755
--- a/python/sglang/srt/layers/quantization/modelopt_quant.py
+++ b/python/sglang/srt/layers/quantization/modelopt_quant.py
@@ -1497,6 +1497,15 @@ class ModelOptFp4Config(ModelOptQuantConfig):
     def get_min_capability(cls) -> int:
         return 80
 
+    def can_fuse_shared_expert(self) -> bool:
+        # A shared-expert body kept BF16 via exclude_modules cannot share the packed
+        # FP4 FusedMoE buffers. The shared_expert_gate is a separate linear (kept
+        # BF16 by e.g. Qwen3-Next NVFP4 checkpoints) and must not veto fusion.
+        return not any(
+            "shared_expert" in name and "shared_expert_gate" not in name
+            for name in self.exclude_modules
+        )
+
     @staticmethod
     def common_group_size(cfg: dict) -> int:
         """Return the unique group_size across the config; raise if missing/mismatched."""
diff --git a/test/registered/unit/layers/quantization/test_modelopt_nvfp4.py b/test/registered/unit/layers/quantization/test_modelopt_nvfp4.py
index 923b886a1..8aeed7ae5 100644
--- a/test/registered/unit/layers/quantization/test_modelopt_nvfp4.py
+++ b/test/registered/unit/layers/quantization/test_modelopt_nvfp4.py
@@ -132,6 +132,28 @@ class TestModelOptNvfp4(CustomTestCase):
                 use_per_token_activation=True,
             )
 
+    def test_shared_expert_fusion_requires_matching_fp4_precision(self):
+        quantized_shared = ModelOptFp4Config(
+            is_checkpoint_nvfp4_serialized=True,
+            group_size=16,
+        )
+        bf16_shared = ModelOptFp4Config(
+            is_checkpoint_nvfp4_serialized=True,
+            group_size=16,
+            exclude_modules=["model.layers.*.mlp.shared_experts*"],
+        )
+
+        gate_only_bf16 = ModelOptFp4Config(
+            is_checkpoint_nvfp4_serialized=True,
+            group_size=16,
+            exclude_modules=["model.layers.*.mlp.shared_expert_gate"],
+        )
+
+        self.assertTrue(quantized_shared.can_fuse_shared_expert())
+        self.assertFalse(bf16_shared.can_fuse_shared_expert())
+        # Only the gate is BF16 (Qwen3-Next NVFP4): the FP4 body still fuses.
+        self.assertTrue(gate_only_bf16.can_fuse_shared_expert())
+
 
 if __name__ == "__main__":
     unittest.main()