[ModelOpt][PP] Keep BF16 shared experts out of the NVFP4 fusion so TP1 pipeline stages can load (#40628)

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."""