[Fix][NPU] Preserve existing packed_modules_mapping when merging model-level fused module mappings (#25755)

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

@@ -131,6 +131,9 @@ class QuantizationConfig(ABC):
         # mapping is updated by models as they initialize
         self.packed_modules_mapping: Dict[str, List[str]] = dict()
 
+    def update_packed_modules_mapping(self, mapping: Dict[str, List[str]]) -> None:
+        self.packed_modules_mapping = mapping
+
     @abstractmethod
     def get_name(self) -> str:
         """Name of the quantization method."""

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

@@ -108,6 +108,9 @@ class ModelSlimConfig(QuantizationConfig):
                     [npu_wrapper_rmsnorm_forward],
                 )
 
+    def update_packed_modules_mapping(self, mapping: Dict[str, List[str]]) -> None:
+        self.packed_modules_mapping.update(mapping)
+
     def get_linear_method(self) -> ModelSlimLinearMethod:
         return ModelSlimLinearMethod(self)
 

python/sglang/srt/models/deepseek_v2.py

@@ -2462,8 +2462,8 @@ class DeepseekV2ForCausalLM(nn.Module, DeepseekV2WeightLoaderMixin):
         # Quant configs like Quark may rely on the model to provide fused-module
         # mappings so exclusion checks can unfuse derived names back to the
         # checkpoint's source layer names.
-        if quant_config is not None and hasattr(quant_config, "packed_modules_mapping"):
+        if quant_config is not None:
-            quant_config.packed_modules_mapping = self.packed_modules_mapping
+            quant_config.update_packed_modules_mapping(self.packed_modules_mapping)
 
         self.pp_group = get_pp_group()
         self.config = config