fix: fix NVFP4 Kimi-K2.5 weight mapping and exclude list (sgl-project#18370)

Author: mmangkad

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

@@ -64,6 +64,7 @@
         CombineInput,
         StandardDispatchOutput,
     )
+    from sglang.srt.models.utils import WeightsMapper
 
 fp4_quantize = None
 try:
@@ -304,6 +305,22 @@ def get_config_filenames(cls) -> List[str]:
     def get_scaled_act_names(self) -> List[str]:
         return []
 
+    def apply_weight_name_mapper(
+        self, hf_to_sglang_mapper: "WeightsMapper"
+    ):  # noqa: B027
+        # Map excluded module patterns from HF layout to sglang layout.
+        # Ref: HF hf_quant_config.json for nvidia/Kimi-K2.5-NVFP4
+        # https://huggingface.co/nvidia/Kimi-K2.5-NVFP4/blob/main/hf_quant_config.json
+        if self.exclude_modules:
+            mapped = hf_to_sglang_mapper.apply_list(self.exclude_modules)
+            expanded: List[str] = []
+            for name in mapped:
+                expanded.append(name)
+                if name.startswith("language_model."):
+                    expanded.append(name.removeprefix("language_model."))
+            # Preserve order, drop duplicates.
+            self.exclude_modules = list(dict.fromkeys(expanded))
+
 
 class ModelOptFp8Config(ModelOptQuantConfig):
     """Configuration for ModelOpt FP8 quantization, including serialization and compatibility checks."""

python/sglang/srt/models/kimi_k25.py

@@ -34,6 +34,7 @@
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.deepseek_v2 import DeepseekV3ForCausalLM
 from sglang.srt.models.kimi_vl_moonvit import MLP2
+from sglang.srt.models.utils import WeightsMapper
 from sglang.srt.utils import add_prefix
 
 KIMIV_VT_INFER_MAX_PATCH_NUM = 16328
@@ -643,6 +644,15 @@ def vision_tower_forward_auto(
 
 
 class KimiK25ForConditionalGeneration(nn.Module):
+    # Support nvidia/Kimi-K2.5-NVFP4 naming: language_model.layers.*.
+    # Ref: HF config.json for nvidia/Kimi-K2.5-NVFP4
+    # https://huggingface.co/nvidia/Kimi-K2.5-NVFP4/blob/main/config.json
+    hf_to_sglang_mapper = WeightsMapper(
+        orig_to_new_prefix={
+            "language_model.layers.": "language_model.model.layers.",
+        }
+    )
+
     def __init__(
         self,
         config: KimiK25Config,
@@ -710,7 +720,9 @@ def forward(
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         """Load weights for the model, separating vision and language weights"""
-        weights = list(weights)
+        mapper = getattr(self, "hf_to_sglang_mapper", None)
+        if mapper is not None:
+            weights = mapper.apply(weights)
 
         # Separate vision tower weights and language model weights
         vision_weights = []