[Bugfix] Support HF sharded weights for Mistral3/Pixtral models (#32673)

Signed-off-by: ricky-chaoju <ricky.chen@infinirc.com>
Signed-off-by: vllm-dev <ricky.chen@infinirc.com>

vllm/model_executor/models/pixtral.py

@@ -502,10 +502,12 @@ class PixtralForConditionalGeneration(
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
         def is_vision_encoder_weights(weight: tuple[str, torch.Tensor]):
-            return weight[0].startswith("vision_encoder")
+            return weight[0].startswith(("vision_encoder", "vision_tower"))
 
         def is_vision_lang_adapter_weights(weight: tuple[str, torch.Tensor]):
-            return weight[0].startswith("vision_language_adapter")
+            return weight[0].startswith(
+                ("vision_language_adapter", "multi_modal_projector")
+            )
 
         def is_patch_merger(weight: tuple[str, torch.Tensor]):
             return weight[0].startswith("patch_merger")
@@ -543,9 +545,10 @@ class PixtralForConditionalGeneration(
                         continue
                     # Load vision encoder weights directly
                     trimmed_name = ".".join(name.split(".")[1:])
-                    param = vision_encoder_dict[trimmed_name]
-                    with torch.no_grad():
-                        default_weight_loader(param, w)
+                    param = vision_encoder_dict.get(trimmed_name)
+                    if param is not None:
+                        with torch.no_grad():
+                            default_weight_loader(param, w)
                 elif is_patch_merger((name, w)):
                     if self.patch_merger is None:
                         continue
@@ -567,12 +570,15 @@ class PixtralForConditionalGeneration(
                         continue
                     # Load vision-language adapter weights directly
                     trimmed_name = ".".join(name.split(".")[1:])
-                    param = vision_lang_adapter_dict[trimmed_name]
-                    with torch.no_grad():
-                        default_weight_loader(param, w)
+                    param = vision_lang_adapter_dict.get(trimmed_name)
+                    if param is not None:
+                        with torch.no_grad():
+                            default_weight_loader(param, w)
                 else:
                     # LLM weights: yield them to be loaded
                     # by language_model.load_weights
+                    # Strip "language_model." prefix if present (HF sharded format)
+                    name = name.removeprefix("language_model.")
                     yield (name, w)
 
         # Now we call the language model load with the generator