[BugFix] qwen3 no load (#179)

qwen3 no load

Author: Bounty-hunter

vllm/model_executor/models/qwen3_vl.py

@@ -82,7 +82,9 @@
 from .qwen2_vl import Qwen2VLProcessingInfo
 from .qwen3 import Qwen3ForCausalLM, Qwen3Model
 from .utils import (AutoWeightsLoader, PPMissingLayer, WeightsMapper,
-                    maybe_prefix, merge_multimodal_embeddings)
+                    maybe_init_language_model, maybe_prefix,
+                    maybe_skip_language_model_prefix,
+                    merge_multimodal_embeddings)
 from .vision import get_vit_attn_backend, run_dp_sharded_mrope_vision_model
 
 logger = init_logger(__name__)
@@ -1138,13 +1140,14 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = "model"):
                 use_data_parallel=self.use_data_parallel,
             )
 
-        self.language_model = Qwen3LLMForCausalLM(vllm_config=vllm_config,
-                                                  prefix=maybe_prefix(
-                                                      prefix,
-                                                      "language_model"))
+        self.language_model = maybe_init_language_model(
+            lambda: Qwen3LLMForCausalLM(vllm_config=vllm_config,
+                                        prefix=maybe_prefix(
+                                            prefix, "language_model")))
 
-        self.make_empty_intermediate_tensors = (
-            self.language_model.make_empty_intermediate_tensors)
+        if self.language_model is not None:
+            self.make_empty_intermediate_tensors = (
+                self.language_model.make_empty_intermediate_tensors)
 
         self.use_deepstack = hasattr(config.vision_config,
                                      'deepstack_visual_indexes')
@@ -1599,6 +1602,7 @@ def load_weights(self, weights: Iterable[tuple[str,
         skip_prefixes = []
         if self.visual is None:
             skip_prefixes.extend(["visual."])
+        maybe_skip_language_model_prefix(self, skip_prefixes, "language_model")
         loader = AutoWeightsLoader(self, skip_prefixes=skip_prefixes)
         return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
 

vllm/model_executor/models/qwen3_vl_moe.py

@@ -46,7 +46,8 @@
 from .qwen3_vl import (Qwen3_VisionTransformer, Qwen3VLDummyInputsBuilder,
                        Qwen3VLForConditionalGeneration,
                        Qwen3VLMultiModalProcessor, Qwen3VLProcessingInfo)
-from .utils import is_pp_missing_parameter, maybe_prefix
+from .utils import (is_pp_missing_parameter, maybe_init_language_model,
+                    maybe_prefix)
 
 logger = init_logger(__name__)
 
@@ -331,13 +332,14 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                 use_data_parallel=self.use_data_parallel,
             )
 
-        self.language_model = Qwen3MoeLLMForCausalLM(vllm_config=vllm_config,
-                                                     prefix=maybe_prefix(
-                                                         prefix,
-                                                         "language_model"))
+        self.language_model = maybe_init_language_model(
+            lambda: Qwen3MoeLLMForCausalLM(vllm_config=vllm_config,
+                                           prefix=maybe_prefix(
+                                               prefix, "language_model")))
 
-        self.make_empty_intermediate_tensors = (
-            self.language_model.make_empty_intermediate_tensors)
+        if self.language_model is not None:
+            self.make_empty_intermediate_tensors = (
+                self.language_model.make_empty_intermediate_tensors)
 
         self.use_deepstack = hasattr(config.vision_config,
                                      'deepstack_visual_indexes')

vllm/model_executor/models/utils.py

@@ -82,6 +82,25 @@ def apply_dict(self, values: dict[str, Any]) -> dict[str, Any]:
         }
 
 
+# Skip language model in Encoder instance
+def maybe_init_language_model(init_fn):
+    if has_ec_transfer() and get_ec_transfer().is_producer:
+        return None
+    return init_fn()
+
+
+# Skiped language model prefix
+def maybe_skip_language_model_prefix(
+    module: nn.Module,
+    skip_prefixes: list[str],
+    language_attr: str = "language_model",
+):
+    if (has_ec_transfer() and get_ec_transfer().is_producer
+            and hasattr(module, language_attr)
+            and getattr(module, language_attr) is None):
+        skip_prefixes.append(f"{language_attr}.")
+
+
 class AutoWeightsLoader:
     """
     Helper class to load weights into a [`torch.nn.Module`][]. It is able