[Bugfix] Skip VllmConfig re-validation for with_hf_config submodel views

tests/test_config.py

@@ -289,6 +289,46 @@ def test_kv_offloading_does_not_skip_dcp_interleave_validation():
         VllmConfig.validate_block_size(config)
 
 
+def test_nixl_dcp_check_skipped_for_submodel_config():
+    # with_hf_config() builds a submodule view of the config (e.g. a
+    # multimodal model's text stack) whose architecture list is empty.
+    # Re-running VllmConfig validation on it is unsafe: the NIXL DCP check
+    # reads use_mla, which resolves the architecture registry.
+    model_config = ModelConfig("Qwen/Qwen2-VL-2B-Instruct", max_model_len=2048)
+    vllm_config = VllmConfig(
+        model_config=model_config,
+        device_config=DeviceConfig(device="cpu"),
+        kv_transfer_config=KVTransferConfig(
+            kv_connector="NixlConnector",
+            kv_role="kv_both",
+        ),
+    )
+    submodel_config = vllm_config.with_hf_config(model_config.hf_text_config)
+    assert submodel_config.model_config.is_submodel_config
+    assert submodel_config.model_config.architectures == []
+    assert submodel_config.model_config.use_mla is False
+
+
+def test_nixl_dcp_check_rejects_non_mla_model_with_dcp(monkeypatch):
+    # Pretend the model has a single KV head so the DCP feasibility checks
+    # pass and the MLA-only assert is what actually fires.
+    monkeypatch.setattr(ModelConfig, "get_total_num_kv_heads", lambda self: 1)
+    with pytest.raises(ValidationError, match="only supported for MLA models"):
+        VllmConfig(
+            model_config=ModelConfig("Qwen/Qwen3-0.6B", max_model_len=2048),
+            device_config=DeviceConfig(device="cpu"),
+            parallel_config=ParallelConfig(
+                tensor_parallel_size=2,
+                decode_context_parallel_size=2,
+                distributed_executor_backend="mp",
+            ),
+            kv_transfer_config=KVTransferConfig(
+                kv_connector="NixlConnector",
+                kv_role="kv_both",
+            ),
+        )
+
+
 def test_compile_config_repr_succeeds():
     # setup: VllmBackend mutates the config object
     config = VllmConfig()

vllm/config/model.py

@@ -187,6 +187,10 @@ class ModelConfig:
     """The Hugging Face config of the model."""
     hf_text_config: PretrainedConfig = field(init=False)
     """The Hugging Face config of the text model (same as hf_config for text models)."""
+    is_submodel_config: bool = field(default=False, init=False)
+    """Whether this is a submodule view derived by `VllmConfig.with_hf_config`
+    (e.g. a multimodal model's text stack). Its architecture list is empty, so
+    deployment-level validation must not run against it."""
     word_embeddings_untied_by_checkpoint: bool = field(default=False, init=False)
     """Whether `tie_word_embeddings` was overridden to `False` because the checkpoint
     contains an `lm_head` of its own. The two may still turn out to be identical, in
@@ -1082,7 +1086,10 @@ def _get_transformers_backend_cls(self) -> str:
         # Check if the architecture we're wrapping has defaults
         runner = None
         task = None
-        if defaults := try_match_architecture_defaults(self.architectures[0]):
+        # architectures is empty for with_hf_config() submodel views.
+        if self.architectures and (
+            defaults := try_match_architecture_defaults(self.architectures[0])
+        ):
             _, (runner, task) = defaults
         # User specified value take precedence
         if self.runner != "auto":

vllm/config/vllm.py

@@ -943,6 +943,7 @@ def with_hf_config(
 
         model_config.hf_config = hf_config
         model_config.model_arch_config = model_config.get_model_arch_config()
+        model_config.is_submodel_config = True
 
         return replace(self, model_config=model_config)
 
@@ -1331,6 +1332,12 @@ def __post_init__(self):
         # To give each torch profile run a unique instance name.
         self.instance_id = f"{time.time_ns()}"
 
+        if self.model_config is not None and self.model_config.is_submodel_config:
+            # with_hf_config() view: the parent config was already validated,
+            # and this view's empty architecture list makes the model-dependent checks
+            # below unsafe (e.g. use_mla resolves the architecture registry).
+            return
+
         self._resolve_mm_encoder_only()
 
         if self.performance_mode != "balanced":