[Bugfix] Exclude language_model_only key from MM AOT compile hash but include in model one (#34508)

Signed-off-by: Roger Wang <hey@rogerw.io>

tests/config/test_multimodal_config.py

@@ -3,6 +3,7 @@
 
 import pytest
 
+from vllm.config.model import ModelConfig
 from vllm.config.multimodal import MultiModalConfig
 from vllm.v1.attention.backends.registry import AttentionBackendEnum
 
@@ -23,3 +24,20 @@ def test_mm_encoder_attn_backend_hash_updates():
         mm_encoder_attn_backend=AttentionBackendEnum.FLASH_ATTN
     ).compute_hash()
     assert base_hash != overridden_hash
+
+
+def test_language_model_only_does_not_affect_mm_hash():
+    """language_model_only does not affect the ViT computation graph,
+    so it should not change the multimodal config hash."""
+    base_hash = MultiModalConfig().compute_hash()
+    lm_only_hash = MultiModalConfig(language_model_only=True).compute_hash()
+    assert base_hash == lm_only_hash
+
+
+def test_language_model_only_affects_model_hash():
+    """language_model_only affects the LM computation graph,
+    so it should change the model config hash."""
+    model = "llava-hf/llava-1.5-7b-hf"
+    base_hash = ModelConfig(model).compute_hash()
+    lm_only_hash = ModelConfig(model, language_model_only=True).compute_hash()
+    assert base_hash != lm_only_hash

vllm/config/model.py

@@ -357,6 +357,12 @@ class ModelConfig:
         from vllm.config.utils import get_hash_factors, hash_factors
 
         factors = get_hash_factors(self, ignored_factors)
+
+        # NOTE: For some models (e.g, Qwen3-VL), whether the MM code path is enabled
+        # affects the computation graph of the language model, therefore we add it
+        # here early.
+        if self.multimodal_config:
+            factors["language_model_only"] = self.multimodal_config.language_model_only
         return hash_factors(factors)
 
     def _update_nested(

vllm/config/multimodal.py

@@ -219,7 +219,6 @@ class MultiModalConfig:
         the final hidden states.
         """
         factors: list[Any] = [
-            self.language_model_only,
             self.mm_encoder_attn_backend.name
             if self.mm_encoder_attn_backend is not None
             else None,