[Bugfix] Always apply MM processor even when no MM items are passed (#26240)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/inputs/preprocess.py

@@ -314,15 +314,19 @@ class InputPreprocessor:
             parsed_content["prompt_token_ids"], tokenization_kwargs)
 
         inputs: Union[TokenInputs, MultiModalInputs]
-        if multi_modal_data := parsed_content.get("multi_modal_data"):
+        if self.model_config.is_multimodal_model:
             inputs = self._process_multimodal(
                 prompt_token_ids,
-                multi_modal_data,
+                parsed_content.get("multi_modal_data", {}),
                 parsed_content.get("mm_processor_kwargs"),
                 tokenization_kwargs=tokenization_kwargs,
                 mm_uuids=mm_uuids,
             )
         else:
+            if parsed_content.get("multi_modal_data"):
+                raise ValueError(
+                    "This model does not support multimodal inputs")
+
             inputs = token_inputs(prompt_token_ids)
 
         if cache_salt := parsed_content.get("cache_salt"):
@@ -340,15 +344,19 @@ class InputPreprocessor:
         prompt_text = parsed_content["prompt"]
 
         inputs: Union[TokenInputs, MultiModalInputs]
-        if multi_modal_data := parsed_content.get("multi_modal_data"):
+        if self.model_config.is_multimodal_model:
             inputs = self._process_multimodal(
                 prompt_text,
-                multi_modal_data,
+                parsed_content.get("multi_modal_data", {}),
                 parsed_content.get("mm_processor_kwargs"),
                 tokenization_kwargs=tokenization_kwargs,
                 mm_uuids=mm_uuids,
             )
         else:
+            if parsed_content.get("multi_modal_data"):
+                raise ValueError(
+                    "This model does not support multimodal inputs")
+
             prompt_token_ids = self._tokenize_prompt(
                 prompt_text,
                 tokenization_kwargs=tokenization_kwargs,

vllm/model_executor/models/phi3v.py

@@ -507,8 +507,8 @@ class Phi3VMultiModalProcessor(BaseMultiModalProcessor[Phi3VProcessingInfo]):
         )
 
         # Keep the behavior in line with HF processor
-        if token_ids[:2] == tokenizer.encode("<s> <|image|>",
-                                             add_special_tokens=False):
+        if len(mm_prompt_updates) and (token_ids[:2] == tokenizer.encode(
+                "<s> <|image|>", add_special_tokens=False)):
             token_ids = [token_ids[0], *token_ids[2:]]
             placeholders = {
                 modality: [

vllm/model_executor/models/qwen2_5_omni_thinker.py

@@ -331,6 +331,7 @@ class Qwen2_5OmniThinkerMultiModalProcessor(
         """
         mm_item_counts = mm_items.get_all_counts()
         self._validate_mm_kwargs(mm_kwargs, mm_item_counts)
+        self._validate_mm_updates(mm_prompt_updates, mm_item_counts)
 
         use_audio_in_video = False
         if "video" in mm_kwargs:

vllm/multimodal/processing.py

@@ -1946,6 +1946,24 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
                     "model (usually arising from an inconsistency between "
                     "`_call_hf_processor` and `_get_mm_fields_config`).")
 
+    def _validate_mm_updates(
+        self,
+        mm_updates: MultiModalPromptUpdates,
+        mm_item_counts: Mapping[str, int],
+    ) -> None:
+        for modality, item_count in mm_item_counts.items():
+            placeholders = mm_updates.get(modality, [])
+
+            if len(placeholders) != item_count:
+                raise RuntimeError(
+                    f"Expected there to be {item_count} prompt updates "
+                    f"corresponding to {item_count} {modality} items, but "
+                    f"instead found {len(placeholders)} prompt updates! "
+                    "This is likely because you forgot to include input "
+                    "placeholder tokens (e.g., `<image>`, `<|image_pad|>`) "
+                    "in the prompt. If the model has a chat template, make "
+                    "sure you have applied it before calling `LLM.generate`.")
+
     def _validate_mm_placeholders(
         self,
         mm_placeholders: Mapping[str, list[PlaceholderFeaturesInfo]],
@@ -1955,17 +1973,12 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
             placeholders = mm_placeholders.get(modality, [])
 
             if len(placeholders) != item_count:
-                # NOTE: If you are a model developer, this can also arise from
-                # an inconsistency between `_call_hf_processor` and
-                # `_get_mm_fields_config` implementations
                 raise RuntimeError(
-                    f"Expected there to be {item_count} prompt updates "
+                    f"Expected there to be {item_count} prompt placeholders "
                     f"corresponding to {item_count} {modality} items, but "
-                    f"instead found {len(placeholders)} prompt updates! "
-                    "This is likely because you forgot to include input "
-                    "placeholder tokens (e.g., `<image>`, `<|image_pad|>`) "
-                    "in the prompt. If the model has a chat template, make "
-                    "sure you have applied it before calling `LLM.generate`.")
+                    f"instead found {len(placeholders)} prompt placeholders! "
+                    "Make sure the implementation of `_call_hf_processor` and "
+                    "`_get_mm_fields_config` are consistent with each other.")
 
     def _maybe_apply_prompt_updates(
         self,
@@ -1977,6 +1990,7 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
     ) -> tuple[list[int], Mapping[str, list[PlaceholderFeaturesInfo]]]:
         mm_item_counts = mm_items.get_all_counts()
         self._validate_mm_kwargs(mm_kwargs, mm_item_counts)
+        self._validate_mm_updates(mm_prompt_updates, mm_item_counts)
 
         if is_update_applied:
             mm_placeholders = self._find_mm_placeholders(