[Misc][V1] Avoid using envs.VLLM_USE_V1 in mm processing

vllm/inputs/preprocess.py

@@ -254,6 +254,7 @@ def _process_multimodal(
         mm_data: MultiModalDataDict,
         mm_processor_kwargs: Optional[Mapping[str, object]],
         lora_request: Optional[LoRARequest],
+        return_mm_hashes: bool = False,
     ) -> MultiModalInputs:
         """
         Apply the model's multi-modal processor to a multi-modal prompt,
@@ -274,14 +275,16 @@ def _process_multimodal(
         if mm_processor_kwargs is None:
             mm_processor_kwargs = {}
 
-        return mm_processor.apply(prompt, mm_data, mm_processor_kwargs)
+        return mm_processor.apply(prompt, mm_data, mm_processor_kwargs,
+                                  return_mm_hashes)
 
     async def _process_multimodal_async(
         self,
         prompt: Union[str, List[int]],
         mm_data: MultiModalDataDict,
         mm_processor_kwargs: Optional[Mapping[str, object]],
         lora_request: Optional[LoRARequest],
+        return_mm_hashes: bool = False,
     ) -> MultiModalInputs:
         """Async version of :meth:`_process_multimodal`."""
         # At the moment on model (PrithviGeoSpatialMAE) requires to be
@@ -299,13 +302,15 @@ async def _process_multimodal_async(
         if mm_processor_kwargs is None:
             mm_processor_kwargs = {}
 
-        return mm_processor.apply(prompt, mm_data, mm_processor_kwargs)
+        return mm_processor.apply(prompt, mm_data, mm_processor_kwargs,
+                                  return_mm_hashes)
 
     def _prompt_to_llm_inputs(
         self,
         prompt: SingletonPrompt,
         request_id: str,
         lora_request: Optional[LoRARequest] = None,
+        return_mm_hashes: bool = False,
     ) -> SingletonInputs:
         """
         Extract the singleton inputs from a prompt.
@@ -315,6 +320,7 @@ def _prompt_to_llm_inputs(
         * request_id
         * prompt: single encoder or decoder input prompt
         * lora_request: this is only valid for decoder prompts
+        * return_mm_hashes: whether to return multimodal hashes
 
         Returns:
 
@@ -349,6 +355,7 @@ def _prompt_to_llm_inputs(
                     multi_modal_data,
                     mm_processor_kwargs,
                     lora_request=lora_request,
+                    return_mm_hashes=return_mm_hashes,
                 )
 
             return token_inputs(
@@ -695,6 +702,7 @@ def _process_decoder_only_prompt(
         request_id: str,
         lora_request: Optional[LoRARequest] = None,
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
+        return_mm_hashes: bool = False,
     ) -> DecoderOnlyInputs:
         """
         For decoder-only models:
@@ -706,6 +714,7 @@ def _process_decoder_only_prompt(
         * request_id
         * lora_request
         * prompt_adapter_request
+        * return_mm_hashes
 
         Returns:
 
@@ -729,6 +738,7 @@ async def _process_decoder_only_prompt_async(
         request_id: str,
         lora_request: Optional[LoRARequest] = None,
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
+        return_mm_hashes: bool = False,
     ) -> DecoderOnlyInputs:
         """Async version of :meth:`_process_decoder_only_prompt`."""
         prompt_comps = await self._prompt_to_llm_inputs_async(
@@ -748,9 +758,13 @@ def preprocess(
         request_id: str,
         lora_request: Optional[LoRARequest] = None,
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
+        return_mm_hashes: bool = False,
     ) -> ProcessorInputs:
         """Preprocess the input prompt."""
         if self.model_config.is_encoder_decoder:
+            assert not return_mm_hashes, (
+                "Multimodal hashes for encoder-decoder models should not be ",
+                "returned until they are supported on vLLM V1.")
             # Encoder-decoder model requires special mapping of
             # input prompts to encoder & decoder
             return self._process_encoder_decoder_prompt(
@@ -768,6 +782,7 @@ def preprocess(
             request_id=request_id,
             lora_request=lora_request,
             prompt_adapter_request=prompt_adapter_request,
+            return_mm_hashes=return_mm_hashes,
         )
 
     async def preprocess_async(
@@ -776,9 +791,13 @@ async def preprocess_async(
         request_id: str,
         lora_request: Optional[LoRARequest] = None,
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
+        return_mm_hashes: bool = False,
     ) -> ProcessorInputs:
         """Async version of :meth:`preprocess`."""
         if self.model_config.is_encoder_decoder:
+            assert not return_mm_hashes, (
+                "Multimodal hashes for encoder-decoder models should not be ",
+                "returned until they are supported on vLLM V1.")
             # Encoder-decoder model requires special mapping of
             # input prompts to encoder & decoder
             return await self._process_encoder_decoder_prompt_async(
@@ -796,4 +815,5 @@ async def preprocess_async(
             request_id=request_id,
             lora_request=lora_request,
             prompt_adapter_request=prompt_adapter_request,
+            return_mm_hashes=return_mm_hashes,
         )

vllm/model_executor/models/llava.py

@@ -767,6 +767,7 @@ def apply(
         prompt: Union[str, list[int]],
         mm_data: MultiModalDataDict,
         hf_processor_mm_kwargs: Mapping[str, object],
+        return_mm_hashes: bool = False,
     ) -> MultiModalInputs:
         hf_config = self.info.get_hf_config()
         image_token_id = hf_config.image_token_index
@@ -777,7 +778,8 @@ def apply(
             image_height=-1,
         )
 
-        result = super().apply(prompt, mm_data, hf_processor_mm_kwargs)
+        result = super().apply(prompt, mm_data, hf_processor_mm_kwargs,
+                               return_mm_hashes)
 
         mm_items = self._to_mm_items(mm_data)
         mm_item_counts = mm_items.get_all_counts()

vllm/model_executor/models/minicpmv.py

@@ -780,6 +780,7 @@ def apply(
         prompt: Union[str, List[int]],
         mm_data: MultiModalDataDict,
         hf_processor_mm_kwargs: Mapping[str, object],
+        return_mm_hashes: bool = False,
     ) -> MultiModalInputs:
         supported_mm_modalities = self.info.get_supported_mm_modalities()
         if isinstance(prompt, list):
@@ -791,7 +792,8 @@ def apply(
                 [index for index, m in enumerate(matches) if m == modality])
             for modality in supported_mm_modalities
         }
-        result = super().apply(prompt, mm_data, hf_processor_mm_kwargs)
+        result = super().apply(prompt, mm_data, hf_processor_mm_kwargs,
+                               return_mm_hashes)
         # Exclude <image_id>x</image_id> from placeholders
         if "image" in result["mm_placeholders"] and \
             self.info.get_model_version() == (2, 6):

vllm/model_executor/models/mllama.py

@@ -175,8 +175,10 @@ def apply(
         prompt: Union[str, list[int]],
         mm_data: MultiModalDataDict,
         hf_processor_mm_kwargs: Mapping[str, object],
+        return_mm_hashes: bool = False,
     ) -> MultiModalEncDecInputs:
-        mm_inputs = super().apply(prompt, mm_data, hf_processor_mm_kwargs)
+        mm_inputs = super().apply(prompt, mm_data, hf_processor_mm_kwargs,
+                                  return_mm_hashes)
 
         # Check that the number of image tokens in the decoder prompt matches
         # the number of images provided in mm_data

vllm/model_executor/models/prithvi_geospatial_mae.py

@@ -93,6 +93,7 @@ def apply(
         prompt: Union[str, list[int]],
         mm_data: MultiModalDataDict,
         hf_processor_mm_kwargs: Mapping[str, object],
+        return_mm_hashes: bool = False,
     ) -> MultiModalInputs:
         mm_kwargs = {}
 

vllm/multimodal/processing.py

@@ -14,7 +14,6 @@
 from transformers import BatchFeature, PretrainedConfig, ProcessorMixin
 from typing_extensions import assert_never
 
-import vllm.envs as envs
 from vllm.inputs import InputProcessingContext
 from vllm.logger import init_logger
 from vllm.transformers_utils.tokenizer import (AnyTokenizer, decode_tokens,
@@ -1435,6 +1434,7 @@ def apply(
         prompt: Union[str, list[int]],
         mm_data: MultiModalDataDict,
         hf_processor_mm_kwargs: Mapping[str, object],
+        return_mm_hashes: bool = False,
     ) -> MultiModalInputs:
         """
         Process multi-modal inputs to be used in vLLM.
@@ -1451,11 +1451,11 @@ def apply(
         """
         mm_items = self._to_mm_items(mm_data)
 
-        # Create MM hashes (only used in V1)
+        # Create MM hashes to be returned (only used in V1)
         # TODO: Use these hash keys for caching operations in apply_hf_processor
         # instead of rehashing.
 
-        if envs.VLLM_USE_V1:
+        if return_mm_hashes:
             model_id = self.info.model_id
             mm_hashes = {
                 modality: [
@@ -1554,6 +1554,7 @@ def apply(
         prompt: Union[str, list[int]],
         mm_data: MultiModalDataDict,
         hf_processor_mm_kwargs: Mapping[str, object],
+        return_mm_hashes: bool = False,
     ) -> MultiModalEncDecInputs:
         """
         Process multi-modal inputs to be used in vLLM.
@@ -1567,6 +1568,7 @@ def apply(
             encoder_prompt,
             mm_data,
             hf_processor_mm_kwargs,
+            return_mm_hashes,
         )
 
         tokenizer = self.info.get_tokenizer()

vllm/v1/engine/processor.py

@@ -131,6 +131,7 @@ def process_inputs(
             request_id=request_id,
             lora_request=lora_request,
             prompt_adapter_request=prompt_adapter_request,
+            return_mm_hashes=self.use_hash,
         )
         eos_token_id = self.input_preprocessor.get_eos_token_id(lora_request)