[V1] Clarify input processing and multimodal feature caching logic (#13211)

vllm/v1/engine/core.py

@@ -20,7 +20,7 @@ from vllm.v1.core.kv_cache_utils import get_kv_cache_configs
 from vllm.v1.core.scheduler import Scheduler
 from vllm.v1.engine import (EngineCoreOutputs, EngineCoreRequest,
                             EngineCoreRequestType)
-from vllm.v1.engine.mm_input_mapper import MMInputMapperServer
+from vllm.v1.engine.mm_input_cache import MMInputCacheServer
 from vllm.v1.executor.abstract import Executor
 from vllm.v1.request import Request, RequestStatus
 from vllm.v1.serial_utils import MsgpackDecoder, MsgpackEncoder
@@ -65,7 +65,7 @@ class EngineCore:
             log_stats=self.log_stats,
         )
 
-        self.mm_input_mapper_server = MMInputMapperServer(
+        self.mm_input_cache_server = MMInputCacheServer(
             vllm_config.model_config)
 
     def _initialize_kv_caches(self,
@@ -102,13 +102,13 @@ class EngineCore:
         """Add request to the scheduler."""
 
         if request.mm_hashes is not None:
-            # Here, if hash exists for an image, then it will be fetched
-            # from the cache, else it will be added to the cache.
-            # Note that the cache here is mirrored with the client side of the
-            # MM mapper, so anything that has a hash must have a HIT cache
-            # entry here as well.
+            # Here, if hash exists for a multimodal input, then it will be
+            # fetched from the cache, else it will be added to the cache.
+            # Note that the cache here is mirrored with the client cache, so
+            # anything that has a hash must have a HIT cache entry here
+            # as well.
             assert request.mm_inputs is not None
-            request.mm_inputs = self.mm_input_mapper_server.process_inputs(
+            request.mm_inputs = self.mm_input_cache_server.get_and_update(
                 request.mm_inputs, request.mm_hashes)
 
         req = Request.from_engine_core_request(request)

vllm/v1/engine/mm_input_mapper.py → vllm/v1/engine/mm_input_cache.py

@@ -10,12 +10,18 @@ from vllm.utils import LRUCache
 
 logger = init_logger(__name__)
 
-# The idea of MM preprocessor caching is based on having a client and a server,
-# where the client executes in the frontend process (=P0) and the server in the
-# core process (=P1).
+# The idea of multimodal preprocessing caching is based on having a client and
+# a server, where the client executes in the frontend process (=P0) and the
+# server in the core process (=P1).
 #
-# -- Client: Executes the MM mapper and performs caching of the results.
-# -- Server: Performs caching of the results
+# -- Client:
+#  - Apply legacy input_mapper (if one exists) to generate MultiModalKwargs.
+#  - Perform caching of the generated MultiModalKwargs.
+#  - This client can be deprecated once all mutimodal models migrate to use
+#    merged preprocessor with built-in caching functionality.
+#
+# -- Server:
+#  - Perform caching of the received MultiModalKwargs.
 #
 # The caching for both client and server is mirrored/similar, and this allows us
 # to avoid the serialization of "mm_inputs" (like pixel values) between
@@ -27,7 +33,9 @@ logger = init_logger(__name__)
 MM_CACHE_SIZE = 256
 
 
-class MMInputMapperClient:
+# TODO(ywang96): Deprecate this class once all multimodal models migrate to use
+# merged preprocessor with built-in caching functionality.
+class MMInputCacheClient:
 
     def __init__(
         self,
@@ -54,7 +62,8 @@ class MMInputMapperClient:
             logger.debug("MMInputMapper: cache_hit_ratio = %.2f ",
                          self.mm_cache_hits / self.mm_cache_total)
 
-    # TODO: Support modalities beyond image.
+    # NOTE: process_inputs only supports image inputs since all multimodal
+    # models with other modalities have migrated to use merged preprocessor.
     def process_inputs(
         self,
         mm_data: MultiModalDataDict,
@@ -95,7 +104,7 @@ class MMInputMapperClient:
                     # Reuse precomputed input (for merged preprocessor)
                     mm_input = precomputed_mm_inputs[input_id]
                 else:
-                    # Apply MM mapper
+                    # Apply legacy input_mapper
                     mm_input = self.multi_modal_input_mapper(
                         {"image": [image_inputs[input_id]]},
                         mm_processor_kwargs=mm_processor_kwargs,
@@ -114,13 +123,13 @@ class MMInputMapperClient:
         return ret_inputs
 
 
-class MMInputMapperServer:
+class MMInputCacheServer:
 
     def __init__(self, model_config):
         self.use_cache = not model_config.disable_mm_preprocessor_cache
         self.mm_cache = LRUCache[str, MultiModalKwargs](MM_CACHE_SIZE)
 
-    def process_inputs(
+    def get_and_update(
         self,
         mm_inputs: List[Optional[MultiModalKwargs]],
         mm_hashes: List[str],

vllm/v1/engine/processor.py

@@ -17,7 +17,7 @@ from vllm.prompt_adapter.request import PromptAdapterRequest
 from vllm.sampling_params import SamplingParams
 from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
 from vllm.v1.engine import EngineCoreRequest
-from vllm.v1.engine.mm_input_mapper import MMInputMapperClient
+from vllm.v1.engine.mm_input_cache import MMInputCacheClient
 
 
 class Processor:
@@ -46,7 +46,7 @@ class Processor:
             model_config)
 
         # Multi-modal (huggingface) input mapper
-        self.mm_input_mapper_client = MMInputMapperClient(model_config)
+        self.mm_input_cache_client = MMInputCacheClient(model_config)
 
         # Multi-modal hasher (for images)
         self.use_hash = (not model_config.disable_mm_preprocessor_cache) or \
@@ -106,16 +106,24 @@ class Processor:
         assert priority == 0, "vLLM V1 does not support priority at the moment."
         assert trace_headers is None, "vLLM V1 does not support tracing yet."
 
-        # Process inputs.
+        # Process inputs, which includes:
+        # 1. Tokenize text prompt, with LoRA request if one exists.
+        # 2. For multimodal models with a merged preprocessor, preprocess
+        #   multimodal data and expand prompt token ids accordingly.
+        # 3. Apply prompt adapter to prompt token ids if one exists.
         preprocessed_inputs = self.input_preprocessor.preprocess(
             prompt,
             request_id=request_id,
             lora_request=lora_request,
             prompt_adapter_request=prompt_adapter_request,
         )
+        eos_token_id = self.input_preprocessor.get_eos_token_id(lora_request)
+
+        # Process prompt and prompt token ids.
+        # Only applicable to multimodal models with legacy input processor.
         processed_inputs = self.input_processor(preprocessed_inputs)
+
         self._validate_model_inputs(processed_inputs)
-        eos_token_id = self.input_preprocessor.get_eos_token_id(lora_request)
 
         if is_encoder_decoder_inputs(processed_inputs):
             decoder_inputs = SingletonInputsAdapter(
@@ -200,8 +208,8 @@ class Processor:
                     key=lambda mm_input: modality_order_dict[list(
                         mm_input.modalities)[0]])
 
-            # Apply mm input cache update (and input mapper if necessary).
-            sorted_mm_inputs = self.mm_input_mapper_client.process_inputs(
+            # Apply mm input cache update and legacy input mapper if one exists.
+            sorted_mm_inputs = self.mm_input_cache_client.process_inputs(
                 mm_data=decoder_mm_data,
                 mm_hashes=sorted_mm_hashes,
                 mm_processor_kwargs=decoder_inputs.mm_processor_kwargs,

vllm/v1/worker/gpu_model_runner.py

@@ -27,7 +27,7 @@ from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
 from vllm.v1.attention.backends.flash_attn import (FlashAttentionBackend,
                                                    FlashAttentionMetadata)
 from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
-from vllm.v1.engine.mm_input_mapper import MMInputMapperClient
+from vllm.v1.engine.mm_input_cache import MMInputCacheClient
 from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
                                         KVCacheSpec)
 from vllm.v1.outputs import LogprobsTensors, ModelRunnerOutput
@@ -95,9 +95,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         self.mm_registry = MULTIMODAL_REGISTRY
         self.uses_mrope = model_config.uses_mrope
 
-        # NOTE: Initialized input mapper is only used for processing dummy
+        # NOTE: Initialized client is only used for processing dummy
         # multimodal data into multimodal kwargs for GPU memory profiling.
-        self.mm_input_mapper_profiling = MMInputMapperClient(self.model_config)
+        # Only applicable to multimodal models with legacy input mapper.
+        self.mm_input_mapper_profiling = MMInputCacheClient(self.model_config)
         self.mm_input_mapper_profiling.use_cache = False
 
         encoder_compute_budget, encoder_cache_size = compute_encoder_budget(