[V1] Use pickle for serializing EngineCoreRequest & Add multimodal in…

…puts to EngineCoreRequest (vllm-project#10245)

Signed-off-by: Woosuk Kwon <[email protected]>

vllm/v1/engine/__init__.py

@@ -1,10 +1,11 @@
 import enum
 from dataclasses import dataclass
-from typing import List, Optional, Union
+from typing import Any, Dict, List, Optional, Union
 
 import msgspec
 
 from vllm.lora.request import LoRARequest
+from vllm.multimodal import MultiModalDataDict, MultiModalPlaceholderDict
 from vllm.sampling_params import RequestOutputKind, SamplingParams
 
 
@@ -22,7 +23,8 @@ class DetokenizerRequest:
     include_stop_str_in_output: bool
 
 
-class EngineCoreRequest(msgspec.Struct, omit_defaults=True):
+@dataclass
+class EngineCoreRequest:
 
     # NOTE: prompt and prompt_token_ids should be DecoderOnlyInput,
     # but this object is currently not playing well with msgspec
@@ -33,6 +35,9 @@ class EngineCoreRequest(msgspec.Struct, omit_defaults=True):
     # always be tokenized?
     prompt: Optional[str]
     prompt_token_ids: List[int]
+    mm_data: Optional[MultiModalDataDict]
+    mm_placeholders: Optional[MultiModalPlaceholderDict]
+    mm_processor_kwargs: Optional[Dict[str, Any]]
     sampling_params: SamplingParams
     eos_token_id: Optional[int]
     arrival_time: float

vllm/v1/engine/core.py

@@ -19,6 +19,7 @@
                             EngineCoreRequest, EngineCoreRequestType)
 from vllm.v1.executor.gpu_executor import GPUExecutor
 from vllm.v1.request import Request, RequestStatus
+from vllm.v1.serial_utils import PickleEncoder
 from vllm.version import __version__ as VLLM_VERSION
 
 logger = init_logger(__name__)
@@ -315,7 +316,7 @@ def process_input_socket(self, input_path: str):
         """Input socket IO thread."""
 
         # Msgpack serialization decoding.
-        decoder_add_req = msgpack.Decoder(EngineCoreRequest)
+        decoder_add_req = PickleEncoder()
         decoder_abort_req = msgpack.Decoder(list[str])
 
         with self.make_socket(input_path, zmq.constants.PULL) as socket:

vllm/v1/engine/core_client.py

@@ -11,6 +11,7 @@
 from vllm.v1.engine import (EngineCoreOutput, EngineCoreOutputs,
                             EngineCoreRequest, EngineCoreRequestType)
 from vllm.v1.engine.core import EngineCore, EngineCoreProc
+from vllm.v1.serial_utils import PickleEncoder
 
 logger = init_logger(__name__)
 
@@ -115,7 +116,7 @@ def __init__(
         **kwargs,
     ):
         # Serialization setup.
-        self.encoder = msgspec.msgpack.Encoder()
+        self.encoder = PickleEncoder()
         self.decoder = msgspec.msgpack.Decoder(EngineCoreOutputs)
 
         # ZMQ setup.

vllm/v1/engine/processor.py

@@ -91,7 +91,10 @@ def process_inputs(
         # Make Request for EngineCore.
         engine_core_request = EngineCoreRequest(
             request_id, processed_inputs.get("prompt"),
-            processed_inputs.get("prompt_token_ids"), sampling_params,
+            processed_inputs.get("prompt_token_ids"),
+            processed_inputs.get("multi_modal_data"),
+            processed_inputs.get("multi_modal_placeholders"),
+            processed_inputs.get("mm_processor_kwargs"), sampling_params,
             eos_token_id, arrival_time, lora_request)
 
         return detokenizer_request, engine_core_request

vllm/v1/serial_utils.py

@@ -0,0 +1,10 @@
+import pickle
+
+
+class PickleEncoder:
+
+    def encode(self, obj):
+        return pickle.dumps(obj)
+
+    def decode(self, data):
+        return pickle.loads(data)