Refactor OmniLLM and MultimodalOutputProcessor initialization

- Removed AsyncOmniLLM from the main entry points as it is not supported yet.
- Updated the constructor of MultimodalOutputProcessor to set a default value for engine_core_output_type.
- Adjusted output type handling in various methods to utilize the new default engine_core_output_type for consistency.

vllm_omni/__init__.py

@@ -14,7 +14,7 @@ __author__ = "vLLM-omni Team"
 __email__ = "hsliuustc@gmail.com"
 
 # Main entry points
-from .entrypoints.omni_llm import OmniLLM, AsyncOmniLLM
+from .entrypoints.omni_llm import OmniLLM
 from .config import (
     OmniStageConfig,
     DiTConfig,
@@ -31,7 +31,7 @@ __all__ = [
     
     # Main components
     "OmniLLM",
-    "AsyncOmniLLM",
+    #AsyncOmniLLM is not supported yet
     
     # Configuration
     "OmniStageConfig",

vllm_omni/engine/output_processor.py

@@ -177,7 +177,7 @@ class MultimodalOutputProcessor(VLLMOutputProcessor):
       produce the correct RequestOutput/PoolingRequestOutput.
     - Allow custom per-modality handlers via register_handler().
     """
-    def __init__(self, tokenizer: TokenizerGroup, log_stats: bool, engine_core_output_type: Optional[str] = None):
+    def __init__(self, tokenizer: TokenizerGroup, log_stats: bool, engine_core_output_type: Optional[str] = "text"):
         super().__init__(tokenizer=tokenizer, log_stats=log_stats)
         self.output_handlers: Dict[str, Callable[[EngineCoreOutput], None]] = {}
         self._reqid_to_mm_type: Dict[str, str] = {}
@@ -223,7 +223,7 @@ class MultimodalOutputProcessor(VLLMOutputProcessor):
     ) -> OutputProcessorOutput:
         self._reqid_to_mm_type.clear()
         for eco in engine_core_outputs:
-            mm_type = (getattr(eco, "output_type", None) or "").lower()
+            mm_type = (self.engine_core_output_type or "").lower()
             if mm_type:
                 self._reqid_to_mm_type[eco.request_id] = mm_type
             self._route_and_normalize(eco)
@@ -262,7 +262,7 @@ class MultimodalOutputProcessor(VLLMOutputProcessor):
 
             # 2.5) Accumulate multimodal tensors in RequestState
             try:
-                mm_type = (getattr(eco, "output_type", self.engine_core_output_type) or "").lower()
+                mm_type = (self.engine_core_output_type or "").lower()
                 if pooling_output is not None and isinstance(req_state, OmniRequestState):
                     req_state.add_multimodal_tensor(pooling_output, mm_type)
             except Exception:
@@ -315,7 +315,7 @@ class MultimodalOutputProcessor(VLLMOutputProcessor):
     
     # ---- routing helpers ----
     def _route_and_normalize(self, eco: EngineCoreOutput) -> None:
-        output_type = (getattr(eco, "output_type", self.engine_core_output_type) or "").lower()
+        output_type = (self.engine_core_output_type or "").lower()
 
         # Custom handler first (if registered)
         if output_type in self.output_handlers:

vllm_omni/entrypoints/omni_llm.py

@@ -138,7 +138,8 @@ class OmniStageLLM(LLM):
         self.llm_engine = LLMEngine.from_engine_args(
             engine_args=engine_args, usage_context=UsageContext.LLM_CLASS)
         self.llm_engine.output_processor = MultimodalOutputProcessor(tokenizer=self.llm_engine.tokenizer, 
-                                                                    log_stats=self.llm_engine.log_stats)
+                                                                    log_stats=self.llm_engine.log_stats,
+                                                                    engine_core_output_type=engine_args.engine_output_type)
         self.llm_engine.processor = OmniProcessor(vllm_config=self.llm_engine.vllm_config,
                                                   tokenizer=self.llm_engine.tokenizer)
         self.engine_class = type(self.llm_engine)