rebase 2025/02/16

Qwen3 Omni offline/online passed

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

docker/Dockerfile.ci

@@ -14,11 +14,11 @@ RUN apt-get update && \
 
 # Install vLLM wheel for the specific commit
 # URL format: https://wheels.vllm.ai/<commit_hash>/
-# Current vLLM wheel commit: 2d5be1dd5ce2e44dfea53ea03ff61143da5137eb
+# Current vLLM wheel commit: d5fe3f702c2f4392515cdfde6fd0442271c74dda
 # Note: --prerelease=allow is required because nightly/commit wheels use
 # pre-release version numbers (e.g. 2.10.0.dev...) which uv ignores by default,
 # causing it to fall back to the older stable release from PyPI.
-RUN uv pip install --system --upgrade --force-reinstall --prerelease=allow vllm --extra-index-url https://wheels.vllm.ai/2d5be1dd5ce2e44dfea53ea03ff61143da5137eb
+RUN uv pip install --system --upgrade --force-reinstall --prerelease=allow vllm --extra-index-url https://wheels.vllm.ai/d5fe3f702c2f4392515cdfde6fd0442271c74dda
 
 
 RUN uv pip uninstall --system nvidia-cublas-cu12

vllm_omni/config/model.py

@@ -97,6 +97,7 @@ class OmniModelConfig(ModelConfig):
     def __post_init__(
         self,
         # Multimodal config init vars
+        language_model_only: bool,
         limit_mm_per_prompt: dict[str, int | dict[str, int]] | None,
         enable_mm_embeds: bool | None,
         media_io_kwargs: dict[str, dict[str, Any]] | None,
@@ -113,6 +114,7 @@ class OmniModelConfig(ModelConfig):
     ) -> None:
         # Call parent's __post_init__ to handle all standard ModelConfig initialization
         super().__post_init__(
+            language_model_only=language_model_only,
             limit_mm_per_prompt=limit_mm_per_prompt,
             enable_mm_embeds=enable_mm_embeds,
             media_io_kwargs=media_io_kwargs,

vllm_omni/engine/arg_utils.py

@@ -110,6 +110,7 @@ class OmniEngineArgs(EngineArgs):
 
         # Keep compatibility when async args are constructed from partial payloads.
         limit_mm_per_prompt = getattr(self, "limit_mm_per_prompt", {})
+        language_model_only = getattr(self, "language_model_only", False)
         enable_mm_embeds = getattr(self, "enable_mm_embeds", False)
         interleave_mm_strings = getattr(self, "interleave_mm_strings", False)
         media_io_kwargs = getattr(self, "media_io_kwargs", {})
@@ -163,6 +164,7 @@ class OmniEngineArgs(EngineArgs):
             skip_tokenizer_init=self.skip_tokenizer_init,
             enable_prompt_embeds=self.enable_prompt_embeds,
             served_model_name=self.served_model_name,
+            language_model_only=language_model_only,
             limit_mm_per_prompt=limit_mm_per_prompt,
             enable_mm_embeds=enable_mm_embeds,
             interleave_mm_strings=interleave_mm_strings,
@@ -177,7 +179,6 @@ class OmniEngineArgs(EngineArgs):
             mm_encoder_tp_mode=mm_encoder_tp_mode,
             mm_encoder_attn_backend=mm_encoder_attn_backend,
             pooler_config=self.pooler_config,
-            logits_processor_pattern=self.logits_processor_pattern,
             generation_config=self.generation_config,
             override_generation_config=self.override_generation_config,
             enable_sleep_mode=self.enable_sleep_mode,
@@ -266,6 +267,7 @@ class AsyncOmniEngineArgs(AsyncEngineArgs):
 
         # Keep compatibility when async args are constructed from partial payloads.
         limit_mm_per_prompt = getattr(self, "limit_mm_per_prompt", {})
+        language_model_only = getattr(self, "language_model_only", False)
         enable_mm_embeds = getattr(self, "enable_mm_embeds", False)
         interleave_mm_strings = getattr(self, "interleave_mm_strings", False)
         media_io_kwargs = getattr(self, "media_io_kwargs", {})
@@ -319,6 +321,7 @@ class AsyncOmniEngineArgs(AsyncEngineArgs):
             skip_tokenizer_init=self.skip_tokenizer_init,
             enable_prompt_embeds=self.enable_prompt_embeds,
             served_model_name=self.served_model_name,
+            language_model_only=language_model_only,
             limit_mm_per_prompt=limit_mm_per_prompt,
             enable_mm_embeds=enable_mm_embeds,
             interleave_mm_strings=interleave_mm_strings,
@@ -333,7 +336,6 @@ class AsyncOmniEngineArgs(AsyncEngineArgs):
             mm_encoder_tp_mode=mm_encoder_tp_mode,
             mm_encoder_attn_backend=mm_encoder_attn_backend,
             pooler_config=self.pooler_config,
-            logits_processor_pattern=self.logits_processor_pattern,
             generation_config=self.generation_config,
             override_generation_config=self.override_generation_config,
             enable_sleep_mode=self.enable_sleep_mode,

vllm_omni/engine/input_processor.py

@@ -90,12 +90,11 @@ class OmniInputProcessor(InputProcessor):
         vllm_config: VllmConfig,
         mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
     ):
-        super().__init__(vllm_config, mm_registry)
+        super().__init__(vllm_config, mm_registry=mm_registry)
         self.input_preprocessor = OmniInputPreprocessor(
-            self.model_config,
-            vllm_config.observability_config,
-            mm_registry,
-            mm_processor_cache=self.mm_processor_cache,
+            vllm_config=vllm_config,
+            renderer=self.renderer,
+            mm_registry=mm_registry,
         )
 
     def process_inputs(
@@ -194,7 +193,7 @@ class OmniInputProcessor(InputProcessor):
             processed_inputs=processed_inputs,
         )
 
-        eos_token_id = self.input_preprocessor.get_eos_token_id()
+        eos_token_id = self.renderer.get_eos_token_id()
 
         encoder_inputs, decoder_inputs = split_enc_dec_inputs(processed_inputs)
         self._validate_model_inputs(encoder_inputs, decoder_inputs)
@@ -281,7 +280,6 @@ class OmniInputProcessor(InputProcessor):
             mm_features=mm_features,
             sampling_params=sampling_params,
             pooling_params=pooling_params,
-            eos_token_id=eos_token_id,
             arrival_time=arrival_time,
             lora_request=lora_request,
             cache_salt=decoder_inputs.get("cache_salt"),

vllm_omni/entrypoints/async_omni_llm.py

@@ -9,7 +9,7 @@ import torch
 from vllm.config import VllmConfig
 from vllm.logger import init_logger
 from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry
-from vllm.tokenizers import cached_tokenizer_from_config
+from vllm.renderers import renderer_from_config
 from vllm.tracing import init_tracer
 from vllm.transformers_utils.config import maybe_register_config_serialize_by_value
 from vllm.usage.usage_lib import UsageContext
@@ -106,21 +106,17 @@ class AsyncOmniLLM(AsyncLLM):
                 "enabling logging without default stat loggers"
             )
 
-        if self.model_config.skip_tokenizer_init:
-            tokenizer = None
-        else:
-            # Tokenizer (+ ensure liveness if running in another process).
-            tokenizer = cached_tokenizer_from_config(model_config=vllm_config.model_config)
-
         # InputProcessor (converts Inputs --> EngineCoreRequests).
         self.input_processor = OmniInputProcessor(
             vllm_config=vllm_config,
             mm_registry=mm_registry,
         )
 
+        self.renderer = renderer_from_config(self.vllm_config)
+
         # OutputProcessor (converts EngineCoreOutputs --> RequestOutput).
         self.output_processor = MultimodalOutputProcessor(
-            tokenizer=tokenizer,
+            tokenizer=self.renderer.tokenizer,
             log_stats=self.log_stats,
             engine_core_output_type=engine_args.engine_output_type,
         )

vllm_omni/entrypoints/omni_llm.py

@@ -173,6 +173,9 @@ class OmniLLM(LLM):
 
         self.supported_tasks = supported_tasks
 
+        # Keep parity with vLLM's LLM initialization fields so inherited
+        # generate/chat preprocessing paths work as expected.
+        self.renderer = self.llm_engine.renderer
         # Load the Input/Output processor plugin if any
         io_processor_plugin = self.llm_engine.model_config.io_processor_plugin
         self.io_processor = get_io_processor(self.llm_engine.vllm_config, io_processor_plugin)

vllm_omni/entrypoints/omni_stage.py

@@ -67,20 +67,16 @@ def _sequential_init_lock(engine_args: dict[str, Any], stage_init_timeout: int =
     """
     from vllm_omni.worker.gpu_memory_utils import is_process_scoped_memory_available
 
-    nvml_available = is_process_scoped_memory_available()
-    pid_host = detect_pid_host()
-
-    if nvml_available and pid_host:
-        logger.info(
+    if is_process_scoped_memory_available() and detect_pid_host():
+        logger.debug(
             "NVML process-scoped memory available and PID host is available — concurrent init is safe, skipping locks"
         )
         yield
         return
     else:
-        logger.info(
-            "Using sequential init locks (nvml_available=%s, pid_host=%s)",
-            nvml_available,
-            pid_host,
+        logger.debug(
+            "NVML unavailable or PID host is not available (usually inside a container, "
+            "--pid=host is not set in docker run command) — using sequential init locks"
         )
 
     from vllm_omni.platforms import current_omni_platform
@@ -205,7 +201,7 @@ def _sequential_init_lock(engine_args: dict[str, Any], stage_init_timeout: int =
 
 
 def _resolve_worker_cls(engine_args: dict[str, Any]) -> None:
-    worker_type = engine_args.get("worker_type", None)
+    worker_type = engine_args.pop("worker_type", None)
     if not worker_type:
         return
     if engine_args.get("worker_cls"):
@@ -1127,7 +1123,7 @@ async def _stage_worker_async(
                 usage_context=usage_context,
                 engine_args=omni_engine_args,
                 disable_log_stats=bool(
-                    engine_args.get("disable_log_stats", True) or getattr(omni_engine_args, "disable_log_stats", True)
+                    engine_args.get("disable_log_stats", False) or getattr(omni_engine_args, "disable_log_stats", False)
                 ),
             )
     omni_stage.set_async_engine(stage_engine)

vllm_omni/entrypoints/openai/serving_chat.py

@@ -13,7 +13,7 @@ import torch
 from fastapi import Request
 from PIL import Image
 from pydantic import TypeAdapter
-from vllm.renderers.protocol import BaseRenderer
+from vllm.renderers import BaseRenderer
 
 from vllm_omni.entrypoints.async_omni import AsyncOmni
 from vllm_omni.entrypoints.openai.protocol.chat_completion import OmniChatCompletionResponse

vllm_omni/request.py

@@ -82,7 +82,6 @@ class OmniRequest(Request):
             mm_features=request.mm_features,
             sampling_params=request.sampling_params,
             pooling_params=request.pooling_params,
-            eos_token_id=request.eos_token_id,
             arrival_time=request.arrival_time,
             lora_request=request.lora_request,
             cache_salt=request.cache_salt,