[bugfix] add api process rank in default multimodal request (#36150)

Signed-off-by: fake0fan <645327136@qq.com>
Signed-off-by: Chenguang ZHENG <645327136@qq.com>

tests/entrypoints/openai/test_chat_error.py

@@ -59,16 +59,22 @@ class MockModelConfig:
         return self.diff_sampling_param or {}
 
 
+@dataclass
+class MockParallelConfig:
+    _api_process_rank: int = 0
+
+
 @dataclass
 class MockVllmConfig:
     model_config: MockModelConfig
+    parallel_config: MockParallelConfig
 
 
 def _build_renderer(model_config: MockModelConfig):
     _, tokenizer_name, _, kwargs = tokenizer_args_from_config(model_config)
 
     return HfRenderer.from_config(
-        MockVllmConfig(model_config),
+        MockVllmConfig(model_config, parallel_config=MockParallelConfig()),
         tokenizer_kwargs={**kwargs, "tokenizer_name": tokenizer_name},
     )
 

tests/entrypoints/openai/test_completion_error.py

@@ -58,9 +58,15 @@ class MockModelConfig:
         return self.diff_sampling_param or {}
 
 
+@dataclass
+class MockParallelConfig:
+    _api_process_rank: int = 0
+
+
 @dataclass
 class MockVllmConfig:
     model_config: MockModelConfig
+    parallel_config: MockParallelConfig
 
 
 def _build_serving_completion(engine: AsyncLLM) -> OpenAIServingCompletion:
@@ -79,7 +85,7 @@ def _build_renderer(model_config: MockModelConfig):
     _, tokenizer_name, _, kwargs = tokenizer_args_from_config(model_config)
 
     return HfRenderer.from_config(
-        MockVllmConfig(model_config),
+        MockVllmConfig(model_config, parallel_config=MockParallelConfig()),
         tokenizer_kwargs={**kwargs, "tokenizer_name": tokenizer_name},
     )
 

tests/entrypoints/openai/test_lora_resolvers.py

@@ -58,9 +58,15 @@ class MockModelConfig:
         return self.diff_sampling_param or {}
 
 
+@dataclass
+class MockParallelConfig:
+    _api_process_rank: int = 0
+
+
 @dataclass
 class MockVllmConfig:
     model_config: MockModelConfig
+    parallel_config: MockParallelConfig
 
 
 class MockLoRAResolver(LoRAResolver):
@@ -97,7 +103,7 @@ def _build_renderer(model_config: MockModelConfig):
     _, tokenizer_name, _, kwargs = tokenizer_args_from_config(model_config)
 
     return HfRenderer.from_config(
-        MockVllmConfig(model_config),
+        MockVllmConfig(model_config, parallel_config=MockParallelConfig()),
         tokenizer_kwargs={**kwargs, "tokenizer_name": tokenizer_name},
     )
 

tests/entrypoints/openai/test_serving_chat.py

@@ -537,16 +537,22 @@ class MockModelConfig:
         return self.diff_sampling_param or {}
 
 
+@dataclass
+class MockParallelConfig:
+    _api_process_rank: int = 0
+
+
 @dataclass
 class MockVllmConfig:
     model_config: MockModelConfig
+    parallel_config: MockParallelConfig
 
 
 def _build_renderer(model_config: MockModelConfig):
     _, tokenizer_name, _, kwargs = tokenizer_args_from_config(model_config)
 
     return HfRenderer.from_config(
-        MockVllmConfig(model_config),
+        MockVllmConfig(model_config, parallel_config=MockParallelConfig()),
         tokenizer_kwargs={**kwargs, "tokenizer_name": tokenizer_name},
     )
 
@@ -797,7 +803,7 @@ async def test_serving_chat_mistral_token_ids_prompt_is_validated():
 
     mock_tokenizer = MagicMock(spec=MistralTokenizer)
     mock_renderer = MistralRenderer(
-        MockVllmConfig(mock_engine.model_config),
+        MockVllmConfig(mock_engine.model_config, parallel_config=MockParallelConfig()),
         tokenizer=mock_tokenizer,
     )
     # Force the Mistral chat template renderer to return token IDs.
@@ -837,7 +843,7 @@ async def test_serving_chat_mistral_token_ids_prompt_too_long_is_rejected():
 
     mock_tokenizer = MagicMock(spec=MistralTokenizer)
     mock_renderer = MistralRenderer(
-        MockVllmConfig(mock_engine.model_config),
+        MockVllmConfig(mock_engine.model_config, parallel_config=MockParallelConfig()),
         tokenizer=mock_tokenizer,
     )
     # prompt_token_ids length == max_model_len should be rejected for

tests/renderers/test_completions.py

@@ -41,9 +41,15 @@ class MockModelConfig:
     is_multimodal_model: bool = False
 
 
+@dataclass
+class MockParallelConfig:
+    _api_process_rank: int = 0
+
+
 @dataclass
 class MockVllmConfig:
     model_config: MockModelConfig
+    parallel_config: MockParallelConfig
 
 
 @dataclass
@@ -78,7 +84,7 @@ def _build_renderer(
     _, tokenizer_name, _, kwargs = tokenizer_args_from_config(model_config)
 
     renderer = HfRenderer(
-        MockVllmConfig(model_config),
+        MockVllmConfig(model_config, parallel_config=MockParallelConfig()),
         tokenizer=(
             None
             if model_config.skip_tokenizer_init

tests/renderers/test_mistral.py

@@ -39,9 +39,15 @@ class MockModelConfig:
     is_multimodal_model: bool = False
 
 
+@dataclass
+class MockParallelConfig:
+    _api_process_rank: int = 0
+
+
 @dataclass
 class MockVllmConfig:
     model_config: MockModelConfig
+    parallel_config: MockParallelConfig
 
 
 @pytest.mark.asyncio
@@ -57,7 +63,7 @@ async def test_async_mistral_tokenizer_does_not_block_event_loop():
     mock_tokenizer = Mock(spec=MistralTokenizer)
     mock_tokenizer.apply_chat_template = mocked_apply_chat_template
     mock_renderer = MistralRenderer(
-        MockVllmConfig(mock_model_config),
+        MockVllmConfig(mock_model_config, parallel_config=MockParallelConfig()),
         tokenizer=mock_tokenizer,
     )
 

vllm/renderers/base.py

@@ -75,6 +75,7 @@ class BaseRenderer(ABC, Generic[_T]):
 
         self.config = config
         self.model_config = config.model_config
+        self.api_process_rank = config.parallel_config._api_process_rank
 
         self.tokenizer = tokenizer
 
@@ -539,7 +540,7 @@ class BaseRenderer(ABC, Generic[_T]):
         from vllm.multimodal.parse import parse_mm_uuids
         from vllm.multimodal.processing import ProcessorInputs as MMProcessorInputs
 
-        mm_req_id = f"renderer-mm-{self._mm_req_counter.inc(1)}"
+        mm_req_id = f"renderer{self.api_process_rank}-mm-{self._mm_req_counter.inc(1)}"
 
         mm_processor = self.get_mm_processor()