[Perf] Isolate MM preprocessing on its own executor (#49524)

Signed-off-by: Guan-Ming (Wesley) Chiu <105915352+guan404ming@users.noreply.github.com>

tests/test_config.py

@@ -1558,20 +1558,31 @@ def test_fault_tolerance_requires_single_api_server():
 
 
 def test_renderer_num_workers_with_mm_cache():
-    """Disallow renderer_num_workers > 1 when mm processor cache is enabled,
-    since neither cache type is thread-safe."""
+    """Disallow renderer_num_workers > 1 with the mm processor cache only for
+    pooling models, whose preprocessing runs on the renderer workers."""
     mm_model = "Qwen/Qwen2-VL-2B-Instruct"
 
-    # Should raise: multi-worker + cache enabled (default cache_gb=4)
+    # Should raise: pooling + multi-worker + cache enabled (default cache_gb=4)
     with pytest.raises(ValueError, match="renderer-num-workers"):
-        ModelConfig(mm_model, renderer_num_workers=4)
+        ModelConfig(mm_model, runner="pooling", renderer_num_workers=4)
 
-    # Should raise: multi-worker + explicit cache size
+    # Should raise: pooling + multi-worker + explicit cache size
     with pytest.raises(ValueError, match="renderer-num-workers"):
-        ModelConfig(mm_model, renderer_num_workers=2, mm_processor_cache_gb=1.0)
+        ModelConfig(
+            mm_model,
+            runner="pooling",
+            renderer_num_workers=2,
+            mm_processor_cache_gb=1.0,
+        )
 
-    # Should pass: multi-worker + cache disabled
-    config = ModelConfig(mm_model, renderer_num_workers=4, mm_processor_cache_gb=0)
+    # Should pass: pooling + multi-worker + cache disabled
+    config = ModelConfig(
+        mm_model, runner="pooling", renderer_num_workers=4, mm_processor_cache_gb=0
+    )
+    assert config.renderer_num_workers == 4
+
+    # Should pass: generate models preprocess on the dedicated mm executor
+    config = ModelConfig(mm_model, renderer_num_workers=4)
     assert config.renderer_num_workers == 4
 
     # Should pass: single worker + cache enabled (default)

vllm/config/model.py

@@ -748,12 +748,13 @@ class ModelConfig:
             if (
                 self.renderer_num_workers > 1
                 and self.multimodal_config.mm_processor_cache_gb > 0
+                and self.runner_type == "pooling"
             ):
                 raise ValueError(
                     "Cannot use --renderer-num-workers > 1 with the "
-                    "multimodal processor cache enabled. The cache is "
-                    "not thread-safe and does not support concurrent "
-                    "renderer workers. Please set "
+                    "multimodal processor cache enabled for pooling models. "
+                    "Pooling preprocessing runs on the renderer workers, and "
+                    "the cache is not thread-safe. Please set "
                     "--renderer-num-workers 1 (the default), or "
                     "disable the cache with --mm-processor-cache-gb 0."
                 )

vllm/renderers/base.py

@@ -79,16 +79,15 @@ class BaseRenderer(ABC, Generic[_T]):
 
         self.tokenizer = tokenizer
 
-        # Shared thread pool executor for blocking tokenizer and
-        # multimodal preprocessing operations.  The multimodal processor
-        # receives a deep-copied tokenizer (see #36557) so it is safe to
-        # run tokenization and MM preprocessing concurrently.
+        # Thread pool executor for blocking tokenizer operations.  The
+        # multimodal processor receives a deep-copied tokenizer (see #36557)
+        # so it is safe to run tokenization and MM preprocessing concurrently.
         pool_workers = config.model_config.renderer_num_workers
         self._executor = ThreadPoolExecutor(max_workers=pool_workers)
 
-        # Multimodal preprocessing is always offloaded to the thread pool
-        # to keep the asyncio event loop responsive under concurrent load.
-        self._mm_executor: Executor = self._executor
+        # Separate single-worker executor so tokenization never queues behind
+        # MM preprocessing; must stay single-worker per #38418 (P0/P1 order).
+        self._mm_executor: Executor = ThreadPoolExecutor(max_workers=1)
 
         # Offload tokenization to the thread pool. The sync
         # ``_tokenize_prompt`` already encapsulates the unified ``__call__``
@@ -103,7 +102,7 @@ class BaseRenderer(ABC, Generic[_T]):
         self._readonly_mm_processor: BaseMultiModalProcessor | None = None
         self._mm_cache_stats: MultiModalCacheStats | None = None
         self._clear_mm_cache_async = make_async(
-            self.clear_mm_cache, executor=self._executor
+            self.clear_mm_cache, executor=self._mm_executor
         )
         self._process_multimodal_async = make_async(
             self._process_multimodal, executor=self._mm_executor
@@ -282,7 +281,7 @@ class BaseRenderer(ABC, Generic[_T]):
                 self._clear_processor_cache(self._readonly_mm_processor)
 
     async def clear_mm_cache_async(self) -> None:
-        """Serialize clear_mm_cache through the shared executor to avoid
+        """Serialize clear_mm_cache through the multimodal executor to avoid
         races with concurrent process_inputs on the mm_processor_cache."""
         await self._clear_mm_cache_async()