[Bugfix] Disable multi-modal preprocessor cache for DP (#21896)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/config.py

@@ -871,6 +871,12 @@ class ModelConfig:
 
         return None
 
+    def set_disable_mm_preprocessor_cache(self, value: bool) -> None:
+        mm_config = self.get_multimodal_config()
+
+        self.disable_mm_preprocessor_cache = value
+        mm_config.disable_mm_preprocessor_cache = value
+
     def _get_encoder_config(self):
         return get_sentence_transformer_tokenizer_config(
             self.model, self.revision)

vllm/engine/arg_utils.py

@@ -1197,6 +1197,18 @@ class EngineArgs:
             enable_multimodal_encoder_data_parallel,
         )
 
+        supports_mm_preprocessor_cache = (self.data_parallel_size == 1
+                                          or data_parallel_external_lb)
+        if (not supports_mm_preprocessor_cache
+                and model_config.is_multimodal_model
+                and not model_config.disable_mm_preprocessor_cache):
+            logger.warning(
+                "Multi-modal preprocessor cache is not compatible "
+                "with data parallelism when there does not exist a "
+                "one-to-one correspondance between API process and "
+                "EngineCore process, so the cache will be disabled.")
+            model_config.set_disable_mm_preprocessor_cache(True)
+
         speculative_config = self.create_speculative_config(
             target_model_config=model_config,
             target_parallel_config=parallel_config,

vllm/entrypoints/cli/serve.py

@@ -167,8 +167,9 @@ def run_multi_api_server(args: argparse.Namespace):
 
         if model_config.is_multimodal_model and not (
                 orig_disable_mm_preprocessor_cache):
-            logger.warning("Multi-model preprocessor cache will be disabled "
-                           "for api_server_count > 1")
+            logger.warning(
+                "Multi-modal preprocessor cache is not compatible "
+                "with api_server_count > 1, so the cache will be disabled.")
 
     executor_class = Executor.get_class(vllm_config)
     log_stats = not engine_args.disable_log_stats