Skip mm feature pool init to avoid EPD OOM (#16388)

python/sglang/srt/disaggregation/encode_receiver.py

@@ -303,7 +303,11 @@ class MMReceiver:
                     else:
                         raise e
                 self.mm_processor = get_mm_processor(
-                    hf_config, server_args, _processor, transport_mode
+                    hf_config,
+                    server_args,
+                    _processor,
+                    transport_mode,
+                    skip_mm_pool=True,
                 )
 
     def create_req(self, recv_req):

python/sglang/srt/managers/multimodal_processor.py

@@ -41,11 +41,13 @@ def import_processors(package_name: str, overwrite: bool = False):
 
 
 def get_mm_processor(
-    hf_config, server_args: ServerArgs, processor, transport_mode
+    hf_config, server_args: ServerArgs, processor, transport_mode, **kwargs
 ) -> BaseMultimodalProcessor:
     for model_cls, processor_cls in PROCESSOR_MAPPING.items():
         if model_cls.__name__ in hf_config.architectures:
-            return processor_cls(hf_config, server_args, processor, transport_mode)
+            return processor_cls(
+                hf_config, server_args, processor, transport_mode, **kwargs
+            )
 
     raise ValueError(
         f"No processor registered for architecture: {hf_config.architectures}.\n"

python/sglang/srt/multimodal/processors/base_processor.py

@@ -234,7 +234,9 @@ class BaseMultimodalProcessor(ABC):
             "input_features",
         ]
 
-        if SGL_USE_CUDA_IPC:
+        skip_mm_pool = kwargs.get("skip_mm_pool", False)
+
+        if SGL_USE_CUDA_IPC and not skip_mm_pool:
             self.cudaipc_mmfeature_pool = MmItemMemoryPool(
                 MM_FEATURE_CACHE_SIZE,
                 MM_ITEM_MEMORY_POOL_RECYCLE_INTERVAL,