[Performance] Add --enable-ep-weight-filter CLI option (#37351)

Signed-off-by: esmeetu <jasonailu87@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
(cherry picked from commit 761e0aa7a0)

vllm/config/parallel.py

@@ -138,6 +138,13 @@ class ParallelConfig:
     """Whether the deployed model is MoE (if known)."""
     enable_expert_parallel: bool = False
     """Use expert parallelism instead of tensor parallelism for MoE layers."""
+    enable_ep_weight_filter: bool = False
+    """Skip non-local expert weights during model loading when expert
+    parallelism is active.  Each rank only reads its own expert shard from
+    disk, which can drastically reduce storage I/O for MoE models with
+    per-expert weight tensors (e.g. DeepSeek, Mixtral, Kimi-K2.5).  Has no
+    effect on 3D fused-expert checkpoints (e.g. GPT-OSS) or non-MoE
+    models."""
     enable_eplb: bool = False
     """Enable expert parallelism load balancing for MoE layers."""
     eplb_config: EPLBConfig = Field(default_factory=EPLBConfig)

vllm/engine/arg_utils.py

@@ -419,6 +419,7 @@ class EngineArgs:
     data_parallel_external_lb: bool = False
     data_parallel_backend: DataParallelBackend = ParallelConfig.data_parallel_backend
     enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel
+    enable_ep_weight_filter: bool = ParallelConfig.enable_ep_weight_filter
     moe_backend: MoEBackend = KernelConfig.moe_backend
     all2all_backend: All2AllBackend = ParallelConfig.all2all_backend
     enable_elastic_ep: bool = ParallelConfig.enable_elastic_ep
@@ -901,6 +902,10 @@ class EngineArgs:
             "-ep",
             **parallel_kwargs["enable_expert_parallel"],
         )
+        parallel_group.add_argument(
+            "--enable-ep-weight-filter",
+            **parallel_kwargs["enable_ep_weight_filter"],
+        )
         parallel_group.add_argument(
             "--all2all-backend", **parallel_kwargs["all2all_backend"]
         )
@@ -1727,6 +1732,7 @@ class EngineArgs:
             data_parallel_hybrid_lb=self.data_parallel_hybrid_lb,
             is_moe_model=model_config.is_moe,
             enable_expert_parallel=self.enable_expert_parallel,
+            enable_ep_weight_filter=self.enable_ep_weight_filter,
             all2all_backend=self.all2all_backend,
             enable_elastic_ep=self.enable_elastic_ep,
             enable_dbo=self.enable_dbo,

vllm/model_executor/model_loader/default_loader.py

@@ -313,7 +313,11 @@ class DefaultModelLoader(BaseModelLoader):
         vllm_config = get_current_vllm_config()
         parallel_config = vllm_config.parallel_config
 
-        if not (model_config.is_moe and parallel_config.enable_expert_parallel):
+        if not (
+            model_config.is_moe
+            and parallel_config.enable_expert_parallel
+            and parallel_config.enable_ep_weight_filter
+        ):
             return
 
         num_experts = model_config.get_num_experts()