[Weight Cache] Enhance test and support EPLB (#36198)

python/sglang/srt/server_args.py

@@ -8467,11 +8467,6 @@ class ServerArgs:
                 "(--weight-cache-mode off) for this configuration."
             )
 
-        if cfg.weight_cache_mode != "off" and cfg.enable_eplb:
-            raise ValueError(
-                "--weight-cache-mode is not supported together with --enable-eplb."
-            )
-
     def _is_mistral_native_format(self) -> bool:
         """True iff the checkpoint requires load_format=mistral.
 

python/sglang/srt/weight_cache/daemon.py

@@ -316,6 +316,7 @@ class WeightCacheDaemon:
         # The initialized groups are the authority for rank identity. This
         # avoids maintaining a second copy of the model-parallel hierarchy.
         self._init_distributed(server_args, model_config)
+        self._initialize_eplb_expert_location_metadata(model_config)
         moe_dp_rank = get_parallel().moe_dp_rank
         moe_ep_rank = get_parallel().moe_ep_rank
         self.config = CacheConfig(
@@ -457,6 +458,23 @@ class WeightCacheDaemon:
             f"metadata size ~{total_bytes / 1024 / 1024:.1f} MB"
         )
 
+    def _initialize_eplb_expert_location_metadata(self, model_config) -> None:
+        """Build the same initial physical expert layout as the engine."""
+        if not self.server_args.enable_eplb:
+            return
+
+        from sglang.srt.eplb.expert_location import (
+            compute_initial_expert_location_metadata,
+            set_global_expert_location_metadata,
+        )
+
+        set_global_expert_location_metadata(
+            compute_initial_expert_location_metadata(
+                model_config=model_config,
+                moe_ep_rank=get_parallel().moe_ep_rank,
+            )
+        )
+
     def serve(self):
         """Block and serve IPC handles over Unix socket."""
         # Do NOT unlink an existing socket here: stale-file cleanup is the launch