[Bugfix] Fix EP weight filter breaking EPLB and NVFP4 accuracy (#37322)

Signed-off-by: Elvir Crncevic <elvircrn@gmail.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Kevin H. Luu <khluu000@gmail.com>

vllm/model_executor/model_loader/default_loader.py

@@ -320,6 +320,13 @@ class DefaultModelLoader(BaseModelLoader):
         ):
             return
 
+        # When EPLB is enabled, redundant physical expert slots may map to
+        # logical experts that belong to other ranks in the default partition.
+        # The weight loader needs to see ALL logical expert weights so it can
+        # populate these redundant slots.  Skip the filter entirely.
+        if parallel_config.enable_eplb:
+            return
+
         num_experts = model_config.get_num_experts()
         if num_experts <= 0:
             return

vllm/model_executor/model_loader/ep_weight_filter.py

@@ -73,4 +73,9 @@ def should_skip_weight(
     if eid is None:
         # Not an expert weight (dense / shared-expert / embedding) → keep.
         return False
+    # Only skip heavy weight tensors, never scale/metadata tensors.
+    # Scale tensors are tiny and some backends need them from ALL experts
+    # (e.g. FlashInfer NVFP4 computes a global max of activation scales).
+    if not weight_name.endswith(".weight"):
+        return False
     return eid not in local_expert_ids