[BugFix] LoRA: Support loading base_layer of experts (#31104)

Signed-off-by: Hollow Man <hollowman@opensuse.org>

vllm/model_executor/layers/fused_moe/layer.py

@@ -2007,6 +2007,7 @@ class FusedMoE(CustomOp):
     @classmethod
     def make_expert_params_mapping(
         cls,
+        model: torch.nn.Module,
         ckpt_gate_proj_name: str,
         ckpt_down_proj_name: str,
         ckpt_up_proj_name: str,
@@ -2025,13 +2026,19 @@ class FusedMoE(CustomOp):
             )
         )
 
+        base_layer = (
+            "base_layer."
+            if any(".base_layer." in name for name, _ in model.named_parameters())
+            else ""
+        )
+
         return [
             # (param_name, weight_name, expert_id, shard_id)
             (
-                "experts.w13_"
+                f"experts.{base_layer}w13_"
                 if weight_name in [ckpt_gate_proj_name, ckpt_up_proj_name]
-                else "experts.w2_",
-                f"experts.{physical_to_logical_map[expert_id]}.{weight_name}.",
+                else f"experts.{base_layer}w2_",
+                f"experts.{physical_to_logical_map[expert_id]}.{weight_name}.{base_layer}",
                 expert_id,
                 shard_id,
             )