[Bugfix] Use FusedMoEWithLoRA for shared MoE LoRAs on 3D models (#60098)

Signed-off-by: ARAVINDHAN T <arvindhant01@gmail.com>
Co-authored-by: linitra24 <renshuang.zhou@daocloud.io>

vllm/config/lora.py

@@ -81,8 +81,9 @@ class LoRAConfig:
     """If True, load MoE expert adapters in the "shared-outer" layout, where the
     gate/up (`w1`/`w3`) lora_A and the down (`w2`) lora_B are shared across all
     experts (stored once with expert-dim 1) instead of per-expert. The shared
-    factors are broadcast to the expert count at kernel time. Only meaningful for
-    MoE models whose adapters use this layout; ignored otherwise."""
+    factors are broadcast to the expert count at kernel time. 3D-weight MoE
+    models use `FusedMoEWithLoRA` for this layout. Only meaningful for MoE
+    models whose adapters use this layout; ignored otherwise."""
     max_lora_cls_labels: int | None = Field(default=None, ge=1)
     """Maximum output size for LoRA classification heads. Defaults to the
     base classification head size."""

vllm/lora/model_manager.py

@@ -144,20 +144,18 @@ class LoRAModelManager:
             else set()
         )
 
-        # When the engine is started with enable_mixed_moe_lora_format=True
-        # we force the universal 2D wrapper (FusedMoEWithLoRA) regardless of
-        # the model's 3D flag, so 2D and 3D adapters can coexist.
+        # Mixed-format and shared-outer adapters both use FusedMoEWithLoRA.
+        # FusedMoE3DWithLoRA only implements the fused gate/up PEFT pair.
         self._enable_mixed_moe_lora_format = (
             is_moe and lora_config.enable_mixed_moe_lora_format
         )
+        self._enable_moe_shared_loras = is_moe and lora_config.enable_moe_shared_loras
         self._is_3d_moe_model = (
             self._is_moe
             and self.model.is_3d_moe_weight
             and not self._enable_mixed_moe_lora_format
+            and not self._enable_moe_shared_loras
         )
-        # Shared MoE adapters: w13 lora_A / w2 lora_B shared across experts,
-        # stored as pre-stacked experts.w{1,2,3} tensors (startup opt-in).
-        self._enable_moe_shared_loras = is_moe and lora_config.enable_moe_shared_loras
         self.packed_modules_mapping = process_packed_modules_mapping(
             self.model,
             force_2d_moe=self._enable_mixed_moe_lora_format,