[BugFix] add select_gemm_impl on CompressedTensorsWNA16MoEMethod to support LoRA (#31453)

Signed-off-by: JartX <sagformas@epdcenter.es>

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -1996,6 +1996,29 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
             block_shape=[0, self.group_size],
         )
 
+    def select_gemm_impl(
+        self,
+        prepare_finalize: mk.FusedMoEPrepareAndFinalize,
+        layer: torch.nn.Module,
+    ) -> mk.FusedMoEPermuteExpertsUnpermute:
+        if self.moe.is_lora_enabled:
+            assert self.moe_quant_config is not None
+            from vllm.triton_utils import HAS_TRITON
+
+            if HAS_TRITON:
+                from vllm.model_executor.layers.fused_moe import TritonExperts
+
+                layer.w13_weight = layer.w13_weight_packed
+                layer.w2_weight = layer.w2_weight_packed
+                return TritonExperts(quant_config=self.moe_quant_config)
+            else:
+                raise NotImplementedError(
+                    "TritonExperts requires Triton. "
+                    "Install triton or disable LoRA for MoE."
+                )
+
+        raise NotImplementedError
+
     def apply(
         self,
         layer: FusedMoE,