Fix CompressedTensorsWNA16MoE with grouped scales (vllm-project#13769)

Author: mgoin

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

@@ -527,7 +527,8 @@ def marlin_moe_permute_scales(s: torch.Tensor, size_k: int,
         replace_tensor("w13_weight_scale", marlin_w13_scales)
         marlin_w2_scales = marlin_moe_permute_scales(
             layer.w2_weight_scale,
-            layer.w2_weight_scale.shape[1] * self.packed_factor,
+            layer.w2_weight_scale.shape[1] *
+            (self.group_size if self.group_size != -1 else self.packed_factor),
             size_k2,
             self.group_size,
             self.num_bits,