diff --git a/vllm/model_executor/layers/fused_moe/cpu_fused_moe.py b/vllm/model_executor/layers/fused_moe/cpu_fused_moe.py index a957aca9ec..c89e5ddc30 100644 --- a/vllm/model_executor/layers/fused_moe/cpu_fused_moe.py +++ b/vllm/model_executor/layers/fused_moe/cpu_fused_moe.py @@ -381,6 +381,12 @@ class CPUFusedMOE: if not (w13_output_size % 32 == 0 and w2_output_size % 32 == 0): return False, "none" + if ( + layer.activation == MoEActivation.SWIGLUOAI + and w2_input_size % _MOE_GROUPED_GEMM_N_TILE != 0 + ): + return False, "none" + supports_neon = current_platform.get_cpu_architecture() == CpuArchEnum.ARM if supports_neon: if (