[Bugfix][CPU] Fall back to torch for unaligned swigluoai on NEON/vec MoE (#49985)
Signed-off-by: oops-oom <73481342@qq.com> Co-authored-by: oops-oom <73481342@qq.com> Co-authored-by: Claude <noreply@anthropic.com>
oops-oom committed
Jul 28, 2026 at 02:35 UTC
ed13deb37622cf9f603596743642807561c7691d
1 file changed
+6
vllm/model_executor/layers/fused_moe/cpu_fused_moe.py
+6
@@ -381,6 +381,12 @@ class CPUFusedMOE:
381
if not (w13_output_size % 32 == 0 and w2_output_size % 32 == 0):
382
return False, "none"
383
384
+ if (
385
+ layer.activation == MoEActivation.SWIGLUOAI
386
+ and w2_input_size % _MOE_GROUPED_GEMM_N_TILE != 0
387
+ ):
388
+ return False, "none"
389
+
390
supports_neon = current_platform.get_cpu_architecture() == CpuArchEnum.ARM
391
if supports_neon:
392
if (