[XPU][CI] fallback to 7-args of moe_align_block_size (#57855)
Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
liuzhenwei committed
Sep 21, 2026 at 13:15 UTC
a15dbbd63daaba913b3302a858ad7e20e6dcc219
1 file changed
+13
vllm/_custom_ops.py
+13
@@ -2274,6 +2274,19 @@ def moe_align_block_size(
2274
expert_map: torch.Tensor | None = None,
2275
scatter_idx: torch.Tensor | None = None,
2276
) -> None:
2277
+ if current_platform.is_xpu():
2278
+ if scatter_idx is not None:
2279
+ raise NotImplementedError("scatter_idx is not supported on XPU")
2280
+ torch.ops._moe_C.moe_align_block_size(
2281
+ topk_ids,
2282
+ num_experts,
2283
+ block_size,
2284
+ sorted_token_ids,
2285
+ experts_ids,
2286
+ num_tokens_post_pad,
2287
+ expert_map,
2288
+ )
2289
+ return
2290
torch.ops._moe_C.moe_align_block_size(
2291
topk_ids,
2292
num_experts,