[AMD] Cache AITER expert mask across decode (#31889)

This commit is contained in:
Chengze Fan
2026-07-22 07:50:23 -07:00
committed by GitHub
parent e8e765b9d6
commit 40b2119b23
@@ -200,7 +200,7 @@ class StandardDispatcher(BaseDispatcher):
)
if self.local_expert_mapping is not None and not self.skip_local_expert_mapping:
if self.use_aiter_moe_runner:
if self.use_aiter_moe_runner and self.expert_mask_gpu is None:
self.expert_mask_gpu = (
(
(self.local_expert_mapping >= 0)
@@ -209,7 +209,7 @@ class StandardDispatcher(BaseDispatcher):
.to(torch.int32)
.to(device="cuda")
)
else:
elif not self.use_aiter_moe_runner:
if TopKOutputChecker.format_is_standard(topk_output):
topk_output = topk_output._replace(
topk_ids=self.local_expert_mapping[topk_output.topk_ids]