EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[Bugfix] Support Qwen3-MOE on aclgraph mode (#1381)

Browse Source 
### What this PR does / why we need it?
Fix the shape of the `npu_moe_init_routing` input parameters to support
aclgraph mode on qwen3-moe

In addition to this PR, resolving the `gatherv3` error might be
necessary. See related PR
https://github.com/vllm-project/vllm-ascend/pull/1297
https://github.com/vllm-project/vllm-ascend/pull/1446

Thanks to @yiz-liu  for providing the idea

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
Tested on Qwen3-30B-A3B

Closes: https://github.com/vllm-project/vllm-ascend/issues/1368

---------

Signed-off-by: ApsarasX <apsarax@outlook.com>
Signed-off-by: Yikun Jiang <yikunkero@gmail.com>
Co-authored-by: Yizhou Liu <liu_yizhou@outlook.com>
Co-authored-by: Yikun Jiang <yikunkero@gmail.com>
This commit is contained in:
ApsarasX
2025-07-06 15:29:36 +08:00
committed by GitHub
parent 14373f65d7
commit c58accc15e
3 changed files with 21 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/ops/fused_moe.py
View File

@@ -655,6 +655,7 @@ def fused_experts(
top_k: int,
expert_map: torch.Tensor = None,
apply_router_weight_on_input: bool = False,
max_num_tokens: Optional[int] = None,
) -> torch.Tensor:
"""
Fused experts with top-k routing.
@@ -748,11 +749,12 @@ def fused_experts(
dtype=torch.int32,
device=device).view(top_k, -1).permute(
1, 0).contiguous())
active_num = max_num_tokens if max_num_tokens is not None else num_tokens
sorted_hidden_states, expanded_row_idx, expanded_expert_idx = torch_npu.npu_moe_init_routing(
hidden_states,
row_idx=row_idx,
expert_idx=topk_ids,
active_num=num_tokens)
active_num=active_num)
expert_tokens = torch_npu.npu_moe_compute_expert_tokens(
expanded_expert_idx, num_experts)