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

fix qwen3next full graph break. (#3812)

Browse Source 
### What this PR does / why we need it?
fix qwen3next full graph break.
linearattention doesnot has aclgraph_support attr，so change to
cudagraph_support to support vllm.
<img width="603" height="120" alt="image"
src="https://github.com/user-attachments/assets/d2de53bb-4147-495a-9129-51d9083749be"
/>

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?


- vLLM version: v0.11.0rc3
- vLLM main:
https://github.com/vllm-project/vllm/commit/releases/v0.11.1

Signed-off-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
Co-authored-by: wangxiaoxin-sherie <wangxiaoxin7@huawei.com>
This commit is contained in:
XiaoxinWang
2025-10-29 10:30:23 +08:00
committed by GitHub
parent c76db627ab
commit 1e31b07fa7
1 changed files with 6 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
vllm_ascend/worker/model_runner_v1.py
View File

@@ -3775,7 +3775,12 @@ class NPUModelRunner(LoRAModelRunnerMixin):
for attn_group in self._attn_group_iterator(): for attn_group in self._attn_group_iterator():
builder = attn_group.get_metadata_builder() builder = attn_group.get_metadata_builder()
if builder.aclgraph_support.value < min_ag_support.value: graph_support = None
if hasattr(builder, 'aclgraph_support'):
graph_support = builder.aclgraph_support.value
else:
graph_support = builder.cudagraph_support.value
if graph_support < min_ag_support.value:
min_ag_support = builder.aclgraph_support min_ag_support = builder.aclgraph_support
min_ag_builder_name = builder.__class__.__name__ min_ag_builder_name = builder.__class__.__name__