Skip to content

[v0.19.1rc1] fix(npu): preserve AFD DBO config during auto worker selection - #194

Merged
jiangkuaixue123 merged 2 commits into
vllm-project:release/v0.19.1rc1from
jiangkuaixue123:agent/pr190-release-import-test
Aug 4, 2026
Merged

[v0.19.1rc1] fix(npu): preserve AFD DBO config during auto worker selection#194
jiangkuaixue123 merged 2 commits into
vllm-project:release/v0.19.1rc1from
jiangkuaixue123:agent/pr190-release-import-test

Conversation

@jiangkuaixue123

Copy link
Copy Markdown
Collaborator

Summary

  • Preserve AFD-owned NPU DBO and ubatching settings during vLLM-Ascend config normalization.
  • Keep upstream normalization unchanged for non-AFD configurations.
  • Make the initial import-failure retry test deterministic under PEP 660 editable installs.

Root cause

vLLM-Ascend resets DBO-related fields while selecting and normalizing the NPU worker. AFD owns its NPU ubatching path and needs those values restored after upstream normalization. The retry test also removed vllm_ascend.platform from sys.modules, which still allowed an editable-install finder to reload the real module.

Validation

  • python -m pytest tests/unit: 328 passed, 2 skipped.
  • NPU E2E excluding async and accuracy: 14 passed, 4 expected DBO skips.
  • DeepSeek-V3.2 8A4F Attention/FFN smoke request returned HTTP 200.

This PR includes the runtime changes from #190 plus the deterministic import-failure test fix.

jiaran-king and others added 2 commits August 3, 2026 18:38
Signed-off-by: zzh <jiaranran2@gmail.com>
Signed-off-by: jiangkuaixue123 <jiangxiaozhou111@163.com>
@jiangkuaixue123
jiangkuaixue123 force-pushed the agent/pr190-release-import-test branch from 2d604cd to d8d9c4c Compare August 4, 2026 07:41
@jiangkuaixue123
jiangkuaixue123 marked this pull request as ready for review August 4, 2026 07:44
@jiangkuaixue123
jiangkuaixue123 merged commit 3436dcb into vllm-project:release/v0.19.1rc1 Aug 4, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants