[BugFix] Fix bugs in /v1/abort_requests interface from PR(#6992)#7176
[BugFix] Fix bugs in /v1/abort_requests interface from PR(#6992)#7176qwes5s5 wants to merge 3 commits intoPaddlePaddle:developfrom
Conversation
|
Thanks for your contribution! |
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## develop #7176 +/- ##
==========================================
Coverage ? 74.39%
==========================================
Files ? 383
Lines ? 53629
Branches ? 8414
==========================================
Hits ? 39898
Misses ? 11012
Partials ? 2719
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
qwes5s5
force-pushed
the
abort_requests_fix
branch
from
81f06aa to
81239a0
Compare
There was a problem hiding this comment.
Pull request overview
该 PR 用于修复 PR(#6992) 引入的 /v1/abort_requests 相关问题，避免在多模态/Completion 场景下触发 500 或 OpenAI 协议 finish_reason 校验失败，并补强 abort 清理路径的健壮性。
Changes:
- 扩展 OpenAI 协议响应的
finish_reason枚举，加入"abort"以通过校验。 - 优化 abort 等待/清理逻辑：对已结束请求做过滤与集合清理，并在资源回收时清理 abort 相关集合，避免残留导致异常。
- 更新/修复相关单测，使其与新的过滤逻辑一致。
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.
| File | Description |
|---|---|
| tests/engine/test_common_engine.py | 调整 _wait_abort_complete 相关用例，补齐 requests 字典以匹配新过滤逻辑 |
| fastdeploy/entrypoints/openai/protocol.py | finish_reason 增加 "abort"，修复协议校验失败 |
| fastdeploy/engine/sched/resource_manager_v1.py | abort/finish 回收路径使用 discard 并补充清理 abort 集合 |
| fastdeploy/engine/common_engine.py | _wait_abort_complete 增强：过滤已结束目标并清理 abort 集合，降低卡住/500 风险 |
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
qwes5s5
force-pushed
the
abort_requests_fix
branch
from
81239a0 to
e9d94e0
Compare
PaddlePaddle-bot
left a comment
PaddlePaddle-bot
left a comment
There was a problem hiding this comment.
📋 Review 摘要
PR 概述：修复 PR #6992 引入的 /v1/abort_requests 接口 bug
变更范围：common_engine.py、resource_manager_v1.py、protocol.py、test_common_engine.py
影响面 Tag：[Engine] [APIServer]
📝 PR 规范检查
PR 标题和描述均符合规范。
问题
| 级别 | 文件 | 概述 |
|---|---|---|
| 🔴 Bug | fastdeploy/entrypoints/openai/v1/serving_base.py |
_calc_finish_reason 方法未处理 abort 情况 |
总体评价
PR 的主要修复点都是合理的：
- 修改
eos_token_ids获取方式，从self.data_processor.eos_token_ids[0]改为getattr(request, "eos_token_ids", [0])，避免了对data_processor的依赖，更加健壮 - 在
_wait_abort_complete中添加已完成请求的清理逻辑，避免等待已完成的请求 - 将
remove()改为discard()，避免 KeyError，这是正确的修改 - 在
finish_requests中添加对 abort 相关集合的清理 - protocol.py 中添加 "abort" 到 finish_reason 类型定义
- 测试文件中添加必要的 mock 请求对象
但存在一个相关问题（不在本次 PR 变更范围内，但与 PR 目标相关）：
fastdeploy/entrypoints/openai/v1/serving_base.py 中的 _calc_finish_reason 方法（第 213-223 行）没有处理 abort 情况。当 request_output.error_msg 包含 "Aborted" 时（在 _control_abort_requests 中设置为 error_msg="Aborted"），finish_reason 不会被正确设置为 "abort"，而是保持默认的 "stop"。这与 protocol.py 中已添加的 "abort" 类型不一致。
对比旧版本的 serving_chat.py 和 serving_completion.py，它们都通过检查 error_msg 是否包含 "Aborted" 来设置 finish_reason = "abort"。v1 版本的 serving 代码使用 _calc_finish_reason 方法，但没有处理这个情况，这可能是导致 PR #6992 中提到的 "finish_reason validation failures" 的原因之一。
建议后续 PR 修复 serving_base.py 中的 _calc_finish_reason 方法，添加 abort 处理逻辑。
Motivation
It was discovered that PR(#6992) introduced bugs that may trigger 500 errors or finish_reason validation failures when using multimodal or completions interfaces.
Modifications
Modified finish_reason to include the "abort" value.
Usage or Command
Accuracy Tests
Checklist
[FDConfig],[APIServer],[Engine],[Scheduler],[PD Disaggregation],[Executor],[Graph Optimization],[Speculative Decoding],[RL],[Models],[Quantization],[Loader],[OP],[KVCache],[DataProcessor],[BugFix],[Docs],[CI],[Optimization],[Feature],[Benchmark],[Others],[XPU],[HPU],[GCU],[DCU],[Iluvatar],[Metax]]pre-commitbefore commit.releasebranch, make sure the PR has been submitted to thedevelopbranch, then cherry-pick it to thereleasebranch with the[Cherry-Pick]PR tag.