[FDConfig] 默认开启 FD_ENABLE_E2W_TENSOR_CONVERT 和 FD_ENGINE_TASK_QUEUE_WITH_SHM#7746
[FDConfig] 默认开启 FD_ENABLE_E2W_TENSOR_CONVERT 和 FD_ENGINE_TASK_QUEUE_WITH_SHM#7746sunlei1024 wants to merge 9 commits intoPaddlePaddle:developfrom
Conversation
|
Thanks for your contribution! |
sunlei1024
changed the title
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
CI报告基于以下代码生成（30分钟更新一次）: 1 任务总览当前有 1 个 Required 任务失败（Approval 审批未通过），3 个 Required 任务运行中，请等待所有 Required 任务完成后再合并。
2 任务状态汇总2.1 Required任务 : 6/10 通过
2.2 可选任务 — 24/26 通过
3 失败详情（仅 required）Approval — 基础设施（置信度: 高）Approval
根因详情: 关键日志: 修复建议:
修复建议摘要: 请 Reviewer 审批此 PR 链接: 查看日志 |
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## develop #7746 +/- ##
==========================================
Coverage ? 72.41%
==========================================
Files ? 396
Lines ? 55708
Branches ? 8706
==========================================
Hits ? 40341
Misses ? 12606
Partials ? 2761
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This comment was marked as outdated.
This comment was marked as outdated.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.
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
PaddlePaddle-bot
left a comment
PaddlePaddle-bot
left a comment
There was a problem hiding this comment.
🤖 Paddle-CI-Agent | pr_review |
2026-05-09 23:20:19
📋 Review 摘要
PR 概述：将 FD_ENABLE_E2W_TENSOR_CONVERT 与 FD_ENGINE_TASK_QUEUE_WITH_SHM 两个环境变量的默认值从 0 改为 1，并为 EngineWorkerQueue 新增队列健康检测方法，同时重构测试工具函数
变更范围：fastdeploy/envs.py、fastdeploy/engine/common_engine.py、fastdeploy/inter_communicator/engine_worker_queue.py、测试文件
影响面 Tag：[FDConfig] [Engine]
📝 PR 规范检查
PR 标题使用了官方 Tag [FDConfig]，格式合规；## Motivation、## Modifications、## Usage or Command、## Accuracy Tests、## Checklist 五个必填段落均已填写，结构符合模板要求。✓
问题
| 级别 | 文件 | 概述 |
|---|---|---|
| 🟡 建议 | tests/ci_use/EB_Lite_with_adapter/test_eblite_serving.py:93 |
rm -rf /dev/shm/* 清除全局 /dev/shm，在 CI 共享环境中可能误删其他进程的 socket 文件 |
| ❓ 疑问 | fastdeploy/inter_communicator/engine_worker_queue.py:859 |
is_broken() 的 except Exception: return False 分支无任何日志，未知异常会被静默吞掉 |
| ❓ 疑问 | fastdeploy/envs.py:178 |
FD_ENGINE_TASK_QUEUE_WITH_SHM=1 默认开启后，若 /dev/shm 空间不足（Docker 默认仅 64 MB），启动时无容量预检或明确报错 |
总体评价
本次变更逻辑清晰，性能优化意图明确，代码质量良好。有两处细节建议关注：测试中的全局 /dev/shm 清理范围过大，以及 is_broken() 的静默异常处理；此外建议在文档或启动日志中补充 SHM 容量要求提示。
| """ | ||
| # 清理/dev/shm中的临时文件 | ||
| try: | ||
| subprocess.run("rm -rf /dev/shm/*", shell=True) |
There was a problem hiding this comment.
🟡 建议 rm -rf /dev/shm/* 会清除 /dev/shm 目录下的所有文件，在多测试并发或 CI 共享环境下可能误删其他进程的 unix socket 文件。
建议改为只清理本测试相关的 socket 文件：
import glob
for sock in glob.glob("/dev/shm/fd_task_queue_*.sock"):
try:
os.remove(sock)
except OSError:
pass或直接复用本 PR 引入的 cleanup_unix_socket 函数，按 PORTS_TO_CLEAN 逐一清理对应 socket。
| except (ConnectionRefusedError, ConnectionResetError, BrokenPipeError, EOFError, OSError): | ||
| llm_logger.error("Failed to connect to engine worker queue") | ||
| return True | ||
| except Exception: |
There was a problem hiding this comment.
❓ 疑问 except Exception: return False 分支无任何日志输出，遇到预期外异常（如 AuthenticationError、MemoryError）时会静默返回 False（表示队列正常），可能掩盖真实故障。
建议补充日志：
except Exception as e:
llm_logger.warning("Unexpected exception in is_broken check: %s", e)
return False| "FD_ENABLE_E2W_TENSOR_CONVERT": lambda: int(os.getenv("FD_ENABLE_E2W_TENSOR_CONVERT", "0")), | ||
| "FD_ENGINE_TASK_QUEUE_WITH_SHM": lambda: int(os.getenv("FD_ENGINE_TASK_QUEUE_WITH_SHM", "0")), | ||
| "FD_ENABLE_E2W_TENSOR_CONVERT": lambda: int(os.getenv("FD_ENABLE_E2W_TENSOR_CONVERT", "1")), | ||
| "FD_ENGINE_TASK_QUEUE_WITH_SHM": lambda: int(os.getenv("FD_ENGINE_TASK_QUEUE_WITH_SHM", "1")), |
There was a problem hiding this comment.
❓ 疑问 FD_ENGINE_TASK_QUEUE_WITH_SHM 默认值改为 1 后，Docker 容器默认 /dev/shm 仅有 64 MB，大模型场景下可能导致 socket 通信失败且错误信息不直观。
建议在 EngineWorkerQueue.__init__ 初始化时加入容量预检（或至少输出 warning）：
import shutil
shm_free = shutil.disk_usage("/dev/shm").free
if shm_free < 1 * 1024 ** 3: # 1 GB
logger.warning("/dev/shm free space %d MB < 1 GB, consider --shm-size=1g", shm_free // 1024**2)3 participants
Motivation
默认开启
FD_ENABLE_E2W_TENSOR_CONVERT和FD_ENGINE_TASK_QUEUE_WITH_SHM，以提升 Engine-to-Worker 的张量传递效率，以及引擎任务队列基于共享内存（SHM）的通信性能。
该优化在大模型推理场景下可以减少序列化/反序列化开销，提高吞吐和延迟表现。
Modifications
fastdeploy/envs.pyFD_ENABLE_E2W_TENSOR_CONVERT默认值由0改为1FD_ENGINE_TASK_QUEUE_WITH_SHM默认值由0改为1行为变更说明：
FD_ENABLE_E2W_TENSOR_CONVERT=0FD_ENGINE_TASK_QUEUE_WITH_SHM=0/dev/shm空间充足（建议 ≥ 1GB，视模型规模而定）Usage or Command
默认无需额外配置，升级后自动生效。
如需关闭相关功能，可通过环境变量控制：
Docker 使用示例（配置共享内存）：
Accuracy Tests
本次修改仅涉及环境变量默认值调整，不涉及模型计算逻辑变更。
验证结果：
Checklist