[Feature] implement log channel separation and request log level system#7190
[Feature] implement log channel separation and request log level system#7190Jiang-Jia-Jun merged 3 commits intoPaddlePaddle:developfrom
Conversation
|
Thanks for your contribution! |
Codecov Report❌ Patch coverage is Additional details and impacted files@@ Coverage Diff @@
## develop #7190 +/- ##
==========================================
Coverage ? 73.75%
==========================================
Files ? 397
Lines ? 54828
Branches ? 8587
==========================================
Hits ? 40439
Misses ? 11682
Partials ? 2707
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This comment was marked as outdated.
This comment was marked as outdated.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This comment was marked as outdated.
This comment was marked as outdated.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This comment was marked as outdated.
This comment was marked as outdated.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This comment was marked as outdated.
This comment was marked as outdated.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.
There was a problem hiding this comment.
Pull request overview
该 PR 在 FastDeploy 现有日志体系上引入“日志通道化（main/request/console）”与“请求日志 L0-L3 分级”的统一机制，并将多处请求生命周期/调度/输出相关日志迁移到 request 通道，同时补充相应单测与文档说明。
Changes:
- 新增
fastdeploy.logger.request_logger（L0-L3 请求日志）与fastdeploy.logger.config（日志级别/请求日志默认配置解析）。 setup_logging支持 main/request/console 三通道 logger 配置，并在大量业务模块中用log_request/log_request_error替换原散落 logger 调用。- 更新与新增日志相关单测，并同步更新中英文日志文档。
Reviewed changes
Copilot reviewed 48 out of 48 changed files in this pull request and generated 4 comments.
Show a summary per file
| File | Description |
|---|---|
| tests/scheduler/test_local_scheduler.py | 将断言从 scheduler_logger 迁移为验证 log_request 调用 |
| tests/scheduler/test_dp_scheduler.py | 通过 mock log_request 适配 DP scheduler 的请求日志断言 |
| tests/output/test_process_batch_output.py | 用 log_request 替换对 llm_logger 的断言，覆盖 preemption 场景 |
| tests/output/test_process_batch_output_use_zmq.py | 用 log_request 替换对 llm_logger 的断言，覆盖 abort recycle 场景 |
| tests/logger/test_setup_logging.py | 更新默认 handler 断言并新增通道 logger/handler 配置测试 |
| tests/logger/test_request_logger.py | 新增 request_logger 的 _should_log/_truncate/log_request 测试 |
| tests/logger/test_logging_config.py | 新增 resolve_log_level 与请求日志默认值解析测试 |
| tests/logger/test_logger.py | 更新 unified logger 命名规则（默认走 main 通道）相关测试 |
| tests/input/test_video_utils.py | 调整 assertLogs 指向 main 通道 logger 名 |
| tests/entrypoints/test_llm.py | 将异常日志断言改为验证 log_request_error 调用 |
| tests/entrypoints/test_engine_client.py | 将参数告警断言改为验证 log_request 调用；修正 getenv mock 签名 |
| tests/entrypoints/test_abort.py | 将 abort 日志断言改为验证 log_request 调用 |
| tests/entrypoints/openai/v1/test_serving_completion_v1.py | 将异常日志断言改为验证 log_request_error 调用 |
| fastdeploy/utils.py | get_logger 增加 channel；全局 logger 默认归入 main/console/comm 等通道 |
| fastdeploy/scheduler/splitwise_scheduler.py | 将多处错误日志改为 log_request_error |
| fastdeploy/scheduler/local_scheduler.py | 将请求入队/出队/完成等日志迁移到 log_request |
| fastdeploy/scheduler/global_scheduler.py | 增加 log_request/log_request_error 并调整部分请求相关日志 |
| fastdeploy/scheduler/dp_scheduler.py | 将请求/结果相关日志迁移到 log_request |
| fastdeploy/output/token_processor.py | 将大量请求生命周期/异常日志迁移到 log_request/log_request_error |
| fastdeploy/logger/setup_logging.py | 默认 dictConfig 改为三通道配置并使用 LazyFileHandler |
| fastdeploy/logger/request_logger.py | 新增请求日志模块（L0-L3、截断、error 日志接口） |
| fastdeploy/logger/logger.py | 支持按 channel 获取 logger，并调整 unified logger 默认落 main 通道 |
| fastdeploy/logger/config.py | 新增日志级别解析与请求日志默认配置解析 |
| fastdeploy/input/tokenizer_client.py | 将轮询/解码异常日志迁移到 request 通道日志接口 |
| fastdeploy/input/qwen3_vl_processor/qwen3_vl_processor.py | 将“Processed request”日志迁移到 request 通道 |
| fastdeploy/input/qwen_vl_processor/qwen_vl_processor.py | 将“Processed request”日志迁移到 request 通道 |
| fastdeploy/input/ernie4_5_vl_processor/ernie4_5_vl_processor.py | 将“Processed request”日志迁移到 request 通道 |
| fastdeploy/input/base_processor.py | 将 token_ids/请求处理过程等日志迁移到 request 通道并分级 |
| fastdeploy/entrypoints/openai/v1/serving_completion.py | 将异常与流式关键点日志迁移到 request 通道 |
| fastdeploy/entrypoints/openai/v1/serving_chat.py | 将异常与流式关键点日志迁移到 request 通道 |
| fastdeploy/entrypoints/openai/tool_parsers/ernie_x1_tool_parser.py | 将错误日志迁移到 log_request_error |
| fastdeploy/entrypoints/openai/tool_parsers/ernie_45_vl_thinking_tool_parser.py | 将错误日志迁移到 log_request_error |
| fastdeploy/entrypoints/openai/serving_reward.py | 将关键输出日志迁移到 log_request |
| fastdeploy/entrypoints/openai/serving_models.py | 将错误日志迁移到 log_request_error |
| fastdeploy/entrypoints/openai/serving_engine.py | 将 semaphore/初始化/错误日志迁移到 request 通道接口 |
| fastdeploy/entrypoints/openai/serving_embedding.py | 将关键输出日志迁移到 log_request |
| fastdeploy/entrypoints/openai/serving_completion.py | 将异常与流式关键点日志迁移到 request 通道接口 |
| fastdeploy/entrypoints/openai/serving_chat.py | 将异常与流式关键点日志迁移到 request 通道接口 |
| fastdeploy/entrypoints/openai/protocol.py | 将 metadata obsolete 警告日志迁移到 log_request |
| fastdeploy/entrypoints/openai/api_server.py | 将接收请求/异常日志迁移到 request 通道接口 |
| fastdeploy/entrypoints/llm.py | 将 _receive_output 异常日志迁移到 log_request_error |
| fastdeploy/entrypoints/engine_client.py | 将校验/接收/abort/异常日志迁移到 request 通道接口 |
| fastdeploy/entrypoints/api_server.py | 将接收请求与异常日志迁移到 request 通道接口 |
| fastdeploy/engine/request.py | 将 from_dict/mm_positions 异常日志迁移到 request 通道接口 |
| fastdeploy/engine/engine.py | 将 add_requests/generate 相关日志迁移到 request 通道接口 |
| fastdeploy/engine/async_llm.py | 将 abort/generate 异常日志迁移到 log_request_error |
| docs/zh/usage/log.md | 补充日志通道与请求日志级别、环境变量说明（中文） |
| docs/usage/log.md | 补充日志通道与请求日志级别、环境变量说明（英文） |
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.
| from fastdeploy.engine.expert_service import start_data_parallel_service | ||
| from fastdeploy.engine.request import Request | ||
| from fastdeploy.inter_communicator import EngineWorkerQueue, IPCSignal | ||
| from fastdeploy.logger.request_logger import log_request, log_request_error |
There was a problem hiding this comment.
error分开打印，这个是一种通用的做法吗？是会形成一个类似request_error.log?
There was a problem hiding this comment.
这个分开是因为， log_request 会分 4 个级别，然后根据当前设置的级别过滤日志。 log_request_error 是不分级别。两个都会在 request.log 里打印，然后 request_error 会同时映射到 终端 和 error.log
| llm_logger.info(f"Cache task with request_id ({request.get('request_id')})") | ||
| llm_logger.debug(f"cache task: {request}") | ||
| log_request( | ||
| level=1, |
There was a problem hiding this comment.
这里的level=1~n，文档提到对应不同程度，是否可以改为枚举变量，代码中更清楚知道含义
There was a problem hiding this comment.
好的，预计改成 4 个 枚举变量，把名称和对应的程度对齐，方便清楚含义
|
|
||
| if not function_call_arr: | ||
| data_processor_logger.error("No valid tool calls found") | ||
| log_request_error(message="No valid tool calls found") |
There was a problem hiding this comment.
这种打印，是针对请求的话，建议加入request_id
|
|
||
| except Exception as e: | ||
| data_processor_logger.error(f"Error in extracting tool call from response: {str(e)}") | ||
| log_request_error(message="Error in extracting tool call from response: {error}", error=str(e)) |
| api_server_logger.info(f"Chat Streaming response last send: {chunk.model_dump_json()}") | ||
| log_request( | ||
| level=0, | ||
| message="Chat Streaming response last send: request_id={request_id}, finish_reason={finish_reason}, completion_tokens={completion_tokens}", |
There was a problem hiding this comment.
如果有logprob，logprob也增加打印，训练中依赖根据这个值判断是否出NaN
| """ | ||
| return { | ||
| "enabled": int(os.getenv("FD_LOG_REQUESTS", "1")), | ||
| "level": int(os.getenv("FD_LOG_REQUESTS_LEVEL", "0")), |
There was a problem hiding this comment.
level=默认值是0，是当前的输出日志没变吗？
There was a problem hiding this comment.
level=2是对齐当前 FastDeploy 的 info 级别的请求相关的输出日志
| scheduler_logger = get_logger("scheduler", "scheduler.log") | ||
| api_server_logger = get_logger("api_server", "api_server.log") | ||
| console_logger = get_logger("console", "console.log", print_to_console=True) | ||
| llm_logger = get_logger("fastdeploy", channel="main") |
There was a problem hiding this comment.
日志的打印目前有多处实现，在这一版升级中，可全部统一，不用在logger下存在，又在utils下有一份类似的了
|
对于error类的打印，都Review下是否针对请求级别的，如果是，增加request_id便于追溯。 |
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.
|
针对 review 意见的修改和说明： 1、log_request 分 4 个级别，根据当前设置的级别过滤日志。 log_request_error 不分级别，打印错误日志。两者都会在 request.log 里打印， log_request_error 还会同时映射到 终端 和 error.log。 2、log_request 的 4 个级别，分别对应有枚举变量：
3、对于error类的打印，针对请求级别的，增加request_id便于追溯。 4、logprob也增加打印 5、统一将日志相关的实现，整合到 fastdeploy/logger，在 utils 做了兼容处理 6、FD_LOG_REQUESTS_LEVEL 默认值设为了 2，该级别对应改造前的 FastDeploy 各相关模块的 info 级别日志 |
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.
…ger implementation from utils to logger module
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.
🤖 AI Code Review |
2026-04-16
📋 Review 摘要
PR 概述：实现日志通道划分和请求日志级别系统，将日志分为 main/request/console 三个通道，支持 L0-L3 四级请求日志详细度。
变更范围：fastdeploy/logger/（新增模块）、entrypoints/、engine/、scheduler/、docs/（文档更新）
影响面 Tag：[Feature] [APIServer] [Engine] [Scheduler] [Docs]
问题
| 级别 | 文件 | 概述 |
|---|---|---|
| 🟡 建议 | docs/usage/environment_variables.md:28 |
FD_LOG_REQUESTS_LEVEL 默认值变更属于 Breaking Change，需要明确说明 |
总体评价
日志系统重构整体设计合理，代码结构清晰，测试覆盖较为完善。实现了日志通道划分和分级控制功能，核心实现质量良好。但默认值变更属于 Breaking Change，建议在文档中增加醒目的迁移说明。
|
|
||
| # Request logging detail level (0-3). Higher level means more verbose output. | ||
| "FD_LOG_REQUESTS_LEVEL": lambda: int(os.getenv("FD_LOG_REQUESTS_LEVEL", "0")), | ||
| "FD_LOG_REQUESTS_LEVEL": lambda: int(os.getenv("FD_LOG_REQUESTS_LEVEL", "2")), |
There was a problem hiding this comment.
🟡 建议 FD_LOG_REQUESTS_LEVEL 默认值从 0 改为 2，这是一个 Breaking Change。
影响说明：
- 原默认值 0 (LIFECYCLE): 只记录请求创建/完成/中止等关键生命周期事件
- 新默认值 2 (CONTENT): 记录请求参数、调度信息、响应内容
潜在影响：
- 现有部署如果不设置
FD_LOG_REQUESTS_LEVEL环境变量，会产生更多的日志输出 - 日志文件大小可能显著增加，可能影响磁盘 I/O 和存储成本
建议：
- 在
log.md文档中增加「重要提示」或「Breaking Change」标记，提醒用户注意日志行为的变化 - 考虑在 PR 描述中增加 Breaking Change 说明
- 或者设置一个迁移期，在下一个版本再调整默认值
Motivation
针对 FastDeploy 的日志系统进行优化，预计分 4 个 pr 完成。
Modifications
Usage or Command
日志通道划分
FastDeploy 将日志分为三个通道：
fastdeploy.main.*fastdeploy.logfastdeploy.request.*request.logfastdeploy.console.*console.log请求日志级别
请求日志 (
request.log) 支持 4 个级别，通过环境变量FD_LOG_REQUESTS_LEVEL控制：默认级别为 L2 (CONTENT)，记录请求参数、调度信息和响应内容。如需更精简的日志，可设置 FD_LOG_REQUESTS_LEVEL=0 只记录关键生命周期事件。
Accuracy Tests
1
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.