Repository navigation
[Core] Model console logging as CLI configuration - #57205
Conversation
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
98d57bf to
8e6db9e
Compare
|
This pull request has merge conflicts that must be resolved before it can be |
|
In the nvidia-l4-distributed-comm-ops suite, |
Thank you, Ben. Spot on 👍 |
|
/ci run |
|
✅ Triggered Buildkite CI #90397 for commit |
|
|
|
/ci run |
|
✅ Triggered Buildkite CI #90588 for commit |
|
|
|
|
|
/ci run |
|
✅ Triggered Buildkite CI #90676 for commit |
|
On the
|
|
/ci run |
|
✅ Triggered Buildkite CI #90942 for commit |
Move vLLM logging settings into VllmConfig.logging_config. Users can supply a LoggingConfig as --logging-config JSON or with dotted arguments, and use --log-level as a convenience shortcut. The configuration includes whether to apply the default dictConfig, the log level, and a path to a custom Python logging dictConfig. Retain the pre-existing --log-config-file option as a legacy alias for --logging-config.pylogging_config_file; it takes precedence when both are provided. Apply the resolved configuration in the parent process and propagate it to Uvicorn and child processes. Co-authored-by: AI Assistant <noreply@openai.com> Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Defer vLLM logging configuration until CLI parsing or engine construction so --logging-config.configure_logging false has parity with VLLM_CONFIGURE_LOGGING=0. Co-authored-by: AI Assistant <noreply@openai.com> Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Co-authored-by: AI Assistant <noreply@openai.com> Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Warn when the legacy logging configuration flag is used and document its replacement and v0.33.0 removal. Co-authored-by: AI Assistant <noreply@openai.com> Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Emit CUDA platform diagnostics after logging configuration is applied, rather than as an import side effect. This lets custom handlers capture the warning and avoids fallback stderr output. Co-authored-by: AI Assistant <noreply@openai.com> Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Assisted-by: Codex Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Configure logging before the offline entrypoint records its arguments and before EngineArgs builds VllmConfig. This preserves initial log output after deferred logging configuration. Reported-by: Ben Browning Assisted-by: Codex Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Capture INFO records from an unconfigured vLLM logger and restore its prior propagation setting after each test. Reported-by: Ben Browning Assisted-by: Codex Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Use a real LoggingConfig in the two profiler unit tests so logging setup does not interpret MagicMock as a file descriptor. Assisted-by: OpenAI Codex Signed-off-by: Mark McLoughlin <markmc@redhat.com>
|
/ci run |
|
✅ Triggered Buildkite CI #91187 for commit |
|
With CI actually passing, and since there was no feedback from other committers when I gave a heads-up on Slack one week ago, I'm going ahead and merging based on Ben's approval |
3 participants
Addresses #57200.
Adds a typed
LoggingConfigtoVllmConfigwith three initial fields:configure_logginglog_levelpylogging_config_fileThe complete object can be supplied as
--logging-configJSON or with dotted arguments such as--logging-config.log_level DEBUG.--log-levelis a convenience override forlog_level, while the pre-existing--log-config-fileis retained as a legacy override forpylogging_config_file.The same model is available through serve YAML and Python configuration.
VLLM_CONFIGURE_LOGGING,VLLM_LOGGING_LEVEL, andVLLM_LOGGING_CONFIG_PATHremain as compatibility defaults. Explicit CLI and YAML configuration takes precedence, and the convenience/legacy CLI arguments override their matching--logging-configfields.The resolved configuration is applied after argument parsing and propagated through
VllmConfigto Uvicorn and API, engine, coordinator, and worker processes. This fixes the asymmetry where a CLI-supplied--log-config-fileconfigured Uvicorn but arrived too late for vLLM's already-initialized root logger. Code that previously inspectedVLLM_LOGGING_LEVELdirectly now uses the effective logger level.The existing default text output and custom Python
dictConfigbehavior are preserved. Tests cover JSON and dotted CLI forms, CLI/YAML/environment precedence, post-import reconfiguration, custom configuration files, and spawned-process propagation.Co-authored-by: AI Assistant noreply@openai.com