Programmatically retrieve full console logs (stdout/stderr) of a W&B run

W&B Help
1

I’m trying to programmatically retrieve the full console logs from one of my runs – essentially the same output shown under the Logs tab (including all stdout/stderr and any wandb: messages). My goal is to save these logs locally for debugging, archival, or post-processing purposes.

What I’ve tried/researched:

  • Using the Python API for run history: I attempted wandb.Api().run("entity/project/run_id").history(), but realized this only returns metrics logged via wandb.log() (the run’s history of metrics), not the console text.

  • Listing run files: I used run.files() and noticed that for a finished run, there is an output.log file (which I believe contains the console output). In the web UI, I also see a “Download logs” button for finished runs. I suspect I can download that file via the API (maybe with run.file("output.log").download() as hinted in [W&B docs/forums]). Before going down that route, I wanted to confirm if this is the recommended approach.

  • Runs not finished: One concern is runs that crash or are still running. In one case, a run crashed and I didn’t see an output.log in the files. The UI still displayed the partial logs, but there was no download button. I came across a GitHub issue suggesting that output.log is only downloadable for finished runs, not for crashes. Is there any workaround or API method to get logs from a crashed or ongoing run? For example, can we fetch the live log stream or force an upload of the log before a run finishes?

What I’m looking for:

  • A supported way (via the W&B Python SDK or CLI) to download the complete console log of a run. If the solution is to use the Python API to grab output.log, I’ll gladly do that – just want to ensure I’m not missing a more direct method.

  • Any best practices for capturing these logs, especially in cases where a run might not finish cleanly. My use case is to automatically archive logs for all runs (successful or not) for later analysis.

I appreciate any guidance! If there’s an existing feature or a known workaround that I missed, please let me know. Thank you in advance for your help :slightly_smiling_face: