How do I download the console log file from a run?

W&B captures your training script’s stdout and stderr output and stores it as output.log in the run’s file storage. Where to retrieve it depends on whether the run finished cleanly, is still active, or crashed. From the UI

Open the run page.
Click the Files tab.
Find output.log in the file list and click the download icon.

By default, output.log is only uploaded when the run finishes — while a run is active the file will not appear in the Files tab. See “While the run is still running” below for the workaround. Programmatically via the API

import wandb

api = wandb.Api()
run = api.run("entity/project/run_id")

# Download output.log to the current directory
run.file("output.log").download(replace=True)

While the run is still running Upload cadence cannot be changed after a run has started, so you must opt in at wandb.init time using multipart console logging. With console_multipart=True (SDK v0.22.3+), the SDK writes console output to timestamped chunks under a logs/ folder and uploads each chunk as soon as it closes, giving you live access to logs while the run is active.

import wandb

with wandb.init(settings=wandb.Settings(
    console_multipart=True,
    console_chunk_max_bytes=1_000_000,   # rotate at ~1MB
    console_chunk_max_seconds=60,        # or every 60s, whichever first
)):
    ...

Both limits can be combined — whichever fires first triggers a new chunk. Setting both to 0 reverts to the default behavior (single upload at run finish). Uploaded chunks are immutable; terminal control sequences that modify previous lines (e.g., progress bars using carriage returns) only affect the current chunk. If the run crashed Without multipart logging enabled, a crashed run produces no output.log in the Files tab and no download button appears. Enabling console_multipart=True means chunks uploaded before the crash are preserved on the server and remain downloadable. If you didn’t have multipart enabled, check your local wandb/ directory — W&B writes a local copy of the log to wandb/run-<timestamp>-<run-id>/logs/output.log before sync. Resuming the output.log across restarts When using wandb.init(resume="allow", id="RUN_ID") without multipart logging, the run’s single output.log is overwritten on resume — earlier content is lost. From SDK v0.20.1+ with console_multipart=True, the original and resumed sessions each get their own chunks under logs/, so nothing is lost.

wandb.init(
    id=run_id,
    resume="allow",
    settings=wandb.Settings(console_multipart=True),
)

Viewing logs longer than 10,000 lines The Logs tab in the UI applies a display limit for performance: for runs with 0–100,000 lines total, it shows the latest 10,000 lines; once a run exceeds 100,000 lines, the tab may show stale lines rather than the latest tail. The underlying output.log (or multipart chunks) always contains the full output — to access all lines, download the file using one of the methods above.

Logs Runs

Documentation Index