Skip to content

tinker checkpoint

tinker checkpoint

Manage checkpoints — list, download, publish, delete, and upload to HuggingFace.

All checkpoint commands take a CHECKPOINT_PATH in the format tinker://run-id/sampler_weights/final.

All commands support -f, --format [table|json] for output format.

tinker checkpoint list

List checkpoints across all runs, or for a specific run.

Options:

  • -f, --format [table|json] — Output format (default: table)

  • --run-id TEXT — List checkpoints for a specific training run

  • --limit INTEGER — Max checkpoints to show (default: 20, use 0 for all)
$ tinker checkpoint list --limit 3
          3 checkpoints (1789 more not shown)
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━┳━━━━━━━┳━━━━━━━┳━━━━━━━┓
┃ Checkpoint ID                  ┃ Type    ┃ Size    ┃ Pub   ┃ Crea  ┃ Expir ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━╇━━━━━━━╇━━━━━━━╇━━━━━━━┩
│ sampler_weights/lifecycle_test │ sampler │ 88.2 MB │ No    │ 3h    │ Never │
│ sampler_weights/final          │ sampler │ 102 MB  │ No    │ 4h    │ Never │
│ weights/final                  │ train   │ 306 MB  │ No    │ 4h    │ Never │
└────────────────────────────────┴─────────┴─────────┴───────┴───────┴───────┘

tinker checkpoint info CHECKPOINT_PATH

Show details of a specific checkpoint.

$ tinker checkpoint info tinker://run-id/sampler_weights/final

Options:

  • -f, --format [table|json] — Output format (default: table)

tinker checkpoint download CHECKPOINT_PATH

Download and extract a checkpoint to a local directory.

Options:

  • -f, --format [table|json] — Output format (default: table)

  • -o, --output PATH — Parent directory for extracted checkpoint (default: current directory)

  • --force — Overwrite existing directory if it exists
# Download to current directory
$ tinker checkpoint download tinker://run-id/sampler_weights/final

# Download to a specific directory
$ tinker checkpoint download tinker://run-id/sampler_weights/final -o ./models/

# Force overwrite
$ tinker checkpoint download tinker://run-id/sampler_weights/final --force

tinker checkpoint publish CHECKPOINT_PATH

Make a checkpoint publicly accessible. Other users can load it by path.

$ tinker checkpoint publish tinker://run-id/sampler_weights/final

Options:

  • -f, --format [table|json] — Output format (default: table)

tinker checkpoint unpublish CHECKPOINT_PATH

Make a published checkpoint private again.

$ tinker checkpoint unpublish tinker://run-id/sampler_weights/final

Options:

  • -f, --format [table|json] — Output format (default: table)

tinker checkpoint set-ttl CHECKPOINT_PATH

Set or remove the time-to-live on a checkpoint. After the TTL expires, the checkpoint is automatically deleted.

Options:

  • -f, --format [table|json] — Output format (default: table)

  • --ttl SECONDS — TTL in seconds (0 = no expiration)

# Expire after 7 days
$ tinker checkpoint set-ttl tinker://run-id/sampler_weights/step-100 --ttl 604800

# Keep forever (remove TTL)
$ tinker checkpoint set-ttl tinker://run-id/sampler_weights/step-100 --ttl 0

tinker checkpoint delete CHECKPOINT_PATH [...]

Permanently delete one or more checkpoints.

$ tinker checkpoint delete tinker://run-id/sampler_weights/step-100

# Delete multiple
$ tinker checkpoint delete \
    tinker://run-id/sampler_weights/step-100 \
    tinker://run-id/sampler_weights/step-200

Options:

  • -f, --format [table|json] — Output format (default: table)

tinker checkpoint push-hf CHECKPOINT_PATH

Upload a checkpoint to the HuggingFace Hub as a PEFT LoRA adapter.

Options:

  • -f, --format [table|json] — Output format (default: table)

  • -r, --repo TEXT — HuggingFace repo ID (e.g., username/my-lora-adapter). If omitted, derived from run.

  • --public — Create or upload to a public repo (default: private)
  • --revision TEXT — Target branch/revision
  • --commit-message TEXT — Commit message
  • --create-pr — Create a pull request instead of pushing to main
  • --allow-pattern TEXT — Only upload files matching this pattern (repeatable)
  • --ignore-pattern TEXT — Skip files matching this pattern (repeatable)
  • --no-model-card — Do not create a README.md model card
$ tinker checkpoint push-hf tinker://run-id/sampler_weights/final --repo my-org/my-model
$ tinker checkpoint push-hf tinker://run-id/sampler_weights/final --repo my-org/my-model --public --create-pr