veri checkpoints - Veri Documentation

The veri checkpoints group reads the intermediate checkpoints a managed job saved while it ran; veri jobs resume continues a finished job from one of them as a new job. Conceptual overview, knobs and retention live in Checkpoints and resume.

Intermediate checkpoints are rolling out behind a feature flag. Until it is enabled for your workspace these commands report the feature as unavailable.

Commands

Command	Purpose
`veri checkpoints list <job_id>`	The job’s checkpoints, newest step first (id, step, payload, size, files, metrics).
`veri checkpoints get <job_id> <checkpoint>`	One checkpoint. `<checkpoint>` is an id or `<job_id>@<step>`.
`veri checkpoints download <job_id> <checkpoint> [--output DIR] [--artifact weights\|resumable\|adapter]`	Download a checkpoint into `DIR/<checkpoint_id>/`.
`veri jobs resume <job_id> [-c <checkpoint>] [--max-steps N] [--num-epochs N] [--output-name NAME] [--provider P] [--region R] [--gpu-type G] [-f]`	Continue a finished job from a checkpoint as a new job.

List and inspect

$ veri checkpoints list job_abc123
┃ id               ┃ step ┃ payload   ┃ size     ┃ files ┃ pinned ┃ created_at           ┃
┃ ckpt_8f2a1c0d9e7 ┃ 400  ┃ resumable ┃ 24.1 GB  ┃ 9     ┃        ┃ 2026-10-09T18:40:02Z ┃
┃ ckpt_5b7e3a2f1c4 ┃ 300  ┃ resumable ┃ 24.1 GB  ┃ 9     ┃        ┃ 2026-10-09T18:12:51Z ┃

payload is resumable for a managed job’s saves (weights plus optimizer, scheduler and trainer state) and weights for a custom script’s directories.

Download

veri checkpoints download job_abc123 job_abc123@400 --output ./ckpt
veri checkpoints download job_abc123 job_abc123@400 --output ./ckpt --artifact resumable

weights (the default) is the model and tokenizer, loadable with from_pretrained. resumable adds the optimizer and trainer state, about two thirds of a full fine-tune checkpoint’s bytes. adapter is the LoRA adapter files only.

Resume

veri jobs resume job_abc123                          # newest ready resumable checkpoint
veri jobs resume job_abc123 -c job_abc123@400 --max-steps 2000
veri jobs resume job_abc123 --provider digitalocean --gpu-type H100-80GB

The new job reuses the source job’s base model, dataset, method, hyperparameters, GPU count and reward functions; the options are the knobs that may change. The API refuses a mismatch with the reason. -f follows the new job’s logs.