Quickstart #

Five commands from zero to a finished training run.

pip install vaultlayer
vaultlayer init                          # one-time authentication
vl run train.py                          # submit your training script
vl logs <job_id> --follow                # stream logs as the job runs
vl credits                               # check your balance

Run a training job #

Submit any Python training script — VaultLayer routes it to GPU capacity, streams logs back, checkpoints as it runs, and auto-recovers on interruption. The command is the same whether you wrote your script in PyTorch, JAX, or Hugging Face (see Frameworks).

The interpreter is optional for a local .py script: vl run train.py and vl run python train.py do the same thing. Explicit launchers — python -u, torchrun, accelerate launch — are always passed through unchanged.

Advanced routing, cost, and safety flags:

Multi-node distributed training (private beta):

Frameworks #

VaultLayer wraps the command you already run — PyTorch, JAX, or Hugging Face — with no code changes required. The run command is identical across all of them:

vl run train.py

The only optional step is checkpoint integration. Add it once and VaultLayer can resume your job from the last saved step if a GPU is reclaimed or the host fails mid-run. vaultlayer init writes the helper files locally, and on a first run vl run offers to auto-insert the right snippet for you — so the lines below are usually added for you, not by hand. The helper module itself is provided automatically on the GPU node; you never upload it.

PyTorch — run it, then add 3 lines for resume-on-interruption:

vl run train.py

# inside train.py (optional — enables auto-resume across migrations):
from vaultlayer_checkpoint import checkpoint, restore, CHECKPOINT_DIR
start_step = restore(CHECKPOINT_DIR, model=model, optimizer=optimizer)
# ...inside your training loop:
checkpoint(step=step, model=model, optimizer=optimizer, save_path=CHECKPOINT_DIR)

Hugging Face (Transformers Trainer / TRL / Accelerate) — checkpoints auto-save and VaultLayer resumes from the last one automatically. Nothing to add:

vl run train.py

# Single-node: checkpoints save to mirrored storage and training
# auto-resumes after an interruption — no code changes.
# Multi-node only (advanced): set output_dir=$VAULTLAYER_CHECKPOINT_DIR, then
# pass trainer.train(resume_from_checkpoint=os.environ.get("VAULTLAYER_RESUME_CHECKPOINT"))

JAX / Flax — run it, then add 3 lines for resume-on-interruption:

vl run train.py

# inside train.py (optional — enables auto-resume across migrations):
from vaultlayer_checkpoint_jax import vl_restore, vl_checkpoint
params, opt_state, start_step = vl_restore(params, opt_state)
# ...inside your training loop:
vl_checkpoint(step, params, opt_state, loss=float(loss))

Bring Your Own Cloud (BYOC) #

Run jobs on your own cloud account — AWS, Azure, GCP, or any provider — instead of the managed pool. VaultLayer provisions on your compute, checkpoints to your bucket, and adds the reliability layer — orchestration, monitoring, and auto-recovery — on top. A BYOC job is never routed to the managed pool, and there is no per-run charge from VaultLayer: your compute is billed by your own cloud provider under a monthly contract.

Watch + inspect a running job #

Status, logs, GPU health, recent jobs — all from the CLI.

When a job fails #

The CLI shows the last 15 lines of error output inline on every failure — no extra command needed. For deeper investigation:

Keep the GPU alive for debugging #

After your script exits, the instance is normally torn down so billing stops. Pass --keep-alive to hold it for a debugging window — inspect logs while the host is still warm, pull artifacts, or fix and re-try without paying a fresh cold-boot.

Stop or restart a job #

Pre-flight estimation #

See what a job will cost before you submit. Estimates are price quotes only — they don't check your credit balance (use vl credits for that).

Datasets #

Two ways to get data to a job: upload local data once with vl sync (then reuse it via r2://), or point straight at cloud storage with --data.

Connect your own storage #

Bring your own buckets — jobs read your datasets and write checkpoints back to storage you control, instead of going through an upload.

Account, credits, and tagging #

Troubleshooting #

Need help? #

Email rahuljain@vaultlayer.cloud for anything — bugs, feature requests, or quick questions on how to use a command. We typically reply within a day.

For one-off feedback or a crash report from the CLI, you can also run vl feedback.