Run Jobs

Submit a Run Job

Start training runs, batch jobs and evaluations without managing a machine.

This page shows how to submit a Run Job with the CLI. Each one needs compute, an image and a command.

Run a job

Install and sign in

Install the CLI and sign in:

featherless-cloud login

Run a command

Put the command after --:

featherless-cloud exec --gpu-count 1 -- python -c "import torch; print(torch.cuda.get_device_name(0))"

The CLI waits for the job, streams its output, and exits with the command's exit code.

Choose the image and data

Without --template or --image, the job runs the default PyTorch environment. Mount network volumes for input and output:

featherless-cloud exec \
  --template featherless/rocm-pytorch \
  --gpu-count 1 \
  --volume my-data:/workspace \
  -- python /workspace/train.py

Use --cpu-cores <n> instead of --gpu-count for CPU compute. See exec for every option.

Run Job options

OptionFlagDefault
Time limit, up to 24 hours--timeout <seconds>24 hours
Retries after an interruption, 0–3--preemption-retries <n>0
Working directory--workdir <path>/workspace
Container disk, 20–2,000 GiB--disk <gb>20
Network volumes, up to 8--volume <name>:<path>None
Environment variables and secrets--env K=V, --secret K=VNone
Collect output--no-retain-logs turns it offOn

The time limit starts when compute is assigned, not when you submit, and covers all retries together. Secrets are files under $FEATHERLESS_SECRETS_PATH, not environment variables; see Secret variables.

Retry safely

Submitting the same job twice with the same --idempotency-key <key> returns the first job instead of starting another, so a script can retry a submission safely.

Run in the background

--detach submits the job and returns at once. Follow it later with featherless-cloud runjob logs <job-id> --follow, using the ID it prints. Pressing Ctrl-C while exec waits also detaches; add --cancel-on-interrupt to cancel the job instead.

Next steps

Monitor and manage

Follow, cancel and delete Run Jobs.

Fine-tune with a Run Job

An end-to-end example with a network volume.