GPU work, from one job file.

Bring your code. Set the hardware and budget. Gipper finds a cloud GPU, runs the job, brings the results home, and checks that the machine is gone.

For training, batch inference, rendering, and other Linux GPU jobs. CLI or Python. Runpod and Vast.ai. No Python runtime dependencies.

MIT licensed. Python 3.12+. Linux client. Alpha software.

job.toml Your workload, your limits
name = "batch-inference"
workload = "my-model"

[gpu]
min_mem_gb = 24
providers = ["vast", "runpod"]

[run]
setup = "bash project/setup.sh"
command = "python project/main.py"

[budget]
max_usd = 2.0
max_hours = 1.0

Upload code and collect outputs with [[inputs]] and [[outputs]]. Write a portable job.

A rental has an end.

  1. EstimateStartup, setup, transfer, compute, fees.
  2. CheckRent, connect, verify GPU and CUDA.
  3. RunExecute, checkpoint, collect results.
  4. CloseTerminate, verify absence, record spend.

Cleanup runs after success or failure. Unconfirmed closure blocks a replacement; the independent reaper retries. Provider outages and a stopped client can still leave a rental billing. Understand the limits.

The hourly rate is only part of the price.

Cheap hardware can be expensive to start, feed, or retry. Gipper estimates the whole run and shows where each assumption came from.

Choose for the work
Filter memory, regions, CUDA requirements, and ports. Rank by cost, time, or latency, with your GPU preferences first.
Learn from your runs
Your local ledger builds compute-time, transfer-speed, billing, and host-reliability history. New installs start from stated defaults.
Keep the application yours
Install the libraries and models your job needs. Plain shell commands and TOML files work from the CLI or existing Python scripts.
Read how placement works

Start with an estimate.

Install from source, add a provider key to your local config, then inspect a bounded smoke job. Estimation and dry runs do not rent a machine.

Actual runs spend money. Keep the cleanup timer available and check the provider console when closure is uncertain.

Configuration and first run
git clone https://github.com/lavallee/gipper.git
cd gipper
uv tool install .

gipper pod estimate examples/smoke.toml \
  --provider vast

gipper pod run examples/smoke.toml \
  --provider vast --dry-run

Go further

Python API and portable jobsPlacement and estimatesOperational notesSecurity and host trustContributeReleases

Runpod and Vast.ai have live-run validation. Shadeform is experimental; owned-machine scheduling is planned. Unit tests establish decisions and parsing, not current provider availability.