SDK (Python) - Compute

The Python package you write against is the same one the CLI uploads.

import compute

app = compute.App("my-job")
image = compute.Image.cuda_pytorch()


@app.function(gpu="H100-80GB", image=image, timeout=600)
def train(steps: int = 40) -> dict:
    # imports that need the GPU image go inside the function
    return {"ok": True, "steps": steps}

Then launch the tested example (source):

compute run simple_mlp.py::train --gpu H100-SXM --wait --yes

Surface that is documented

Symbol	Role
`compute.App(name)`	Names the app in the upload.
`compute.Image.cuda_pytorch()`	NVIDIA image (RunPod H100 and Vast.ai CUDA SKUs). Installs `torch==2.5.1+cu124` and `numpy` at run start unless you pin a `torch*` package.
`compute.Image.rocm_pytorch()`	AMD / MI300X image. Installs `torch==2.7.1+rocm6.3` and `numpy` at run start unless you pin a `torch*` package.
`Image.<curated>().pip_install(...)`	Extra pip packages on that image.
`@app.function(...)`	Marks the entrypoint. `gpu`, `image`, `timeout`.
`fn.remote(**kwargs)`	In-script call shape. Args and return must be JSON.

Put GPU-only imports inside the function so local dry-runs and packaging stay honest.

Not in this manual

Do not treat local process spawning, detached handles you wait on later, persistent disks, or extra decorators as supported customer API. If it is not in the table above, do not build a workflow on it.

Timeout

Decorator timeout= is seconds on the function. The CLI default is 10 minutes if neither the decorator nor --timeout is set. The maximum is 24 hours. See Limits.