> ## Documentation Index
> Fetch the complete documentation index at: https://docs.replicas.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# GPU sandboxes

> Run training, inference, and other GPU workloads on cloud GPUs from a workspace.

Workspaces have no GPU. GPU sandboxes give agents scoped access to cloud GPUs, powered by [Modal](https://modal.com), through the `replicas gpu` CLI.

## What it enables

The CLI is installed in every workspace and authenticated automatically. Agents can start GPU sandboxes, copy code in, run commands with streamed output, run long jobs in the background, and copy results back.

Each workspace can only see and control its own GPU sandboxes, and provider credentials are never exposed inside workspaces.

## Example

```bash theme={null}
replicas gpu run --gpu H100 --pull outputs -- python train.py
```

This starts an H100 sandbox, copies the current directory in (respecting `.gitignore`), streams the command's output, copies `outputs/` back, and stops the sandbox.

For longer sessions:

```bash theme={null}
replicas gpu create --gpu A100-80GB --timeout 3h
replicas gpu push
replicas gpu exec --detach -- python train.py
replicas gpu logs <job-id> -f
replicas gpu pull /workspace/checkpoints
replicas gpu delete
```

Run `replicas gpu --help` for every option, including `--image` for any public registry image and `--port` to expose a port as a public HTTPS URL.

## GPUs

T4, L4, A10G, L40S, A100, A100-80GB, H100, H200, and B200, with up to 8 GPUs per sandbox (for example `--gpu H100:4`). The default is an L4 with a PyTorch image.

## Limits

* 3 running sandboxes per workspace and 10 per organization
* 1-hour lifetime by default, up to 24 hours with `--timeout`
* Stops after 10 minutes with no running command by default, up to 1 hour with `--idle-timeout`
* 1 GiB per `push` or `pull`; download larger datasets inside the sandbox
* Sandboxes stop when their workspace sleeps, is archived, or is deleted

## Billing

GPU sandboxes are billed at cost and count toward included usage. See [Billing](/admin/billing).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.