CLI
Sandbox
Pick Local or Modal compute from the CLI — same presets as the Web UI.
sandwalk
# then:
/sandboxInteractive picker
Bare /sandbox opens a numbered list (same style as /model):
| Choice | What it sets |
|---|---|
| Local Sandbox | Runs on this machine (agent.cli_sandbox=builtin) |
| Modal · 2 CPU · 8 GB (and other CPU presets) | Modal CPU compute |
| Modal · T4 · 16GB (and other GPU presets) | Modal GPU + CUDA base image |
Presets match the Web UI composer (CPU cores/RAM and GPU class).
After a Modal choice you can edit the base image inline. The current image is prefilled; press Enter to keep it, or type a Docker/registry ref.
| Mode | Default base image |
|---|---|
| CPU | python:3.12-slim |
| GPU | pytorch/pytorch:2.4.0-cuda12.4-cudnn9-runtime |
CUDA tooling comes from the GPU base image (not from pip alone). Restart / start the Modal sandbox after changing compute or image so the new profile is used.
Other commands
/sandbox status
/sandbox builtin|off
/sandbox modal|on
/sandbox toggle
/sandbox helpFrom the shell (non-interactive):
sandwalk config set agent.cli_sandbox modal
sandwalk config set sandbox.compute_type gpu
sandwalk config set sandbox.gpu_class T4
sandwalk config set sandbox.gpu_image pytorch/pytorch:2.4.0-cuda12.4-cudnn9-runtime
sandwalk sandbox probe # optional live Modal probeConfig keys
| Key | Role |
|---|---|
agent.cli_sandbox | CLI-only Local vs Modal (builtin / modal) |
sandbox.backend | Web dashboard default (independent of CLI) |
sandbox.compute_type | cpu or gpu |
sandbox.gpu_class | e.g. T4, A100 |
sandbox.cpu_cores / sandbox.ram_gb | Modal size |
sandbox.cpu_image / sandbox.gpu_image | Base images |
Related
Last updated on