Run

compute run <file.py>::<function> --gpu <sku|policy> [options]

Example (source):

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

Pass a GPU SKU to pin the machine, or pass auto, fastest, or cheapest to route the workload. --yes skips the interactive spend confirmation.

Automatic routing

compute run simple_mlp.py::train --gpu auto
compute run simple_mlp.py::train --gpu fastest
compute run simple_mlp.py::train --gpu cheapest

Policy	Selection objective
`auto`	Balance predicted execution time and provider cost by minimizing the worst relative tradeoff.
`fastest`	Choose the compatible candidate with the lowest predicted execution time. (`fast` is an alias.)
`cheapest`	Choose the compatible candidate with the lowest predicted provider execution cost. (`cheap` is an alias.)

For a policy run, the CLI packages and uploads the source closure and attaches a typed Torch graph when it can capture one. Compute analyzes the workload, filters out incompatible candidates, and scores each remaining provider/GPU pair against a target-specific hardware profile. The preflight quote shows the selected provider and GPU, locked rate, execution estimate, analysis warnings, and projected billed total. The CLI creates no run until you approve that quote or pass --yes. Routing uses execution time and provider execution cost. Queue and machine startup time are not part of the current score. If the analysis cannot establish facts such as accelerator compatibility or a safe memory bound, Compute returns an analysis error instead of guessing. If the API marks a quote as a selector fallback, the CLI refuses it unless you pass --allow-selector-fallback. Automatic routing is opt-in. With no --gpu, Compute uses the decorator’s GPU selection or defaults to hotaisle/MI300X.

Flags

Flag	Meaning
`--gpu`	GPU SKU, `PROVIDER/SKU`, or policy: `auto`, `fastest`, `cheapest` (`fast` and `cheap` accepted as aliases).
`--provider`	Pin a SKU or scope a policy to `runpod`, `hotaisle`, or `vastai`. Required when a SKU name exists on more than one provider.
`--allow-selector-fallback`	Accept a quote that the selector marks as a fallback.
`--wait`	Stream until terminal; print JSON result. Default is detach after create.
`--timeout`	Function timeout in seconds. Default 600 (10 minutes).
`--args`	JSON object for function kwargs.
`--dry-run`	Print the local source and upload plan. No upload, remote analysis, quote, or machine.
`--yes`	Skip interactive confirmation.
`--json`	(global) machine-readable envelope.

compute run --help is the live flag list.

What happens

The CLI analyzes and packages the target file and its local imports.
It uploads the payload and requests a rate-locked quote.
For a routing policy, Compute analyzes the workload and selects a compatible provider/GPU pair.
The CLI shows the preflight quote and asks for confirmation.
The API creates the run if admission and credit allow it.
A worker provisions a new machine, installs the image, runs the function, streams logs, then terminates the machine.
You get a JSON return value, a receipt, and an empty compute machines list.

Provider vs SKU

An explicit SKU pins the hardware. Compute resolves its provider unless the SKU is ambiguous, in which case you must pass --provider or use PROVIDER/SKU. A provider can also scope a routing policy:

compute run simple_mlp.py::train --gpu cheapest --provider vastai

Use --provider for a policy rather than writing vastai/cheapest. You do not pick a region or an instance template.

Detached vs wait

# prints run_id and exits
compute run simple_mlp.py::train --gpu H100-SXM --yes

# later
compute runs status <run_id>
compute logs <run_id> -f
compute runs result <run_id>
compute runs receipt <run_id>

simple_mlp.py is in First run. The homepage workloads are Fine-tune, Reinforcement learning, and Batch inference. Or do it in one shot with --wait.

Inspect

compute runs list
compute runs status <run_id>
compute runs result <run_id>
compute runs receipt <run_id>

result is the function return value. receipt is billed minutes and the platform fee (see Billing).

Overview

Get started

Tutorials

Guides

CLI

Reference

Tinker training

Automatic routing

Flags

What happens

Provider vs SKU

Detached vs wait

Inspect

​Automatic routing

​Flags

​What happens

​Provider vs SKU

​Detached vs wait

​Inspect

Automatic routing

Flags

What happens

Provider vs SKU

Detached vs wait

Inspect