Skip to main content
Compute runs a Python function on a newly provisioned cloud GPU, streams output to your terminal, returns JSON, and confirms the machine is gone. You do not manage images, keep-alive pools, or a login to the box. The machine exists for one run.

​
What you do

Save simple_mlp.py first, or copy it from First run. --yes skips the interactive spend confirmation; the run still burns prepaid credit. Use --dry-run in place of --wait --yes to print the upload plan without creating a machine. --gpu H100 is an alias for reserved H100-SXM (RunPod Secure). Cheaper interruptible cards: --provider vastai --gpu RTX-4090. Live names: compute gpu list. The three-command funnel (install → setup → run) also lives on compute.cx/quickstart. This site is the manual.

Use with an agent

Point Cursor, Claude Code, or Codex at https://compute.cx/SKILL.md, or add the MCP server. --dry-run creates no machine; a real run spends prepaid credit.

​
What is live

Compute is in early access. Up to two active runs per account and spend limits may refuse a run; provider capacity is not guaranteed. Different accounts are not serialized by a Compute-wide machine slot.

Tinker-compatible training API

Fine-tune Qwen3-4B with the official Tinker SDK on a dedicated H100 or MI300X. The private-alpha guide covers the one-command lifecycle, safety caps, and checkpoint harvest.

​
What this manual does not cover

These exist in design or in the binary and are not documented as customer-ready here:
  • pip install compute (v0.2)
  • Detached SDK handles that you wait on later
  • Persistent disks
  • Checkpoint resume
  • Community (RunPod) SKUs beyond the reserved Secure H100
  • Multi-GPU shapes
Rates and the platform fee are on Pricing. Use compute gpu list for the locked list rate on a SKU. Do not hard-code a fee percentage from this page.

​
Next

Install

Isolated venv, compute on PATH.

First run

Tiny MLP on H100-SXM.

Agent Skill

Spend-aware workflow for coding agents. MCP: connect the server.

​
Workload guides

These match the three cards on compute.cx. Each page has the file the command expects.

Fine-tune a model

Run LoRA inside a Compute function, or use the managed Tinker API.

Reinforcement learning

rl.py::train on MI300X.

Batch inference

batch_infer.py::generate on MI300X.