TunePlane

TunePlane — a control plane for LLM post-training: environments, train, evaluate, playground, deploy, with reflow closing the loop

uv tool install tuneplane
tp init my-lab && cd my-lab
tp new my-grpo --method nemo-rl/grpo
tp submit my-grpo --profile h200:8

Four commands from an empty directory to a GRPO run on a GPU cluster. The loss curve streams into a web console. No cluster credential ever leaves the control plane.

Quickstart

Install, log in, submit, and watch the first point land on the chart.

Glossary

Experiment, run, Fleet, profile, Resource Pool — the nouns the CLI and the console use.

What it does

You declare an experiment in YAML. TunePlane validates it on your machine, sends it to a cluster you never log in to, and brings the results back.


Submit	`tp submit` packs the working tree and queues a job. Quota, method version and image allowlist are checked before it reaches hardware.
Watch	Loss, reward, KL, GPU utilisation, live logs and validation samples, in the console while the job runs.
Score	GSM8K, MMLU, C-Eval and others through standard harnesses. Or your own benchmark pack, or an LLM judge scoring against a rubric your team wrote.
Serve	Chat with a checkpoint in the Playground, or promote a version to a Model Deployment with a stable endpoint.
Govern	Quota, schedule windows, an image allowlist, an audit trail, and Volumes a job can read but nobody can download.

What it does not do

TunePlane does not own what runs inside a job. The training framework does, and so does your code.

Trains the model	Runs the job
NeMo-RL, verl, TRL, OpenRLHF, PyTorch, or a `train.sh` of your own	Docker on one host, a pool of registered machines, KubeRay, or Slurm

One JobSpec expresses the intent. The same experiment runs on any of the four backends without a change to your config.

Pick a path

Train a model

Quickstart, then submitting, sweeps and multi-stage pipelines.

Extend the platform

Add a method, a benchmark, an agent environment, a rubric, or a runtime patch.

Run a deployment

Install the console, register a Fleet, set quota, keep it running.

What runs where

	Your machine	The control plane	The cluster
Config validation	✓	✓
Cluster credentials		✓
Object-storage keys		✓
Training code			✓
Plugin code			✓
Dataset contents			✓

The control plane never executes third-party code, and a training job never receives the raw contents of a protected Volume. See extension points for what that means when you add your own code.

Get started

Core concepts

Train a model

Evaluate and serve

Agent environments

Extend the platform

Collaborate

Integrations

Reference

Help

Quickstart

Glossary

What it does

What it does not do

Pick a path

Train a model

Extend the platform

Run a deployment

What runs where

Quickstart

Glossary

​What it does

​What it does not do

​Pick a path

Train a model

Extend the platform

Run a deployment

​What runs where

What it does

What it does not do

Pick a path

What runs where