Skip to main content

​
Why not just run ray job submit, kubectl or sbatch?

You can, if you have the credentials. TunePlane holds them instead, and in exchange the run gets quota enforcement, an audit trail, one place its metrics land, and a record of the commit it came from. The only thing the proxied path takes away is the ability to skip a quota.

​
Why does a submission need a clean working tree?

Every run records its commit, a config snapshot and the recipe digest. Three months later, that is the only reliable way to reproduce the good result.
is there for when you accept the cost. A project with no git repository is not blocked at all: it submits with no provenance recorded. The clean-tree rule is about a repository that could answer the question and was not asked to.

​
Can the lock file upgrade itself?

No. An upgrade can change entrypoints, defaults and images, so it has to be a decision somebody made.
compresses that decision into one flag.

​
Which framework versions can I use?

Each method lists the versions your deployment has published. A new upstream release enters the catalogue as a manifest and an image, usually with no platform code change — see upgrades.

​
How do I run my own framework or private code?

custom/custom with an allowlisted image and a train.sh in the experiment. The image needs Python, your framework and your dependencies — not the CLI. The published custom recipe declares external observability, so pass --observability-url and install the tuneplane wheel if you want console charts. See custom training.

​
Where should training data live?

Three options, in order of preference: For material a job may read but nobody may download, use a Volume.

​
How is GPU time counted?

Every job records gpu_seconds, which is wall time holding cards times the card count. Queued time is free; counting starts when the container is up. It shows on the console’s Usage page.

​
Can I work on a machine with no GPU?

Yes, for everything except training. tp init, tp new, tp validate, tp methods and tp job logs are all local or read-only. A whole submit-to-ingest loop runs without a GPU too, with scheduling and quota semantics intact — see smoke-testing.

​
The console does not match this documentation

Several features are switched on per deployment: the Hugging Face integration, the Playground, the plugin shelf, MCP, teams and approvals. A missing sidebar entry means your deployment has not enabled it. Ask an administrator.

​
Why is my job’s team not the team I am in?

A run’s team comes from its project, not from you. A person can be in several teams, so charging a run to all of them would count it several times. See teams.

​
Next

Troubleshooting · Glossary · Job states