Serve

Publishing adapters to Hugging Face

Upload a trained adapter and a model card to your own Hugging Face repository, private by default.

You can upload a training run's adapter to a model repository on your organization's Hugging Face account. Anyone who can read the repository can then load the adapter with PEFT on the same base model, in any tool.

Only the adapter and a short model card are uploaded. Nothing about your dataset leaves Tensorant: no dataset name, no examples, no scores.

To serve the adapter to your applications through Tensorant instead, publish it as a model.

Before you start

A run with a verified adapter, showing Adapter ready or Completed. See Training runs.
A Hugging Face token with write access, added by an owner under Hugging Face in Connections. It must be allowed to write to the account or Hugging Face organization you publish to. A read-only token is refused.
The member or owner role. See roles.

Publish an adapter

Open the run in Runs and choose Publish to Hugging Face.
Enter the Repository as owner/name. It suggests your token's account and the base model's name with -lora, for example acme/Qwen2.5-1.5B-Instruct-lora.
Keep Private repository ticked for a private repository, or untick it for a public one.
Choose Publish.

The upload runs in the background: follow it with the Activity button. When it completes, the run shows Published · repository, linking to it on Hugging Face.

Before uploading, Tensorant checks that the adapter on your network volume is still the one the run verified. A changed or missing adapter is not published.

What is uploaded

One commit adds:

File	What it is
`adapter_config.json`	The adapter's configuration
`adapter_model.safetensors`	The adapter's weights
`README.md`	A model card, only if the repository has none

The model card is tagged lora and peft. It links the base model and the exact revision trained on, lists the method (LoRA or QLoRA), LoRA rank and alpha, epochs, learning rate and context length, and ends with a "Use it" example that loads the adapter with PEFT.

The repository

Name: owner/name. Each part has up to 96 letters, digits, underscores, dots or hyphens, starts with a letter, digit or underscore, and does not end with a dot or hyphen. --, .. and a trailing .git are not allowed.
A new repository is created as a model repository, private if Private repository is ticked.
An existing repository keeps its visibility. If it is public and Private repository is ticked, publishing is refused, so you never publish a private adapter to a public repository by mistake. Untick Private repository to publish there on purpose.
Publishing again to the same repository adds a new commit.

Limits

Limit	Value
Repository name	3 to 200 characters, up to 96 per part
Uploads of one run at a time	1
Attempts	Up to 3 when Hugging Face fails. A refused token or repository is not retried.

When something goes wrong

A problem found before the upload shows in the dialog. Otherwise it shows on the upload in Activity.

Message	What to do
"Enter a Hugging Face model repository such as your-name/adapter-name."	Enter `owner/name`. See The repository.
"Add a Hugging Face token with write access…"	An owner adds a token in Connections.
"The Hugging Face token was refused…"	The token was deleted or expired. Create a new one and replace it in Connections.
"This Hugging Face token can only read…"	Replace it with a token that has write access.
"This token cannot write to repository…"	Choose a repository under the token's account, or give the token write access on Hugging Face.
"repository is a public repository…"	Choose another repository, or untick Private repository.
"This adapter is being published to repository…"	Wait for that upload to finish in Activity.
"Publishing needs your organization's own Hugging Face token…"	Your organization uses connections managed by Tensorant. Download the adapter from the run and upload it yourself.
"The adapter is missing from storage…" or "The adapter in storage is not the one this run verified…"	The adapter file on your volume is gone or changed. Train again, or upload a copy you downloaded earlier yourself.
"Storage configuration changed…"	Connect the network volume the run was trained on.
"Hugging Face did not accept the adapter (HTTP code)…"	Hugging Face had a problem. Try again later.
"Your role in this organization cannot make changes"	You are a viewer. Ask an owner for the member role.