LoRA adapters

A LoRA adapter is a small file that changes how a base model answers: your task, your tone, your format. It is far smaller than the model itself, so one endpoint can serve the base model and several adapters at once.

What we accept

• Base model: Qwen3 8B. In adapter_config.json, base_model_name_or_path must be Qwen/Qwen3-8B (or a copy such as unsloth/Qwen3-8B).
• Plain LoRA with rank (r) up to 64, on any of q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj and down_proj. No DoRA, no prompt tuning, no extra trained layers (modules_to_save), no per-layer settings.
• Files: adapter_config.json and adapter_model.safetensors at the top of the repo, up to 1 GB. Adapters saved only as .bin files are refused; PEFT's save_pretrained writes safetensors by default.

Register an adapter

1. Open Adapters in the dashboard and select Register adapter.
2. Choose where it is:
   • On Hugging Face: type hf:owner/repo or paste the repo page.
   • Upload a folder: choose the folder PEFT saved. Only adapter_config.json and adapter_model.safetensors are sent; notes such as README.md stay on your computer. A folder with .bin or other weight files is refused before anything uploads.
3. Add a label if you like, then select Register adapter (or Upload and register). We check the config first, then copy the files into your storage, usually in under a minute. The row shows Ready when it is done.

We use the repo's latest version at that moment and keep it fixed. Private and gated repos need your Hugging Face token, see Gated and private repos. An adapter counts toward your storage like an upload, see Model storage. An account holds up to 10 adapters.

Add adapters to an endpoint

Adapters run on your own endpoint deployed from hf:Qwen/Qwen3-8B-FP8 or hf:Qwen/Qwen3-8B (served by vLLM).

• New endpoint: after Check model, choose up to 4 ready adapters under Adapters.
• Existing endpoint: select Performance, choose adapters under Settings, then Review changes and Apply settings. The endpoint restarts on its next request with the new set.

Their memory comes out of the endpoint's memory length, so a long context may get a little shorter. See Performance settings.

Call an adapter

Send the adapter's id (it starts with lora-) as model to your endpoint's URL. Any other model name answers with the base Qwen3 8B.

curl https://<endpoint-id>.endpoint.bitsrun.com/v1/chat/completions \
  -H "Authorization: Bearer $BITSRUN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "lora-<adapter-id>", "messages": [{"role": "user", "content": "Hello"}]}'

The Adapters page shows this sample with your own endpoint URL and adapter id.

Price

Adapters on your endpoint cost nothing extra: you pay the endpoint's GPU price by the second while it runs, as always.

Delete an adapter

Remove it from every endpoint that uses it first (in Performance), then select Delete on its row. Its files leave your storage at once.

What can go wrong