Deploy a LoRA

Serve a PEFT LoRA adapter as a private model through the Boardwalk inference API.

Repository requirements

Publish your exported adapter to a Hugging Face repository. Include adapter_config.json and adapter_model.safetensors. Pickle weight files are not supported.

The configuration must identify the base repository and a positive rank. Pin the base commit in revision when available. Boardwalk checks the adapter against that base before activation.

{
  "peft_type": "LORA",
  "task_type": "CAUSAL_LM",
  "base_model_name_or_path": "Qwen/Qwen3-0.6B",
  "r": 16,
  "lora_alpha": 32,
  "target_modules": ["q_proj", "v_proj"],
  "bias": "none",
  "use_dora": false
}

Keep the configuration produced by training; the example shows the expected format. Editing metadata does not make incompatible weights compatible.

Compatibility

Use peft_type: "LORA". Supported target modules are q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, and down_proj.
Use bias: "none", disable DoRA, and leave modules_to_save empty or unset.
Keep vocabulary and special tokens compatible with the base model. Include the intended chat template when it differs.
Base architecture, adapter rank, and runtime compatibility are checked during deployment. A listed public base does not guarantee every adapter configuration is supported.

If your fine-tune changes unsupported modules or token embeddings, export a merged full checkpoint using your training tools.

Deploy the adapter

As an organization owner or admin, open Models → Deploy a model in the console.
Enter the adapter repository and optional branch, tag, or commit. Connect Hugging Face if the adapter or base is private or gated; accept any required license first.
Inspect the repository, choose a model name, and review the context and serving configuration.
Select Deploy. Boardwalk loads the adapter, verifies the configuration, and benchmarks it before activating it.

Deployment qualification is free. The activated model is private to your organization, even when its source repository is public.

Call your model

Copy the model’s Call as value from the console. Use that value as model; no adapter metadata is needed. This example assumes you named it support-agent.

curl https://api.boardwalk.cloud/v1/chat/completions \
  -H "Authorization: Bearer $BOARDWALK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"support-agent","messages":[{"role":"user","content":"Hello"}],"max_completion_tokens":256}'

Use a key with inference:write and add credits before calling it. The receipt records the served revision and rates.

Troubleshooting

Repository access: verify your connected Hugging Face account can read both the adapter and base.
Compatibility check failed: check the base commit, tensor shapes, target modules, rank, and tokenizer. Review the failed check in the deployment details.
Unexpected output: confirm the training base and chat template, then compare a request against your local export.

To publish an update, redeploy the new repository revision. The current revision keeps serving until its replacement passes qualification. See deployment updates.