Run a LoRA fine-tune with LLaMA Factory on CCI

LLaMA Factory is an open-source low-code framework for LLM fine-tuning that ships with most mainstream techniques. This tutorial walks through deploying LLaMA Factory inside an Alaya NeW CCI instance and running LoRA SFT on llama3-8b-instruct.

Prerequisites

• Alaya NeW enterprise account; if not registered, see account registration.
• Sufficient balance to cover an H800A × 1 CCI instance. For pricing details, contact us.

Step 1: Create a CCI instance

1. Sign in to Alaya NeW and go to Product → Compute → Cloud Container Instance.

2. Click New Cloud Container and configure the instance:

   • Resource type: pick Cloud Container Instance — GPU — H800A — 1 card
   • Reference table:

   Field	Description	Requirement	Required
   Instance name	Unique identifier	Letter-prefixed; letters / digits / - / _; 4–20 chars	Yes
   Description	Free-form notes	—	No
   Region	Data center	e.g. Beijing-3, Beijing-5	Yes
   Billing	Currently pay-as-you-go only	—	Yes
   Resources	Resource type / GPU / CPU / disk	As needed	Yes
   Storage	Mount NAS	Optional	No
   Image	Public (base / app) or private	—	Yes
   Other	Env vars, auto-stop, auto-release	—	No

3. Click Activate, confirm in the dialog, and click OK.

4. Wait until the instance state becomes Running.

Step 2: Install LLaMA Factory

1. In the instance list, click the Web Connect icon to open a shell.

   

2. In the shell:

   git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
   cd LLaMA-Factory
   pip install -e ".[torch,metrics]"

3. Verify:

   llamafactory-cli version

   A version banner means the install succeeded.

   

Step 3: Launch the LLaMA Factory Webui

1. Open a Web Connect shell again (or use SSH; see SSH access).

   

2. Start the Webui:

   cd /app
   GRADIO_SERVER_PORT=9001 llamafactory-cli webui

   

3. In the instance list, click the Open ports icon to grab the external address mapped to port 9001.

   

4. Paste the external address into a browser to open the Webui — pick a model, prepare a dataset, and run LoRA fine-tuning end to end.