Fine-tuning the Model

Single-Card Training

weclone-cli train-sft

For single-card training in a multi-card environment, you first need to run export CUDA_VISIBLE_DEVICES=0

The training script will read the configuration from settings.jsonc and start fine-tuning. Pay attention to the terminal output to see if the loss is decreasing normally.

Multi-Card Training

If you have multiple NVIDIA GPUs and want to perform multi-card training:

1. Install Deepspeed:

   uv pip install deepspeed

2. Configure Deepspeed
   In settings.jsonc, find the deepspeed configuration item and uncomment it or fill in the path to the Deepspeed JSON configuration file as needed.

3. Start multi-card training:

   deepspeed --num_gpus=<number_of_gpus_to_use> weclone/train/train_sft.py

After the training is complete, the fine-tuned LoRA adapter weights will be saved in the output_dir you specified in settings.jsonc.

Enable QLoRA (Optional Configuration)

If you want to further reduce video memory consumption, you can enable QLoRA quantized training.

WARNING

The weights fine-tuned using the quantized model cannot be merged back into the original model.

Add the following configuration to the common_args field in settings.jsonc:

"quantization_bit": 4,
"quantization_type": "nf4",
"double_quantization": true,
"quantization_method": "bitsandbytes"

NOTE

• quantization_bit supports values: 2 / 4 / 8. The lower the value, the more video memory is saved, but the inference speed and effect may be slightly reduced.

• If you encounter the error ImportError: Please install bitsandbytes>=0.45.3, you can try reinstalling bitsandbytes:

  uv pip install bitsandbytes>=0.39.0