Skip to content

Required Hardware Configuration

Running WeClone (especially during the model fine-tuning phase) has high requirements for video memory. It is recommended to use a dedicated GPU device or a cloud GPU rental service. Running on integrated graphics or with only a CPU is not recommended.

The project uses the Qwen2.5-7B-Instruct model by default and employs the LoRA method for fine-tuning, which requires approximately 16GB of video memory.
Additionally, please reserve more than 20GB of hard disk space to store model files, intermediate results, and cached data.

The following table lists the estimated video memory requirements for different model sizes and fine-tuning methods. Multimodal models will occupy more video memory depending on the size and number of images:

Fine-tuning MethodPrecision (bits)7B Model14B Model30B Model70B ModelxB Model
Full (bf16 / fp16)32120GB240GB600GB1200GB18x GB
Full (pure_bf16)1660GB120GB300GB600GB8x GB
Freeze / LoRA / GaLore / APOLLO1616GB32GB64GB160GB2x GB
QLoRA810GB20GB40GB80GBx GB
QLoRA46GB12GB24GB48GBx/2 GB
QLoRA24GB8GB16GB24GBx/4 GB

If you wish to enable the QLoRA fine-tuning method, please refer to the subsequent chapter "Enable QLoRA (Optional Configuration)" to learn how to switch the fine-tuning strategy.

WARNING

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

TIP

Video Memory ≥16GB: The default LoRA fine-tuning scheme is recommended. Video Memory <16GB: Consider switching to QLoRA or choosing a model with a smaller number of parameters.