WeClone's hardware requirements depend on the specific usage scenario:
- basic operation: at least 16GB video memory GPU (e.g. RTX3090) is required when using the default ChatGLM3-6B model + LoRA fine-tuning
- low-profile program: With QLoRA (4-bit quantization), the 7B model requires a minimum of 6GB of video memory, which can be adapted by adjusting parameters such as per_device_train_batch_size
- multi-card program: Supports DeepSpeed multi-card training, 70B large model requires 8 A100s (80GB)
- Substitute program: Try cloud platforms such as Colab or use the API services of the Magic Hitch community when you are low on video memory!
Real-world data: it takes about 4 hours to train 20,000 pieces of data on a consumer RTX4080 (16GB). If doing only inference (no training), 8GB of video memory can barely run the basic functions.
This answer comes from the articleWeClone: training digital doppelgangers with WeChat chats and voicesThe
May not be reproduced without permission:AI productivity tools " Can I run WeClone on a regular home computer? What kind of hardware configuration is required?
