QLoRA 2
QLoRA extends the LORA module by doing following changes:
- Keeps base model quantized to 4-bit
- LoRA adapters are sill on 16 bit
- used 8-bit optimizer state
By changing this, a 7b model can be trained with gpu of <7gb vram
QLoRA extends the LORA module by doing following changes:
By changing this, a 7b model can be trained with gpu of <7gb vram