|
|
|
|
|
|
by danielhanchen
806 days ago
|
|
|
A bit late, but Unsloth makes LoRA / QLoRA finetuning 2x faster and reduces VRAM by 80% with 0% degradation in accuracy! (no approximations are done!) Mistral 7b is 2x faster than HuggingFace + Flash Attention 2.
Gemma 7b is 2.4x faster than HF + FA2. Check out https://github.com/unslothai/unsloth for full benchmarks! |
|
|