4 ms·
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
by danielhanchen 3y 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 https://github.com/unslothai/unsloth for full benchmarks!