Record 084 · Tooling · 2024
50%24.84.F
Unsloth
Unsloth AI
“A fine-tune stack that made QLoRA fast enough to iterate on a single box.”
Unsloth's kernels cut VRAM and time for local adapter training. The homelab fine-tune default of 2024–25.
Made local fine-tuning an afternoon, not a weekend.
- Fused kernels
- Lower-VRAM SFT
- Notebook-friendly
Faster LoRA, less VRAM.
