WikiofAuto
frameworksUpdated: 2026-11-01

Unsloth

আনস্লথ (Unsloth)

Unsloth is an open-source LLM fine-tuning engine written in OpenAI Triton kernels that speeds up LoRA/QLoRA and GRPO training by 2x while reducing GPU VRAM usage by up to 68%.

Overview

Unsloth replaces standard PyTorch autograd passes for RoPE embeddings, RMSNorm, cross-entropy loss, and MLP layers with hand-written OpenAI Triton kernels. This allows developers to fine-tune 8B–14B models on a single free 15GB Google Colab T4 GPU and export quantized Q4_K_M GGUF files directly to Ollama.