Training a LLaMA in your Backyard: Fine-tuning Very Large...
summary
Shows how to fine-tune LLaMA 7B on consumer GPUs using LoRA, QLoRA and Hugging Face PEFT/Transformers to cut memory: full fine-tune needs ~112 GB, LoRA reduces that to ~14 GB by training only 0.29% of parameters and saving a ~300 MB adapter, and 4‑bit quantization (QLoRA) drops it further to ~4.5 GB so 12–16 GB or Colab GPUs suffice. Demonstrations include RLHF via DPO producing StackLama, CodeLlama adapted to a private API, and multimodal Elicit 9B improvements.
AI-written summary · about
chapters (17)
- 0:00Fine-tuning LLMs on consumer hardware
- 1:48Memory costs of Llama 7B
- 3:56Scaling in fine-tuning LLMs
- 6:27Parameter-efficient fine-tuning methods
- 7:10LoRA deep dive
- 8:56LoRA memory savings for Llama 7B
- 10:03Memory savings with quantization
- 11:06How 4-bit quantization works
- 13:44QLoRA: LoRA + quantization
- 14:47QLoRA memory savings for Llama 7B
- 15:52LoRA/QLoRA performance impact
- 16:58Real-world Llama fine-tuning apps
- 18:02RLHF fine-tuning with StackLlama
- 19:06Ads generation with Llama 2 7B
- 19:49CodeLlama for proprietary codebases
- 20:50Fine-tuning Pix2Struct
- 21:34Acknowledgments and resources
related talks
29mLeveraging open-source LLMs for production
25mFinetuning Large Language Models (LLMs) for Security Log Detections
32mFine-tune LLMs from Kaggle Models using (Q)LoRA
31mBuilding LLaMA: Meta's Director of GenAI
30mUnderstanding the end-to-end LLM training and inference pipeline
29mRunning Large Language Models on Laptops...
39mThe Future is Fine-Tuned: Deploying Task-specific LLMs
20mSelf-hosting AI LLMs - a beginners guide