bestpredicts / TPU-Alignment

Fully fine-tune large models like Mistral, Llama-2-13B, or Qwen-14B completely for free

TPU-Alignment

Fully fine-tune large models like Mistral-7B, Llama-2-13B,Phi-2 or Qwen-14B completely for free.

The code in this repository have powered some larger models on my Hugging Face profile. I recommend using TPUs provided for free by Kaggle, they are strong enough to train models of up to 7 billion parameters without freezing parameters. SPMD is used as a parallelization technique for high MXU efficiency while training.

Not every model architecture is supported by this code, here's a complete list of models supported:

• llama
• mistral
• Phi
• gpt2? (untested but should work)
• gptneox
• qwen2
• t5
• mixtral (untested but should work)

I'm open to contributions that propose additional model architectures.

Getting started

1. Head on over to Kaggle, make sure you have verified your account with a phone number, and create a new notebook. Select TPU VM v3-8 as the accelerator.
2. Import the notebook Fine-Tuning LLM on TPU.ipynb into the kaggle notebook.
3. Modify the notebook to your needs. Make sure to provide your hugging face write access token in the Kaggle secrets.

Wandb Support

Model:Phi2

Dataset:- Magicoder-oss-instruct

Effective batch size = 64(8*8)