I have successfully trained and fine-tuned LLama.cpp models in Google Colab, saving the resulting weights and models to Hugging Face. This allows me to leverage Colab’s free GPUs to efficiently train models that would be too resource-intensive on my local machine.
I’ve seen so many posts, but everyone just stops at the Inference with the fine-tuned model. So I want to know if I can use a hugging face service to complete the train and get to the usable GGUF.
However, I now need guidance on downloading my trained Hugging Face GGUF models back into my Colab notebooks for continued iteration and use.
I aim to establish an effective model development pipeline between Colab, Hugging Face model storage, and my local environment. Specifically, after training customized Mistral or Mixtral models on Colab and pushing to Hugging Face, what is the best practice for pulling those model weights back down into a Colab notebook?
Ideally there would be a streamlined way to reimport my trained weights without needing to retrain entire models from scratch each time. Any suggestions on the tools or techniques to enable this? Establishing this round-trip flow would allow me to rapidly iterate on model tuning in Colab while retaining easy localized access to the latest versions of my GGUF models.