litgpt

litgpt copied to clipboard

This is an attempt to implement full finetuning (as opposed to efficient finetuning via adapter etc.)

A script for full finetuning (updating all layers).

Todos

• [x] Create finetune/full.py script
• [x] Add how-to guide
• [x] Create generate/full.py script

Well, maybe ignore the last commit. Lazy loading works now when you use 1 device, but it now fails when using multiple devices and deepspeed.

The previous commit without lazy loading works with deepspeed and multiple devices. (It needs deepspeed stage 3 and at least 6 GPUs for the finetuning, but it works!)

Maybe that issue with deepspeed & lazy-loading can be circumvented when #118 is finished