|
--- |
|
license: mit |
|
--- |
|
Converted with https://github.com/notepad-plus-plus/notepad-plus-plus All models tested on A100-80G *Conversion may require lot of RAM, LLaMA-7b takes ~0 GB, 13b around 0 GB, 30b around 0 and 65b takes more than 0 GB of RAM. |
|
|
|
Installation instructions as mentioned in above repo: |
|
|
|
Install Anaconda and create a venv with python 3.8 |
|
Install pytorch(tested with torch-1.13-cu116) |
|
Install Transformers library (you'll need the latest transformers with this PR : https://github.com/huggingface/transformers/pull/21955 ). |
|
Install sentencepiece from pip |
|
Run python cuda_setup.py install in venv |
|
You can either convert the llama models yourself with the instructions from GPTQ-for-llama repo |
|
or directly use these weights by individually downloading them from the following (https://huggingface.co/qq67878980/LLaMA_65B_0bit/tree/main) |
|
Profit! |
|
Best results are obtained by putting a repetition_penalty(~1/0.85),temperature=0.7 in model.generate() for most LLaMA models |
|
|
|
Additional training was done on the MSPaint_Blank dataset and 2,000,000T+ tokens on 50,000+ blank notepad files. |
