nanogptrs

nanogpt (RS)

Ceci est une mise en œuvre de la rouille du modèle Nanogpt de la vidéo YT d'Andrej Karpathy: https://www.youtube.com/watch?v=kcc8fmeb1ny&t=12s

Avec l'aide de: https://github.com/laurentmazare/tch-rs/blob/main/examples/min-gpt/main.rs et https://github.com/karpathy/nanogpt/blob/master/model.py.

Installation

Créer un environnement Micromamba (ou conda):

micromamba env create -f environment.yml

Activer l'environnement:

micromamba activate nanogptrs

 export LD_LIBRARY_PATH=/opt/conda/lib/python3.10/site-packages/torch/lib/: $LD_LIBRARY_PATH

Run - Générer à partir de GPT-2 pré-entraîné

./data/download.sh
./models/download.sh gpt2
cargo run --release -- --device=cuda --restore-from models/gpt2/model.safetensors generate --max-len 32 --prompt " Once upon a time " gpt2

Run - Train Nano-GPT

cargo run --release -- --device=cuda train --n-epochs=3 --final-checkpoint-path=models/nanogptrs.safetensors nano-gpt

Devrait éventuellement (~ 5h sur mon Titan XP) produire quelque chose comme ceci:

 DUCHESS OF YORK:
Here comes already.

EXTOLY:
O, by the means of your crown?

KING HENRY VI:
Brother, that my lord, change thou givest queen.

KING RICHARD II:
Mine honour, because I am advertised
The queen our is not your voice. Would thy sight
Next Rome, among, insible express to dictliffe:
For ere for goings
Abova drunking redel her food pain soul to every it.

QUEEN MARGARET:
I took! O, if you so, good and the Montague of slave,
That he's breathing which holy a holy brats.