Which GPU provider do you use? And what's your setup script?

[ragulpr]

I'm running on prime intellect, it works great but been struggling to re-create the timings as the cuda 12.4 (pytorch/pytorch:2.5.1-cuda12.4.1-cudnn8-runtime ) is the latest available default image. Anyone tried to bring your own image there? Or what do you recommend?

As I can't easily reuse the default image (is it published on dockerhub somewhere?) I'm using the img above with the following setup:

apt-get update
apt-get install nvtop nano vim tree tmux python3-pip python-is-python3 -y
python -m pip install --upgrade pip

git clone https://github.com/ragulpr/modded-nanogpt.git
cd modded-nanogpt
git checkout my-branch
pip install -r requirements.txt
pip install --pre torch==2.7.0.dev20250110+cu124 --index-url https://download.pytorch.org/whl/nightly/cu124 --upgrade
python data/cached_fineweb10B.py 8

nohup ./run.sh > ./logs/my-named-run.log 2>&1 &

Thank you!

[KellerJordan]

lambda labs has 12.6