Skip to content
This repository was archived by the owner on Feb 22, 2026. It is now read-only.

DEVELOP.md

Latest commit

 

History

History
136 lines (96 loc) · 3.26 KB

DEVELOP.md

File metadata and controls

136 lines (96 loc) · 3.26 KB

PR first

git checkout -b feature/awesome-feature
git checkout -b fix/awesome-fix

Philosophies for Better Engineering

  • The lingua folder should contain only commonly used classes and functions, avoiding application-specific elements.
  • Each model or pipeline should have its own dedicated folder within the apps directory.
  • Each Pull Request should focus on a specific new feature (e.g., adding CFG support, implementing a new architecture, or introducing new configurations) and include a detailed test plan to facilitate effective code review.

Create User

Example:

sudo adduser mczhuge
sudo usermod -aG sudo mczhuge

Mount JFS

sudo curl -L https://juicefs.com/static/juicefs -o /usr/local/bin/juicefs && sudo chmod +x /usr/local/bin/juicefs && sudo /usr/local/bin/juicefs mount world-model /jfs

For Conda Installation

su - mczhuge
sudo chmod -R u+w $Pollux
sudo chown -R mczhuge $Pollux
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh
source /home/mczhuge/miniconda3/bin/activate

Setup Github SSH

Example:

ssh-keygen -t ed25519 -f ~/.ssh/mczhuge -C "mczhuge@gmail.com"
cat ~/.ssh/mczhuge.pub
chmod 600 ~/.ssh/mczhuge
eval "$(ssh-agent -s)"
ssh-add ~/.ssh/mczhuge
ssh-keyscan -t ed25519 github.com >> ~/.ssh/known_hosts
ssh -T git@github.com

NVIDIA-Fabric Issue

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/nvidia-fabricmanager-535_535.183.01-1_amd64.deb
sudo dpkg -i nvidia-fabricmanager-535_535.183.01-1_amd64.deb
sudo systemctl daemon-reload
sudo systemctl start nvidia-fabricmanager

Login Setup

sudo mkdir -p /home/mczhuge/.ssh
sudo chmod 700 /home/mczhuge/.ssh
sudo chown mczhuge:mczhuge /home/mczhuge/.ssh
sudo cp /home/ubuntu/.ssh/authorized_keys /home/mczhuge/.ssh/
sudo chmod 600 /home/mczhuge/.ssh/authorized_keys
sudo chown mczhuge:mczhuge /home/mczhuge/.ssh/authorized_keys
sudo vim /etc/ssh/sshd_config
# add
PermitRootLogin no
PubkeyAuthentication yes
PasswordAuthentication no
AllowUsers ubuntu mczhuge
sudo systemctl restart sshd

Local Run

export PYTHONPATH=/jfs/mczhuge/Pollux:$PYTHONPATH
python -m apps.main.train config=apps/main/configs/pollux_v0.5.yaml

Huggingface Credentials

To access models like Llama or FLUX.1-dev, authenticate with Huggingface by following the steps below:

huggingface-cli login
git config --global credential.helper store

Wandb Integration

If you need access to the Wandb team, please contact Mingchen for an invitation.

  1. Obtain your Wandb API key from this link.
  2. Log in using the following command:
wandb login
  1. Update your configuration file (e.g., apps/main/configs/LLAMA_Baseline_1B.yaml) with the relevant Wandb settings:
# Update the `name` field for each new run. The `dump_dir` will be auto-generated.
name: "ImageNet_1B_BaseLine_256_Flux_LLAMA_Pre_Train_MC" # Ensure to modify this for each run

Training

Use the following command to train the model on 4 GPUs:

torchrun --standalone --nnodes 1 --nproc-per-node 4 -m apps.main.train config=apps/main/configs/pollux_v0.5.yaml