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

Latest commit

 

History

History
65 lines (53 loc) · 1.51 KB

Install.md

File metadata and controls

65 lines (53 loc) · 1.51 KB

Install

This guide assumes Linux + NVIDIA GPUs.

Prerequisites

  • CUDA 12.8 (see FA3.md)
  • Conda
  • Python 3.12.9

Environment

conda create -n pollux python=3.12.9
conda activate pollux

PyTorch (CUDA 12.8)

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

Dependencies

pip install xformers ninja packaging
pip install -r requirements.txt

FlashAttention (optional / recommended)

  • For FlashAttention v3 build instructions, see FA3.md.
  • Quick install:
MAX_JOBS=128 python -m pip -v install flash-attn --no-build-isolation

COSMOS Tokenizer (optional)

cd apps/Cosmos-Tokenizer
pip3 install -e .
cd ../..
python apps/main/test_vae.py

CLIP (optional)

pip install git+https://github.com/openai/CLIP.git

Extra deps (optional)

pip install timm torchmetrics

MongoDB tools (optional)

If you need mongoexport / mongoimport, install the MongoDB database tools via your distro package manager.

Example (Ubuntu):

sudo apt-get update
sudo apt-get install mongodb-database-tools

Use environment variables (see .env.sample) and avoid hardcoding credentials:

mongoexport --uri="mongodb+srv://${MONGODB_USER}:${MONGODB_PASSWORD}@${MONGODB_URI}" --db=world_model --collection=COLLECTION --out=out.json --jsonArray
mongoimport --uri="mongodb+srv://${MONGODB_USER}:${MONGODB_PASSWORD}@${MONGODB_URI}" --db=world_model --collection=COLLECTION --file=in.json --jsonArray