GLAM: Global-Local Variation Awareness in Mamba-based World Model

1. Prerequisites

• Linux (recommended)
• Conda (Miniconda or Anaconda)
• NVIDIA GPU + CUDA driver (recommended for training)

2. Create Conda Environment

conda create -n glam python=3.10 -y
conda activate glam

3. Install PyTorch

Choose one command based on your CUDA version.

CUDA 11.8:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

4. Install GLAM Dependencies

Run from the project root:

pip install -r requirements.txt

If mamba-ssm installation fails, first confirm that your PyTorch/CUDA versions match, then retry.

5. Quick Start

Training example:

python train.py \
  -suite atari \
  -env_name RoadRunner \
  -seed 1 \
  -base_model Glam \
  -version 1_1 \
  -config_path config_files/Glam.yaml \
  -cuda_device 0

6. Open-Source Notes

• The current codebase uses both gymnasium and gym; for compatibility, keep both installed.
• If you need a strict lock file for reproducibility, export one separately:

pip freeze > requirements-lock.txt