Installation¶
Prerequisites¶
- NVIDIA GPU and driver support for CUDA 13.
condaor miniforge,git, and a Hugging Face account/token.- Cosmos3-Super 64B uses 4xB200. LTX-2.3 and SANA-Video use 1xB200.
Environment¶
git clone https://github.com/NVlabs/Sol-Video-Inference-Engine.git Sol-Video-Inference-Engine
cd Sol-Video-Inference-Engine
PYTHON_VERSION=3.12 bash scripts/create_code_conda_env.sh
source "$PWD/scripts/use_code_storage_env.sh"
conda activate "$PWD/.conda/ltx23"
uv pip install -e "$PWD/python[diffusion]" --prerelease=allow
PYTHON_BIN=.conda/ltx23/bin/python bash scripts/postinstall_cuda_jit.sh
scripts/create_code_conda_env.sh creates .conda/ltx23 by default, installs
pip and uv, and installs conda activation hooks for scripts/use_code_storage_env.sh.
The script defaults to Python 3.11 when PYTHON_VERSION is unset; the agent guide
sets PYTHON_VERSION=3.12 for this stack.
Add --with-te to scripts/postinstall_cuda_jit.sh when using the NVFP4 path for Cosmos3-Super or LTX-2.3.
Verify¶
.conda/ltx23/bin/python -c "import torch, diffusers, sglang; print(torch.__version__, diffusers.__version__, torch.cuda.is_available())"
Expected versions are torch 2.11.0+cu130 and diffusers 0.38.0. torch.cuda.is_available() is false on CPU-only setup hosts.
Model downloads¶
export HF_HOME="$PWD/.hf_cache"
huggingface-cli login
huggingface-cli download nvidia/Cosmos3-Super
huggingface-cli download Lightricks/LTX-2.3
huggingface-cli download Efficient-Large-Model/SANA-Video_2B_480p_diffusers
Convenience scripts are available under scripts/cosmos/, scripts/ltx/, and scripts/sana/ for resumable or Slurm-based downloads.