# Packed `unvideo` environment

`unvideo-conda.tar.gz` is produced with `conda-pack` from
`/rczhang/rczhang/conda_packs/unvideo`. It is relocatable and is unpacked into
`.env/unvideo` by `scripts/bootstrap.sh`.

Key validated packages:

- Python 3.10
- PyTorch 2.8.0+cu128
- torchvision 0.23.0+cu128
- DeepSpeed 0.18.9
- Transformers 5.6.0.dev0 (bundled editable source is reinstalled)
- Diffusers 0.38.0.dev0 (bundled editable source is reinstalled)
- PEFT 0.19.1
- Accelerate 1.13.0
- flash-attn 2.8.3
- qwen-vl-utils 0.0.14
- SwanLab 0.9.2

The target host still needs a Linux x86_64 system and an NVIDIA driver new
enough for the CUDA 12.8 runtime. No system CUDA toolkit is required for the
normal prebuilt PyTorch path.