Implementation:Zai org CogVideo SAT Requirements Install
Metadata
| Field | Value |
|---|---|
| Page Type | Implementation (External Tool Doc) |
| Knowledge Sources | CogVideo, SwissArmyTransformer |
| Domains | Environment, Training_Infrastructure |
| Last Updated | 2026-02-10 00:00 GMT |
Overview
Concrete tool for installing SAT framework dependencies via pip provided by the project requirements file at sat/requirements.txt.
Description
The sat/requirements.txt file defines all Python package dependencies required for the SAT-based CogVideoX fine-tuning pipeline. Running pip install -r sat/requirements.txt installs these packages into the active Python environment, establishing all necessary libraries for data loading, model construction, distributed training, and experiment tracking.
The requirements file uses minimum version constraints (>=) to ensure compatibility while allowing newer patch releases that do not break the API contract.
Usage
Run this installation step once per environment before any SAT-based training or inference workflow. It must be executed from a Python environment that already has PyTorch and CUDA properly installed.
Code Reference
Source Location
sat/requirements.txt:L1-11
Command
pip install -r sat/requirements.txt
I/O Contract
Inputs
| Input | Type | Description |
|---|---|---|
sat/requirements.txt |
File | The pip requirements file containing dependency specifications with minimum version constraints. |
Outputs
| Output | Type | Description |
|---|---|---|
| Installed packages | Python environment | All SAT dependencies installed and available for import: SwissArmyTransformer, omegaconf, pytorch_lightning, kornia, beartype, fsspec, safetensors, scipy, decord, wandb, deepspeed. |
Key Dependencies
| Package | Minimum Version | Purpose |
|---|---|---|
SwissArmyTransformer |
0.4.12 | Core training framework with model parallelism and DeepSpeed integration. |
omegaconf |
2.3.0 | Hierarchical YAML configuration parsing and merging. |
pytorch_lightning |
2.4.0 | Training utility abstractions. |
kornia |
0.7.3 | Differentiable computer vision transforms. |
beartype |
0.19.0 | Runtime type checking. |
fsspec |
2024.2.0 | Unified filesystem interface for local and remote access. |
safetensors |
0.4.5 | Safe tensor serialization for LoRA weight export. |
scipy |
1.14.1 | Scientific computing utilities. |
decord |
0.6.0 | High-performance video frame decoding. |
wandb |
0.18.5 | Experiment tracking and visualization logging. |
deepspeed |
0.15.3 | Distributed training optimization with ZeRO partitioning. |
Usage Examples
Standard Installation
# From the repository root
pip install -r sat/requirements.txt
Installation in a Virtual Environment
python -m venv cogvideo_env
source cogvideo_env/bin/activate
pip install torch torchvision # Install PyTorch first with appropriate CUDA version
pip install -r sat/requirements.txt