decord2 — Efficient Video Loader for Deep Learning
Published:
decord2 is a modernized, efficient video loader for deep learning with smart shuffling — a maintained evolution of the deprecated decord library.
Key Features
- Updated for FFMPEG 8 and CUDA 13.2 compatibility
- Full aarch64/SBSA architecture support (Jetson, DGX Spark, ARM servers)
- Handles random access patterns common in neural network training
- Supports both FFMPEG/LibAV and NVIDIA hardware-accelerated video codecs (NVDEC)
- Drop-in replacement for the original decord with modern build system
Adoption
decord2 has been adopted in NVIDIA Isaac-GR00T for improved video loading in robot learning pipelines, replacing the deprecated original library with full ARM and SBSA support.
Impact
- 47+ GitHub stars and growing
- Enables efficient video-based training on edge devices
- Critical infrastructure for Vision-Language-Action (VLA) model training on diverse hardware