DALI

Why we included this project

When you train or serve models on NVIDIA GPUs, the data pipeline is often what actually holds things back: CPU-side decoding, resizing, and augmentation can't feed the GPU fast enough. DALI moves that work onto the GPU, using a set of optimized operators and its own execution engine to handle prefetching, parallel execution, and batching. It drops into PyTorch, TensorFlow, PaddlePaddle, and JAX as a replacement for their built-in data loaders, so you can describe a pipeline declaratively and swap it into an existing training loop without rewriting model code. Teams whose image, video, or audio workloads stall on I/O or augmentation will find it a practical way to raise throughput, and it also connects to Triton Inference Server for serving. If your training runs are starved for data, test it against your own pipeline before adding more CPU workers.