mvp-dataset is a minimal, high-performance data loading library for multimodal training pipelines, with local tar and JSONL support, deterministic sharding, and PyTorch loader integration.
mvp-orbit is a small HTTP-based remote execution loop for preparing code on one machine, sending it to another, executing commands there, and streaming output back immediately.
codec-patch-stream provides native video decoding and energy-based patch selection, with separate decode and process backends for CPU and GPU pipelines.
