This is a fork of Microsoft/DCVC focused on a deterministic INT16 runtime, bit-exact CUDA kernel optimizations, and a production-oriented video archive pipeline for DCVC-RT.
The maintained implementation is in DCVC-family/DCVC-RT.
Highlights:
- INT16 feature and weight execution designed for deterministic cross-device coding
- Optimized bit-exact CUDA convolution kernels
- File-level multi-GPU encoding and decoding with one independent worker per GPU
- Direct yt-dlp/FFmpeg streaming without storing source video containers
- Unified encode, decode, viewer, and validated-cleanup CLI
- Multi-channel scheduling with a tmux-friendly terminal dashboard
- Non-blocking cleanup approvals and optional
--auto-delete - Persistent archive manifests, SHA-256 hashes, and restart-safe deletion auditing
- Portable native-extension build script for each target computer
The rest of the repository is retained from upstream so its history, license, and the other DCVC-family implementations remain intact. The original upstream root README is preserved as UPSTREAM_ROOT_README.md.
- Original DCVC/DCVC-RT research and implementation: Microsoft Research and the authors listed in the DCVC-RT paper.
- Fork owner and maintainer: Salatiel Jordão (@QLaHPD).
- Primary implementation contributor for this fork: OpenAI Codex, operating as an AI coding agent under the direction and review of the repository owner.
Codex is credited transparently as an implementation contributor; it is not represented as a GitHub account, legal copyright holder, or independent project maintainer. This fork is not affiliated with or endorsed by Microsoft or OpenAI.
The upstream project is distributed under the MIT License. Existing file-level notices remain in place. Contributions in this fork are distributed under the same repository license.