Dragonfly, a CNCF Graduated P2P file distribution system, now natively supports hf:// and modelscope:// URL protocols for downloading AI models directly from Hugging Face and ModelScope hubs. Instead of each GPU node independently downloading large models (e.g., a 130 GB DeepSeek-V3 across 200 nodes = 26 TB of origin traffic), Dragonfly's seed peer fetches the model once and distributes pieces across the P2P mesh, reducing origin traffic by ~99.5%. The new backends are implemented in Rust using a pluggable Backend trait, support authentication tokens, revision pinning, recursive repository downloads, and single-file downloads. Key use cases include multi-node GPU cluster deployments, CI/CD ML pipelines, air-gapped environments, and dataset distribution for training jobs. The feature is available via dfget CLI and integrates natively with Kubernetes via Helm.

12m read timeFrom cncf.io
Post cover image
Table of contents
The problem: AI model distribution is broken at scaleWhat Is Dragonfly?Introducing native model hub protocols in DragonflyThe hf:// Protocol — Hugging Face hubThe modelscope:// Protocol — ModelScope hubUnder the hood: Technical deep diveReal-world impact: Where this mattersComparison: Why not just use platform CLIs?Getting startedWhat’s nextContributingConclusion

Sort: