Skip to content
OrbitInfraPublic

Repository files navigation

OrbiTE

A transfer engine for LLM inference

Documentation | Quick Start | Architecture

CI Crates.io PyPI Docs

OrbiTE moves KV bytes between GPU memory, DRAM, local SSD and peer hosts over RDMA/GPUDirect, NVLink/MNNVL, TCP, io_uring and GDS, in Rust with a Python API. Memory, registrations and staging stay retained until physical retirement proves the hardware has stopped accessing them. OrbitKV is its first user.

OrbiTE network transports and the implemented DRAM and GPU paths to local SSD

Supported paths

Transport Memory Current qualification
TCP DRAM; CUDA staging DRAM: Linux CI; CUDA: unqualified
HP-TCP DRAM Linux loopback CI
SHM Same-host DRAM Linux CI; CXL unqualified
RDMA DRAM; CUDA direct or staged Unqualified
NVLink / MNNVL CUDA; pinned or host-NUMA DRAM Unqualified
io_uring / GDS DRAM/CUDA ↔ local SSD DRAM io_uring: Linux CI; CUDA/GDS: unqualified

Quick Start

On Linux (Rust 1.85+, or CPython 3.10–3.14), after beta.1 publication:

cargo install orbite --version 0.1.0-beta.1 --locked --features io-uring --example orbite-quickstart
orbite-quickstart
# Or Python, in a virtual environment:
python -m pip install orbite==0.1.0b1
python -I -m orbite.quickstart

Both programs verify two-process TCP READ/WRITE and DRAM↔file io_uring bytes and guards, without a GPU. Linux must permit io_uring; an unavailable kernel is an error. Source and prerequisites.

Status

Beta.1 preparation: CPU paths qualified on Linux CI. Wheels include GPU/RDMA/GDS paths, unqualified on hardware; those paths need the host’s NVIDIA driver, rdma-core and cuFile (limits). APIs may break between betas; the 0.1.x API and wire freeze begins at 0.1.0.

License

Apache-2.0. Contributing. Mooncake TENT is the design reference.

Releases

Packages

Used by

Contributors

Languages