A Rust-native transfer engine for AI infrastructure.
::: ## Overview OrbitTE provides owned, batched data movement across CPU memory, NVIDIA host staging and network transports. Its design follows Mooncake TENT's resource and runtime boundaries, implemented in Rust without a Mooncake C++ dependency. The engine owns registration, batches and completion. TCP, RDMA and opt-in CUDA IPC are separate payload transports, with a shared control service for peer discovery, grants and receipts. Allocations remain owned until native work has drained. :::{image} _static/architecture.svg :alt: OrbitTE engine and sibling TCP/RDMA transports; dashed paths are planned :class: dark-light ::: ## Getting started Run the [CPU example](getting-started.md), select optional providers with the [build guide](build.md), and use the [Rust API guide](rust-api.md) to register memory and submit transfers. The [RDMA/CUDA guide](rdma-cuda.md) explains device selection, persistent resources, completion and shutdown. ## Development status CPU/TCP, persistent host RDMA, multirail slicing and synchronized NVIDIA host staging with owned copy streams/events, and explicit GPU direct RDMA are implemented. Native qualification applies to the exact profiles in [qualification](qualification.md). The published alpha reserves the crate name; the source implementation remains under development. The host RDMA implementation and declared installed cleanup fixtures are complete. [Physical rail accounting](qualification.md#installed-physical-rail-accounting) verifies concurrent parents, shared QP limits and exact native payload bytes. [Automatic discovery](qualification.md#installed-automatic-discovery) now passes two-host READ/WRITE against a separately captured PCI, NUMA and route inventory. The [P4 exit queue](completion-plan.md#p4-remaining-delivery) tracks physical-fault, fabric qualification before full acceptance. Source and the bounded RoCE profile passed independent review with four fresh live repeats; P5 CUDA work now proceeds with available hardware. The [GPU Direct profile](qualification.md#gpudirect-rdma) passes 72 installed H20/RoCE cases with an explicit host-wide context inventory limit. The cooperative [CUDA IPC/NVLink path](cuda-ipc.md) is implemented, with 48 installed H20 cases passing. Bounded subbatch import reuse passes 28 additional installed cases on two H20 hosts. OrbitKV integration remains deferred. [Capabilities](capabilities.md) distinguishes current paths from later work, including broader GPU/IPC qualification and full TENT compatibility. For code navigation, use the [architecture reading order](architecture.md#source-reading-order). :::{toctree} :caption: Getting Started :maxdepth: 1 getting-started build capabilities ::: :::{toctree} :caption: User Guides :maxdepth: 1 cpu-tcp async-batches rdma-cuda cuda-ipc staging ::: :::{toctree} :caption: Design Documents :maxdepth: 1 architecture tent-parity ::: :::{toctree} :caption: API Reference :maxdepth: 1 rust-api ::: :::{toctree} :caption: Troubleshooting :maxdepth: 1 troubleshooting ::: :::{toctree} :caption: Development and Validation :maxdepth: 1 contributing qualification installed-host native-cleanup cpu-oracle completion-plan :::