All current OrbiTE GPU, RDMA, GDS and installed profiles are unqualified. CPU/TCP, io_uring and CUDA-shim controls are CI evidence. Historical TENT acceptances retain their original releases and scope in performance; none carries over to the current data path.
| Engine | Current package/source pin | Implemented boundary |
|---|---|---|
| vLLM | 0.33.0 | KV Connector, V1 runner by default; single attention group or Full + SWA; V2 opt-in admits uniform Full attention only and remains unqualified; recurrent, zeroing and V2 multi-group rejected |
| SGLang | 0.5.23 | UnifiedRadixCache external linker, Full/SWA and static Mamba checkpoints; MLA per-rank; replay, draft, moving pools and combined Full/SWA/Mamba serving rejected; private pre-capture and admission hooks remain |
The pins select adapter source contracts; installed GPU conformance remains queued in M1. Follow adapter contracts and model recipes, not upstream model support alone, when selecting a layout.
| Profile | Implementation and limit |
|---|---|
| Local DRAM / SSD | Implemented; default SSD is recreated at startup. Opt-in persistent SSD has CPU crash controls; installed NVMe restart serving is unqualified |
| Shared Manager | Local engines can share capacity; fairness and tail isolation remain open |
| Peer cache | Inventory, grants and OrbiTE TCP/RDMA; current hardware recovery and cache HA unqualified |
| Metadata scaling | Four-owner scope; historical isolation FAIL and sixteen-owner UNQUALIFIED remain unchanged |
| Graph / TP / PP / EP | Some layout, readiness and rank-control paths exist; each installed shape/rank profile needs its own evidence |
| Native P/D plus cache | vLLM NIXL/MultiConnector or SGLang native disaggregation; cancellation, peer loss, delayed ACK and page reuse unqualified |
| Containers | Shared socket, IPC resources, credentials and process visibility required; separate-container use unqualified |
| GDS | Implemented opt-in; io_uring or compatibility-mode bytes do not prove native GDS |
Use deployment, fault gates and P/D composition. Manager sharing uses OrbiTE; live P/D uses the engine's native transport. Immutable local static LoRA participates in cache identity; dynamic adapter mutation and live weight replacement remain unsupported. Different engines do not automatically share compatible KV bytes.
| Runtime | Distribution | Import |
|---|---|---|
| CUDA 13 | orbitkv |
orbitkv |
CUDA 13 is the only release runtime; normally use NVIDIA R580 or newer with toolkit-compatible PTX/NVRTC support. Supported older data-center drivers need the explicit forward-compatibility runtime profile. Version 0.1.0b1 is being prepared for release; build the complete wheel using the quickstart. Upgrade native clients and Manager together. The wheel supplies native binaries; the host supplies compatible Python, CUDA and selected RDMA libraries, and the Manager also requires compatible PyTorch. See release artifact requirements.
By default, SSD files are recreated on Manager startup. --ssd-persistent enables
durable commits and verified recovery. CPU process-crash
controls do not qualify real NVMe serving or power-loss durability.