Documentation
Index of docs.tractionlayer.ai plus the LTS-1 compatibility matrix: accelerators, CUDA lanes, Kubernetes, OpenShift, operating systems and model sources.
Index of docs.tractionlayer.ai plus the LTS-1 compatibility matrix: accelerators, CUDA lanes, Kubernetes, OpenShift, operating systems and model sources.
Hardening checklist
[Description]
Hardening checklist[docs URL][Doc title]
[Description]
[Doc title][docs URL][Doc title]
[Description]
[Doc title][docs URL][Doc title]
[Description]
[Doc title][docs URL]
LTS-1, confirmed at GA
| Area | LTS-1 |
|---|---|
| Upstream base | SGLang v0.5.20, with published backports |
| Accelerators validated | NVIDIA H100, L40S, A100, B200 |
| Other accelerators | AMD ROCm builds on request under a paid engagement |
| CUDA | 12.x lane and 13 lane, both maintained for the life of the branch |
| Platforms | Amazon EKS; Red Hat OpenShift; self-managed Kubernetes; bare metal (RHEL 9, Ubuntu 22.04 and 24.04); Docker Compose for single node |
| Air-gapped | Offline bundle with mirrored registry and signatures; license verified offline with a grace period |
| APIs | OpenAI-compatible and native SGLang HTTP APIs, unchanged from upstream |
| Identity | API keys on admin routes; Inference Reliability behind your OIDC provider with viewer, operator and admin rolesIn development |
| Observability | Upstream Prometheus metrics; Inference Reliability derived metrics bridged to Prometheus and OpenTelemetryIn development |
| Distribution | Authenticated registry, public mirror, offline bundle |
| Kubernetes | [Kubernetes versions] |
| OpenShift | [OpenShift versions] |
| Model sources | [Model sources] |