SGLang that tunes itself, with long-term releases you can run anywhere.
Pinned long-term-support branches of the SGLang inference engine, with security fixes backported to the branch you run, a hardened configuration, signed reproducible builds and vulnerability response. A runtime tuner that adjusts the scheduler to your live workload so the GPUs you already own stay busy. Support from the engineers who build the branches. It runs where your data lives: your cloud account, your data center, or a site with no network at all.
Built on upstream SGLang (Apache License 2.0). Traction Runtime is an independent downstream distribution; Traction Layer AI is not affiliated with, sponsored by or endorsed by the SGLang project. We send our scheduler and security work upstream and never fork the API; your applications keep talking to the same OpenAI-compatible endpoint.
In development. LTS-1, cut from SGLang v0.5.20, reaches general availability in [Q4 2026]. Evaluation builds on the public mirror at launch.
- distribution
- traction-runtime/sglang
- release
- lts-1 (0.5.20-1)
- upstream base
- sglang v0.5.20, pinned
- backports
- security and stability fixes, 12 months
- cuda
- 12.x and 13 lanes
- validated on
- H100, L40S, A100, B200
- profile
- hardened: pickle off, admin routes keyed, zmq on localhost, non-root
- sbom
- spdx, sha256:[digest]
- build
- reproducible, vulnerability-scanned
- signature
- cosign, keyless (Sigstore)
- backports until
- [LTS-1 EOL date, published at GA]
- advisories
- backported to this branch; tracker status per advisory
verify: cosign verify [registry URL]/runtime/sglang:lts-1
Upstream SGLang ships every few weeks. Your change board meets every quarter.
SGLang is one of the fastest open inference engines, and its maintainers move quickly: by the project's own release history, a new 0.x release lands roughly every two weeks, and fixes land only on the newest one. That cadence is right for a research project. It is unworkable for a team that qualifies a version once and has to run it for a year.
Nothing to pin to.
Upstream fixes go forward only. The version you qualified in March gets no security patches in September. Teams on CUDA 12 have a further problem: upstream retired its CUDA 12 wheels after v0.5.19.
Traction Runtime cuts a long-term-support branch every quarter, backports security and stability fixes for twelve months, publishes the end-of-life date up front, and keeps a CUDA 12 lane alive alongside CUDA 13.
A research-grade security surface.
Open inference engines have carried a steady run of remote-code-execution-class issues, and the classes recur: deserialization of untrusted data, internal sockets reachable from the network, administrative routes with no authentication.
We ship a hardened profile that closes those classes by default, backport upstream fixes to the branch you run, and carry our own patch when upstream is silent.
GPUs that sit idle while requests queue.
The scheduler settings that decide how many requests run, how prefill is chunked and how deep the queue goes are fixed at launch. The workload is not.
The Traction Tuner reads live scheduler signals and adjusts those settings against your workload, with verify-or-revert on every change and a recommend-only mode for teams that want a human in the loop.
What is in a subscription
The engine is upstream SGLang. Support is the wrapper. Security and tuning are the value. Three packages, Base, Enhanced and Premium, build on each other.
| Name | What you get | Package |
|---|---|---|
Enterprise support | Severity-based cases worked by the engineers who build the branches; redacted offline diagnostics bundle; business, extended or round-the-clock hours by package | All |
Hardened profile and signed builds | Pickle off, every admin route keyed, ZeroMQ on localhost, non-root image, append-only audit log, published checklist; cosign-signed images with SBOM and vulnerability scan from an authenticated registry; offline bundles | Enhanced, Premium |
Traction Tuner Early access | Live adjustment of the running-request cap, chunked-prefill size and queue limit; verify-or-revert, canary, recommend-only, freeze, dry-run; operator with a TuningPolicy resource; no egress. Early access. | Basic in Enhanced; advanced in Premium |
LTS branch | One branch per quarter, twelve months of backports, published end of life; upstream fixes backported to your branch; public advisory tracker; CUDA 12.x and 13 lanes | Enhanced, Premium |
Vulnerability response | Our own patch when upstream is silent; day-zero advisory stating hardened-profile exposure; extended lifecycle and source escrow | Premium |
Traction Inference Reliability In development | Per-step and per-request telemetry that attributes latency to its cause; no prompt text captured; no egress. In development. | Premium |
Runs where your data already lives
Your GPUs, your network, your rules. Nothing in the subscription requires egress.
Your cloud account.
Kubernetes on GPU instances in your own AWS, Azure or Google Cloud account, or a dedicated cluster from a GPU cloud operator. Amazon EKS validated; Helm chart and operator.
Red Hat OpenShift and self-managed Kubernetes.
Operator packaging alongside the NVIDIA GPU Operator and your admission controls. Works inside llm-d SGLang pools.
Bare metal.
Single-node and multi-node installs on DGX, HGX and OEM GPU servers.
Air-gapped and sovereign.
Offline bundles with mirrored registry and signatures that verify without a network. License verified offline with a grace period. Source escrow available.
Validated on NVIDIA H100, L40S, A100 and B200
Measured, not marketed
Every performance number we publish comes from an open-source benchmark harness you can run from a clean checkout, with the configuration, the hardware and the method attached. The metric is goodput: tokens per second delivered inside your latency targets. Until the harness publishes, this site carries no performance numbers. When it does, each result comes with the command that reproduces it.
How the tuner and harness workPerformance figures on this site, when present, are produced by the open-source Traction Layer benchmark harness and published with model, hardware, driver, CUDA and container digests and the command that reproduces them. We publish no performance figure the harness has not produced.
How teams adopt it
- 1
Runtime Assessment.
AvailableA fixed-fee engagement on your cluster: we run the harness on your workload, size it, apply the hardened profile, and leave you a measured baseline and a written findings report. The fee is credited against a subscription. Available now.
- 2
Base or Enhanced.
BaseIn developmentavailable from LTS-1 launchEnhancedIn developmentBase puts a vendor on the hook for the SGLang you already run. Enhanced adds the LTS branch with backported fixes, the hardened profile, signed builds and the tuner in recommend-only or basic control mode.
- 3
Premium, after sixty days of measured results.
PremiumRoadmapVulnerability response with our own patches, extended lifecycle, the advanced tuner, Traction Inference Reliability and round-the-clock coverage.
- AWS Partner
- Microsoft Azure Partner
- NVIDIA Inception member
- Red Hat Partner
Partnerships and programs, not customers.
Read the policies before the pitch
The pages enterprise buyers ask for first.