SGLang hits v1.0 as NVIDIA, AMD and Google join new consortium

LMSYS describes the consortium as a standing forum. Members can discuss emerging inference trends with SGLang's maintainers and turn shared priorities into joint engineering work. The focus LMSYS names is frontier model support, reasoning and multimodal workloads, distributed inference and hardware optimisation.[1][2]

Each founding member will designate a technical representative and join two roadmap meetings a year. LMSYS says a member contributes through engineering expertise, deployment feedback, testing resources or joint development. Working groups may collaborate between those meetings. LMSYS says there will be public updates on roadmap direction and on progress.[1][2]

The maintainers keep the decision. LMSYS says they remain responsible for technical decisions and for accepting code, and that the project stays open to the whole community.[1][2]

LMSYS calls SGLang a full-stack inference engine for frontier AI, and part of the PyTorch ecosystem. It says the engine runs on hundreds of thousands of GPUs and powers some of the largest production inference deployments in the world, for companies including Meta, Microsoft, Google, Amazon, Thinking Machines Lab, Cloudflare, Reflection and Modal.[1][2]

That sentence is not a membership list. Microsoft, Amazon, Thinking Machines Lab and Reflection are named as production users and are not listed as founding members. Intel and RadixArk are founding members and are not named in that production sentence. Google, Meta, Cloudflare and Modal appear in both.[1][2]

LMSYS cites Deloitte for the size of the workload. According to Deloitte, as relayed by LMSYS, inference is expected to account for roughly two-thirds of all AI compute in 2026, up from about half in 2025. LMSYS says the inference engine determines how fast, and how cost-effectively, that compute runs on each company's hardware.[1][2]

Ying Sheng, co-author of the original SGLang paper and co-founder and CEO of RadixArk, spoke on behalf of the SGLang team. "The teams running SGLang at the largest scale see what inference needs next before anyone else does," Sheng said. "With the challenges in building new data centers, software approaches to improve AI efficiency will be incredibly important. These lessons will benefit every SGLang user and advance the AI ecosystem."[1][2]

Ian Buck, vice president of hyperscale and high-performance computing at NVIDIA, said: "Better software helps developers get more performance from NVIDIA compute."[1][2]

Ramine Roane, corporate vice president of AI applications at AMD, said: "SGLang gives developers a fast, open inference engine that runs across hardware, including AMD Instinct GPUs."[1]

The v1.0 post is separate from the consortium release. SGLang says the version brings three changes together. On speed, it lists zero-overhead scheduling, performance-critical paths rewritten in Rust, speculative decoding, and megakernels. On memory, it lists unified memory for hybrid models, plus hierarchical caching to keep long agent histories reusable beyond GPU memory. On scaling, it lists prefill/decode disaggregation and flexible parallelism, so a serving setup can be matched to the workload.[3]

SGLang says adoption spans 330-plus organisations, with day-zero support for 217 models, and that the project has grown through contributors and through teams running it in production. The post publishes no benchmark numbers for v1.0. It says the details of the release are coming soon.[3]

LMSYS has put chip, model and infrastructure companies in a standing forum with the maintainers who accept SGLang's code. It cites Deloitte's expectation that inference will be roughly two-thirds of AI compute in 2026, up from about half in 2025, and it says the engine decides how fast and how cost-effectively that compute runs.[1][2]

Membership is not control of the repository. LMSYS has described two meetings a year, a named technical representative, and a contribution of engineering, deployment feedback, testing resources or joint development. Maintainers still take the technical decisions and still accept the code, and the project stays open to people outside the consortium.[1][2]

v1.0 is a separate post from the same summit, not a line in the consortium release. The scheduling, Rust, memory and scaling claims, and the figures of 330-plus organisations and 217 models with day-zero support, are SGLang's own. The post gives no benchmark numbers, and it says the release details are still to come.[3]