SGLang

license

Apache-2.0(OSI)

source

public

core features withheld

yes — python/pyproject.toml lists NVIDIA's proprietary nvidia-cutlass-dsl and nvidia-mathdx among its required dependencies

closed-dependency

nvidia-cutlass-dsl and nvidia-mathdx(both under NVIDIA proprietary licenses and required by the default install)

SGLang's own code is Apache-2.0, the repository has no enterprise directory or paid tier, and what the project sells is consulting and sponsorship. The default install is not open end to end, though: the package's required dependencies include NVIDIA's nvidia-cutlass-dsl and nvidia-mathdx, both under proprietary licenses. A closed component that the default build requires makes this open core rather than open source.

• https://github.com/sgl-project/sglang recorded 2026-08-11

  Repo page for SGLang, "a high-performance serving framework for large language models and multimodal models". Licensed Apache-2.0. Top-level directories are .claude, .devcontainer, .github, 3rdparty/amd, assets, benchmark, docker, docs, examples, experimental/sgl-router, proto, python, rust, scripts, sgl-model-gateway, test -- no ee/ or enterprise directory. Nothing on the page names a closed package, a license key, or a paid tier; the only commercial line is an invitation to "enterprises interested in adopting SGLang at scale, including technical consulting, sponsorship opportunities".

• https://pypi.org/pypi/nvidia-cutlass-dsl/json recorded 2026-10-08

  PyPI metadata for nvidia-cutlass-dsl carries the classifier "License :: Other/Proprietary License".

• https://pypi.org/pypi/nvidia-mathdx/json recorded 2026-10-08

  PyPI metadata for nvidia-mathdx is licensed under the "LICENSE AGREEMENT FOR NVIDIA MATH LIBRARIES SOFTWARE DEVELOPMENT KITS" and classified "License :: Other/Proprietary License".

• https://raw.githubusercontent.com/sgl-project/sglang/main/LICENSE recorded 2026-10-08

  The LICENSE file is the Apache License text, "Apache License" and "Version 2.0, January 2004".

• https://raw.githubusercontent.com/sgl-project/sglang/main/python/pyproject.toml recorded 2026-10-08

  python/pyproject.toml declares "License :: OSI Approved :: Apache Software License" and lists "nvidia-cutlass-dsl[cu13]==4.8.0" and "nvidia-mathdx==25.6.0" in its required dependencies.

• https://raw.githubusercontent.com/sgl-project/sglang/main/README.md recorded 2026-10-08

  README: "a non-profit open-source organization" hosts SGLang, and the only commercial line is a contact for "enterprise adoption and deployment, technical consulting, sponsorship, or partnership inquiries".

SGLang's own GitHub page and a third-party deployment guide both describe it as running at very large scale across major operators' GPU fleets, with xAI, NVIDIA, AMD, LinkedIn, Cursor and major cloud providers named as users. That is a heavy but operator-concentrated footprint rather than a broad public figure, since no daily-user or install-base count is published for it.

• https://github.com/sgl-project/sglang recorded 2026-08-09

  SGLang has become the de facto industry standard, with deployments running on over 400,000 GPUs worldwide.

• https://inference.net/content/sglang-complete-guide/ recorded 2026-08-09

  SGLang is deployed on over 400,000 GPUs worldwide, generating trillions of tokens daily. Major technology companies have built their inference infrastructure on SGLang: xAI uses SGLang for Grok inference

An independent H100 benchmark on Llama-3.3-70B put SGLang and vLLM within a few percent of each other on throughput and latency. vLLM remains the broader-purpose default, while SGLang is the stronger choice for shared-prefix workloads.

• https://github.com/sgl-project/sglang recorded 2026-08-09

  Fast Runtime: Provides efficient serving with RadixAttention for prefix caching, a zero-overhead CPU scheduler, prefill-decode disaggregation, speculative decoding, continuous batching, paged attention

• https://raw.githubusercontent.com/sgl-project/sglang/main/README.md recorded 2026-09-29

  README feature list: "Provides efficient serving with RadixAttention for prefix caching, a zero-overhead CPU scheduler, prefill-decode disaggregation, speculative decoding" and "Runs on NVIDIA GPUs (GB200/B300/H100/A100/Spark/5090), AMD GPUs (MI355/MI300), Intel Xeon CPUs, Google TPUs, Ascend NPUs, and more."

• https://www.spheron.network/blog/vllm-vs-tensorrt-llm-vs-sglang-benchmarks/ recorded 2026-09-29

  Benchmark table: vLLM "General use, broad model support" at "1,850 tok/s" and "120 ms"; SGLang "Shared-prefix workloads, low latency" at "1,920 tok/s" and "112 ms".