sgl-kernel

What it is and what it does

sgl-kernel is a compiled CUDA kernel library designed to accelerate large language model inference by providing optimized compute primitives. It exposes custom kernel operations that LLM inference engines like SGLang and LightLLM use to speed up matrix operations and other compute-intensive tasks during model execution. The package ships as pre-built wheels for x86_64 and aarch64 architectures, requiring Python ≥3.10 and torch == 2.9.1.

The library is intended for developers building or extending LLM inference systems who need lower-level kernel control and optimization. It has no Python runtime dependencies beyond torch, making it a thin wrapper around CUDA code. Installation is straightforward via pip, though the strict torch version requirement means you must align your environment carefully. The project is actively maintained with recent commits and has accumulated significant adoption in the LLM inference ecosystem.

Use it for:

• Accelerate matrix operations in custom LLM inference engines by replacing generic CUDA kernels with optimized primitives.
• Build vision-language model inference pipelines that require specialized compute kernels for efficient token processing.
• Integrate into SGLang or LightLLM deployments to improve throughput and latency of model serving.
• Develop custom LLM serving frameworks that need fine-grained control over GPU compute without writing raw CUDA code.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

sgl-kernel provides optimized CUDA compute primitives for LLM inference, enabling efficient execution of custom kernel operations for large language models and vision-language models.

Yes, if you are building or extending an LLM inference engine and your environment uses torch == 2.9.1. The library is actively maintained, has no Python runtime dependencies, and is already proven in production LLM systems. The strict torch version requirement is a real constraint—verify compatibility before committing. No known security vulnerabilities.

Install

sgl-kernel on PyPI

pip install sgl-kernel

uv add sgl-kernel

poetry add sgl-kernel

Installing sgl-kernel

Before you install

Medium install friction due to compiled wheel distribution (x86_64 and aarch64 manylinux2014 wheels available). Requires torch == 2.9.1 as a hard dependency. Active maintenance with recent commits and no runtime Python dependencies simplifies integration once the torch version constraint is satisfied.

License in practice

Apache License 2.0 is permissive; you may use, modify, and distribute this package freely in commercial and private projects, provided you include the license and attribution notices.

Quickstart

pip install sgl-kernel

import sgl_kernel
# Use optimized kernels via sgl_kernel module in your inference engine

Requires torch == 2.9.1 and Python ≥3.10; NVIDIA GPU with CUDA support required for runtime execution.

Verify before relying

• Whether torch == 2.9.1 is pinned strictly or if compatible newer/older versions work in practice
• GPU compute capability requirements beyond what the description implies
• Performance characteristics relative to other LLM inference kernel libraries

Package facts

Evidence: sgl_kernel-0.3.21-cp310-abi3-manylinux2014_aarch64.whl; sgl_kernel-0.3.21-cp310-abi3-manylinux2014_x86_64.whl

Tags

More Artificial Intelligence packages

Further reading