guqiong96Community·8 Agent Skills Included

Lsglang

GPU and NUMA hybrid inference for large MoE models

Runs large MoE language models with GPU plus NUMA dual-parallel hybrid inference, splitting work between CPU memory and VRAM. Eliminates the need for expensive multi-GPU setups by letting one GPU and system RAM jointly hold and compute huge models. Includes ready-made skills for performance tuning, quantization, benchmarking, testing, and debugging SGLang-based serving workloads.

105|9

↗

npx skills add guqiong96/Lsglang --all -g -y

Available:

Tells the agent how the diffusion subsystem is structured, how to run and extend it, and routes performance questions to the right sub-skill, while the docs guide sets rules for writing accurate Mintlify documentation pages.

All Skills in This Repository (8)

Pure Emerald Level Indicators

📦 In Repo

guqiong96

add-new-diffusion-model

Guide adding new diffusion models to the SGLang framework.

diffusion-optimal-perf

Optimize diffusion model inference with lossless and lossy techniques in SGLang-Diffusion.

diffusion-kernel

Develop and optimize Triton and CUDA JIT kernels for diffusion models.

diffusion-perf

Benchmark and profile diffusion models for inference performance analysis.

write-sglang-test

Write SGLang CI and unit tests following project conventions.

add-jit-kernel

Add custom CUDA kernels to the SGLang JIT kernel module.

add-sgl-kernel

Add custom CUDA/C++ kernels to the sgl-kernel library with CMake and Python API integration.

SGLang Bisect CI Regression

Automate SGLang CI regression debugging via temporal bisection and remote GPU reproduction.

Community

Advanced

Frequently Asked Questions

FAQPage Schema

How to install Lsglang?▼

Run `npx skills add guqiong96/Lsglang --all -g -y` in your terminal to install all skills in this suite globally.

What problem does Lsglang solve?▼

It lets you run very large MoE models by splitting inference between GPU VRAM and CPU memory with NUMA-aware scheduling, so huge models fit on affordable hardware.

Which models does Lsglang support?▼

It supports tested MoE models from the Qwen3, GLM, MiniMax, Kimi, and DeepSeek families, in BF16, FP16, FP8, NVFP4, MXFP4, and AWQ 4-bit formats.

What hardware do I need for Lsglang?▼

You need an x86 CPU with AVX2 or newer instructions and an NVIDIA GPU with sm80 or newer architecture, plus enough combined VRAM and system RAM to hold the model.

Do these skills work with Claude Code and other agents?▼

Yes. All skills follow the standard SKILL.md format and work in Claude Code, OpenClaw, Cursor, and other compatible coding agents.