skills.rest
air-gapped
air-gappedOFFICIAL
Explore all skills by this author →
GitHub

vllm-performance-tuning

Optimize vLLM MoE performance tuning workflows across GPU hardware configurations.

5|1|Updated Apr 19, 2026
Search Tags:#performance#benchmarking#vllm#gpu-optimization#moe#cuda-graphs#nvlink
One-click install
npx skills add https://github.com/air-gapped/skills --skill vllm-performance-tuning
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: vllm-performance-tuning
Source: https://github.com/air-gapped/skills/tree/main/.claude/skills/vllm-performance-tuning
Command: npx skills add https://github.com/air-gapped/skills --skill vllm-performance-tuning

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

vLLM performance tuning for MoE and hardware configurations reduces time-to-value in deploying large-scale LLMs by enabling structured, repeatable optimization workflows.

Core Features & Use Cases

  • MoE kernel autotuning with benchmark_moe.py to generate tuned configurations and compare performance across token shapes.
  • Comprehensive guidance for tensor-parallel, data-parallel, expert-parallel, and pipeline-parallel setups, plus disaggregation patterns (Nixl/Mooncake/LMCache) for scaling.
  • End-to-end workflow from baseline to re-bench, including CUDA graphs, compile cache, and scheduler knob tuning to hit SLOs.

Quick Start

Tune MoE kernels on the target GPU and load the generated configs before serving.

Frequently Asked Questions about vllm-performance-tuning

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I tune vLLM MoE kernels for specific GPU hardware?▼

Run benchmark_moe.py to generate tuned MoE configurations across token shapes on your target GPU, then load the generated configs via VLLM_TUNED_CONFIG_FOLDER before serving to optimize expert-parallel performance.

What parallelism strategies are supported for vLLM performance scaling?▼

vLLM performance tuning supports tensor-parallel, data-parallel, expert-parallel, and pipeline-parallel setups, alongside disaggregation patterns like Nixl, Mooncake, and LMCache for scaling across multiple GPUs.

Does vLLM performance tuning work with H100, H200, GB200, and MI300X GPUs?▼

Yes, the tuning workflow applies to operators deploying models on new hardware, specifically evaluating parallelism and disaggregation strategies across H100, H200, GB200, and MI300X SKUs.

How do I enable CUDA graphs and compile caching in vLLM?▼

Enable CUDA graphs and compile caching as part of the end-to-end vLLM tuning workflow from baseline to re-bench, adjusting scheduler knobs to hit your serving SLOs.

What environment variables are needed for vLLM MoE auto-tuning?▼

Use VLLM_TUNED_CONFIG_FOLDER and VLLM_ALL2ALL_BACKEND environment variables alongside auto_tune.sh to generate and apply tuned MoE configurations for your specific hardware setup.

When should I use PD disaggregation strategies in vLLM?▼

Use PD disaggregation strategies like Nixl, Mooncake, or LMCache when scaling large-scale LLMs across multiple GPUs to reduce time-to-value and hit strict serving SLOs.

Related Skills

Explore More Software Engineering Skills→
LancerLabLancerLab

cursor-croq-tune

Orchestrate GPU kernel autotuning cycles for CUDA and DSL-targeted kernels on NVIDIA GPUs.

Official
Advanced
belokonmbelokonm
8

machine-learning-engineer

Deploy ML models to production with infrastructure-as-code templates and monitoring.

Community
Advanced
air-gappedair-gapped
5

lmcache-mp

Run LMCache MP to share KV caches among vLLM pods in Kubernetes clusters.

Official
Advanced
zhuangbiaoweizhuangbiaowei
2

moe-training

Train Mixture of Experts models using DeepSpeed and HuggingFace Transformers.

Community
Advanced