skills.rest
FutureMLS-Lab
FutureMLS-LabOFFICIAL
Explore all skills by this author →
GitHub

sglang-diffusion-performance

Select SGLang Diffusion performance flags for generation speed, VRAM, and quality.

556|81|Updated May 19, 2026
Search Tags:#performance tuning#quantization#diffusion#sglang#gpu optimization#vram reduction#multi-gpu inference
One-click install
npx skills add https://github.com/FutureMLS-Lab/OSCAR --skill sglang-diffusion-performance-futuremls-lab
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: sglang-diffusion-performance
Source: https://github.com/FutureMLS-Lab/OSCAR/tree/main/sglang-research/python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-performance
Command: npx skills add https://github.com/FutureMLS-Lab/OSCAR --skill sglang-diffusion-performance-futuremls-lab

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

This Skill helps you choose effective SGLang Diffusion performance flags without guesswork across different models, GPUs, VRAM limits, and quality requirements.

Core Features & Use Cases

  • Lossless Tuning: Select torch compilation, warmup, multi-GPU parallelism, offloading, attention backends, and FSDP settings to improve speed or reduce memory while preserving output quality.
  • Lossy Optimization: Evaluate approximate attention, Cache-DiT, quantized weights, precision overrides, and reduced inference steps when speed or VRAM savings justify quality tradeoffs.
  • Workload-Specific Recipes: Build practical configurations for image generation, video generation, low-VRAM systems, Wan2.2, LTX-2, and multi-GPU deployments.
  • Use Case: Given a Wan2.2 model, eight GPUs, and a requirement for maximum throughput, use this Skill to compare Ulysses and CFG parallelism and select appropriate compilation and warmup options.

Quick Start

Ask the skill to recommend the fastest SGLang Diffusion command for your model, GPU configuration, VRAM budget, and required image or video quality.

Frequently Asked Questions about sglang-diffusion-performance

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

FAQPage Schema
What is the best way to reduce VRAM usage when running large diffusion models like Wan2.2?▼

To reduce VRAM usage for diffusion models, apply offloading settings, FSDP, approximate attention, Cache-DiT, and quantized weights to trade output quality for lower memory consumption on low-VRAM systems.

How does multi-GPU parallelism affect diffusion generation throughput in SGLang?▼

Multi-GPU parallelism affects diffusion generation throughput by allowing you to compare strategies like Ulysses and CFG parallelism, selecting appropriate compilation and warmup options to maximize speed across multiple GPUs.

Can I use quantization and reduced inference steps to speed up SGLang diffusion without breaking model compatibility?▼

You can use quantization, precision overrides, and reduced inference steps for faster SGLang diffusion, but you must ensure matching model support, compatible GPU hardware, and installed optimization packages to maintain stability.

What are the tradeoffs between lossless tuning and lossy optimization for SGLang diffusion workloads?▼

Lossless tuning preserves output quality using torch compilation and attention backends, while lossy optimization applies approximate attention and quantized weights to maximize speed and VRAM savings when quality tradeoffs are justified.

Related Skills

Explore More Software Engineering Skills→
QingHuan-6QingHuan-6

sglang-diffusion-performance

Identify fastest SGLang Diffusion flags for a model, GPU, and VRAM budget.

Community
Advanced
spacecat2002spacecat2002

sglang-diffusion-performance

Select SGLang Diffusion configuration flags for generation speed and GPU memory usage.

Community
Advanced
hey-konghey-kong

sglang-diffusion-performance

Benchmark SGLang and Diffusers backends to optimize diffusion performance on multi-GPU setups.

Community
Advanced
moirai-internalmoirai-internal

diffusion-optimal-perf

Optimize SGLang-Diffusion inference with CLI flags and environment variables.

Official
Advanced