skills.rest
sqjian
sqjianCOMMUNITY
Explore all skills by this author →
GitHub

sglang-diffusion-modelopt-quant

Quantize diffusion transformer models with NVIDIA ModelOpt FP8 and NVFP4 workflows.

Updated May 14, 2026
Search Tags:#quantization#diffusion#transformer#fp8#sglang#nvfp4#modelopt
One-click install
npx skills add https://github.com/sqjian/sglang --skill sglang-diffusion-modelopt-quant-sqjian
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: sglang-diffusion-modelopt-quant
Source: https://github.com/sqjian/sglang/tree/main/python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-modelopt-quant
Command: npx skills add https://github.com/sqjian/sglang --skill sglang-diffusion-modelopt-quant-sqjian

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Quantization and bridge tooling for diffusion models require a smooth workflow from NVIDIA ModelOpt exports to SGLang Diffusion, including validation and benchmarking.

Core Features & Use Cases

  • End-to-end ModelOpt FP8/NVFP4 quantization for DiT backbones
  • Convert and load quantized checkpoints into SGLang Diffusion with compatibility checks
  • Validate trajectory similarity and performance using the included tools

Quick Start

Run the end-to-end ModelOpt quantization workflow to produce an SGLang-ready diffusion transformer and verify its performance

Frequently Asked Questions about sglang-diffusion-modelopt-quant

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

FAQPage Schema
How do I quantize diffusion transformer models using NVIDIA ModelOpt for SGLang?▼

To quantize diffusion transformer models, use the official ModelOpt quantize.py script to apply FP8 or NVFP4 workflows, then adapt the exported checkpoints for SGLang Diffusion using provided helper tools.

Can I use mixed BF16 and NVFP4 builds for diffusion transformer backbones?▼

Yes, you can build mixed BF16 and NVFP4 configurations. The workflow supports benchmarking performance and verifying quality retention across per-component backbones and mixed precision builds.

How do I verify trajectory similarity after FP8 quantization for diffusion models?▼

You verify trajectory similarity using the included helper tools designed for FP8 and NVFP4 conversion. These tools compare trajectories to ensure quality retention in the SGLang runtime.

Do I need to maintain SGLang runtime loader paths when adapting ModelOpt exports?▼

Yes, maintaining the SGLang runtime loader paths is a requirement. The workflow bridges ModelOpt exports to SGLang Diffusion by ensuring compatibility checks and preserving these loader paths.

What is the best way to benchmark performance of NVFP4 quantized DiT backbones?▼

The best way to benchmark NVFP4 quantized DiT backbones is using the end-to-end ModelOpt workflow, which validates trajectory similarity and tests performance across per-component builds.

Why does my ModelOpt export fail compatibility checks in SGLang Diffusion?▼

ModelOpt exports fail compatibility checks if the SGLang runtime loader paths are not maintained. The workflow requires using helper tools to correctly convert and load quantized checkpoints.

Related Skills

Explore More Software Engineering Skills→
annealing-inversionannealing-inversion

sglang-diffusion-modelopt-quant

Quantize diffusion DiT models with NVIDIA ModelOpt for SGLang Diffusion.

Community
Advanced
OpenMOSSOpenMOSS
410

sglang-diffusion-modelopt-quant

Quantize diffusion DiT models with NVIDIA ModelOpt to FP8 or NVFP4 checkpoints for SGLang Diffusion.

Official
Advanced
hey-konghey-kong

sglang-diffusion-modelopt-quant

Quantize diffusion transformers with NVIDIA ModelOpt to generate FP8 or NVFP4 checkpoints for SGLang Diffusion.

Community
Advanced
hung-phanhung-phan
1

attention

Analyze attention mechanism performance, memory usage, and implementation trade-offs for transformer architectures.

Community
Advanced