sglang-diffusion-modelopt-quant

What problem does it solve?

Enable end-to-end quantization of diffusion DiT models with NVIDIA ModelOpt, producing FP8 or NVFP4 checkpoints that are loadable, verifiable, and benchmarkable in SGLang Diffusion.

Core Features & Use Cases

• Quantize diffusion transformers using ModelOpt's FP8 or NVFP4 paths.
• Convert and adapt ModelOpt exports to SGLang Diffusion format.
• Validate accuracy and benchmark performance of the quantized model.
• Maintain a clear separation between general converter logic and model-specific branching.

Quick Start

Quantize a diffusion DiT with ModelOpt, convert the exported checkpoint to SGLang Diffusion, and validate its accuracy and speed.