sglang-diffusion-modelopt-quant

To quantize a diffusion DiT model with NVIDIA ModelOpt, run the workflow to export FP8 or NVFP4 checkpoints, then load and verify them within the SGLang Diffusion runtime.

Yes, SGLang supports loading FP8 and NVFP4 quantized diffusion checkpoints. You can bridge ModelOpt exports to SGLang, verify model quality, and benchmark performance across both paths.

Before quantizing DiT with ModelOpt, you need access to NVIDIA ModelOpt tooling, a diffusion transformer implementation, and the SGLang runtime with FP8/NVFP4 support to ensure successful checkpoint preparation.

Yes, you can benchmark quantized FP8 vs NVFP4 performance in SGLang Diffusion. The workflow includes verifying model quality and benchmarking performance across both precision paths.

If your ModelOpt quantized checkpoint is not working with SGLang Diffusion, verify that your SGLang runtime supports FP8/NVFP4 and that the checkpoint was properly adapted and verified during the export workflow.