sglang-diffusion-modelopt-quant

To quantize a diffusion model to FP8 or NVFP4, this Skill automates the NVIDIA ModelOpt workflow to convert models, adapt checkpoints, and verify compatibility for optimized inference.

Yes, you can validate and benchmark diffusion checkpoints post-quantization to measure speed improvements and evaluate model quality compared to the baseline performance.

Yes, this Skill transforms exported checkpoints to be directly loadable and verifiable within SGLang Diffusion for deployment and performance evaluation.

Checkpoint adaptation involves transforming exported NVIDIA ModelOpt checkpoints into a loadable format, ensuring they are verifiable and compatible within the SGLang Diffusion environment.

Yes, Python and SGLang are required dependencies to execute the scripts for quantization, checkpoint adaptation, and performance benchmarking of diffusion transformers.

Use NVFP4 quantization for diffusion transformers when you need optimized deployment and efficient inference, provided you validate model accuracy and speed improvements against the baseline.