sglang-diffusion-modelopt-quant

To quantize a diffusion DiT model, this Skill applies NVIDIA ModelOpt's FP8 or NVFP4 paths to generate quantized checkpoints, which are then adapted to load cleanly in SGLang Diffusion.

Yes, you can convert ModelOpt exports to SGLang Diffusion format. This Skill adapts the exported checkpoints, maintaining a clear separation between general converter logic and model-specific branching.

SGLang Diffusion supports mixed BF16 and NVFP4 workflows. This Skill enables end-to-end quantization pipelines, including per-component workflows that preserve provenance and provide validated performance metrics.

To validate accuracy and benchmark performance of a quantized diffusion model, this Skill provides compatibility checks and validated performance metrics across ModelOpt exports and SGLang runtimes.

Using ModelOpt's quantize.py for SGLang runtime compatibility involves applying the quantization script, adapting outputs to SGLang Diffusion format, and validating the resulting checkpoints for clean loading.

Format conversion is needed after quantizing a diffusion transformer because ModelOpt exports must be adapted to the SGLang Diffusion format to ensure the checkpoints load cleanly and verifiably in the target runtime.