sglang-diffusion-modelopt-quant

To quantize diffusion transformers using ModelOpt for SGLang, apply this skill to run official ModelOpt scripts, producing FP8 or NVFP4 checkpoints. You then adapt and convert these checkpoints for direct loading into the SGLang Diffusion runtime.

Yes, ModelOpt FP8 quantization works with multi-transformer pipelines. This skill supports both single-transformer and multi-transformer workflows, allowing per-component overrides and validation tooling to ensure proper calibration data alignment across all components.

Yes, you can use NVFP4 checkpoints in SGLang Diffusion. This skill bridges NVIDIA ModelOpt NVFP4 exports to the SGLang runtime, generating loadable quantized models while ensuring proper calibration data alignment and layout adaptation.

To verify accuracy after FP8 quantization of DiT models, this skill provides a structured workflow to validate quality and benchmark speed. It ensures parity or gains across quantized paths by comparing outputs against the base transformer layout.

ModelOpt FP8 quantization may fail to load in SGLang if calibration data is misaligned or the base transformer layout is incorrect. This skill ensures per-component overrides and proper checkpoint conversion to generate loadable SGLang runtime models.

Yes, you need official ModelOpt scripts for NVFP4 quantization. This skill requires these official scripts, a base transformer layout, and SGLang runtime utilities to generate, convert, and load the quantized models successfully.