sglang-diffusion-modelopt-quant

To quantize a diffusion DiT model for SGLang Diffusion, you use ModelOpt to produce FP8 or NVFP4 checkpoints and adapt the component paths for loader compatibility. This workflow relies on ModelOpt's quantize.py as the PTQ source of truth.

Yes, you can validate quantized FP8 diffusion checkpoints against a BF16 baseline. The quantization workflow includes validation guidance to verify model quality and ensure fidelity before deployment.

Yes, ModelOpt quantization supports multi-transformer setups for diffusion models. The workflow applies ModelOpt-based quantization across FP8 and NVFP4 configurations while preserving a generic workflow with isolated fallbacks.

You benchmark quantized diffusion checkpoints by comparing their performance and throughput against BF16 baselines. The workflow provides guidance for benchmarking to measure fidelity and runtime metrics after adapting checkpoints.

The best way to adapt exported ModelOpt checkpoints is to adjust component paths to maintain SGLang Diffusion runtime compatibility. This ensures the FP8 or NVFP4 checkpoints are loadable and verifiable within the runtime environment.