vllm-omni/docs/user_guide/diffusion/quantization/overview.md

Quantization for Diffusion Transformers

vLLM-Omni supports quantization of DiT linear layers to reduce memory usage and accelerate inference.

Supported Methods

Method	Guide
FP8	FP8

Device Compatibility

GPU Generation	Example GPUs	FP8 Mode
Ada/Hopper (SM 89+)	RTX 4090, H100, H200	Full W8A8 with native hardware

Kernel selection is automatic.