Views
No views yet
QwenImageEditPipeline.QwenImageTransformer2DModel are quantized to
e4m3_e4m3_dynamic — float8_e4m3fn weights with dynamically scaled
float8_e4m3fn activations. The Qwen2.5-VL text encoder, the processor, the
VAE and the transformer's non-block tensors are untouched and stay in bf16. The
transformer drops from ~40.9 GB to ~20.5 GB.xelerate.ao.quantize. Weights are
stored as pickled .bin shards, so loading requires use_safetensors=False.
FP8 matmul needs Hopper (H100/H200) or newer to actually be faster than bf16.1import torch
2from diffusers import QwenImageEditPipeline
3from diffusers.utils import load_image
4
5pipe = QwenImageEditPipeline.from_pretrained(
6 "wavespeed/Qwen-Image-Edit-e4m3",
7 torch_dtype=torch.bfloat16,
8 use_safetensors=False,
9).to("cuda")
10
11image = load_image("input.png")
12out = pipe(image=image, prompt="make it a winter scene").images[0]wavespeed/Qwen-Image-Edit-l8v1.1-e4m3
— the same quantization with an 8-step Lightning LoRA fused in.