Qwen-Image-e4m3

FP8 (e4m3) dynamically-quantized Qwen-Image, saved as a complete QwenImagePipeline.

What was changed

All 60 blocks of the QwenImageTransformer2DModel are quantized to e4m3_e4m3_dynamic — float8_e4m3fn weights with dynamically scaled float8_e4m3fn activations. The Qwen2.5-VL text encoder, 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.

Quantization was done with WaveSpeed's 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 be faster than bf16; on older GPUs the weights are dequantized on the fly and you only get the memory saving.

One caveat worth repeating from our own testing: on Qwen-Image, running fp8 weight-and-activation matmul under a fully fused fast path produces visible quality loss. The configuration published here — dynamic per-tensor activation scaling with bf16 accumulation — is the one that holds up.

Usage

import torch
from diffusers import QwenImagePipeline

pipe = QwenImagePipeline.from_pretrained(
    "wavespeed/Qwen-Image-e4m3",
    torch_dtype=torch.bfloat16,
    use_safetensors=False,
).to("cuda")

image = pipe("a chalkboard menu written in neat cursive").images[0]

License

Apache-2.0, inherited from Qwen-Image.