addlabsviral/Qwen-Image-2.1-Turbo-diffuser-8bit — torchao int8 weight-only

**8 bit torchao Qwen/Qwen-Image-2.1-Turbo for text-to-image generation in 8 steps. Transformer + text encoder are int8, VAE stays bfloat16. Weights saved as safetensors, ready to run with torchao.

Quantized from Qwen/Qwen-Image-2.1-Turbo — same QwenImage21Pipeline, same saved 8-step sample_sigmas, CFG=1.

Install

pip install git+https://github.com/huggingface/diffusers.git
pip install "transformers>=5.17.0" accelerate torchao pillow sentencepiece

Requires diffusers with pipeline-configured sampling sigmas (PR #14950).

Use

import torch
from diffusers import QwenImage21Pipeline

pipe = QwenImage21Pipeline.from_pretrained(
    "addlabsviral/Qwen-Image-2.1-Turbo-torchao-int8",
    dtype=torch.bfloat16,
).to("cuda")

image = pipe(
    prompt="Cinematic portrait of an astronaut on Mars, golden hour lighting",
    width=1024,
    height=1024,
    use_kv_cache=True,
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]
image.save("output.png")

Needs torchao installed at load time.

Samples (generated on HF Job, rtx-pro-6000, quantized weights)

Notes

• int8 weight-only — ~50% VRAM vs BF16, minimal quality drift.
• Keep use_kv_cache=True, default 8-step schedule, CFG=1.
• License follows the base model (Qwen Research License).