qwen-image-edit load error

[Luwill6]

You mentioned on hf that qwen-image-edit supports 64GB of memory, but I encountered a memory error during loading. What could be the reason?

[mingyi456]

What is the code you used, as well as your error message?

[Luwill6]

exit when load qwen-tranformer. error message: 0xc0000005。
I guess it means out of memory.

What is the code you used, as well as your error message?

[mingyi456]

Sorry, but this is still too vague for me. Can you paste the code that you were using to run the model, and the full error logs?

[Luwill6]

Sorry, but this is still too vague for me. Can you paste the code that you were using to run the model, and the full error logs?

code:
`import argparse
import torch
from diffusers.utils import load_image
from diffusers import QwenImageTransformer2DModel, QwenImageEditPipeline
from transformers.modeling_utils import no_init_weights
from dfloat11 import DFloat11Model
import torch
print(torch.version, torch.version.cuda)

def parse_args():
parser = argparse.ArgumentParser(description='Edit images using Qwen-Image-Edit model')
parser.add_argument('--cpu_offload', action='store_true',default = True, help='Enable CPU offloading')
parser.add_argument('--cpu_offload_blocks', type=int, default=12, help='Number of transformer blocks to offload to CPU')
parser.add_argument('--no_pin_memory', action='store_true', help='Disable memory pinning')
parser.add_argument('--image', type=str, default="input.jpg",
help='Path to input image or URL')
parser.add_argument('--prompt', type=str, default='Add a hat to the rabbit.',
help='Text prompt for image editing')
parser.add_argument('--negative_prompt', type=str, default=' ',
help='Negative prompt for image editing')
parser.add_argument('--num_inference_steps', type=int, default=50,
help='Number of denoising steps')
parser.add_argument('--true_cfg_scale', type=float, default=4.0,
help='Classifier free guidance scale')
parser.add_argument('--seed', type=int, default=42,
help='Random seed for generation')
parser.add_argument('--output', type=str, default='qwen_image_edit.png',
help='Output image path')
return parser.parse_args()

args = parse_args()
model_id = "D:/software/hfd/Qwen-Image-Edit/transformer"
config_tf = QwenImageTransformer2DModel.load_config(
model_id,
)
print(config_tf)
with no_init_weights():
transformer = QwenImageTransformer2DModel.from_config(
config_tf
).to(torch.bfloat16)
# print(transformer)
print("start")
DFloat11Model.from_pretrained(
"D:/software/hfd/Qwen-Image-Edit-DF11",
device="cpu",
cpu_offload=True,
cpu_offload_blocks=20,
pin_memory=False,
bfloat16_model=transformer,
)

pipeline = QwenImageEditPipeline.from_pretrained(
model_id, transformer=transformer, torch_dtype=torch.bfloat16,
)
pipeline.enable_model_cpu_offload()
pipeline.set_progress_bar_config(disable=None)

image = load_image(args.image)
inputs = {
"image": image,
"prompt": args.prompt,
"generator": torch.manual_seed(args.seed),
"true_cfg_scale": args.true_cfg_scale,
"negative_prompt": args.negative_prompt,
"num_inference_steps": args.num_inference_steps,
}

with torch.inference_mode():
output = pipeline(**inputs)
output_image = output.images[0]
output_image.save(args.output)

max_gpu_memory = torch.cuda.max_memory_allocated()
print(f"Max GPU memory allocated: {max_gpu_memory / 1000 ** 3:.2f} GB")
`
error info:
Loading DFloat11 safetensors (offloaded to CPU): 0%| | 0/1 [00:00<?, ?it/s]
进程已结束，退出代码为 -1073741819 (0xC0000005)

[Luwill6]

accelerate | 1.10.1 |  
bzip2 | 1.0.8 |  
ca-certificates | 2025.9.9 |  
certifi | 2025.8.3 |  
charset-normalizer | 3.4.3 |  
colorama | 0.4.6 |  
cupy-cuda12x | 13.6.0 |  
dahuffman | 0.4.2 |  
dfloat11 | 0.5.0 |  
diffusers | 0.35.1 |  
expat | 2.7.1 |  
fastrlock | 0.8.3 |  
filelock | 3.19.1 |  
fsspec | 2025.9.0 |  
huggingface-hub | 0.35.0 |  
idna | 3.10 |  
importlib-metadata | 8.7.0 |  
jinja2 | 3.1.6 |  
libffi | 3.4.4 |  
libzlib | 1.3.1 |  
markupsafe | 3.0.2 |  
mpmath | 1.3.0 |  
networkx | 3.2.1 |  
numpy | 2.0.2 |  
openssl | 3.0.17 |  
packaging | 25.0 |  
pillow | 11.3.0 |  
pip | 25.2 |  
psutil | 7.0.0 |  
python | 3.9.23 |  
pyyaml | 6.0.2 |  
regex | 2025.9.1 |  
requests | 2.32.5 |  
safetensors | 0.6.2 |  
setuptools | 78.1.1 |  
sqlite | 3.50.2 |  
sympy | 1.14.0 |  
tk | 8.6.15 |  
tokenizers | 0.22.0 |  
torch | 2.8.0+cu126 |  
tqdm | 4.67.1 |  
transformers | 4.56.1 |  
typing-extensions | 4.15.0 |  
tzdata | 2025b |  
ucrt | 10.0.22621.0 |  
urllib3 | 2.5.0 |  
vc | 14.3 |  
vc14-runtime | 14.44.35208 |  
vs2015-runtime | 14.44.35208 |  
wheel | 0.45.1 |  

env

[Luwill6]

I am using it on Windows 10. Has anyone tested this demo on Windows 10?

[mingyi456]

You mentioned you have 64gb of system RAM. How much VRAM do you have?

[Luwill6]

24g/3090. At first, I couldn't load the qwen tranformer model, but later I upgraded to 128GB of memory, and now it crashes when loading the FLOAT11 model.

[mingyi456]

Hmm, 64gb of system RAM and 24GB of VRAM should already be more than enough. This is weird.

Does it work if you try using a GGUF? Personally, I use this Q6_K quant from QuantStack with my 4090 because I only have 48GB of system RAM (which means my system RAM might also start overflowing if I use DF11), and not having the full transformer in VRAM is really slow. Even with the GGUF, it is already takes me about 5 minutes per edit.

The other 2 things I can think of are:

1. I have trouble with pytorch 2.8 and running even an SDXL pipeline in diffusers, so I downgraded to 2.7.1, Not sure if downgrading will solve your problem or not.
2. I have flash attention and triton-windows installed. Maybe not having flash attention is causing the discrepancy in VRAM/RAM usage?

Anyway, here is the output of pip freeze in my diffusers venv:

accelerate==1.10.0
aiofiles==24.1.0
aiohappyeyeballs==2.6.1
aiohttp==3.12.15
aiosignal==1.4.0
annotated-types==0.7.0
anyio==4.10.0
asttokens==3.0.0
asyncio==4.0.0
attrs==25.3.0
Authlib==1.6.3
bitsandbytes==0.47.0
Brotli==1.1.0
certifi==2025.8.3
cffi==1.17.1
charset-normalizer==3.4.3
click==8.2.1
cloudpickle==3.1.1
colorama==0.4.6
comm==0.2.3
cryptography==45.0.7
cupy-cuda12x==13.5.1
cyclopts==3.23.1
dahuffman==0.4.2
debugpy==1.8.16
decorator==5.2.1
dfloat11==0.5.0
diffusers @ git+https://github.com/huggingface/diffusers@4fcd0bc7ebb934a1559d0b516f09534ba22c8a0d
disutils==1.4.32.post2
dnspython==2.7.0
docstring_parser==0.17.0
docutils==0.22
einops==0.8.1
email-validator==2.3.0
exceptiongroup==1.3.0
executing==2.2.1
fastapi==0.116.1
fastmcp==2.12.2
fastrlock==0.8.3
ffmpy==0.6.1
filelock==3.18.0
flash-attn==2.7.4.post1
frozenlist==1.7.0
fsspec==2025.7.0
gguf==0.17.1
gradio==5.42.0
gradio_client==1.11.1
groovy==0.1.2
h11==0.16.0
hf-xet==1.1.8
httpcore==1.0.9
httpx==0.28.1
httpx-sse==0.4.1
huggingface-hub==0.34.4
idna==3.10
imageio==2.37.0
imageio-ffmpeg==0.6.0
importlib_metadata==8.7.0
inquirerpy==0.3.4
ipykernel==6.30.1
ipython==8.37.0
ipywidgets==8.1.7
isodate==0.7.2
jedi==0.19.2
Jinja2==3.1.4
jsonargparse==4.40.2
jsonschema==4.25.1
jsonschema-path==0.3.4
jsonschema-specifications==2025.4.1
jupyter_client==8.6.3
jupyter_core==5.8.1
jupyterlab_widgets==3.0.15
lark==1.2.2
lazy-object-proxy==1.12.0
markdown-it-py==4.0.0
MarkupSafe==2.1.5
matplotlib-inline==0.1.7
mcp==1.13.1
mdurl==0.1.2
more-itertools==10.8.0
mpmath==1.3.0
multidict==6.6.4
nag @ git+https://github.com/ChenDarYen/Normalized-Attention-Guidance.git@8bb34a16e517692743fab99bc60838c20f975ded
nest-asyncio==1.6.0
networkx==3.3
ninja==1.13.0
numpy==2.2.6
openapi-core==0.19.5
openapi-pydantic==0.5.1
openapi-schema-validator==0.6.3
openapi-spec-validator==0.7.2
opencv-python==4.12.0.88
optimum-quanto==0.2.7
orjson==3.11.1
packaging==25.0
pandas==2.3.1
parse==1.20.2
parso==0.8.5
patch-conv==0.0.1b0
pathable==0.4.4
peft==0.17.0
pfzy==0.3.4
pillow==11.3.0
platformdirs==4.4.0
prompt_toolkit==3.0.51
propcache==0.3.2
protobuf==6.31.1
psutil==7.0.0
pure_eval==0.2.3
pycparser==2.22
pydantic==2.11.7
pydantic-settings==2.10.1
pydantic_core==2.33.2
pydub==0.25.1
Pygments==2.19.2
pyperclip==1.9.0
PySide6==6.9.1
PySide6_Addons==6.9.1
PySide6_Essentials==6.9.1
python-dateutil==2.9.0.post0
python-dotenv==1.1.1
python-multipart==0.0.20
pytz==2025.2
pywin32==311
PyYAML==6.0.2
pyzmq==27.0.2
referencing==0.36.2
regex==2025.7.34
requests==2.32.4
rfc3339-validator==0.1.4
rich==14.1.0
rich-rst==1.3.1
rpds-py==0.27.1
ruff==0.12.8
safehttpx==0.1.6
safetensors==0.6.2
scipy==1.16.1
sd_embed @ git+https://github.com/xhinker/sd_embed.git@4a47f71150a22942fa606fb741a1c971d95ba56f
semantic-version==2.10.0
sentencepiece==0.2.1
shellingham==1.5.4
shiboken6==6.9.1
six==1.17.0
sniffio==1.3.1
spyder-kernels==2.5.2
sse-starlette==3.0.2
stack-data==0.6.3
starlette==0.47.2
sympy==1.13.3
tokenizers==0.21.4
tomlkit==0.13.3
torch==2.7.1+cu128
torchaudio==2.7.1+cu128
torchdiffeq==0.2.5
torchvision==0.22.1+cu128
tornado==6.5.2
tqdm==4.67.1
traitlets==5.14.3
transformers==4.55.0
triton-windows==3.4.0.post20
typer==0.16.0
typing-inspection==0.4.1
typing_extensions==4.14.1
tzdata==2025.2
urllib3==2.5.0
uvicorn==0.35.0
wcwidth==0.2.13
websockets==15.0.1
Werkzeug==3.1.1
widgetsnbextension==4.0.14
xformers==0.0.31.post1
yarl==1.20.1
zipp==3.23.0

[Luwill6]

Thank you! I was able to use it normally after changing the torch version 2.7.1. By the way, which parameters affect the memory usage in this demo?

[mingyi456]

I guess it is just the --cpu-offload-blocks and --no_pin_memory parameters. I have not tried these myself, honestly, so I only know as much as you do regarding them.

But regarding the line pipeline.enable_model_cpu_offload(), it basically tells the pipeline to keep all the components in system RAM first, and move them to VRAM only when each component needs to be executed. So you will need enough system RAM to hold the entire pipeline (basically the text encoder, unet/DiT, and the VAE) at once, but you only need enough VRAM to hold one component at a time (in this case you are using the cpu offload option in the dfloat11 model settings so that requirement is overridden, at the cost of speed).
If you have enough VRAM to hold the entire pipeline then you should replace that line with pipeline.to("cuda"), which moves the entire pipeline to VRAM, as this eliminates the time needed to shuffle individual components in and out of VRAM.