not support load Qwen-Image-Edit-2511 gguf or fp8 safetensors？

[zhaoyun0071]

import torch
from diffusers import QwenImageTransformer2DModel, GGUFQuantizationConfig

ckpt_id = r"Qwen/Qwen-Image-Edit-2511"

transformer1 = QwenImageTransformer2DModel.from_single_file(
    r"qwen-image-edit-2511-Q2_K.gguf",
    quantization_config=GGUFQuantizationConfig(compute_dtype=torch.bfloat16),
    torch_dtype=torch.bfloat16,
    config=ckpt_id,
    subfolder="transformer",
)

transformer2 = QwenImageTransformer2DModel.from_single_file(
    r"qwen_image_edit_2511_fp8mixed.safetensors",
    config=ckpt_id,
    subfolder="transformer",
    torch_dtype=torch.bfloat16,
)

but get error below：

diffusers\loaders\single_file_model.py", line 449, in from_single_file
    diffusers_format_checkpoint = checkpoint_mapping_fn(
TypeError: <lambda>() got an unexpected keyword argument 'config'

[mbalabanski]

Hello. What version of diffusers are you using? This script ran for me.

For reference, I pulled the two quantizations from https://huggingface.co/xms991/Qwen-Image-Edit-2511-fp8-e4m3fn (for FP8) and https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF (for GGUF)

My environment:

• Python 3.13.11
• diffusers==0.36.0
• torch==2.9.1, CUDA 13.0

[zhaoyun0071]

Hello. What version of diffusers are you using? This script ran for me.

For reference, I pulled the two quantizations from https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models (for FP8) and https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF (for GGUF)

My environment:

• Python 3.13.11
• diffusers==0.36.0
• torch==2.9.1, CUDA 13.0

Hello. What version of diffusers are you using? This script ran for me.

For reference, I pulled the two quantizations from https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models (for FP8) and https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF (for GGUF)

My environment:

• Python 3.13.11
• diffusers==0.36.0
• torch==2.9.1, CUDA 13.0

https://github.com/huggingface/diffusers/tree/f6b6a7181eb44f0120b29cd897c129275f366c2a

[Zhzet9611]

@mbalabanski hi, When I use the following code, I am unable to map the model parameters. Could you show me your code? thanks a lot !

ckpt_id = r"Qwen/Qwen-Image-Edit-2511"
transformer = QwenImageTransformer2DModel.from_single_file(
r"/data/lab/Models/Qwen/Qwen_Image/Qwen_image_edit_diffuser/2511_fp8/qwen_image_edit_2511_fp8mixed.safetensors",
config=ckpt_id,
subfolder="transformer",
torch_dtype=torch.bfloat16,
)

pipe = QwenImageEditPlusPipeline.from_pretrained(
"/data/lab/Models/Qwen/Qwen_Image/Qwen_image_edit_diffuser/2511_bf16",
transformer = transformer,
torch_dtype=torch.bfloat16)

[mbalabanski]

@Zhzet9611 Sorry, I made a mistake posting the links to the models. The FP8 quantization I posted was from a ComfyUI pipeline and I think has a different parameter mapping. I pulled from https://huggingface.co/xms991/Qwen-Image-Edit-2511-fp8-e4m3fn. The original link for GGUF quantization should map correctly and works in the original issue's script.

Here is my code:

import torch
from huggingface_hub import hf_hub_download
from diffusers import QwenImageTransformer2DModel, QwenImageEditPlusPipeline

ckpt_id = r"Qwen/Qwen-Image-Edit-2511"

quanitzation_checkpoint_id = "xms991/Qwen-Image-Edit-2511-fp8-e4m3fn"
quantization_filename = "qwen_image_edit_2511_fp8_e4m3fn.safetensors"

model_path = hf_hub_download(
    repo_id = quanitzation_checkpoint_id,
    filename = quantization_filename,
)

transformer = QwenImageTransformer2DModel.from_single_file(
    model_path,
    config = ckpt_id,
    subfolder = "transformer",
    torch_dtype = torch.float16,
)

pipe = QwenImageEditPlusPipeline.from_pretrained(
    ckpt_id,
    transformer = transformer,
    torch_dtype = torch.float16
)

[Zhzet9611]

@mbalabanski Thank you so much!

[yzl1014]

@Zhzet9611 Sorry, I made a mistake posting the links to the models. The FP8 quantization I posted was from a ComfyUI pipeline and I think has a different parameter mapping. I pulled from https://huggingface.co/xms991/Qwen-Image-Edit-2511-fp8-e4m3fn. The original link for GGUF quantization should map correctly and works in the original issue's script.

Here is my code:

import torch
from huggingface_hub import hf_hub_download
from diffusers import QwenImageTransformer2DModel, QwenImageEditPlusPipeline

ckpt_id = r"Qwen/Qwen-Image-Edit-2511"

quanitzation_checkpoint_id = "xms991/Qwen-Image-Edit-2511-fp8-e4m3fn"
quantization_filename = "qwen_image_edit_2511_fp8_e4m3fn.safetensors"

model_path = hf_hub_download(
repo_id = quanitzation_checkpoint_id,
filename = quantization_filename,
)

transformer = QwenImageTransformer2DModel.from_single_file(
model_path,
config = ckpt_id,
subfolder = "transformer",
torch_dtype = torch.float16,
)

pipe = QwenImageEditPlusPipeline.from_pretrained(
ckpt_id,
transformer = transformer,
torch_dtype = torch.float16
)

Why isn't the speed of fp8 faster, and why isn't the GPU resource usage reduced

[DN6]

@yzl1014 Diffusers doesn't automatically handle FP8 casting. You will need to enable it in order to see memory savings with FP8 checkpoints.

transformer = QwenImageTransformer2DModel.from_single_file(
    model_path,
    config = ckpt_id,
    subfolder = "transformer",
    torch_dtype = torch.float16,
)
transformer.enable_layerwise_casting(storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16)
transformer.to("cuda")

[ReynardLin]

@yzl1014 Diffusers doesn't automatically handle FP8 casting. You will need to enable it in order to see memory savings with FP8 checkpoints.

transformer = QwenImageTransformer2DModel.from_single_file(
model_path,
config = ckpt_id,
subfolder = "transformer",
torch_dtype = torch.float16,
)
transformer.enable_layerwise_casting(storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16)
transformer.to("cuda")

Thank you very much for your answer. I would like to ask further, does the model loaded in this way only reduce the amount of GPU memory used, and still cannot be accelerated during inference?