unsloth

/

Qwen-Image-Edit-2511-GGUF

like 276

Follow

Unsloth AI 12k

Confirmed this works, got the same output as the old file using the workaround.

The F16 file downloads correctly (SHA256 matches 16999927cc...), but llama.cpp fails to load it:

gguf_init_from_file_impl: tensor '__index_timestep_zero__' has offset 0, expected 40871966976

This suggests the __index_timestep_zero__ tensor added in the ComfyUI compatibility patch has its offset incorrectly set to 0 in the GGUF metadata, rather than the actual offset where the tensor data was appended.
Verified with both browser download and huggingface-cli. Same issue may affect BF16.gguf.

The expected offset (40,871,966,976) also doesn't match either the pre-patch size (40,872,114,720) or post-patch size (40,872,114,784), suggesting the index_timestep_zero tensor metadata may not have been written correctly during the ComfyUI compatibility patch.

The F16 file downloads correctly (SHA256 matches 16999927cc...), but llama.cpp fails to load it:

gguf_init_from_file_impl: tensor '__index_timestep_zero__' has offset 0, expected 40871966976

This suggests the __index_timestep_zero__ tensor added in the ComfyUI compatibility patch has its offset incorrectly set to 0 in the GGUF metadata, rather than the actual offset where the tensor data was appended.
Verified with both browser download and huggingface-cli. Same issue may affect BF16.gguf.

The expected offset (40,871,966,976) also doesn't match either the pre-patch size (40,872,114,720) or post-patch size (40,872,114,784), suggesting the index_timestep_zero tensor metadata may not have been written correctly during the ComfyUI compatibility patch.

Should now be fixed! You may need to redownload. Thanks for letting us know!

NEW UPDATE: ComfyUI looks for a special key in the Qwen Image Edit 2511 model data. We added a 0-byte dummy tensor but used the wrong offset. llama.cpp validates offsets and fails, while ComfyUI’s GGUF loaders don’t check and accept it anyway.

NEW UPDATE: ComfyUI looks for a special key in the Qwen Image Edit 2511 model data. We added a 0-byte dummy tensor but used the wrong offset. llama.cpp validates offsets and fails, while ComfyUI’s GGUF loaders don’t check and accept it anyway.

Is llama.cpp issue fixed now ? I use sd.cpp. 20Gb is big amount to download on my slow connection so I want to check before downloading. I cannot understand from your statement if it was fixed or now this issue remains.

Thanks Daniel.
Should it in fact work with llama.cpp at all? I patched the offset and it then ran but still didn't like the qwen_image and refused to load the model.

Loading model... |llama_model_load: error loading model: error loading model architecture: unknown model architecture: 'qwen_image'
llama_model_load_from_file_impl: failed to load model
llama_model_load: error loading model: error loading model architecture: unknown model architecture: 'qwen_image'
llama_model_load_from_file_impl: failed to load model
common_init_from_params: failed to load model 'models/qwen-image-edit-2511-F16.gguf'
srv    load_model: failed to load model, 'models/qwen-image-edit-2511-F16.gguf'
 
Failed to load the model

I have slightly lost the plot regarding whether llama.cpp and associated GGUF files and so llama-server and llama-cli support visual models or not.
It works with ComfyUI.

Thanks Daniel.
Should it in fact work with llama.cpp at all? I patched the offset and it then ran but still didn't like the qwen_image and refused to load the model.

Loading model... |llama_model_load: error loading model: error loading model architecture: unknown model architecture: 'qwen_image'
llama_model_load_from_file_impl: failed to load model
llama_model_load: error loading model: error loading model architecture: unknown model architecture: 'qwen_image'
llama_model_load_from_file_impl: failed to load model
common_init_from_params: failed to load model 'models/qwen-image-edit-2511-F16.gguf'
srv    load_model: failed to load model, 'models/qwen-image-edit-2511-F16.gguf'
 
Failed to load the model

I have slightly lost the plot regarding whether llama.cpp and associated GGUF files and so llama-server and llama-cli support visual models or not.
It works with ComfyUI.

You cannot do image generation with llama.cpp. sd.cpp is correct application for that. I believe it is buit on top of some components of llama.cpp

https://github.com/leejet/stable-diffusion.cpp/releases

Is llama.cpp issue fixed now ? I use sd.cpp. 20Gb is big amount to download on my slow connection so I want to check before downloading. I cannot understand from your statement if it was fixed or now this issue remains.

@engrtipusultanit it is fixed for stablediffusion.cpp but make sure you've built the most recent source. Otherwise it may not have the latest qwen image edit 2511 updates.