special_lora_error

[taiczhi]

I successfully installed it but still encountered the same issue as before. I can guarantee that this model works in the normal dev mode and performs better than the original model. I used a special reverse distillation dev model for training, but it seems that the quantization is not compatible with it.
NunchakuFluxLoraLoader
Incompatible keys detected:
lora_unet_img_in.alpha, lora_unet_img_in.lora_down.weight, lora_unet_img_in.lora_up.weight, lora_unet_txt_in.alpha, lora_unet_txt_in.lora_down.weight, lora_unet_txt_in.lora_up.weight
11:39

lora_unet_img_in.alpha, lora_unet_img_in.lora_down.weight, lora_unet_img_in.lora_up.weight, lora_unet_txt_in.alpha, lora_unet_txt_in.lora_down.weight, lora_unet_txt_in.lora_up.weight
2025-04-08T11:38:43.325874 - Traceback (most recent call last):
File "/root/autodl-tmp/ComfyUI/execution.py", line 327, in execute
output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/autodl-tmp/ComfyUI/execution.py", line 202, in get_output_data
return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/autodl-tmp/ComfyUI/execution.py", line 174, in _map_node_over_list
process_inputs(input_dict, i)
File "/root/autodl-tmp/ComfyUI/execution.py", line 163, in process_inputs
results.append(getattr(obj, func)(**inputs))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/autodl-tmp/ComfyUI/custom_nodes/ComfyUI-nunchaku/nodes/lora/flux.py", line 61, in load_lora
sd = to_diffusers(lora_path)
^^^^^^^^^^^^^^^^^^^^^^^
File "/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/nunchaku/lora/flux/diffusers_converter.py", line 17, in to_diffusers
new_tensors, alphas = FluxLoraLoaderMixin.lora_state_dict(tensors, return_alphas=True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/huggingface_hub/utils/_validators.py", line 114, in _inner_fn
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/diffusers/loaders/lora_pipeline.py", line 1449, in lora_state_dict
state_dict = _convert_kohya_flux_lora_to_diffusers(state_dict)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/diffusers/loaders/lora_conversion_utils.py", line 561, in _convert_kohya_flux_lora_to_diffusers
return _convert_sd_scripts_to_ai_toolkit(state_dict)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/diffusers/loaders/lora_conversion_utils.py", line 523, in _convert_sd_scripts_to_ai_toolkit
raise ValueError(f"Incompatible keys detected: \n\n {', '.join(remaining_keys)}")
ValueError: Incompatible keys detected:

lora_unet_img_in.alpha, lora_unet_img_in.lora_down.weight, lora_unet_img_in.lora_up.weight, lora_unet_txt_in.alpha, lora_unet_txt_in.lora_down.weight, lora_unet_txt_in.lora_up.weight

Originally posted by @taiczhi in #262 (comment)

Okay, this is the model I trained. Thank you very much
https://modelscope.cn/models/ewq32213/test_model_nunchaku/file/view/master?fileName=models%252Fxin-000001.safetensors&status=2

[lmxyy]

Yeah, your LoRA format was not detected by diffusers, so it caused the error. I will take a look later.

[renqqc]

我也有同样的问题，看来是使用了这个蒸馏模型训练Lora：https://huggingface.co/ashen0209/Flux-Dev2Pro

[taiczhi]

我也有同样的问题，看来是使用了这个蒸馏模型训练Lora：https://huggingface.co/ashen0209/Flux-Dev2Pro

me too

[renqqc]

我也有同样的问题，看来是使用了这个蒸馏模型训练Lora：https://huggingface.co/ashen0209/Flux-Dev2Pro

me too

我尝试用原版的flux模型重新训练了一个Lora，同样出现了这个错误，看来不是Flux-Dev2Pro这个蒸馏模型的问题，应该是lora训练过程中其它参数设置的问题。
我之前使用的青龙圣者的脚本训练的lora会出现这个问题，换成秋叶的脚本就没问题了，他们都是使用的kohya的训练器，所以一定是某个参数的问题，但我现在不知道究竟是哪个参数出现问题了。😢

[secret-deus]

我也出现了这个问题，用原版flux工作流正常，使用量化nunchaku则输出字典key不匹配，且输出图片为噪点

[renqqc]

我也出现了这个问题，用原版flux工作流正常，使用量化nunchaku则输出字典key不匹配，且输出图片为噪点

我训练时，将save_precision从bf16修改成fp16、将timestep_sampling 从shift修改成sigmoid、关闭flux_block_dim之后，解决了key不匹配的问题，我不知道是哪项导致的，但现在它看起来是没问题的。

[lmxyy]

I successfully installed it but still encountered the same issue as before. I can guarantee that this model works in the normal dev mode and performs better than the original model. I used a special reverse distillation dev model for training, but it seems that the quantization is not compatible with it. NunchakuFluxLoraLoader Incompatible keys detected: lora_unet_img_in.alpha, lora_unet_img_in.lora_down.weight, lora_unet_img_in.lora_up.weight, lora_unet_txt_in.alpha, lora_unet_txt_in.lora_down.weight, lora_unet_txt_in.lora_up.weight 11:39

lora_unet_img_in.alpha, lora_unet_img_in.lora_down.weight, lora_unet_img_in.lora_up.weight, lora_unet_txt_in.alpha, lora_unet_txt_in.lora_down.weight, lora_unet_txt_in.lora_up.weight 2025-04-08T11:38:43.325874 - Traceback (most recent call last): File "/root/autodl-tmp/ComfyUI/execution.py", line 327, in execute output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/root/autodl-tmp/ComfyUI/execution.py", line 202, in get_output_data return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/root/autodl-tmp/ComfyUI/execution.py", line 174, in _map_node_over_list process_inputs(input_dict, i) File "/root/autodl-tmp/ComfyUI/execution.py", line 163, in process_inputs results.append(getattr(obj, func)(**inputs)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/root/autodl-tmp/ComfyUI/custom_nodes/ComfyUI-nunchaku/nodes/lora/flux.py", line 61, in load_lora sd = to_diffusers(lora_path) ^^^^^^^^^^^^^^^^^^^^^^^ File "/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/nunchaku/lora/flux/diffusers_converter.py", line 17, in to_diffusers new_tensors, alphas = FluxLoraLoaderMixin.lora_state_dict(tensors, return_alphas=True) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/huggingface_hub/utils/_validators.py", line 114, in _inner_fn return fn(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^ File "/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/diffusers/loaders/lora_pipeline.py", line 1449, in lora_state_dict state_dict = _convert_kohya_flux_lora_to_diffusers(state_dict) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/diffusers/loaders/lora_conversion_utils.py", line 561, in _convert_kohya_flux_lora_to_diffusers return _convert_sd_scripts_to_ai_toolkit(state_dict) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/diffusers/loaders/lora_conversion_utils.py", line 523, in _convert_sd_scripts_to_ai_toolkit raise ValueError(f"Incompatible keys detected: \n\n {', '.join(remaining_keys)}") ValueError: Incompatible keys detected:

lora_unet_img_in.alpha, lora_unet_img_in.lora_down.weight, lora_unet_img_in.lora_up.weight, lora_unet_txt_in.alpha, lora_unet_txt_in.lora_down.weight, lora_unet_txt_in.lora_up.weight

Originally posted by @taiczhi in #262 (comment)

Okay, this is the model I trained. Thank you very much https://modelscope.cn/models/ewq32213/test_model_nunchaku/file/view/master?fileName=models%252Fxin-000001.safetensors&status=2

What prompt do you use?

[taiczhi]

我成功安装了它，但仍然遇到了之前相同的问题。我可以保证这个模型在正常的开发模式下运行，并且比原始模型表现更好。我使用了一种特殊的反向蒸馏开发模型进行训练，但似乎量化与它不兼容。检测到不兼容的键：lora_unet_img_in.alpha、lora_unet_img_in.lora_down.weight、lora_unet_img_in.lora_up.weight、lora_unet_txt_in.alpha、lora_unet_txt_in.lora_down.weight、lora_unet_txt_in.lora_up.weight 11:39
lora_unet_img_in.alpha, lora_unet_img_in.lora_down.weight, lora_unet_img_in.lora_up.weight, lora_unet_txt_in.alpha, lora_unet_txt_in.lora_down.weight, lora_unet_txt_in.lora_up.weight 2025 年 4 月 8 日 11 时 38 分 43 秒 325874 - 回溯（最近一次调用最后）：文件“/root/autodl-tmp/ComfyUI/execution.py”，第 327 行，在执行中 output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 文件“/root/autodl-tmp/ComfyUI/execution.py”，第 202 行，在 get_output_data 中 return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 文件“/root/autodl-tmp/ComfyUI/execution.py”，第 174 行，在 _map_node_over_list 中 process_inputs(input_dict, i) 文件“/root/autodl-tmp/ComfyUI/execution.py”，第 163 行，在 process_inputs 中 results.append(getattr(obj, func)(**inputs)) 文件“/root/autodl-tmp/ComfyUI/custom_nodes/ComfyUI-nunchaku/nodes/lora/flux.py”，第 61 行，在加载 lora 时，sd = to_diffusers(lora_path)
文件“/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/nunchaku/lora/flux/diffusers_converter.py”，第 17 行，在转换为 diffusers 时，new_tensors, alphas = FluxLoraLoaderMixin.lora_state_dict(tensors, return_alphas=True)
文件“/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/huggingface_hub/utils/_validators.py”，第 114 行，在内部函数中，返回 fn(*args, **kwargs)
文件“/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/diffusers/loaders/lora_pipeline.py”，第 1449 行，在 lora_state_dict 中，state_dict = _convert_kohya_flux_lora_to_diffusers(state_dict)
文件“/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/diffusers/loaders/lora_conversion_utils.py”，第 561 行，在转换 kohya flux lora 为 diffusers 时，返回 _convert_sd_scripts_to_ai_toolkit(state_dict) "/root/autodl-tmp/ComfyUI/ENV_ComFyUI/lib/python3.11/site-packages/diffusers/loaders/lora_conversion_utils.py", 第 523 行, 在 _convert_sd_scripts_to_ai_toolkit 中引发 ValueError(f"检测到不兼容的键:\n\n {', '.join(remaining_keys)}") ValueError: 检测到不兼容的键:
lora_unet_img_in.alpha, lora_unet_img_in.lora_down.weight, lora_unet_img_in.lora_up.weight, lora_unet_txt_in.alpha, lora_unet_txt_in.lora_down.weight, lora_unet_txt_in.lora_up.weight
最初由 @taiczhi 在 #262 (comment) 中发布
好的，这是我训练的模型。非常感谢 https://modelscope.cn/models/ewq32213/test_model_nunchaku/file/view/master?fileName=models%252Fxin-000001.safetensors&status=2

你使用什么提示？

you can try this

original, digital artwork, this digital artwork portrays a delicate, ethereal young woman with fair skin and long, dark hair adorned with pink and gold floral accessories.