Loading Qwen Base Model Error

[lankyscorp]

Hi,

I am trying to train a Qwen Lora and receive an error when press the button to start training ("could not load model: " + model_names.base_model). Under the model tab I have the "Base Model" field set to qwen_image_fp8_e4m3fn.safetensors. Is this model compatible? I see a note in the QwenModelLoader.py that I should use a diffuser model. If this is not the correct model can someone point me to the correct model I should use with 16GB of RAM.

Thank you for your help.

[thecat99z]

You should leave, the first time you use the Qwen preset, the base model path Qwen/Qwen-Image . This will download from Huggingface the diffusers model compatible with OneTrainer. After that the model will be loaded from the hugging face cache each time you train with Qwen.
This is the full release precision model. You will have to override the transformers Dtype to float8. And if you have 16GB of Vram you will still need to turn on CPU_OFFLOADING in the gradient checkpointing options. The Qwen 16GB preset should normally set it all.