Full fine-tuning fails with quantization (int W8A8/float W8A8/GGUF) - RuntimeError: only Tensors of floating point dtype can require gradients

[knoic]

When performing full fine-tuning on z-image models, selecting specific Transformer DataType options related to quantization triggers a runtime error.
RuntimeError: only Tensors of floating point dtype can require gradients
The error occurs when setting transformer DataType to any of the following quantization modes:
int W8A8
float W8A8
GGUF
Notably, the float (W8) mode works normally during full fine-tuning – is there any fundamental difference between float (W8) and the other modes (float W8A8/int W8A8/GGUF) that explains this compatibility?

[dxqb]

float W8 doesn't work for full finetuning either. If it doesn't give you an error, it will fail to learn anything.
you need a real floating point dtype to full finetune

[knoic]

Understood, thanks!