Garbled up image using IPEX (Intel Extension for PyTorch)
Using notebook from inference/text_to_image.ipynb and replacing mentions of cuda to xpu made it possible to generate images on Intel Arc GPU but stage C images are garbled and are being upscaled either to black image or same garbled mess.
It is working on CPU.
@17314642 Are the weights loaded in FP16 when using the GPU? There's a bug where the FP16 generation is bugged, on the CPU it's probably forced to use FP32 so that's why it works fine.
@17314642 Are the weights loaded in FP16 when using the GPU? There's a bug where the FP16 generation is bugged, on the CPU it's probably forced to use FP32 so that's why it works fine.
No, I'm using stage_b_lite_bf16.safetensors and stage_c_lite_bf16.safetensors using BFloat16 dtype.