Not working on Mac MPS with float32 inference, Latents contain NaN
#27
by Yingding - opened
I tried model to generate image from text prompt using the sample from the model card on my Mac MPS.
The torch data type must be switched to float32, since no bfloat16 or float16 are supported by MPS.
Now I got only black image generated. The latents contain NaNs. The model is not numerically stable on MPS with float32.
Can you provide any hint how to generate successfully using float32 on MPS accelerator?
My inference code is available at
https://github.com/yingding/model-inference-samples/blob/main/text2image/glm_mps_dev.py
Many thanks in advance.