"LayerNormKernelImpl" not implemented for 'Half'
Closed this issue · 5 comments
After pulling the latest commit, error ocurred:
"LayerNormKernelImpl" not implemented for 'Half'
are you running with mps accelerator?
@ParticleDog On the latest commit, everything is loaded on cpu to allow debuging, the issue you mentioned will be fixed once the weights will be properly loaded on GPU
are you running with mps accelerator?
I am not sure what mps accelerator is ... Is it a kind of cpu accelerator for pytorch?
I changed the code in pipeline_loader.py from "weight_dtype = torch.float16" to "weight_dtype = torch.float32", it seems that the error was bypassed.
But another error occured, which also appeared before I pulled the latest commit:
!!! Exception during processing!!! Given groups=1, weight of size [128, 3, 3, 3], expected input[1, 1024, 1024, 768] to have 3 channels, but got 1024 channels instead
yes, running on cpu needs float32 but float16 works on gpu.
Concerning the last error you mentioned, it will be fixed very soon in the next commit
Thanks for your precious feedbacks!!
@ParticleDog dimension issue should be fix now