When we are training a pytorch model, we may get this error: RuntimeError: “fill_cuda” not implemented for ‘ComplexHalf’. In this tutorial, we will introduce you how to fix it.
This error may like:
As to us, when we are running this code, this error is reported.
scaler = GradScaler(enabled=hps.train.fp16_run)
with autocast(enabled=hps.train.fp16_run):
pass
In order to understand how to use GradScaler, you can read:
Implement Mixed Precision Training with GradScaler in PyTorch – PyTorch Tutorial
How to fix this error?
The simplest way is set enabled = False, it means:
with autocast(enabled=False):
pass
Then, you will find this error is fixed.