FastaiV1 RuntimeError: "bernoulli_scalar_cpu_" not implemented for 'Half'

fastai fastai dev
1

I didn’t see this in any Forum Topic, so I apologize if hidden somewhere else. I am running v1 on a Paperspace GPU instance. The line of code I am running is simply:

learn.fit_one_cycle(1, 1e-2)

My learner is from:
learn = language_model_learner(data_lm, AWD_LSTM, metrics=[accuracy, Perplexity()], path=PATH, wd=0.1).to_fp16()

This has all been working fine all week up until this morning.

Any ideas? Thank you!

2

UPDATE. I just realized that my PaperSpace instance this morning was generated as a CPU machine for some reason. I usually keep my previous machine/notebook in the dashboard and simply hit Run which should create another instance of the same, but today it didn’t. I will create a GPU instance now and see if I get the same error. Thanks.

3

UPDATE: Worked fine once again with a GPU machine.