RuntimeError: “add_cpu/sub_cpu” not implemented for ‘Half’ when using Float16/Half

jit
1

Hello, I am testing out different types in PyTorch and noticed when calling torch.jit.trace where the input is float16/half, I am getting a runtime error (RuntimeError: “add_cpu/sub_cpu” not implemented for ‘Half’). Does tracing not work for float16? I believe I also tested with int32 inputs which work.

2

I think this might be more about operations that PyTorch supports on GPU than the types.
Does the same code run in plain PyTorch?

Best regards

Thomas

3

You are right in that this doesn’t seem to work for eager mode (guessing this is plain PyTorch) as just calling model(input) causes the same runtime error. But I am not running on a GPU right now (just a macbook). I guess I can probably change the category and rename the question. I guess Half is just not supported for CPU?

4

Indeed. (That’d be a fun weekend project for you! :stuck_out_tongue: )