Arithmetic (add, sub, mul, div) is not implemented for float16 on CPU. (Not just in-place ops).
Pretty much only conversions are implemented. These ops are implemented for GPU. You can use
torch.bfloat16 on CPU.
Note bfloat16 is supported on CPU.
We should update TorchSHarp to include both float16 and bfloat16 support
158 added interim float16 support.
However PyTorch only implements float16 on the GPU.
Note
bfloat16
is supported on CPU.We should update TorchSHarp to include both float16 and bfloat16 support