Implement torch.pow for float16 and bfloat16 on CPU #50789
Description
🚀 Feature
Add support for torch.pow with float16 and bfloat16 on CPU
Motivation
Currently, these types are not supported.
>>> torch.rand(10, dtype=torch.float16).pow(1.5)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
RuntimeError: "pow" not implemented for 'Half'
>>> torch.rand(10, dtype=torch.bfloat16).pow(1.5)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
RuntimeError: "pow" not implemented for 'BFloat16'
float16 and bfloat16 are supported for CUDA, however:
>>> torch.rand(10, dtype=torch.float16, device='cuda').pow(1.5)
tensor([0.4597, 0.6592, 0.6777, 0.0105, 0.7349, 0.0492, 0.5186, 0.1809, 0.4202,
0.3423], device='cuda:0', dtype=torch.float16)
>>> torch.rand(10, dtype=torch.bfloat16, device='cuda').pow(1.5)
tensor([5.7861e-02, 1.5234e-01, 8.7500e-01, 6.4373e-05, 7.0703e-01, 2.5977e-01,
3.6133e-01, 6.7578e-01, 1.4648e-01, 8.4839e-03], device='cuda:0',
dtype=torch.bfloat16)