Bug fixes: torch::tensor(floating-point values) -> default dtype, and torch::tensor(integer values) ->at::kLong (#32367)
Summary:
Some of the `torch::tensor` behavior is updated to better match Python API. Fixes https://github.com/pytorch/pytorch/issues/32234.
This PR is BC-breaking in the following way:
- `torch::tensor({1.0f, 2.0f})`: float -> default dtype
- `torch::tensor(at::ArrayRef<int>({1, 2, 3}))`: int -> at::kLong
- `torch::tensor(std::vector<int>({1, 2, 3}))`: int -> at::kLong
- `torch::tensor(at::ArrayRef<float>({1.f, 2.f, 3.f}))`: float -> default dtype
- `torch::tensor(std::vector<float>({1.f, 2.f, 3.f}))`: float -> default dtype
- `torch::tensor(at::ArrayRef<double>({1., 2., 3.}))`: double -> default dtype
- `torch::tensor(std::vector<double>({1., 2., 3.}))`: double -> default dtype
Pull Request resolved: https://github.com/pytorch/pytorch/pull/32367
Differential Revision: D19498484
Pulled By: yf225
fbshipit-source-id: 19c8dc2a56476266153cff4c404e7f84d309eb12
