cuda devices should have same dtype

[jfc4050]

addresses #25465

was passing in two tensors of different dtypes for a check making sure the two tensors were on the same device.

[mrshenli]

Ah, yes, I misread the error message, it's the data type not device type.

[facebook-github-bot]

@mrshenli has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[facebook-github-bot]

@mrshenli merged this pull request in e26305e.

[pietern]

@jfc4050 This test is skipped on CPU-only builds. Can you post a follow up change to break it out into a non-CUDA and CUDA test case with checks? I would like to run all tests associated with allreduce_coalesced regardless of CUDA being available or not. Then for any tests that require CUDA, we can have a separate test case (or guard the check in some other way).

[jfc4050]

@jfc4050 This test is skipped on CPU-only builds. Can you post a follow up change to break it out into a non-CUDA and CUDA test case with checks? I would like to run all tests associated with allreduce_coalesced regardless of CUDA being available or not. Then for any tests that require CUDA, we can have a separate test case (or guard the check in some other way).

sure! we were actually talking about doing this in another thread but I hadn't gotten around to it yet. Made a PR(#25555) and CC'd you