Masaki Kozuki

Hi, thank you for your fantastic work! I have a question about this repository. The LICENSE file says this repository's license is MIT, but pypi says this repository's is apache....

Use `int64_t` for nll_loss with cuda inputs

1

Related #85005 ### TODO - [ ] Check where the other implementations of nll loss are used and add tests

Remove `optim_mt` from `test/test_optim.py`

7

As per title, this updates `test_optim.py` so that `foreach` optimizers are constructed using the `foreach` keyword argument of `torch.optim` optimizers. Also, this makes some cosmetic changes to remove `torch.autograd.Variable`, `.data`...

we have an increasing number of install options and custom C++/CUDA modules. it's high time we had a list. @eqy @ptrblck

As it's used by a couple of tests other than `apex.transformer` Signed-off-by: Masaki Kozuki

**Describe the Bug** **Minimal Steps/Code to Reproduce the Bug** ```console $ pytest test_transducer_joint.py ... FAILED transducer/test_transducer_joint.py::TransducerJointTest::test_transducer_joint_pack_relu_dropout - RuntimeError: one of the variables needed for gradient compu... FAILED transducer/test_transducer_joint.py::TransducerJointTest::test_transducer_joint_relu_dropout - RuntimeError:...

BUILD: Change `prefix` to a relative path with `CMAKE_CURRENT_LIST_DIR`

4

## What _Describe what this PR is doing._ Set `prefix` to a relative path with `CMAKE_CURRENT_LIST_DIR` so that users can move the UCX library to a path different from prefix....

[mta] Backward of unary foreach functions

4

as per title, this PR defines backward of those. This doesn't implement forward-mode automatic differentiation as [the current codegen](https://github.com/pytorch/pytorch/blob/a747326423ed4731996769e3b8eb73eecbdee2d4/tools/autograd/gen_variable_type.py#L1513) doesn't seem to handle `ArrayRef`. Rel: - https://github.com/pytorch/pytorch/issues/53796 - https://github.com/pytorch/pytorch/issues/58833 cc...

[mta] implement `_foreach_pow`

1

Mainly for foreach path of `Adam` and `AdamW`