Masaki Kozuki comments

Results 167 comments of


                                            Masaki Kozuki

Small learning rate with `capturable=True` causes Adam optimizer to blow up model paraemters.

Thank you for the snippet. naively it'd need to another overload of `addcdiv` whose `value` argument being Tensor according to https://github.com/pytorch/pytorch/blob/e8be6d813bd093fdb587996662bb060445f70706/torch/optim/adam.py#L367-L370

Change 1D Tensor of 1 element to 0D Tensor

@pytorchbot label "module: optimizer","module: cuda graphs","module: mta"

Change 1D Tensor of 1 element to 0D Tensor

@pytorchbot merge

Update codegen for in-place foreach to return `List[Tensor]`

@pytorchbot rebase

Update codegen for in-place foreach to return `List[Tensor]`

> the simplest fix here is at the python or python binding level. I'm not sure how I'm supposed to avoid the current diff in aten. IIUC this would mean...

Update codegen for in-place foreach to return `List[Tensor]`

> @crcrpar you should have write access to the repo, due to the recent security incident CI won't run automatically unless your PR is on a branch in pytorch/pytorch repo,...

Update codegen for in-place foreach to return `List[Tensor]`

@ezyang thank you! I'll reopen this pull request after pushing this branch to pytorch/pytorch

Convert `ForeachFuncInfo` to `dataclass`

The number of test cases doesn't change with this pr.

Convert `ForeachFuncInfo` to `dataclass`

@pytorchbot merge -r

Convert `ForeachFuncInfo` to `dataclass`

hmm, I thought `dtypesIfROCM` was set to `dtypes` if it's not specified but it doesn't appear to be so. filed https://github.com/pytorch/pytorch/issues/125295