David Chiu
David Chiu
I also noticed that the implementation of `_single_tensor_asgd` and `_multi_tensor_asgd` seems different. I am not sure which one is correct. 1. `_single_tensor_asgd` update eta by `lr / ((1 + lambd...
> Hi, @janeyx99, I have tried to fix the code and add a test. However, the test passed on both the original code and the fixed code on my device...
@pytorchbot drci
> ( @david20571015 I'm not from the maintainers' team, so can't add a more meaningful review ) Oh, l'm sorry for mistakenly request a review from you.
> > I found that the eta between steps are numerically equal by `torch.as_tensor`. (Is this acceptable?) > > ```python > > lr, lambd, alpha = 1e-2, 1e-4, 0.75 #...
@pytorchbot merge
@pytorchbot drci
@janeyx99, should I try to fix this new fail (in C++ code?) or just skip these test case? Isn't aten::_foreach_pow_.Scalar for cuda [here](https://github.com/pytorch/pytorch/blob/31372fa84293f5bdd0fb06acceb098a5c8187a07/aten/src/ATen/native/cuda/ForeachBinaryOpScalar.cu#L175)?
@pytorchbot rebase -b main
@pytorchbot merge