Edgar Andrés Margffoy Tuay
Edgar Andrés Margffoy Tuay
Some of the errors are related to the return type mismatch between NumPy and CuPy. Since now `argmin/max` is returning an array, it voids the int return type produced by...
@asi1024, I think everything is ready to run on the actual CIs
For some reason, the Python version on the CI didn't declare `math.prod`, it was replaced by `operator.mul` used with `functools.reduce`. I think it should pass this time on the CIs
> Does this implementation work with arrays whose ndim > 1? No indeed, the returned indices need to be normalized w.r.t their axis and reshaped, thanks for noticing it
The errors on the CI are related to NCCL, not to the changes of this PR
No worries! Please have a nice holiday time!
I've just extended the `NaN` comparison (lexicographical sorting element-wise on pairs of rows of the axis, looking if `left_row < right_row`) to consider the following cases: ### `equal_nan=False` and non-complex...
@toslunar, thanks for your review! I've just updated the implementation to use `trim_zeros` instead of elementwise scan, I'm just missing a performance comparison between calling `unique` on CPU vs the...
Here are the profiling results where the current implementation (using `trim_zeros` + subtraction) is compared against pre/post-processing on GPU and then calling NumPy's `unique` function as proposed by @toslunar on...
Hi, and thanks for your very thoroughly answer. Here there are some replies and additional questions: > The major change to the mater is commit bcc6cde , in which a...