Severin Dicks

It looks really good. I would love to have a closer look next week

The `to_dense` function only works for csr matrices. I think we need another kernel that handles `csc` or just `.T` the resulting array if csc. I also get performance warnings...

That kernel technically works but is way slower than the default scipy operation due to inefficient memory layout.

I think these should be parallel numba kernels with mask arguments for major and minor axis. I have a [working implementation](https://github.com/scverse/rapids_singlecell/blob/main/src/rapids_singlecell/tools/_kernels/_nan_mean_kernels.py) in rapids-singlecell that doesnt need to copy at all....

The kernels already look good. However my main point I was trying to make is that you can create the means without subsetting. That means you wouldnt need to use...

Do you want to use `conda` or `pypi`?

To install rapids-singlecell with pip, use the following command: ``` pip install 'rapids-singlecell[rapids12]' --extra-index-url=https://pypi.nvidia.com ``` This works for all CUDA 12 versions. If you prefer to install with conda, copy...

@cjnolet @ivirshup I think in general, it's a good idea to support cudf in the long term. cudf is insanely fast when it comes to correlations (significantly faster than cupy)...

I agree that it's not necessarily a bug. The behaviour is nevertheless unwanted and doesn't throw any meaningful errors. ``` autodoc_mock_imports = ["cudf", "cuml", "cugraph", "cupy", "cupyx"] ``` This is...

Benchmarks on an RTX 3090: X.shape = (200,000, 27,998) X_new.shape = (200,000, 19,214) Old implementation = 500 ms New implementation = 120 ms Additionally, the shortcut is significantly more memory...