Dominik Klein

Hi @fafa92 , Thanks for reporting this! There are two options: 1. you move the solution to the cpu (here you seem to get a CUDA OOM). Then, the sparsification...

Hi @AlinaKurjan , Happy it worked for you in the end! The preprocessing, e.g. the PCA computation all happens on CPU, hence, if you had a GPU OOM error, this...

Yes, whenever the solver hasn't converged, it's better to rerun it. In general, doing a pure GW as proposed in the SCOT paper is often not good enough to match...

also mention that `set_xy/x/y`, and `set_graph_xy` exist, and point to the relevant examples.

main message: "show where we can fetch data from our anndata object" - fetch data from `.X` (e.g. if you wanna have raw counts) - fetch data from `.layers` (e.g....

Thanks @selmanozleyen , good catch. It happens because of this: https://github.com/theislab/moscot/blob/3151da7b324672fab6757b8a84bcd1d0bcb9f027/src/moscot/problems/_utils.py#L26 Now I am wondering whether this scenario is realistic, as the joint attribute in the case you describe would...

Great, thanks. what is `batch_size_sparse`?

So seems like `solve` and `apply` require the same memory? Could you please check by setting `batch_size_sparse` to 1?

yeah but seems like `apply` requires just `batch_size_sparse` times more memory, which means that `apply` and `solve` requires equally much when we `apply` only to one vector. Hence, `vmap` is...

okay, now the question is whether it's faster when we decrease the batch size in ott-jax (i.e. PointCloud), and hence to increase the batch size in `sparsify`. Any chance you...