Olivier Grisel

While reviewing this PR and testing it on some data, I realized that the reported number of iterations was always 0 whenever l-BFGS-b would kick in. I pushed a quick...

#8750 should address this by exposing a dedicated `array_api` submodule instead using the `dask.array` module directly. However, as of now, #8750 does not expose an implementat of `isdtype`.

In scikit-learn we do use the isdtype function but we now also depend on array-api-compat so this should not be a problem. The main integration problems are related to value...

We only test with `array-api-compat` wrapping in our array API test harness. However we do not include dask yet. @lithomas1 started a PR to run tests with dask on our...

>> this would open the possibility to train machine learning models (e.g. with scikit-learn or Pytorch) directly via Ibis > This is an interesting path for us to go down;...

Also: related question, does the dask backend support (vector) UDAFs? It should be quite natural to implement however the documentation on UDFs is lacking.

Reading the clickhouse docs, it seems that it would be natural to implement Python vector UDFs with clickhouse as records are batched by default: https://clickhouse.com/docs/en/sql-reference/functions/#executable-user-defined-functions Both clickhouse and pandas can...

@cpcloud I drafted a proposal for Python UDAFs API in a duckdb issue if you are interested: https://github.com/duckdb/duckdb/discussions/5117.

> We now have vectorized scalar UDFs in the DuckDB backend. Check out the [related blog post](https://ibis-project.org/blog/rendered/torch/). Thanks, this is great. A few remarks: - I cannot find those in...

> At what point do you want to control the chunking? When you call to_pandas(), or earlier, such as when you define the function? Not sure what you mean by...