As an example,

@numba.cuda.jit
def foo():
    x = numba.cuda.local.array(shape=(2, 0), dtype=numba.int64)

foo()

gives ValueError: array length <= 0.

Currently these lines contain a workaround because of this issue:

https://github.com/numba/numba/blob/3399d618b334c75a7c5c1ab85ec02769fb3b4c3b/numba/cuda/cudadrv/devicearray.py#L432-L455

Problem:
catboost version: 0.23.2
Operating System: all
Tutorial: https://github.com/catboost/tutorials/blob/master/custom_loss/custom_metric_tutorial.md

Impossible to use custom metric (С++).

Code example

from catboost import CatBoost
train_data = [[1, 4, 5, 6],
 

Current implementation of join can be improved by performing the operation in a single call to the backend kernel instead of multiple calls.

This is a fairly easy kernel and may be a good issue for someone getting to know CUDA/ArrayFire internals. Ping me if you want additional info.

There are (at least) two ways in which the reinterpret_cast is misused:

1. Used instead of static_cast:
   https://github.com/rapidsai/cudf/blob/f7fbc1160b17b969db2708e0cf6033d3db9fc1cf/cpp/src/io/avro/avro_gpu.cu#L95
   static cast should be used to cast from void*.
2. The use causes undefined behavior:
   https://github.com/rapidsai/cudf/blob/f7fbc1160b17b969db2708e0cf6033d3db9fc1cf/cpp/src/io

All dev environments should install faiss from conda. The 10.2 environment is currently missing it. It's also missing ucx-proc. Should follow the other env styles.

See: https://github.com/rapidsai/cuml/blob/branch-0.15/conda/environments/cuml_dev_cuda10.2.yml

I often use -v just to see that something is going on, but a progress bar (enabled by default) would serve the same purpose and be more concise.

We can just factor out the code from futhark bench for this.

Thank you for this fantastic work!

Could it be possible the fit_transform() method returns the KL divergence of the run?

Thx!