Ivan Tsybulin

BUG: ma.mean, ma.var and ma.std are not consistent with numpy for float32 dtype

4

Issue #21023 was resolved by #20914, but I cannot see how the fix affects the problem. Since the tests from issue were not included in the codebase, I've decided to...

UCM/CUDA: Correct region size for pitched CUDA allocation hooks

4

## What? For pitched CUDA memory allocations (cuMemAllocPitch, cuMemAllocPitch_v2, cudaMallocPitch, cudaMalloc3D) the allocated region may be bigger than requested. The UCM code incorrectly marked those regions to have `width *...

### Describe the bug When transferring memory allocated by cudaMalloc3D via MPI a segfault is possible. Probably, UCX fails to identify the address as device memory and tries to access...