🐛 Bug

As discussed here, calling kthvalue like torch.kthvalue(a, k, out=(a, indices)) produces a buffer overflow. Natalia suggested disallowing outputs that overlap with the input.

cc @VitalyFedyunin

Improve navigation of the I/O how-to by prepending a decision tree/flowchart using an image or graphviz, as suggested by @mattip.

PR #6447 adds a public API to get the maximum number of registers per thread (numba.cuda.Dispatcher.get_regs_per_thread()). There are other attributes that might be nice to provide - shared memory per block, local memory per thread, const memory usage, maximum block size.

These are all available in the FuncAttr named tuple: https://github.com/numba/numba/blob/master/numba/cuda/cudadrv/drive

Is there a way to train a bidirectional RNN (like LSTM or GRU) on trax nowadays?

Our instructions only cover installation, not actually running Orange. People unfamiliar with pip/conda need this.

The instructions should be fixed in README.md and our webpage (they should be the same). Please test throughly.

#5111 is an example of what can go wrong.

Description

While using tokenizers.create with the model and vocab file for a custom corpus, the code throws an error and is not able to generate the BERT vocab file

Error Message

ValueError: Mismatch vocabulary! All special tokens specified must be control tokens in the sentencepiece vocabulary.

To Reproduce

from gluonnlp.data import tokenizers
tokenizers.create('spm', model_p

Support Series.median()

MCVE Code Sample

# Your code here
import numpy as np
import xarray as xr

data = np.zeros((10, 4))
example_xr = xr.DataArray(data, coords=[range(10), ["