Machine learning

Current implementation of Go binding can not specify options.

GPUOptions struct is in internal package. And go generate doesn't work for protobuf directory. So we can't specify GPUOptions for NewSession.

Related to #5142, AlbertTokenizer (which uses SentencePiece) doesn't decode special tokens (like [CLS], [MASK]) properly. This issue was discovered when adding the Nystromformer model (#14659), which uses this tokenizer.

To reproduce (Transformers v4.15 or below):

!pip install -q transformers sentencepiece

from transformers import AlbertTokenizer

tokenizer = AlbertTokenizer.from

🐛 Describe the bug

Usage of RRefContext::handleException in torch/csrc/distributed/rpc/rref_context.cpp is wrong when the future has an error.

RRefContext::handleException uses TORCH_CHECK which throws.

Callers of RRefContext::handleException don't expect that and run code after it without any guarding.

Versions

master

cc @pietern @mrshenli @pritamdamania87

These examples take quite a long time to run, and they make our documentation CI fail quite frequently due to timeout. It'd be nice to speed the up a little bit.

To contributors: if you want to work on an example, first have a look at the example, and if you think you're comfortable working on it and have found a potential way to speed-up execution time while preserving the educational message

When zipping with a product iterator with no arguments (which produces only one output, the empty tuple), it seems the check for zip completion does not handle this case:

julia> for (z,) in zip(Iterators.product())
       @show z
       end
z = ()
ERROR: ArgumentError: tuple must be non-empty
Stacktrace:
 [1] first(#unused#::Tuple{})
   @ Base ./tuple.jl:140
 [2] _pisdone
   @ ./ite

Currently many more Python projects like dask and optuna are using Python type hints. With the Python package of xgboost gaining more and more features, we should also adopt mypy as a safe guard against some type errors and for better code documentation.