Deep 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.

As reported by some people (see NielsRogge/Transformers-Tutorials#53 and on the forum), the generate() method currently does not take into account config.decoder.eos_token_id, only config.eos_token_id to properly stop generation.

Hence, models that are made using EncoderDecoderModel/`VisionEnco

🐛 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

Problem: Currently JsonLoggerCallback.handle_result will load in the entirety of the existing results, append the new result, and then rewrite the entire file. This may not scale when running long-running jobs or jobs with large results.

https://github.com/ray-project/ray/blob/4e8f90aca20aa7bb87a4e84039889444824382ca/python/ray/train/callbacks/logging.py#L138-L142

Potential Fix: