Theodore Vasiloudis

Results 45 comments of Theodore Vasiloudis

I will try with `nccl` so far we've only used `gloo` in GraphStorm AFAIK. Is there a reason to avoid `nccl` @classicsong ?

Hi @Rhett-Ying I ran the repro example that @jalencato posted with the code from #7409 and it works fine now. I think we can close this once that PR is...

If we merge this we can run our automated integration tests with the daily pip, I don't have the bandwidth to run more manual tests on the PR code.

There's an issue on the board about `num_feature` which I guess was used to subsample features? This is taken care of by `colsample*` now so I guess this parameter should...

Note that this will get you incompatible versions of zeromq and protobuf. For example libprotobuf-dev on Ubuntu will install 2.6.1 while ps-lite now uses 3.5.1. It's not yet clear to...

Hello @tqchen, I started some work on bringing the code up-to-date with current EC2 standards/software [here](https://github.com/thvasilo/yarn-ec2/commit/65b6927c1d0bcc25378e04dcb7d1b2908484f61c) but I've run into trouble getting the HDFS/YARN distribution to work. My recommendation to...

Hello @clumsy took a closer look at the PR and had a couple of requests. We'd like to maintain a similar level of requirements between GSProcessing and this library. *...

Yes, I think the outcome would be the same in this case. I think for this to be used in XGBoost it would require real-valued weights. I found the paper...

@DanielTing The XGBoost use-case would be for batch training, so we can assume to know all weights in advance and normalize them so they sum to 1. I'm having a...