Tatiana Likhomanenko

Thanks for asking! We actually didn't try them yet. But there should be no any problems to try it. So you need to train lm following instructions in the lm...

Also BERT-style model is not supposed for the sequence generation, so it cannot be used for the decoder.

cc @avidov @vineelpratap @jacobkahn

cc @xuqiantong @padentomasello

Please see solution here https://github.com/flashlight/flashlight/issues/535, let us know if this solves the issue.

Did you install https://github.com/flashlight/flashlight/blob/master/.docker/Dockerfile-CUDA-Base#L45?

Could you try first to install via `pip install -e .` and trying to add the dir into sys.path and see if it works? Next I would try to build...

Hey! @Jibril2 - Python bindings support only featurization, ASG criterion and beam search decoder. For the rest we have only c++ implementation. - flashlight.scripts.colab.record is not included into flashlight bindings....

I am agree with @lunixbochs. Anyone can use any language / tools to prepare tokens/lexicon and lm data. I would really not include pre-processing stuff in C++. Also it will...

You can try something similar to https://github.com/flashlight/flashlight/blob/master/.circleci/config.yml#L297 as there are sometimes problems with intel packages.