[Tutorial] Demo showing how to run a pruned 🤗 model.

[jwfromm]

This tutorial demonstrates how to load and run a sparse model from the popular transformers module from Hugging Face (🤗). Very recently a 95% sparse version of BERT was made publicly available however 🤗 was unable to achieve speedups using existing frameworks. Using this script, TVM enables a 2-3X speedup by converting appropriate dense layers to sparse dense layers. I think this will be a useful tutorial for user's interested in sparse networks and may be good PR for TVM as a small collaboration with 🤗. Thanks @antinucleon for helping getting this all working!

[jwfromm]

@masahi, @vinx13, @binarybana can you take a look and let me know what you think?

[tqchen]

cc @antinucleon @junrushao1994

[masahi]

I liked emojis in the PR:)

How about adding a sample output, with avx2 or 512?

[binarybana]

Looks great! Made some edits.

[merrymercy]

How long does this tutorial take to run?
If it takes a lot of time, it is better to provide some sample outputs, so readers can know what is expected.
Otherwise, it is better to let it run on the CI server, so we can get output from the web server and make sure the tutorial is always runnable.

[jwfromm]

@merrymercy it's fairly quick, I commented out the run command due to dependencies rather than the run time. This tutorial requires tensorflow 2.2 (our servers currently use 2.1) and transformers. If we think its worth updating the server build then we can run this for real.

[u99127]

@merrymercy it's fairly quick, I commented out the run command due to dependencies rather than the run time. This tutorial requires tensorflow 2.2 (our servers currently use 2.1) and transformers. If we think its worth updating the server build then we can run this for real.

A +1 for updating TF versions and keeping this running out of the box.

[tqchen]

Thanks everyone, this is merged, will ping the thread again once we have TF 2.2 landed in the CI

[Wheest]

Note that the .ipynb version of the tutorial doesn't work when running benchmark(), since it uses the __file__ variable in import_graphdef(), which is not defined in most notebook environments. Alternative approach to getting path may be needed.

There are also a lot of dependencies for the tutorial (e.g. transformers, tensorflow) which may not be in a user's environment. Should an Install dependencies section be added à la?