Notebook: ThunderFX tutorial

[kiya00]

Before submitting

• Was this discussed/approved via a Github issue? (no need for typos and docs improvements)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure to update the docs?
• Did you write any new necessary tests?

What does this PR do?

Add tutorial for thunder as torch.compile backend

cc @Borda @apaz-cli

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[mruberry]

Great stuff, @kiya00! Notebooks like this are invaluable. I was asked for a notebook explaining more about ThunderFX just the other day.

I made some comments using ReviewNB. I'm curious to hear your thoughts! They also propose changing some of the current UX to make the notebook easier to understand.

[kiya00]

Hi all, I modified it to a simple introduction of thunderfx, and ignored the #1532 for now

[mruberry]

Hi all, I modified it to a simple introduction of thunderfx, and ignored the #1532 for now

That's awesome, and #1535 is merged now, too!

[mruberry]

Updates look great, @kiya00! I thought the story was really straightforward and a great start to help practitioners learn about ThunderFX. I made some wording suggestions, and I have one question about the absolute difference in the model example.

It would be cool, but is not necessary, if we could include something interactive in the demo so that instead of relying on comparing output tensor discrepancies readers felt like they could compare the quality of the thunderfx compiled model vs. the eager model. Offering an interactive chat prompt is probably too computationally intensive, but maybe in the future we should think about an image generation model. (Doesn't have to be part of this PR.)

This XLA notebook uses stable diffusion, as an example:

https://github.com/pytorch/xla/blob/master/contrib/kaggle/pytorch-xla-2-0-on-kaggle.ipynb

and I think that's visually interesting. Again, something to consider for the future, not today.

[mruberry]

This notebook may also be causing an error in the CI:

https://dev.azure.com/Lightning-AI/Tutorials/_build/results?buildId=221933&view=logs&j=7e5b35f3-25b5-5ac7-f3c6-076eb2046208&t=93498b9b-50d0-54cd-9f79-979d23ce2520

[kiya00]

This notebook may also be causing an error in the CI:

https://dev.azure.com/Lightning-AI/Tutorials/_build/results?buildId=221933&view=logs&j=7e5b35f3-25b5-5ac7-f3c6-076eb2046208&t=93498b9b-50d0-54cd-9f79-979d23ce2520

Yes and it's strange that Torch nightly has no problem with it, Torch 2.4 does. I'll fix it tomorrow

It would be cool, but is not necessary, if we could include something interactive in the demo so that instead of relying on comparing output tensor discrepancies readers felt like they could compare the quality of the thunderfx compiled model vs. the eager model. Offering an interactive chat prompt is probably too computationally intensive, but maybe in the future we should think about an image generation model. (Doesn't have to be part of this PR.)

That's good idea, we could have another notebook or maybe demos would be more appropriate to show using Thunder to run interactive chat and image generation model

[mruberry]

This is a great first notebook introducing ThunderFX. I made one last suggestion for you review, @kiya00, to explain the deviation between ThunderFX and PyTorch's default results.

[kiya00]

Hi @mruberry , thanks for the review! I added the comment to explain t deviation

[mruberry]

Hi @mruberry , thanks for the review! I added the comment to explain t deviation

Oh sorry, would you add it as a text cell, not a comment in the code?

[kiya00]

ah, sure, I got it wrong, let me add it in the text cell