Add Multimodal heading and Document question answering in task_summary.mdx

[y3sar]

What does this PR do?

From #18926
This PR creates a new Multimodal heading in task_summary and add Document question answering task example inside of it.

Who can review?

@stevhliu

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[amyeroberts]

Really nice - thanks for adding!

[amyeroberts]

@y3sar - The current CI tests are failing because this PR triggers tests checking code snippets in the docs, and the doctest CI environment doesn't have ffmpeg installed. I've opened a PR #23329 to resolve this. Once I've confirmed everything works, we can re-run for this PR and if all green then merge :)

[stevhliu]

Looks great, thanks again for your contribution and for adding the new multimodal section! 🤗

[amyeroberts]

@y3sar Sorry for the delay in this. There's been a lot of changes in how doctests are retrieved and run. Could you rebase onto main to get the latest updates?

[y3sar]

@amyeroberts sure thing I'll rebase and commit again

[amyeroberts]

@y3sar Did you force push after rebasing the branch? The current commit history looks like what I get if I push but don't force.

[y3sar]

@amyeroberts no I did not force push. I rebased and pull pushed. What should I do to solve this problem?

[amyeroberts]

@y3sar To rebase onto main you need to force push, as rebasing is a form of history rewrite. The steps are - running on this branch:

• Get the most recent version of main: git fetch upstream main
• Rebase: git rebase upstream/main
• Push changes: git push -f

[y3sar]

@amyeroberts ffmpeg bug still remains. What can I do to solve this?

[amyeroberts]

@y3sar You'll want to modify this line to:

"sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev espeak-ng time ffmpeg",

[y3sar]

@amyeroberts should I commit in this branch or should I create another pull request?

[amyeroberts]

@y3sar Commit on this branch :)

[amyeroberts]

@y3sar Apologies for the delay with this PR. Could you rebase again and push. We've been having issues with timeouts on the CI which should now be resolved. Additionally, could you update the extension for the .mdx to .md please?

[y3sar]

@amyeroberts thank you for remembering this pull request 😄😄😄🙏
Yes ma'am.

[y3sar]

@amyeroberts looks like the timeout issue remains. Should I change the code example?

[amyeroberts]

@y3sar Let's draft in help from the king of tests @ydshieh :)

I did a test run of this example on a CPU and it took 40s, so not fats but not super slow either. I think we could try:

• Another checkpoint. Are there any other smaller DocQA checkpoints we could use?
• Forcing this example to be skipped in the tests

[ydshieh]

Let me see what's happening on the CI runner.

[ydshieh]

Hi, I checked. That task_summary.md is considered as a single test by pytest , but it has multiple examples (and the checkpoints are not always small).

I can change the environment variable to avoid this situation. Once that change is merged, you can rebase and the CI should be green.

[ydshieh]

The PR is opened

#24753

[ydshieh]

That PR is merged into main. If you pull the latest main and rebase on it, we should be good to merge this PR.

[y3sar]

Thank you the king of tests and @amyeroberts. Should I check for a smaller checkpoint?

[ydshieh]

Should I check for a smaller checkpoint?

It would be always great to use a small(er) checkpoint for testing (if there is any) 🙏 Thank you @y3sar

[y3sar]

@ydshieh @amyeroberts I have found some small(er) models. The model that is being used currently is 803 mb. I have found a checkpoint that is 500 mb. Which is still very big but smaller. And also I have found an even smaller checkpoint that is used for testing by the huggingface internals. But the results are not that reliable.

[amyeroberts]

@y3sar Thanks for looking into this! The tiny models are just for internal testing and probably not something we want to have in the docs (we want to be able to adapt to our needs without considering breaking changes). I'd say go for the 500MB one :)

[amyeroberts]

@y3sar Thanks again for iterating and adding this!