Optimize html tag replace regex (Fixes #331) #373

xenova · 2023-05-20T16:47:57Z

Building/previewing the docs for Transformers.js takes >12 minutes on my system and ~15 minutes with GitHub actions:

After investigating further, its due to this line of code, which consumes >99% of build time:

doc-builder/src/doc_builder/convert_md_to_mdx.py

Lines 85 to 87 in c59346e

    
           _re_lt_html = re.compile(r"<(\S+)([^>]*>)(((?!</\1>).)*)<(/\1>)", re.DOTALL) 
        
           while _re_lt_html.search(text): 
        
               text = _re_lt_html.sub(r"LTHTML\1\2\3LTHTML\5", text)

The regex is very strict, and requires that the start and end tag match. However, in practice, this is not necessary at all (and in fact, will break on slightly malformed html tags). A better (and much faster approach) is to just replace the < character of any valid start/end tag (regardless of whether they have a matching pair). That is the purpose of this PR.

Before:

Building the MDX files: 100%|██████████| 30/30 [10:35<00:00, 22.71s/it]

After:

Building the MDX files: 100%|██████████| 30/30 [00:00<00:00, 124.46it/s]

which is a ~3000x speedup.

This fixes #331, which reported the same issue.

xenova · 2023-05-20T16:48:45Z

Although it passes the unit tests (except of course the timm docstring one which we know of), I think we should run before and after this change on the transformers docs and perform a diff to make sure nothing crazy happens.

xenova · 2023-05-20T17:49:02Z

Unrelated failing tests fixed here

testing huggingface/doc-builder#373

xenova · 2023-05-21T10:03:08Z

@mishig25 I see it fails on the case:

 122 |  <p>TP is almost always used within a single node. That is TP size <= gpus per node.</p>

Should be simple to fix. I'll also add this as a test case.

Mostly edge-cases Note: HTML is quite forgiving and will still parse/render tags even when some tags are mismatched or missing. The previous regex would barf on this.

sgugger

This looks simpler to me, thanks! Still would like for @mishig25 to double-check since he originally wrote this part, to make sure we are not forgetting any other use-case.

xenova · 2023-05-22T13:23:13Z

This looks simpler to me, thanks! Still would like for @mishig25 to double-check since he originally wrote this part, to make sure we are not forgetting any other use-case.

Agreed! :) Other than huggingface/transformers#23646, everything seems to convert correctly. To be sure, we can perhaps run a diff between the two?

sgugger · 2023-05-22T13:29:01Z

Note that the tool is used across all Hugging Face repos, so there might be some issues specific to other repos too.

xenova · 2023-05-22T13:38:32Z

Note that the tool is used across all Hugging Face repos, so there might be some issues specific to other repos too.

Right - it's probably best to do what @mishig25 did here for all the repos listed here.

xenova · 2023-05-29T11:31:41Z

@mishig25 @sgugger Now that huggingface/transformers#23646 and #374 are both merged, could we re-run the gh actions so we can preview the documentation? 😇

Once they pass, I will do some more tests for the other repos.

Closes #23625 Testing huggingface/doc-builder#373

xenova · 2023-05-30T16:19:37Z

Failing tests related to #374 (which have been fixed)

mishig25 · 2023-05-30T16:25:06Z

oh I see, @xenova could you update this PR branch with the main branch ?

xenova · 2023-05-30T16:29:48Z

Done 👍 . Looks like the docs built fine (https://moon-ci-docs.huggingface.co/docs/transformers/pr_23867)

xenova · 2023-05-30T16:35:40Z

I imagine that since HF is a large org, GH probably provisions pretty decent hardware for actions, so the speedup isn't as dramatic as it is for Transformers.js (3000x speedup).

But here it is for the Transformers:

Before:

After:

I'm not sure what the rest of the 8 min difference is, but the difference for building mdx files is pretty noticeable too.

radames · 2023-08-08T14:59:17Z

I came to this issue while trying to preview xenova/transformers.js docs. Is there a workaround to speed up the preview using docs-preview ? right now it taking considering amount of time just to preview a couple of docs. thanks

sgugger · 2023-08-08T15:06:00Z

@mishig25 friendly ping.

xenova · 2023-08-21T20:03:08Z

My docs on transformers.js builds are getting quite long (e.g., the latest one is ~21 mins), so it would be nice to get this merged 😅 @mishig25

I could switch to use this branch in the meantime, but I think other repos will also benefit from this.

sgugger · 2023-08-22T05:57:04Z

We can address any of Mishig's comments in a followup PR if needed. Let's not block everyone as doc-building is quite long on all Transformers projects!

xenova · 2023-08-22T13:59:51Z

Thanks so much for merging @sgugger! Just did a test now and it sped up the building for transformers.js by 7000x 🤯 (0.0355it/s-> 246.61it/s)

Before:

Building the MDX files: 100%|██████████| 34/34 [15:56<00:00, 28.14s/it]

After:

Building the MDX files: 100%|██████████| 34/34 [00:00<00:00, 246.61it/s]

and brought down the entire process from 21 mins to 4 mins (with 3 mins spent initializing containers)

sgugger · 2023-08-22T15:07:27Z

Went from 30 minutes to 12 minutes for Transformers :-)

radames · 2023-08-22T18:40:03Z

amazing @xenova super fast now!!

Wauplin · 2023-08-29T10:05:26Z

Hey, I just found out about this change as it just broke huggingface_hub's CI. I have a docstring here in which both a "<" and a ">" are used but not as HTML tag:

...
... Small files (<5MB) are duplicated in ...
...
... README.md -> ../../blobs/d7edf6bd2a681fb0175f7735299831ee1b22b812 ....

The new regex considers it as an HTML tag which ends up breaking the doc build as the < char is not escaped properly.

@xenova Given how great the speed improvement of your change is, I'll try to update the regex but still keep your approach. I'll keep you updated :)

EDIT: Opened a new PR to fix the regex: #394.

Optimize html tag replace regex

01dacdb

mishig25 requested review from mishig25 and sgugger May 21, 2023 09:11

mishig25 added a commit to huggingface/transformers that referenced this pull request May 21, 2023

[wip: testing doc-builder]

c9f3cff

testing huggingface/doc-builder#373

mishig25 mentioned this pull request May 21, 2023

[wip: testing doc-builder] huggingface/transformers#23625

Closed

xenova added 5 commits May 21, 2023 12:31

Simplify replacement regex

4895cfb

Add additional unit tests

21237ea

Mostly edge-cases Note: HTML is quite forgiving and will still parse/render tags even when some tags are mismatched or missing. The previous regex would barf on this.

Fix quality checks

5310f3b

Improve special character replacement regex

4280f07

Add more edge-case unit tests

32a900b

xenova mentioned this pull request May 22, 2023

[wip: testing doc-builder] huggingface/transformers#23643

Closed

sgugger approved these changes May 22, 2023

View reviewed changes

mishig25 added a commit to huggingface/transformers that referenced this pull request May 30, 2023

[wip: test doc-builder]

9d1487a

Closes #23625 Testing huggingface/doc-builder#373

mishig25 mentioned this pull request May 30, 2023

[wip: test doc-builder] huggingface/transformers#23867

Closed

Merge branch 'huggingface:main' into optimize

625fa9c

sgugger merged commit 2d60a3d into huggingface:main Aug 22, 2023

Wauplin mentioned this pull request Aug 29, 2023

FIX html escape regex #394

Merged

Wauplin mentioned this pull request Sep 20, 2023

Fix regex to catch HTML tags #398

Merged

mishig25 mentioned this pull request Sep 23, 2023

Fix svelte special chars escaping problem #401

Merged

1 task

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Optimize html tag replace regex (Fixes #331) #373

Optimize html tag replace regex (Fixes #331) #373

xenova commented May 20, 2023

xenova commented May 20, 2023 •

edited

Loading

xenova commented May 20, 2023

xenova commented May 21, 2023 •

edited

Loading

sgugger left a comment •

edited

Loading

xenova commented May 22, 2023

sgugger commented May 22, 2023

xenova commented May 22, 2023

xenova commented May 29, 2023 •

edited

Loading

xenova commented May 30, 2023

mishig25 commented May 30, 2023

xenova commented May 30, 2023 •

edited

Loading

xenova commented May 30, 2023 •

edited

Loading

radames commented Aug 8, 2023

sgugger commented Aug 8, 2023

xenova commented Aug 21, 2023

sgugger commented Aug 22, 2023

xenova commented Aug 22, 2023 •

edited

Loading

sgugger commented Aug 22, 2023

radames commented Aug 22, 2023

Wauplin commented Aug 29, 2023 •

edited

Loading

	_re_lt_html = re.compile(r"<(\S+)([^>]>)(((?!</\1>).))<(/\1>)", re.DOTALL)
	while _re_lt_html.search(text):
	text = _re_lt_html.sub(r"LTHTML\1\2\3LTHTML\5", text)

Optimize html tag replace regex (Fixes #331) #373

Optimize html tag replace regex (Fixes #331) #373

Conversation

xenova commented May 20, 2023

xenova commented May 20, 2023 • edited Loading

xenova commented May 20, 2023

xenova commented May 21, 2023 • edited Loading

sgugger left a comment • edited Loading

Choose a reason for hiding this comment

xenova commented May 22, 2023

sgugger commented May 22, 2023

xenova commented May 22, 2023

xenova commented May 29, 2023 • edited Loading

xenova commented May 30, 2023

mishig25 commented May 30, 2023

xenova commented May 30, 2023 • edited Loading

xenova commented May 30, 2023 • edited Loading

radames commented Aug 8, 2023

sgugger commented Aug 8, 2023

xenova commented Aug 21, 2023

sgugger commented Aug 22, 2023

xenova commented Aug 22, 2023 • edited Loading

sgugger commented Aug 22, 2023

radames commented Aug 22, 2023

Wauplin commented Aug 29, 2023 • edited Loading

xenova commented May 20, 2023 •

edited

Loading

xenova commented May 21, 2023 •

edited

Loading

sgugger left a comment •

edited

Loading

xenova commented May 29, 2023 •

edited

Loading

xenova commented May 30, 2023 •

edited

Loading

xenova commented May 30, 2023 •

edited

Loading

xenova commented Aug 22, 2023 •

edited

Loading

Wauplin commented Aug 29, 2023 •

edited

Loading