perf(layout): Render only the requested page when hotloading
KR
1mo ago+53+53+01234567890123456789-3-3-0123456789
Summary
PageLayout._get_image hotloads a page image for annotate() when the page has neither an image nor an image path. It rendered every page of the PDF to PNG in a temp dir and then opened one of them, so its cost grew with document length. It now passes the page to convert_pdf_to_image as first_page/last_page, so only that page is rendered.
Impact
sample-docs/pdf2image-memory-error-test-400p.pdf, one call:
| Before | After | |
|---|---|---|
| Page 1 | 19.9 s | 0.07 s |
| Page 400 | 16.3 s | 0.05 s |
| Temporary PNGs at peak | ~165 MiB | one page |
Behavior preserved
- Page numbers below 1 index from the end of the document (
image_paths[page_number - 1]); page 0 returns the last page, whichtest_get_imagerelies on. Those still render every page. - A page past the end raises the same
ValueError. - Pixels, size, mode and rotation metadata (
pdf_rotation,pdf_rotation_correction) match the previous output for every page number from −n−1 to n+1. Checked onrotated-page-90.pdf,loremipsum_multipage.pdf,empty-document.pdf,layout-parser-paper.pdf,form-field.pdf, and a generated 5-page document with pages rotated 0/180/90/270/90. That covers rotation corrections that are looked up by page index when only part of the document is rendered.
Verification
- New tests in
test_unstructured_inference/inference/test_layout.py:- only the requested page is rendered, and its pixels match a full render (fails on
main) - page numbers 0, −1 and −9 still index from the end
- a page past the end still raises
ValueError
- only the requested page is rendered, and its pixels match a full render (fails on
make test: 386 passed, 1 skipped.ruff check/ruff format --checkclean.
No overlap with #524, which adds pre-rendered image_paths to from_file/process_file_with_model and doesn't touch _get_image. Version bumped to 1.6.17 with a CHANGELOG entry; #530–#532 also target 1.6.17, so later merges need the next patch version.