feat: use block matrix to reduce peak memory usage for matmul (#3947)

This PR targets the most memory expensive operation in partition pdf and
images: deduplicate pdfminer elements. In large pages the number of
elements can be over 10k, which would generate multiple 10k x 10k square
double float matrices during deduplication, pushing peak memory usage
close to 13Gb
![Screenshot 2025-03-06 at 3 22
52 PM](https://github.com/user-attachments/assets/fdc26806-947b-4b5a-9d8e-4faeb0179b9f)


This PR breaks this computation down by computing partial IOU. More
precisely it computes IOU for each 2000 elements against all the
elements at a time to reduce peak memory usage by about 10x to around
1.6Gb.

![image](https://github.com/user-attachments/assets/e7b9f149-2b6a-4fc9-83c7-652e20849b76)


The block size is configurable based on user preference for peak memory
usage and it is set by changing the env `UNST_MATMUL_MEMORY_CAP_IN_GB`.

CHANGELOG.md

@@ -1,4 +1,4 @@
-## 0.16.24-dev5
+## 0.16.24
 
 ### Enhancements
 
@@ -6,6 +6,7 @@
   in unstructured and `register_partitioner` to enable registering your own partitioner for any file type.
 
 - **`extract_image_block_types` now also works for CamelCase elemenet type names**. Previously `NarrativeText` and similar CamelCase element types can't be extracted using the mentioned parameter in `partition`. Now figures for those elements can be extracted like `Image` and `Table` elements
+- **use block matrix to reduce peak memory usage for pdf/image partition**.
 
 ### Features
 

unstructured/__version__.py

@@ -1 +1 @@
-__version__ = "0.16.24-dev5"  # pragma: no cover
+__version__ = "0.16.24"  # pragma: no cover

unstructured/partition/pdf_image/pdfminer_processing.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import os
 from typing import TYPE_CHECKING, Any, BinaryIO, Iterable, List, Optional, Union, cast
 
 import numpy as np
@@ -708,10 +709,16 @@ def remove_duplicate_elements(
 ) -> TextRegions:
     """Removes duplicate text elements extracted by PDFMiner from a document layout."""
 
-    iou = boxes_self_iou(elements.element_coords, threshold)
-    # this is equivalent of finding those rows where `not iou[i, i + 1 :].any()`, i.e., any element
-    # that has no overlap above the threshold with any other elements
-    return elements.slice(~np.triu(iou, k=1).any(axis=1))
+    coords = elements.element_coords
+    # experiments show 2e3 is the block size that constrains the peak memory around 1Gb for this
+    # function; that accounts for all the intermediate matricies allocated and memory for storing
+    # final results
+    memory_cap_in_gb = os.getenv("UNST_MATMUL_MEMORY_CAP_IN_GB", 1)
+    n_split = np.ceil(coords.shape[0] / 2e3 / memory_cap_in_gb)
+    splits = np.array_split(coords, n_split, axis=0)
+
+    ious = [~np.triu(boxes_iou(split, coords, threshold), k=1).any(axis=1) for split in splits]
+    return elements.slice(np.concatenate(ious))
 
 
 def aggregate_embedded_text_by_block(