Fix: Remove device_map to prevent meta tensor errors in table detection

CHANGELOG.md

@@ -1,3 +1,11 @@
+## 1.1.9
+
+- Fix: Remove manual pypdfium2 close() calls to prevent finalization AssertionError during concurrent PDF processing. Let pypdfium2 finalizers handle resource cleanup.
+
+## 1.1.8
+
+- Fix: Remove `device_map` parameter from TableTransformerForObjectDetection to prevent meta tensor errors during multi-threaded table detection. Added device normalization and explicit dtype specification for robustness.
+
 ## 1.1.7
 
 - Update OpenCV-Python to 4.13.0.90 to squash ffmpeg vulnerability CVE-2023-6605

unstructured_inference/inference/layout.py

@@ -426,24 +426,21 @@ def convert_pdf_to_image(
             dpi = inference_config.PDF_RENDER_DPI
         scale = dpi / 72.0
         for i, page in enumerate(pdf, start=1):
-            try:
-                if first_page is not None and i < first_page:
-                    continue
-                if last_page is not None and i > last_page:
-                    break
-                bitmap = page.render(
-                    scale=scale,
-                    no_smoothtext=False,
-                    no_smoothimage=False,
-                    no_smoothpath=False,
-                    optimize_mode="print",
-                )
-                try:
-                    images[i] = bitmap.to_pil()
-                finally:
-                    bitmap.close()
-            finally:
-                page.close()
+            if first_page is not None and i < first_page:
+                continue
+            if last_page is not None and i > last_page:
+                break
+            bitmap = page.render(
+                scale=scale,
+                no_smoothtext=False,
+                no_smoothimage=False,
+                no_smoothpath=False,
+                optimize_mode="print",
+            )
+            # Convert to PIL immediately, then let pypdfium2 finalizers handle cleanup
+            images[i] = bitmap.to_pil()
+            # No manual close - prevents finalization race with parent PDF close
+
         if not output_folder:
             return list(images.values())
         else:

unstructured_inference/models/tables.py

@@ -72,7 +72,7 @@ def initialize(
     ):
         """Loads the donut model using the specified parameters"""
         self.device = device
-        self.feature_extractor = DetrImageProcessor.from_pretrained(model, device_map=self.device)
+        self.feature_extractor = DetrImageProcessor.from_pretrained(model)
         # value not set in the configuration and needed for newer models
         # https://huggingface.co/microsoft/table-transformer-structure-recognition-v1.1-all/discussions/1
         self.feature_extractor.size["shortest_edge"] = inference_config.IMG_PROCESSOR_SHORTEST_EDGE
@@ -82,10 +82,8 @@ def initialize(
             logger.info("Loading the table structure model ...")
             cached_current_verbosity = logging.get_verbosity()
             logging.set_verbosity_error()
-            self.model = TableTransformerForObjectDetection.from_pretrained(
-                model,
-                device_map=self.device,
-            )
+            self.model = TableTransformerForObjectDetection.from_pretrained(model)
+            self.model.to(self.device)
             logging.set_verbosity(cached_current_verbosity)
             self.model.eval()