nv-ingest-api 2025.7.17.dev20250717__py3-none-any.whl → 2025.7.19.dev20250719__py3-none-any.whl

nv_ingest_api/internal/extract/image/chart_extractor.py

@@ -13,7 +13,6 @@ from typing import Tuple
 import numpy as np
 import pandas as pd
 
-from nv_ingest_api.internal.primitives.nim.model_interface.helpers import get_version
 from nv_ingest_api.internal.schemas.extract.extract_chart_schema import ChartExtractorSchema
 from nv_ingest_api.internal.schemas.meta.ingest_job_schema import IngestTaskChartExtraction
 from nv_ingest_api.util.image_processing.table_and_chart import join_yolox_graphic_elements_and_ocr_output
@@ -79,10 +78,13 @@ def _run_chart_inference(
 
     future_yolox_kwargs = dict(
         data=data_yolox,
-        model_name="yolox",
+        model_name="yolox_ensemble",
         stage_name="chart_extraction",
-        max_batch_size=8,
+        input_names=["INPUT_IMAGES", "THRESHOLDS"],
+        dtypes=["BYTES", "FP32"],
+        output_names=["OUTPUT"],
         trace_info=trace_info,
+        max_batch_size=8,
     )
     future_ocr_kwargs = dict(
         data=data_ocr,
@@ -211,24 +213,7 @@ def _create_clients(
     ocr_protocol: str,
     auth_token: str,
 ) -> Tuple[NimClient, NimClient]:
-    # Obtain yolox_version
-    # Assuming that the grpc endpoint is at index 0
-    yolox_http_endpoint = yolox_endpoints[1]
-
-    try:
-        yolox_version = get_version(yolox_http_endpoint)
-        if not yolox_version:
-            logger.warning(
-                "Failed to obtain yolox-page-elements version from the endpoint. Falling back to the latest version."
-            )
-            yolox_version = None  # Default to the latest version
-    except Exception:
-        logger.warning(
-            "Failed to get yolox-page-elements version after 30 seconds. Falling back to the latest version."
-        )
-        yolox_version = None  # Default to the latest version
-
-    yolox_model_interface = YoloxGraphicElementsModelInterface(yolox_version=yolox_version)
+    yolox_model_interface = YoloxGraphicElementsModelInterface()
     ocr_model_interface = OCRModelInterface()
 
     logger.debug(f"Inference protocols: yolox={yolox_protocol}, ocr={ocr_protocol}")

nv_ingest_api/internal/extract/image/image_helpers/common.py

@@ -33,7 +33,6 @@ from PIL import Image
 from nv_ingest_api.internal.enums.common import AccessLevelEnum
 from nv_ingest_api.internal.primitives.nim.model_interface.yolox import (
     YoloxPageElementsModelInterface,
-    get_yolox_model_name,
 )
 from nv_ingest_api.internal.schemas.extract.extract_image_schema import ImageConfigSchema
 from nv_ingest_api.util.image_processing.transforms import crop_image, numpy_to_base64
@@ -202,11 +201,8 @@ def extract_page_elements_from_images(
 
     # Obtain yolox_version
     # Assuming that the http endpoint is at index 1
-    yolox_http_endpoint = config.yolox_endpoints[1]
-    yolox_model_name = get_yolox_model_name(yolox_http_endpoint)
-
     try:
-        model_interface = YoloxPageElementsModelInterface(yolox_model_name=yolox_model_name)
+        model_interface = YoloxPageElementsModelInterface()
         yolox_client = create_inference_client(
             config.yolox_endpoints,
             model_interface,
@@ -220,8 +216,11 @@ def extract_page_elements_from_images(
         # Perform inference in a single call. The NimClient handles batching internally.
         inference_results = yolox_client.infer(
             data,
-            model_name="yolox",
+            model_name="yolox_ensemble",
             max_batch_size=YOLOX_MAX_BATCH_SIZE,
+            input_names=["INPUT_IMAGES", "THRESHOLDS"],
+            dtypes=["BYTES", "FP32"],
+            output_names=["OUTPUT"],
             trace_info=trace_info,
             stage_name="pdf_extraction",
         )

nv_ingest_api/internal/extract/image/table_extractor.py

@@ -77,9 +77,12 @@ def _run_inference(
         data_yolox = {"images": valid_arrays}
         future_yolox_kwargs = dict(
             data=data_yolox,
-            model_name="yolox",
+            model_name="yolox_ensemble",
             stage_name="table_extraction",
             max_batch_size=8,
+            input_names=["INPUT_IMAGES", "THRESHOLDS"],
+            dtypes=["BYTES", "FP32"],
+            output_names=["OUTPUT"],
             trace_info=trace_info,
         )
 
@@ -102,11 +105,10 @@ def _run_inference(
         )
 
     with ThreadPoolExecutor(max_workers=2) as executor:
+        future_ocr = executor.submit(ocr_client.infer, **future_ocr_kwargs)
         future_yolox = None
         if enable_yolox:
             future_yolox = executor.submit(yolox_client.infer, **future_yolox_kwargs)
-        future_ocr = executor.submit(ocr_client.infer, **future_ocr_kwargs)
-
         if enable_yolox:
             try:
                 yolox_results = future_yolox.result()

nv_ingest_api/internal/extract/pdf/engines/nemoretriever.py

@@ -258,6 +258,9 @@ def nemoretriever_parse_extractor(
                     nemoretriever_parse_config.yolox_endpoints,
                     nemoretriever_parse_config.yolox_infer_protocol,
                     nemoretriever_parse_config.auth_token,
+                    input_names=["INPUT_IMAGES", "THRESHOLDS"],
+                    dtypes=["BYTES", "FP32"],
+                    output_names=["OUTPUT"],
                     execution_trace_log=execution_trace_log,
                 )
                 futures.append(future_yolox)

nv_ingest_api/internal/extract/pdf/engines/pdfium.py

@@ -29,9 +29,8 @@ from nv_ingest_api.internal.primitives.nim.default_values import YOLOX_MAX_BATCH
 from nv_ingest_api.internal.primitives.nim.model_interface.yolox import (
     YOLOX_PAGE_IMAGE_PREPROC_WIDTH,
     YOLOX_PAGE_IMAGE_PREPROC_HEIGHT,
-    YOLOX_PAGE_IMAGE_FORMAT,
-    get_yolox_model_name,
     YoloxPageElementsModelInterface,
+    YOLOX_PAGE_IMAGE_FORMAT,
 )
 from nv_ingest_api.internal.schemas.extract.extract_pdf_schema import PDFiumConfigSchema
 from nv_ingest_api.internal.enums.common import TableFormatEnum, TextTypeEnum, AccessLevelEnum
@@ -58,7 +57,6 @@ logger = logging.getLogger(__name__)
 def _extract_page_elements_using_image_ensemble(
     pages: List[Tuple[int, np.ndarray, Tuple[int, int]]],
     yolox_client,
-    yolox_model_name: str = "yolox",
     execution_trace_log: Optional[List] = None,
 ) -> List[Tuple[int, object]]:
     """
@@ -72,8 +70,6 @@ def _extract_page_elements_using_image_ensemble(
         and optional padding offset information.
     yolox_client : object
         A pre-configured client instance for the YOLOX inference service.
-    yolox_model_name : str, default="yolox"
-        The name of the YOLOX model to use for inference.
     execution_trace_log : Optional[List], default=None
         List for accumulating execution trace information.
 
@@ -106,8 +102,11 @@ def _extract_page_elements_using_image_ensemble(
         # Perform inference using the NimClient.
         inference_results = yolox_client.infer(
             data,
-            model_name="yolox",
+            model_name="yolox_ensemble",
             max_batch_size=YOLOX_MAX_BATCH_SIZE,
+            input_names=["INPUT_IMAGES", "THRESHOLDS"],
+            dtypes=["BYTES", "FP32"],
+            output_names=["OUTPUT"],
             trace_info=execution_trace_log,
             stage_name="pdf_extraction",
         )
@@ -317,19 +316,7 @@ def _extract_page_elements(
 
     try:
         # Default model name
-        yolox_model_name = "yolox"
-
-        # Get the HTTP endpoint to determine the model name if needed
-        yolox_http_endpoint = yolox_endpoints[1]
-        if yolox_http_endpoint:
-            try:
-                yolox_model_name = get_yolox_model_name(yolox_http_endpoint)
-            except Exception as e:
-                logger.warning(f"Failed to get YOLOX model name from endpoint: {e}. Using default.")
-
-        # Create the model interface
-        model_interface = YoloxPageElementsModelInterface(yolox_model_name=yolox_model_name)
-
+        model_interface = YoloxPageElementsModelInterface()
         # Create the inference client
         yolox_client = create_inference_client(
             yolox_endpoints,
@@ -340,7 +327,7 @@ def _extract_page_elements(
 
         # Extract page elements using the client
         page_element_results = _extract_page_elements_using_image_ensemble(
-            pages, yolox_client, yolox_model_name, execution_trace_log=execution_trace_log
+            pages, yolox_client, execution_trace_log=execution_trace_log
         )
 
         # Process each extracted element based on extraction flags

nv_ingest_api/internal/primitives/nim/model_interface/helpers.py

@@ -14,6 +14,7 @@ from nv_ingest_api.internal.primitives.nim.model_interface.decorators import mul
 from nv_ingest_api.util.image_processing.transforms import pad_image, normalize_image
 from nv_ingest_api.util.string_processing import generate_url, remove_url_endpoints
 
+cv2.setNumThreads(1)
 logger = logging.getLogger(__name__)
 
 

nv_ingest_api/internal/primitives/nim/model_interface/yolox.py

@@ -12,15 +12,14 @@ from typing import List
 from typing import Optional
 from typing import Tuple
 
-import cv2
+import backoff
 import numpy as np
-import packaging
+import json
 import pandas as pd
-import torch
-import torchvision
 
 from nv_ingest_api.internal.primitives.nim import ModelInterface
-from nv_ingest_api.internal.primitives.nim.model_interface.helpers import get_model_name
+import tritonclient.grpc as grpcclient
+from nv_ingest_api.internal.primitives.nim.model_interface.decorators import multiprocessing_cache
 from nv_ingest_api.util.image_processing import scale_image_to_encoding_size
 from nv_ingest_api.util.image_processing.transforms import numpy_to_base64
 
@@ -35,15 +34,6 @@ YOLOX_PAGE_IMAGE_PREPROC_HEIGHT = 1024
 YOLOX_PAGE_IMAGE_PREPROC_WIDTH = 1024
 YOLOX_PAGE_IMAGE_FORMAT = os.getenv("YOLOX_PAGE_IMAGE_FORMAT", "PNG")
 
-# yolox-page-elements-v1 contants
-YOLOX_PAGE_V1_NUM_CLASSES = 4
-YOLOX_PAGE_V1_FINAL_SCORE = {"table": 0.48, "chart": 0.48}
-YOLOX_PAGE_V1_CLASS_LABELS = [
-    "table",
-    "chart",
-    "title",
-]
-
 # yolox-page-elements-v2 contants
 YOLOX_PAGE_V2_NUM_CLASSES = 4
 YOLOX_PAGE_V2_FINAL_SCORE = {"table": 0.1, "chart": 0.01, "infographic": 0.01}
@@ -63,11 +53,6 @@ YOLOX_GRAPHIC_MIN_SCORE = 0.1
 YOLOX_GRAPHIC_FINAL_SCORE = 0.0
 YOLOX_GRAPHIC_NIM_MAX_IMAGE_SIZE = 512_000
 
-# TODO(Devin): Legacy items aren't working right for me. Double check these.
-LEGACY_YOLOX_GRAPHIC_IMAGE_PREPROC_HEIGHT = 1024
-LEGACY_YOLOX_GRAPHIC_IMAGE_PREPROC_WIDTH = 1024
-YOLOX_GRAPHIC_IMAGE_PREPROC_HEIGHT = 1024
-YOLOX_GRAPHIC_IMAGE_PREPROC_WIDTH = 1024
 
 YOLOX_GRAPHIC_CLASS_LABELS = [
     "chart_title",
@@ -111,8 +96,6 @@ class YoloxModelInterfaceBase(ModelInterface):
 
     def __init__(
         self,
-        image_preproc_width: Optional[int] = None,
-        image_preproc_height: Optional[int] = None,
         nim_max_image_size: Optional[int] = None,
         num_classes: Optional[int] = None,
         conf_threshold: Optional[float] = None,
@@ -126,8 +109,6 @@ class YoloxModelInterfaceBase(ModelInterface):
         Parameters
         ----------
         """
-        self.image_preproc_width = image_preproc_width
-        self.image_preproc_height = image_preproc_height
         self.nim_max_image_size = nim_max_image_size
         self.num_classes = num_classes
         self.conf_threshold = conf_threshold
@@ -199,6 +180,7 @@ class YoloxModelInterfaceBase(ModelInterface):
 
         # Helper functions to chunk a list into sublists of length up to chunk_size.
         def chunk_list(lst: list, chunk_size: int) -> List[list]:
+            chunk_size = max(1, chunk_size)
             return [lst[i : i + chunk_size] for i in range(0, len(lst), chunk_size)]
 
         def chunk_list_geometrically(lst: list, max_size: int) -> List[list]:
@@ -206,29 +188,28 @@ class YoloxModelInterfaceBase(ModelInterface):
             chunks = []
             i = 0
             while i < len(lst):
-                chunk_size = min(2 ** int(log(len(lst) - i, 2)), max_size)
+                chunk_size = max(1, min(2 ** int(log(len(lst) - i, 2)), max_size))
                 chunks.append(lst[i : i + chunk_size])
                 i += chunk_size
             return chunks
 
         if protocol == "grpc":
-            logger.debug("Formatting input for gRPC Yolox model")
-            # Resize images for model input (Yolox expects 1024x1024).
-            resized_images = [
-                resize_image(image, (self.image_preproc_width, self.image_preproc_height)) for image in data["images"]
-            ]
-            # Chunk the resized images, the original images, and their shapes.
-            resized_chunks = chunk_list_geometrically(resized_images, max_batch_size)
+            logger.debug("Formatting input for gRPC Yolox Ensemble model")
+            b64_images = [numpy_to_base64(image, format=YOLOX_PAGE_IMAGE_FORMAT) for image in data["images"]]
+            b64_chunks = chunk_list_geometrically(b64_images, max_batch_size)
             original_chunks = chunk_list_geometrically(data["images"], max_batch_size)
             shape_chunks = chunk_list_geometrically(data["original_image_shapes"], max_batch_size)
 
             batched_inputs = []
             formatted_batch_data = []
-            for r_chunk, orig_chunk, shapes in zip(resized_chunks, original_chunks, shape_chunks):
-                # Reorder axes from (B, H, W, C) to (B, C, H, W) as expected by the model.
-                input_array = np.einsum("bijk->bkij", r_chunk).astype(np.float32)
-                batched_inputs.append(input_array)
+            for b64_chunk, orig_chunk, shapes in zip(b64_chunks, original_chunks, shape_chunks):
+                input_array = np.array(b64_chunk, dtype=np.object_)
+                current_batch_size = input_array.shape[0]
+                single_threshold_pair = [self.conf_threshold, self.iou_threshold]
+                thresholds = np.tile(single_threshold_pair, (current_batch_size, 1)).astype(np.float32)
+                batched_inputs.append([input_array, thresholds])
                 formatted_batch_data.append({"images": orig_chunk, "original_image_shapes": shapes})
+
             return batched_inputs, formatted_batch_data
 
         elif protocol == "http":
@@ -337,32 +318,20 @@ class YoloxModelInterfaceBase(ModelInterface):
         list[dict]
             A list of annotation dictionaries for each image in the batch.
         """
-        original_image_shapes = kwargs.get("original_image_shapes", [])
-
         if protocol == "http":
             # For http, the output already has postprocessing applied. Skip to table/chart expansion.
             results = output
 
         elif protocol == "grpc":
+            results = []
             # For grpc, apply the same NIM postprocessing.
-            pred = postprocess_model_prediction(
-                output,
-                self.num_classes,
-                self.conf_threshold,
-                self.iou_threshold,
-                class_agnostic=False,
-            )
-            results = postprocess_results(
-                pred,
-                original_image_shapes,
-                self.image_preproc_width,
-                self.image_preproc_height,
-                self.class_labels,
-                min_score=self.min_score,
-            )
-
+            for out in output:
+                if isinstance(out, bytes):
+                    out = out.decode("utf-8")
+                if isinstance(out, dict):
+                    continue
+                results.append(json.loads(out))
         inference_results = self.postprocess_annotations(results, **kwargs)
-
         return inference_results
 
     def postprocess_annotations(self, annotation_dicts, **kwargs):
@@ -396,22 +365,15 @@ class YoloxPageElementsModelInterface(YoloxModelInterfaceBase):
     An interface for handling inference with yolox-page-elements model, supporting both gRPC and HTTP protocols.
     """
 
-    def __init__(self, yolox_model_name: str = "nemoretriever-page-elements-v2"):
+    def __init__(self):
         """
         Initialize the yolox-page-elements model interface.
         """
-        if yolox_model_name.endswith("-v1"):
-            num_classes = YOLOX_PAGE_V1_NUM_CLASSES
-            final_score = YOLOX_PAGE_V1_FINAL_SCORE
-            class_labels = YOLOX_PAGE_V1_CLASS_LABELS
-        else:
-            num_classes = YOLOX_PAGE_V2_NUM_CLASSES
-            final_score = YOLOX_PAGE_V2_FINAL_SCORE
-            class_labels = YOLOX_PAGE_V2_CLASS_LABELS
+        num_classes = YOLOX_PAGE_V2_NUM_CLASSES
+        final_score = YOLOX_PAGE_V2_FINAL_SCORE
+        class_labels = YOLOX_PAGE_V2_CLASS_LABELS
 
         super().__init__(
-            image_preproc_width=YOLOX_PAGE_IMAGE_PREPROC_WIDTH,
-            image_preproc_height=YOLOX_PAGE_IMAGE_PREPROC_HEIGHT,
             nim_max_image_size=YOLOX_PAGE_NIM_MAX_IMAGE_SIZE,
             num_classes=num_classes,
             conf_threshold=YOLOX_PAGE_CONF_THRESHOLD,
@@ -478,22 +440,11 @@ class YoloxGraphicElementsModelInterface(YoloxModelInterfaceBase):
     An interface for handling inference with yolox-graphic-elemenents model, supporting both gRPC and HTTP protocols.
     """
 
-    def __init__(self, yolox_version: Optional[str] = None):
+    def __init__(self):
         """
         Initialize the yolox-graphic-elements model interface.
         """
-        if yolox_version and (
-            packaging.version.Version(yolox_version) >= packaging.version.Version("1.2.0-rc5")  # gtc release
-        ):
-            image_preproc_width = YOLOX_GRAPHIC_IMAGE_PREPROC_WIDTH
-            image_preproc_height = YOLOX_GRAPHIC_IMAGE_PREPROC_HEIGHT
-        else:
-            image_preproc_width = LEGACY_YOLOX_GRAPHIC_IMAGE_PREPROC_WIDTH
-            image_preproc_height = LEGACY_YOLOX_GRAPHIC_IMAGE_PREPROC_HEIGHT
-
         super().__init__(
-            image_preproc_width=image_preproc_width,
-            image_preproc_height=image_preproc_height,
             nim_max_image_size=YOLOX_GRAPHIC_NIM_MAX_IMAGE_SIZE,
             num_classes=YOLOX_GRAPHIC_NUM_CLASSES,
             conf_threshold=YOLOX_GRAPHIC_CONF_THRESHOLD,
@@ -551,8 +502,6 @@ class YoloxTableStructureModelInterface(YoloxModelInterfaceBase):
         Initialize the yolox-graphic-elements model interface.
         """
         super().__init__(
-            image_preproc_width=YOLOX_TABLE_IMAGE_PREPROC_HEIGHT,
-            image_preproc_height=YOLOX_TABLE_IMAGE_PREPROC_HEIGHT,
             nim_max_image_size=YOLOX_TABLE_NIM_MAX_IMAGE_SIZE,
             num_classes=YOLOX_TABLE_NUM_CLASSES,
             conf_threshold=YOLOX_TABLE_CONF_THRESHOLD,
@@ -600,144 +549,6 @@ class YoloxTableStructureModelInterface(YoloxModelInterfaceBase):
         return inference_results
 
 
-def postprocess_model_prediction(prediction, num_classes, conf_thre=0.7, nms_thre=0.45, class_agnostic=False):
-    # Convert numpy array to torch tensor
-    prediction = torch.from_numpy(prediction.copy())
-
-    # Compute box corners
-    box_corner = prediction.new(prediction.shape)
-    box_corner[:, :, 0] = prediction[:, :, 0] - prediction[:, :, 2] / 2
-    box_corner[:, :, 1] = prediction[:, :, 1] - prediction[:, :, 3] / 2
-    box_corner[:, :, 2] = prediction[:, :, 0] + prediction[:, :, 2] / 2
-    box_corner[:, :, 3] = prediction[:, :, 1] + prediction[:, :, 3] / 2
-    prediction[:, :, :4] = box_corner[:, :, :4]
-
-    output = [None for _ in range(len(prediction))]
-
-    for i, image_pred in enumerate(prediction):
-        # If no detections, continue to the next image
-        if not image_pred.size(0):
-            continue
-
-        # Ensure image_pred is 2D
-        if image_pred.ndim == 1:
-            image_pred = image_pred.unsqueeze(0)
-
-        # Get score and class with highest confidence
-        class_conf, class_pred = torch.max(image_pred[:, 5 : 5 + num_classes], 1, keepdim=True)
-
-        # Confidence mask
-        squeezed_conf = class_conf.squeeze(dim=1)
-        conf_mask = image_pred[:, 4] * squeezed_conf >= conf_thre
-
-        # Apply confidence mask
-        detections = torch.cat((image_pred[:, :5], class_conf, class_pred.float()), 1)
-        detections = detections[conf_mask]
-
-        if not detections.size(0):
-            continue
-
-        # Apply Non-Maximum Suppression (NMS)
-        if class_agnostic:
-            nms_out_index = torchvision.ops.nms(
-                detections[:, :4],
-                detections[:, 4] * detections[:, 5],
-                nms_thre,
-            )
-        else:
-            nms_out_index = torchvision.ops.batched_nms(
-                detections[:, :4],
-                detections[:, 4] * detections[:, 5],
-                detections[:, 6],
-                nms_thre,
-            )
-        detections = detections[nms_out_index]
-
-        # Append detections to output
-        output[i] = detections
-
-    return output
-
-
-def postprocess_results(
-    results, original_image_shapes, image_preproc_width, image_preproc_height, class_labels, min_score=0.0
-):
-    """
-    For each item (==image) in results, computes annotations in the form
-
-     {"table": [[0.0107, 0.0859, 0.7537, 0.1219, 0.9861], ...],
-      "figure": [...],
-      "title": [...]
-      }
-    where each list of 5 floats represents a bounding box in the format [x1, y1, x2, y2, confidence]
-
-    Keep only bboxes with high enough confidence.
-    """
-    out = []
-
-    for original_image_shape, result in zip(original_image_shapes, results):
-        annotation_dict = {label: [] for label in class_labels}
-
-        if result is None:
-            out.append(annotation_dict)
-            continue
-
-        try:
-            result = result.cpu().numpy()
-            scores = result[:, 4] * result[:, 5]
-            result = result[scores > min_score]
-
-            # ratio is used when image was padded
-            ratio = min(
-                image_preproc_width / original_image_shape[0],
-                image_preproc_height / original_image_shape[1],
-            )
-            bboxes = result[:, :4] / ratio
-
-            bboxes[:, [0, 2]] /= original_image_shape[1]
-            bboxes[:, [1, 3]] /= original_image_shape[0]
-            bboxes = np.clip(bboxes, 0.0, 1.0)
-
-            labels = result[:, 6]
-            scores = scores[scores > min_score]
-        except Exception as e:
-            raise ValueError(f"Error in postprocessing {result.shape} and {original_image_shape}: {e}")
-
-        for box, score, label in zip(bboxes, scores, labels):
-            # TODO(Devin): Sometimes we get back unexpected class labels?
-            if (label < 0) or (label >= len(class_labels)):
-                logger.warning(f"Invalid class label {label} found in postprocessing")
-                continue
-            else:
-                class_name = class_labels[int(label)]
-
-            annotation_dict[class_name].append([round(float(x), 4) for x in np.concatenate((box, [score]))])
-
-        out.append(annotation_dict)
-
-    return out
-
-
-def resize_image(image, target_img_size):
-    w, h, _ = np.array(image).shape
-
-    if target_img_size is not None:  # Resize + Pad
-        r = min(target_img_size[0] / w, target_img_size[1] / h)
-        image = cv2.resize(
-            image,
-            (int(h * r), int(w * r)),
-            interpolation=cv2.INTER_LINEAR,
-        ).astype(np.uint8)
-        image = np.pad(
-            image,
-            ((0, target_img_size[0] - image.shape[0]), (0, target_img_size[1] - image.shape[1]), (0, 0)),
-            mode="constant",
-            constant_values=114,
-        )
-
-    return image
-
-
 def expand_table_bboxes(annotation_dict, labels=None):
     """
     Additional preprocessing for tables: extend the upper bounds to capture titles if any.
@@ -1383,14 +1194,16 @@ def get_bbox_dict_yolox_table(preds, shape, class_labels, threshold=0.1, delta=0
     return bbox_dict
 
 
-def get_yolox_model_name(yolox_http_endpoint, default_model_name="nemoretriever-page-elements-v2"):
+@multiprocessing_cache(max_calls=100)  # Cache results first to avoid redundant retries from backoff
+@backoff.on_predicate(backoff.expo, max_time=30)
+def get_yolox_model_name(yolox_grpc_endpoint, default_model_name="yolox"):
     try:
-        yolox_model_name = get_model_name(yolox_http_endpoint, default_model_name)
-        if not yolox_model_name:
-            logger.warning(
-                "Failed to obtain yolox-page-elements model name from the endpoint. "
-                f"Falling back to '{default_model_name}'."
-            )
+        client = grpcclient.InferenceServerClient(yolox_grpc_endpoint)
+        model_index = client.get_model_repository_index(as_json=True)
+        model_names = [x["name"] for x in model_index.get("models", [])]
+        if "yolox_ensemble" in model_names:
+            yolox_model_name = "yolox_ensemble"
+        else:
             yolox_model_name = default_model_name
     except Exception:
         logger.warning(

nv_ingest_api/internal/primitives/nim/nim_client.py

@@ -8,7 +8,7 @@ import time
 from concurrent.futures import ThreadPoolExecutor
 from typing import Any
 from typing import Optional
-from typing import Tuple
+from typing import Tuple, Union
 
 import numpy as np
 import requests
@@ -90,6 +90,10 @@ class NimClient:
 
     def _fetch_max_batch_size(self, model_name, model_version: str = "") -> int:
         """Fetch the maximum batch size from the Triton model configuration in a thread-safe manner."""
+
+        if model_name == "yolox_ensemble":
+            model_name = "yolox"
+
         if model_name in self._max_batch_sizes:
             return self._max_batch_sizes[model_name]
 
@@ -178,7 +182,7 @@ class NimClient:
             max_requested_batch_size = kwargs.pop("max_batch_size", batch_size)
             force_requested_batch_size = kwargs.pop("force_max_batch_size", False)
             max_batch_size = (
-                min(batch_size, max_requested_batch_size)
+                max(1, min(batch_size, max_requested_batch_size))
                 if not force_requested_batch_size
                 else max_requested_batch_size
             )
@@ -233,7 +237,9 @@ class NimClient:
 
         return all_results
 
-    def _grpc_infer(self, formatted_input: np.ndarray, model_name: str, **kwargs) -> np.ndarray:
+    def _grpc_infer(
+        self, formatted_input: Union[list, list[np.ndarray]], model_name: str, **kwargs
+    ) -> Union[list, list[np.ndarray]]:
         """
         Perform inference using the gRPC protocol.
 

{nv_ingest_api-2025.7.17.dev20250717.dist-info → nv_ingest_api-2025.7.19.dev20250719.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nv-ingest-api
-Version: 2025.7.17.dev20250717
+Version: 2025.7.19.dev20250719
 Summary: Python module with core document ingestion functions.
 Author-email: Jeremy Dyer <jdyer@nvidia.com>
 License:                                  Apache License

{nv_ingest_api-2025.7.17.dev20250717.dist-info → nv_ingest_api-2025.7.19.dev20250719.dist-info}/RECORD

@@ -20,19 +20,19 @@ nv_ingest_api/internal/extract/docx/engines/docxreader_helpers/docxreader.py,sha
 nv_ingest_api/internal/extract/html/__init__.py,sha256=wQSlVx3T14ZgQAt-EPzEczQusXVW0W8yynnUaFFGE3s,143
 nv_ingest_api/internal/extract/html/html_extractor.py,sha256=I9oWfj6_As4898GDDh0zsSuKxO3lBsvyYzhvUotjzJI,3282
 nv_ingest_api/internal/extract/image/__init__.py,sha256=wQSlVx3T14ZgQAt-EPzEczQusXVW0W8yynnUaFFGE3s,143
-nv_ingest_api/internal/extract/image/chart_extractor.py,sha256=KLCSkLUczIlaqIQxyXzKQnwm-HQqTJKMPafnTobOEQs,13873
+nv_ingest_api/internal/extract/image/chart_extractor.py,sha256=gk-O-9wjZBoaLVE_6Erb4gMwsSFk4UtPQ2QLpMCW4H4,13212
 nv_ingest_api/internal/extract/image/image_extractor.py,sha256=4tUWinuFMN3ukWa2tZa2_LtzRiTyUAUCBF6BDkUEvm0,8705
 nv_ingest_api/internal/extract/image/infographic_extractor.py,sha256=i7zt_ow1gytU4hK2JCRg7T1wlbokaeuUpXX69LIQkzY,9687
-nv_ingest_api/internal/extract/image/table_extractor.py,sha256=djTRYgGpwhqWBaVFOerh8J6kVH-xGbUtIelcOFecx4o,13641
+nv_ingest_api/internal/extract/image/table_extractor.py,sha256=O0m3N2Tz9W6X7TBI4o-rbBXc8dFOf9zSZq1v9qC1U4M,13780
 nv_ingest_api/internal/extract/image/image_helpers/__init__.py,sha256=wQSlVx3T14ZgQAt-EPzEczQusXVW0W8yynnUaFFGE3s,143
-nv_ingest_api/internal/extract/image/image_helpers/common.py,sha256=P8rcl4YPyeWeMJg7u1yejD3k9EnDVEbJgfYEnJ4WO5c,15025
+nv_ingest_api/internal/extract/image/image_helpers/common.py,sha256=80jRhGzisHvQ9Ky3MKUMM7soKUmvZ5LqRVzwNYjgdPY,14988
 nv_ingest_api/internal/extract/pdf/__init__.py,sha256=wQSlVx3T14ZgQAt-EPzEczQusXVW0W8yynnUaFFGE3s,143
 nv_ingest_api/internal/extract/pdf/pdf_extractor.py,sha256=CxtWaD6mql9MEqSdk2CfSQ9T-Bn87beBkCOuGGjxGt8,2934
 nv_ingest_api/internal/extract/pdf/engines/__init__.py,sha256=u4GnAZmDKRl0RwYGIRiozIRw70Kybw3A72-lcKFeoTI,582
 nv_ingest_api/internal/extract/pdf/engines/adobe.py,sha256=VT0dEqkU-y2uGkaCqxtKYov_Q8R1028UQVBchgMLca4,17466
 nv_ingest_api/internal/extract/pdf/engines/llama.py,sha256=PpKTqS8jGHBV6mKLGZWwjpfT8ga6Fy8ffrvL-gPAf2c,8182
-nv_ingest_api/internal/extract/pdf/engines/nemoretriever.py,sha256=IVbNcH_phMiRSxnkZ04pGfQrPJ-x1zVR3hXyhxv7juc,22977
-nv_ingest_api/internal/extract/pdf/engines/pdfium.py,sha256=hUqLzQgCJOZIVXrP0JLH4jXLKDHQkXmLXh1Nc4KI3nI,23494
+nv_ingest_api/internal/extract/pdf/engines/nemoretriever.py,sha256=VrIfhDXzYVOwvuoQ9dkiYX-y0CHYhy_Ludadq0YpHqY,23132
+nv_ingest_api/internal/extract/pdf/engines/pdfium.py,sha256=SKmias2iZmAE6Q8WXxmFEjvLOZy-vXRoaRIPpi7Tuhs,22962
 nv_ingest_api/internal/extract/pdf/engines/tika.py,sha256=6GyR2l6EsgNZl9jnYDXLeKNK9Fj2Mw9y2UWDq-eSkOc,3169
 nv_ingest_api/internal/extract/pdf/engines/unstructured_io.py,sha256=jrv2B4VZAH4PevAQrFz965qz8UyXq3rViiOTbGLejec,14908
 nv_ingest_api/internal/extract/pdf/engines/pdf_helpers/__init__.py,sha256=4bvN6LsPksLicI6jM0JqbJFiOZNHEcuc8MVVW4XfgV8,5875
@@ -48,19 +48,19 @@ nv_ingest_api/internal/primitives/control_message_task.py,sha256=nWVB3QsP6p8BKwH
 nv_ingest_api/internal/primitives/ingest_control_message.py,sha256=rvipBiiUaHuRhupFCFDCG8rv0PylSJibCiJ7rDeb98A,8514
 nv_ingest_api/internal/primitives/nim/__init__.py,sha256=i_i_fBR2EcRCh2Y19DF6GM3s_Q0VPgo_thPnhEIJUyg,266
 nv_ingest_api/internal/primitives/nim/default_values.py,sha256=W92XjfyeC6uuVxut6J7p00x1kpNsnXIDb97gSVytZJk,380
-nv_ingest_api/internal/primitives/nim/nim_client.py,sha256=rtzqwHpYsEPuzL7aGIdke5P3_Gu1Z8MbxKDIuJA-L8I,16336
+nv_ingest_api/internal/primitives/nim/nim_client.py,sha256=3jXCNIjVTILsATooxA5Yj0EIyyFo5PrzUzmhT4iYxNI,16481
 nv_ingest_api/internal/primitives/nim/nim_model_interface.py,sha256=wMEgoi79YQn_4338MVemkeZgM1J-vnz0aZWpvqDhib4,2392
 nv_ingest_api/internal/primitives/nim/model_interface/__init__.py,sha256=wQSlVx3T14ZgQAt-EPzEczQusXVW0W8yynnUaFFGE3s,143
 nv_ingest_api/internal/primitives/nim/model_interface/cached.py,sha256=b1HX-PY1ExW5V6pXC1ZiHdobeG_BmbPr3rBbVJef13s,11003
 nv_ingest_api/internal/primitives/nim/model_interface/decorators.py,sha256=qwubkHs4WjnexM6rI0wkjWCsrVNEbA4Wjk2oKL9OYCU,1499
 nv_ingest_api/internal/primitives/nim/model_interface/deplot.py,sha256=TvKdk6PTuI1WNhRmNNrvygaI_DIutkJkDL-XdtLZQac,10787
-nv_ingest_api/internal/primitives/nim/model_interface/helpers.py,sha256=wgcwZJnIn3gKsqe4qhUa9V0gYp3NkIFV8R1qW7Zag1w,11571
+nv_ingest_api/internal/primitives/nim/model_interface/helpers.py,sha256=iyGxAr4tG2UZ7LtXXoWO_kF-KsObhPrmZ46Nl0Mi-Ag,11592
 nv_ingest_api/internal/primitives/nim/model_interface/nemoretriever_parse.py,sha256=WysjDZeegclO3mZgVcGOwzWbr8wSI4pWRiYD4iC2EXo,7098
 nv_ingest_api/internal/primitives/nim/model_interface/ocr.py,sha256=Vhim3py_rc5jA0BoKubwfekEqOwxUUePzcmc59pRuOk,21458
 nv_ingest_api/internal/primitives/nim/model_interface/parakeet.py,sha256=5PqD2JuHY2rwd-6SSB4axr2Dd79vm95sAEkcmI3U7ME,12977
 nv_ingest_api/internal/primitives/nim/model_interface/text_embedding.py,sha256=lFhppNqrq5X_fzbCWKphvZQMzaJd3gHrkWsyJORzFrU,5010
 nv_ingest_api/internal/primitives/nim/model_interface/vlm.py,sha256=qJ382PU1ZrIM-SR3cqIhtY_W2rmHec2HIa2aUB2SvaU,6031
-nv_ingest_api/internal/primitives/nim/model_interface/yolox.py,sha256=nsfDQgeupBe9Tdf3S5sfNpYcObEwVlzCZdfg1ObAW88,49584
+nv_ingest_api/internal/primitives/nim/model_interface/yolox.py,sha256=zpfEZIPctWhNfREnP6e77zffU8vs_RfnMprBj-2jXXk,42847
 nv_ingest_api/internal/primitives/tracing/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nv_ingest_api/internal/primitives/tracing/latency.py,sha256=5kVTeYRbRdTlT_aI4MeS20N_S7mqCcLqZR6YHtxhXkY,2215
 nv_ingest_api/internal/primitives/tracing/logging.py,sha256=SSzIgS7afLH-e1C7VagYDmkkA6rTXmQ-bmtLjoEguhg,3851
@@ -153,8 +153,8 @@ nv_ingest_api/util/service_clients/rest/rest_client.py,sha256=dZ-jrk7IK7oNtHoXFS
 nv_ingest_api/util/string_processing/__init__.py,sha256=mkwHthyS-IILcLcL1tJYeF6mpqX3pxEw5aUzDGjTSeU,1411
 nv_ingest_api/util/system/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nv_ingest_api/util/system/hardware_info.py,sha256=ORZeKpH9kSGU_vuPhyBwkIiMyCViKUX2CP__MCjrfbU,19463
-nv_ingest_api-2025.7.17.dev20250717.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-nv_ingest_api-2025.7.17.dev20250717.dist-info/METADATA,sha256=NkhXN1NC8IKy8OWMV5uJHFdcHY8BOj0dQ2IwnvidETk,13947
-nv_ingest_api-2025.7.17.dev20250717.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-nv_ingest_api-2025.7.17.dev20250717.dist-info/top_level.txt,sha256=abjYMlTJGoG5tOdfIB-IWvLyKclw6HLaRSc8MxX4X6I,14
-nv_ingest_api-2025.7.17.dev20250717.dist-info/RECORD,,
+nv_ingest_api-2025.7.19.dev20250719.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+nv_ingest_api-2025.7.19.dev20250719.dist-info/METADATA,sha256=SnVoo7ElFj94vFFFOFpqqPZWpq35LEkZuFZk0rhpxRw,13947
+nv_ingest_api-2025.7.19.dev20250719.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+nv_ingest_api-2025.7.19.dev20250719.dist-info/top_level.txt,sha256=abjYMlTJGoG5tOdfIB-IWvLyKclw6HLaRSc8MxX4X6I,14
+nv_ingest_api-2025.7.19.dev20250719.dist-info/RECORD,,