PaddleOCR

Root cause: SealRecognition enables layout detection by default (use_layout_detection=True), and the built-in PP-DocLayout-L layout model classifies red circular seal stamps as invalid regions, truncating the downstream seal detection pipeline. Fix: disable layout detection when initializing the pipeline:

from paddleocr import SealRecognition
ocr = SealRecognition(device="gpu", use_layout_detection=False)
result = ocr.predict(input="testpic.png")

This restores seal detection. Note: circular/arc text on ring-shaped seals is still not fully supported by the official pipeline (rectangular cropping cannot handle curved text, and detection boxes may be offset). As of PaddleOCR 3.7.0 there is no official fix; the suggested community approach is to extract the seal polygon and apply polar-coordinate unwrapping to straighten the arc text before recognition, or fine-tune a custom seal detection model.

This occurs with llama.cpp b9637 when PaddleOCR-VL output containing curly/smart quotes (U+201C/U+201D) is decoded into invalid UTF-8 or U+FFFD bytes, causing llama-server's chat-template parser to return HTTP 500. PaddleOCR 3.7.0 then cancels every pending VLM block and fails the whole predict() call.

Preferred fix: update or rebuild llama-server to a build containing the upstream fix titled "chat: sanitize invalid UTF-8 before peg-native parsing", then restart the server and rerun predict().

Temporary PaddleOCR workaround: patch paddlex/inference/models/doc_vlm/predictor.py in DocVLMGenAIClientPredictor._genai_client_process() so one failed block returns empty content instead of cancelling all futures and re-raising:

import logging

contents = [""] * len(futures)

for idx, future in enumerate(futures):
    try:
        contents[idx] = future.result().choices[0].message.content or ""
    except Exception as exc:
        logging.warning("Skipping failed VLM block %s: %s", idx, exc)
        contents[idx] = ""

Remove or replace the global handler that cancels pending futures and calls raise. The failed paragraph will be blank, but the remaining blocks and page output will be generated. Locate the installed file with:

python -c "import paddlex, os; print(os.path.dirname(paddlex.__file__))"

Restart Python after patching. Updating llama.cpp is preferable because it prevents the parser failure itself.

This usually happens because validation images have variable H/W sizes after DetResizeForTest, so PaddlePaddle's caching allocator keeps blocks for new shapes; training/optimizer memory also remains resident on the T4. Reduce or fix the evaluation input size in your PP-OCRv3 det config:

Eval:
  loader:
    batch_size_per_card: 1
  dataset:
    transforms:
      - DetResizeForTest:
          limit_type: max
          limit_side_len: 736   # try 640 if it still OOMs

The default is often 960. To make memory reuse deterministic, replace it with a fixed 32-divisible shape, for example:

- DetResizeForTest:
    image_shape: [3, 640, 960]

If you modified the evaluation loop, make sure inference runs without autograd:

model.eval()
with paddle.no_grad():
    for batch in eval_loader:
        preds = model(images)

As a temporary workaround, enable CUDA managed memory before training—use true, not the false shown in the old error message:

export FLAGS_use_cuda_managed_memory=true
python tools/train.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_distill_train.yml

If validation during training still fails, run evaluation separately so optimizer states and training cache are not resident:

python tools/eval.py \
  -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_distill_train.yml \
  -o Global.checkpoints=<output_dir>/best_accuracy

Reducing Train.loader.batch_size_per_card to 1 or using a larger GPU are further options.