Bug: `NotImplementedError` in PIR OneDNN instruction converter on CPU — released v3.3.0/v3.3.1 still affected

[minh-vt]

Severity: Critical (blocks all CPU inference for PP-OCRv5 models)

Description

PaddlePaddle 3.3.0 and 3.3.1 crash immediately on CPU inference with PP-OCRv5 models:

(Unimplemented) ConvertPirAttribute2RuntimeAttribute not support
[pir::ArrayAttribute<pir::DoubleAttribute>]
(at /paddle/paddle/fluid/framework/new_executor/instruction/onednn/onednn_instruction.cc:116)

This happens because the PIR executor's OneDNN instruction handler (onednn_instruction.cc) doesn't handle ArrayAttribute<DoubleAttribute> — an attribute type used by PP-OCRv5 detection models.

Root cause

PR #77430 added this missing handler and was merged to develop on April 16, 2026. However, the fix was never cherry-picked to the v3.3.0 or v3.3.1 release tags. As of July 2026, the latest released versions still crash.

Environment

• OS: Arch Linux (x86_64)
• PaddlePaddle: 3.3.1 (CPU-only, installed from PyPI)
• PaddleOCR: 3.4.1 → downgraded to 3.3.3 (workaround)
• Model: PP-OCRv5 (both mobile and server variants)
• Device: CPU (no CUDA)

Reproduction

import numpy as np
from paddleocr import PaddleOCR

ocr = PaddleOCR(lang='en')
img = np.zeros((100, 300, 3), dtype=np.uint8)
result = list(ocr.predict_iter(img))  # NotImplementedError

Workaround

Downgrade to PaddlePaddle 3.2.0 + PaddleOCR 3.3.x:

pip install paddlepaddle==3.2.0
pip install "paddleocr>=3.3.0,<3.4.0"

This works because PaddlePaddle 3.2.0 either doesn't use PIR for inference, or its OneDNN handler doesn't hit the unimplemented code path.

What we need

Backport PR #77430 to the v3.3.x release branch and publish a patch release (v3.3.2). The fix already exists in develop — it just needs to reach released versions.

Related

• PR #77430 (merged to develop, April 2026)
• Issue #77340 (original bug report)
• Discussion PaddleOCR#17350 (community workaround: downgrade to 3.2.0)