Inference speed in CPU

[mitchou10]

It's that normal i got really slow inference speed in CPU ?
I got like ~3s on average for this code in docker container :

from paddleocr import PaddleOCR
from time import perf_counter


def create_model():
    return PaddleOCR(
        use_doc_orientation_classify=False,
        use_doc_unwarping=False,
        use_textline_orientation=False,
        ocr_version="PP-OCRv3",
        device="cpu",
        cpu_threads=4,
        # text_det_limit_side_len=960,
        text_det_limit_type="max",
        enable_mkldnn=True,
        lang="fr",
    )


for _ in range(8):
    model = create_model()
    t = perf_counter()
    model.predict(input="tests/data/valid/large.png")
    print(perf_counter() - t)

but when I run it locally I got < 0.1s. Does anyone have any idea ?

Paddle version :

python version : 3.10
"paddleocr==3.0.1"
"paddlepaddle==3.0.0",

Dockerfile:

FROM python:3.10-bullseye AS paddle_model
COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /bin/
ENV DEBIAN_FRONTEND=noninteractive

RUN apt-get update -qy && \
    apt-get install -qy --no-install-recommends \
    ffmpeg \
    libsm6 \
    libxext6 \
    libgl1 \
    libglib2.0-0 \
    poppler-utils \
    libopenblas-openmp-dev \
    liblapack-dev \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/*

as follow the image :

[mitchou10]

I resolve it by using only.

cpu_threads %2 == 0

from paddleocr import PaddleOCR
from time import perf_counter

def create_model():
    return PaddleOCR(
        use_doc_orientation_classify=False,
        use_doc_unwarping=False,
        use_textline_orientation=False,
        ocr_version="PP-OCRv5",
        device="cpu",
        cpu_threads=4,
        text_det_limit_type="max",
        enable_mkldnn=True,
    )