OCR Cpu与Gpu推理结果不一致，Cpu错误

温馨提示：根据社区不完全统计，按照模板提问，可以加快回复和解决问题的速度

环境

• 【FastDeploy版本】： fastdeploy-develop
• 【系统平台】: Linux x64(Ubuntu 18.04) / Windows x64(Windows10)
• 【硬件】： 说明具体硬件型号，如 Nvidia GPU 3090， CUDA 11.7 CUDNN 8.9
• 【编译语言】： C++
• 【PaddleOCR版本】：v2.8

【模型来源】：https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_rec_train.tar
【模型导出】： python3 tools/export_model.py -c configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml -o Global.pretrained_model="pretrain/en_PP-OCRv3_rec_train/best_accuracy" Global.save_inference_dir="pretrain/en_PP-OCRv3_rec_train/exportold/"
【python版本CPU推理正确】：python tools/infer/predict_rec.py --rec_model_dir="pretrain/en_PP-OCRv3_rec_train/exportold/" --image_dir="testimg/3/" --rec_char_dict_path="ppocr/utils/en_dict.txt" --rec_image_shape="3,48,320" --use_gpu=false

[2024/09/03 16:12:09] ppocr INFO: Predicts of testimg/3/2C_105_58_52_0088288__train1_1.jpg:('70', 0.997565746307373)
[2024/09/03 16:12:09] ppocr INFO: Predicts of testimg/3/2C_105_58_52_0088288__train1_2.jpg:('61', 0.8275187015533447)
[2024/09/03 16:12:09] ppocr INFO: Predicts of testimg/3/long_00009204_error_train_1.jpg:('T3O', 0.2974623739719391)
【python版本GPU推理正确】python tools/infer/predict_rec.py --rec_model_dir="pretrain/en_PP-OCRv3_rec_train/exportold/" --image_dir="testimg/3/" --rec_char_dict_path="ppocr/utils/en_dict.txt" --rec_image_shape="3,48,320" --use_gpu=true

[2024/09/03 16:13:21] ppocr INFO: Predicts of testimg/3/2C_105_58_52_0088288__train1_1.jpg:('70', 0.9975684881210327)
[2024/09/03 16:13:21] ppocr INFO: Predicts of testimg/3/2C_105_58_52_0088288__train1_2.jpg:('61', 0.8278857469558716)
[2024/09/03 16:13:21] ppocr INFO: Predicts of testimg/3/long_00009204_error_train_1.jpg:('T3O', 0.2983874976634979)

【C++FastDeployGPU推理正确-ubuntu/windows】保证recshape设置正确，dict也正确，模型一致
结果
testimg/3/2C_105_58_52_0088288__train1_1.jpg:text=70 rec_prob=0.997568
Predicts of testimg/3/2C_105_58_52_0088288__train1_2.jpg:text=61 rec_prob=0.827548
Predicts of testimg/3/long_00009204_error_train_1.jpg:text=T3O rec_prob=0.297802

【C++FastDeployCPU推理错误-ubuntu/windows】保证recshape设置正确，dict也正确，模型一致
结果
testimg/3/2C_105_58_52_0088288__train1_1.jpg:text=70 rec_prob=0.745858
Predicts of testimg/3/2C_105_58_52_0088288__train1_2.jpg:text= rec_prob=0
Predicts of testimg/3/long_00009204_error_train_1.jpg:text= rec_prob=0

请问CPU推理如何改进可以正确推理呢？