PaddleOCR服务化部署仅能识别样例图片，其他任何图片都报错无法正常识别

• 【FastDeploy版本】fastdeploy:1.0.4-gpu-cuda11.4-trt8.5-21.10
• 【编译命令】参考(https://github.com/PaddlePaddle/FastDeploy/tree/develop/examples/vision/ocr/PP-OCR/serving/fastdeploy_serving)
• 【系统平台】: Linux x64
• 【硬件】： Nvidia GPU Tesla P4， CUDA 11.6 CUDNN 8.5
• 【编译语言】：Python(3.9）

问题日志及出现问题的操作流程

• 【无法正常识别样例以外的图片】
• • 先执行examples下的部署示例，包括使用examples提供的模型，确认可以正确执行
• • 按照教程，服务端运行后，在客户端运行python3 client.py，对于样例图片12.jpg可以正常识别，识别结果：
    -(ocr) [topsci@localhost fastdeploy_serving]$ python3 client.py
    text= 上海斯格威铂尔大酒店 score= 0.9804123044013977 bbox= [ 42 413 483 391 484 428 43 450]
    text= 打浦路15号 score= 0.9606184363365173 bbox= [187 456 399 448 400 480 188 488]
    text= 绿洲仕格维花园公寓 score= 0.9806857705116272 bbox= [ 23 507 513 488 515 529 24 548]
    text= 打浦路252935号 score= 0.9498506784439087 bbox= [ 74 553 427 542 428 571 75 582]

-- 如果更改client.py文件中的图片，例如将im = cv2.imread("12.jpg")改为im = cv2.imread("002.jpg")，“002.jpg”为另一张图片，则会报错，报错内容：
-(ocr) [topsci@localhost fastdeploy_serving]$ python3 client.py
Traceback (most recent call last):
File "/home/topsci/code/OCR/fastdeploy_serving/client.py", line 100, in
result = runner.Run([im, ])
File "/home/topsci/code/OCR/fastdeploy_serving/client.py", line 71, in Run
results = self._client.infer(
File "/home/topsci/anaconda3/envs/ocr/lib/python3.9/site-packages/tritonclient/grpc/init.py", line 1446, in infer
raise_error_grpc(rpc_error)
File "/home/topsci/anaconda3/envs/ocr/lib/python3.9/site-packages/tritonclient/grpc/init.py", line 76, in raise_error_grpc
raise get_error_grpc(rpc_error) from None
tritonclient.utils.InferenceServerException: [StatusCode.INTERNAL] in ensemble 'pp_ocr', Failed to process the request(s) for model instance 'det_postprocess_0', message: TritonModelException: in ensemble 'cls_pp', softmax_0.tmp_0: failed to perform CUDA copy: invalid argument

At:
/ocr_serving/code/OCR/fastdeploy_serving/models/det_postprocess/1/model.py(176): execute

想问一下这是怎么回事。

Start with the serving example and client.py, comparing the successful 12.jpg request with 002.jpg. Then inspect models/det_postprocess/1/model.py around line 176 and the Triton pp_ocr/cls_pp execution path. Done means identifying and resolving why non-sample images trigger the CUDA copy error, with both image cases completing successfully.