shouxieai / tensorRT_Pro

C++ library based on tensorrt integration
MIT License
2.62k stars 549 forks source link

win10 cuda11.6 cudnn8.4 tensorrt8.4.2.4 编译报错如下 #161

Open hanyong-max opened 2 years ago

hanyong-max commented 2 years ago

[2022-10-11 15:42:17][info][app_yolo.cpp:134]:===================== test YoloV5 FP32 yolov5s6 ================================== [2022-10-11 15:42:17][info][trt_builder.cpp:474]:Compile FP32 Onnx Model 'yolov5s6.onnx'. [2022-10-11 15:42:21][info][trt_builder.cpp:558]:Input shape is -1 x 3 x 640 x 640 [2022-10-11 15:42:21][info][trt_builder.cpp:559]:Set max batch size = 1 [2022-10-11 15:42:21][info][trt_builder.cpp:560]:Set max workspace size = 1024.00 MB [2022-10-11 15:42:21][info][trt_builder.cpp:561]:Base device: [ID 0]<NVIDIA GeForce RTX 2060 SUPER>[arch 7.5][GMEM 6.83 GB/8.00 GB] [2022-10-11 15:42:21][info][trt_builder.cpp:564]:Network has 1 inputs: [2022-10-11 15:42:21][info][trt_builder.cpp:570]: 0.[images] shape is -1 x 3 x 640 x 640 [2022-10-11 15:42:21][info][trt_builder.cpp:576]:Network has 5 outputs: [2022-10-11 15:42:21][info][trt_builder.cpp:581]: 0.[output] shape is -1 x 136000 x 8 [2022-10-11 15:42:21][info][trt_builder.cpp:581]: 1.[911] shape is -1 x 4 x 160 x 160 x 8 [2022-10-11 15:42:21][info][trt_builder.cpp:581]: 2.[954] shape is -1 x 4 x 80 x 80 x 8 [2022-10-11 15:42:21][info][trt_builder.cpp:581]: 3.[997] shape is -1 x 4 x 40 x 40 x 8 [2022-10-11 15:42:21][info][trt_builder.cpp:581]: 4.[1040] shape is -1 x 4 x 20 x 20 x 8 [2022-10-11 15:42:21][info][trt_builder.cpp:585]:Network has 1524 layers: [2022-10-11 15:42:21][info][trt_builder.cpp:652]:Building engine... [2022-10-11 15:42:35][error][trt_builder.cpp:30]:NVInfer: 2: [ltWrapper.cpp::nvinfer1::rt::CublasLtWrapper::setupHeuristic::349] Error Code 2: Internal Error (Assertion cublasStatus == CUBLAS_STATUS_SUCCESS failed. ) [2022-10-11 15:42:35][error][trt_builder.cpp:656]:engine is nullptr [2022-10-11 15:42:35][error][yolo.cpp:192]:Engine yolov5s6.FP32.trtmodel load failed [2022-10-11 15:42:35][error][app_yolo.cpp:54]:Engine is nullptr

麻烦帮忙看看

lhehejunl commented 1 year ago

我遇到了和你相似的问题,我的是3050TI显卡,cuda10.2,cudnn8.2.2.26. tensorRT8.0.1.6,单独跑tensorRT的sample demo,也报错,这样的配置1080TI和1650显卡都没问题。然后我换成cuda11.5,cudnn8.4.0.27,tensorRT8.2.1.8. tensorRT的demo可以跑通,但是onnx转trtmodel报错,转不成功,engine is nullptr。还报找不到cudnn的dll的错,很费解,最后换了cuda11.4,cudnn8.2.2.26。跑通了。所以要不然你尝试换换cuda版本?