How to use deconvolution layer correctly?

[freedenS]

Description

I tried to use deconvolution with the following code, but there's something wrong like this.
How should i correctly use the deconvolution layer?

[03/25/2021-16:25:01] [F] [TRT] Assertion failed: cublasStatus == CUBLAS_STATUS_SUCCESS
C:\source\rtSafe\cublas\cublasLtWrapper.cpp:279
Aborting...
[03/25/2021-16:25:01] [E] [TRT] C:\source\rtSafe\cublas\cublasLtWrapper.cpp (279) - Assertion Error in nvinfer1::CublasLtWrapper::getCublasLtHeuristic: 0 (cublasStatus == CUBLAS_STATUS_SUCCESS)

Environment

TensorRT Version: 7.2.1.6
NVIDIA GPU: 2080Ti
NVIDIA Driver Version: 441.87
CUDA Version: 10.2
CUDNN Version: 8.0.4
Operating System: win10
Python Version (if applicable):
Tensorflow Version (if applicable):
PyTorch Version (if applicable):
Baremetal or Container (if so, version):

Relevant Files

Steps To Reproduce

IBuilder* builder = createInferBuilder(gLogger);
IBuilderConfig* config = builder->createBuilderConfig();
INetworkDefinition* network = builder->createNetworkV2(0U);

ITensor* input = network->addInput("input", DataType::kFLOAT, Dims4{ 10,256,14,14 });
Weights emptywts{ DataType::kFLOAT, nullptr, 0 };
std::vector<float> deval(262144,1.0);
Weights deconvwtsl{ DataType::kFLOAT, deval.data(), 262144 };
auto deconv = network->addDeconvolutionNd(*input, 256, DimsHW{ 2,2 }, deconvwtsl, emptywts);
deconv->setStrideNd(DimsHW{ 2,2 });
network->markOutput(*deconv->getOutput(0));
builder->setMaxBatchSize(maxBatchSize);
config->setMaxWorkspaceSize(16 * (1 << 20));
ICudaEngine* engine = builder->buildEngineWithConfig(*network, *config);
network->destroy();
IHostMemory* modelStream = engine->serialize();
std::ofstream p ("test.engine", std::ios::binary);
p.write(reinterpret_cast<const char*>(modelStream->data()), modelStream->size());
modelStream->destroy();

[pranavm-nvidia]

@freedenS There's a known cuBLAS LT bug in CUDA 10.2. You can fix it either by upgrading to a newer patch version of 10.2, or using the work-around mentioned here. Since you're using the API you can use config->setTacticSources() to disable cuBLAS LT.

[freedenS]

Thank you very much for the comprehensive solutions!
I have updated cuda and it worked!