TensorRT 5.0.3 failed when deserializing engines (use DLA)

Hi all:
I’m using Drive AGX Xavier with SDK 9 for neural networks inference. I trained MobileNet v2 with Caffe. I want to split the network into pieces because I implemented a better kernel of Depthwise Convolution.
I managed to serialize the pieces networks into serilized engines. But when deserializing in runtime I got these errors:

NvMapMemAllocInternalTagged: 1074810371 error 12
NvMapMemHandleAlloc: error 12
NVMEDIA_DLA : 928, ERROR: load from memory failed.
ERROR: dla/dlaUtils.cpp (472) - DLA Error in deserialize: 7 (Failure to load program.)
terminate called after throwing an instance of ‘nvinfer1::DlaError’
what(): std::exception
Aborted (core dumped)

PS. I want to utilize DLA cores in Xavier.



You don’t need to separate the network but adding the depthwise convolution as a plugin.

struct dwDNNPluginConfiguration
Data Fields
    const dwDNNCustomLayer* customLayers: Array of custom layers.
    size_t numCustomLayers: Number of custom layers.