Assertion `!transp_src_ten->is_mod()' failed while building TensorRT engine with batch = 1


I have a model with some attention layers, but an error occured when i try to convert the onnx to tensorrt engine.
My build script is as follows:

trtexec --onnx=./nognn_v2_part_nvidia.onnx --workspace=1000 --explicitBatch --verbose

I am facing the following error while building the engine:

trtexec: /root/gpgpu/MachineLearning/myelin/src/compiler/optimizer/tactic.cpp:1749: bool myelin::ir::tactic_selection_t::global_tactic_constraint_analysis(): Assertion `!transp_src_ten->is_mod()' failed.

I’m not familiar with libmyelin and unable to find more doc about it, but i found the onnx model can be converted successfully while the batch size is static and greater than 1.
what’s more, the model what i need is the dynamic version, so i ask for help here.

Additionally I have attached the onnx model files below.


TensorRT Version: 8.1.0
GPU Type: Jetson orin-x platform
CUDA Version: 11.4
CUDNN Version: 8.2.6
Operating System + Version: Ubuntu 20.04.2 LTS aarch64
Python Version (if applicable) : 3.8.10
PyTorch Version (if applicable): 1.10.1

Relevant Files

nognn_v2_part_nvidia.onnx (48.7 KB)

nognn_v2_part_nvidia_2b.onnx (48.7 KB)

nognn_v2_part_nvidia_dynamic.onnx (48.8 KB)


We could reproduce the same error. Our team will look into this.

Thank you.


It has been passed almost a month… Is there any update?


Currently, we do not have a simple workaround to share. This issue will be fixed in future releases.

Thank you.