Error loading TRTModelFileDocument form JSON

[naoki.tamemoto]

Hi, I’m trying to visualize the TensorRT engine graph using Nsight Deep Learning Designer, but it failed to load trt.json. I’m still not familiar with how to use DL Designer, so I don’t know what’s wrong.

What I did is as below:

• Model conversion
  • Configuration
    • Hardware Platform : Jetson AGX Orin 64GB
    • Jetpack : 6.2.1
    • TensorRT : 10.7.0
  • use a model in Jetson DLA tutorial as a simple model (GitHub - NVIDIA-AI-IOT/jetson_dla_tutorial: A tutorial for getting started with the Deep Learning Accelerator (DLA) on NVIDIA Jetson)
  • First, export onnx model
    • python3 export.py model_gn model_gn.onnx
  • Then, convert model without DLA
    • /usr/src/tensorrt/bin/trtexec --onnx=model_gn.onnx --shapes=input:32x3x32x32 --saveEngine=model_gn.engine --exportLayerInfo=model_gn.trt.json
• Visualization
  • Configuration
    • Nsight Deep Learning Designer 2025.3 (on Windows11)
  • I can open onnx file (model_gn.onnx) with DL Designer
  • However, I can’t open trt.json file (model_gn.trt.json)
    • model_gn.trt.json is as below

      {"Layers": ["/cnn/cnn.0/Conv"
      ,"dummy_shape_call__mye1647_0_myl1_0"
      ,"__myl_Gno_myl1_1"
      ,"__myl_MulAddRel_myl1_2"
      ,"Reformatting CopyNode for Output Tensor 0 to {ForeignNode[/cnn/cnn.1/Constant_1_output_0 + ONNXTRT_unsqueezeTensor.../cnn/cnn.2/Relu]}"
      ,"/cnn/cnn.3/Conv"
      ,"dummy_shape_call__mye1644_0_myl4_0"
      ,"__myl_Gno_myl4_1"
      ,"__myl_MulAddRel_myl4_2"
      ,"Reformatting CopyNode for Input Tensor 0 to /cnn/cnn.6/Conv"
      ,"/cnn/cnn.6/Conv"
      ,"dummy_shape_call__mye1647_0_myl7_0"
      ,"__myl_Gno_myl7_1"
      ,"__myl_MulAddRel_myl7_2"
      ,"Reformatting CopyNode for Input Tensor 0 to /cnn/cnn.9/Conv"
      ,"/cnn/cnn.9/Conv"
      ,"dummy_shape_call__mye1647_0_myl10_0"
      ,"__myl_Gno_myl10_1"
      ,"__myl_MulAddRel_myl10_2"
      ,"/pool/GlobalAveragePool"
      ,"dummy_shape_call__mye594_0_myl12_0"
      ,"/linear/Gemm_myl12_1"
      ],
      "Bindings": ["input"
      ,"output"
      ]}
      

    • error is as below
      
      
    • If I make Layers array empty (rewrite to "Layers": []), the model_gn.trt.json is successfully loaded and visualized as an empty graph (so, I think it is not the issue of a filepath) .

In TensorRT document (Advanced Topics — NVIDIA TensorRT Documentation), it is written that we can visualize TensorRT engine’s computation graph with Nsight Deep Learning Designer, when detailed TensorRT engine layer information is exported to a JSON file with the --exportLayerInfo option, so I think the way to export a JSON file is not wrong.

I think there might be some format inconsistencies, but I don’t know how I can correct them.

I would really appreciate if I get any suggestions or solutions.

[cvanderknyff]

Thanks for reporting this! Add --profilingVerbosity=detailed to your trtexec command line and you should get the necessary JSON format.

The default profiling verbosity level from trtexec is layer_names_only, which saves the list of layers and input/output tensors as in your example, but does not contain sufficient information for DL Designer to visualize meaningfully.

[naoki.tamemoto]

Thank you so much for the answer!

with the option --profilingVerbosity=detailed, I can get .trt.json file including detail information of each layer and it is successfully loaded by Nsight Deep Learning Designer.