NVIDIA
operation.cpp:203: DCHECK(!i->is_use_only()) failed.
Description
When use trtexec to build a onnx, then raise
[05/21/2024-21:01:40] [V] [TRT] Fastest Tactic: 0x0000000000000000 Time: 0.0205211
[05/21/2024-21:01:40] [V] [TRT] >>>>>>>>>>>>>>> Chose Runner Type: Padding Tactic: 0x0000000000000000
[05/21/2024-21:01:40] [V] [TRT] =============== Computing costs for
[05/21/2024-21:01:40] [V] [TRT] *************** Autotuning format combination: Float(96,16,4,1), Float(1440,45,9,1), Float(3360,105,15,1), Float(10368,324,27,1), Float(36608,1144,52,1) -> Float(73728,73728,8,2,1), Float(55296,9216,1,1), Float(9216,1,1), Float(2359296,256,1), Float(2359296,256,1), Float(294912,9216,96,1) ***************
[05/21/2024-21:01:40] [V] [TRT] --------------- Timing Runner: {ForeignNode[1440...Transpose_1934 + Reshape_1941]} (Myelin)
operation.cpp:203: DCHECK(!i->is_use_only()) failed.
Aborted
Similar case https://forums.developer.nvidia.com/t/tensorrt-conversion-fails-with-dcheck-i-is-use-only/237282
My onnx has some large gemm op, it affect this error ?
Does trt has any restrictions on matrix multiplication operations ?
Environment
TensorRT Version: 8.5.10
NVIDIA GPU:rtx2000
CUDA Version:11.4
CUDNN Version:11.6
Operating System:ubuntu20.04
Python Version (if applicable):3.8
Steps To Reproduce
trtexec --onnx=bevf2_simp.onnx --verbose
Hi, are you open to sharing the corresponding onnx file? I tried reproducing the problem with the onnx file from the other issue but that onnx file runs smoothly with trtexec.
I understand. If you know what part of the network is causing the problem, you could isolate the subgraph and see if that can serve as a minimal example producing the same issue. Then you can consider sharing this minimal subgraph instead of the whole model, if possible.
Happy to help, but I'll need some way to reproduce the issue. Cheers!