Building engine failed of TensorRT 8.6.13 when running on GPU H20

[XuDeshengCat]

Description

I tried to run trtexec on GPU H20, but it fails with the error below：

[12/23/2024-09:56:35] [E] Error[2]: Assertion !layerImpls.empty() failed. /motion_head/Atan has no runners that implement it
[12/23/2024-09:56:35] [E] Error[2]: [standardTacticSupply.cpp::getSupportedFormats::37] Error Code 2: Internal Error (Assertion !layerImpls.empty() failed. /motion_head/Atan has no runners that implement it)
[12/23/2024-09:56:35] [E] Engine could not be created from network
[12/23/2024-09:56:35] [E] Building engine failed
[12/23/2024-09:56:35] [E] Failed to create engine from model or file.
[12/23/2024-09:56:35] [E] Engine set up failed

But I can run the same program on the GPU A10. The software versions are exactly the same（I use the same docker）. I want to ask what caused this problem? Is it because TensorRT 8613 has different support for the Ampere and Hopper architecture GPUs?

Environment

TensorRT Version:8.6.13

NVIDIA GPU:H20

[asfiyab-nvidia]

assigned to @zhenhuaw-me

[zhenhuaw-me]

@XuDeshengCat Thank you for reporting this issue! Can you firstly upgrade to TRT 10.7 and then reproduce? TRT 8.6 is a pretty old release.

[XuDeshengCat]

@XuDeshengCat Thank you for reporting this issue! Can you firstly upgrade to TRT 10.7 and then reproduce? TRT 8.6 is a pretty old release.

Thank you for your reply.
For some hardware and software requirements in our project, we had to use TensorRT 8.6.13. I have observed that the production time of H20 is after 8613TensorRT. I would like to know whether this version of TensorRT is not strong enough to support H20. If this problem does exist, it doesn't matter, we can continue to use the A10. But before I do, I want to make sure : )

[zhenhuaw-me]

I see. TRT 8.6 is the best to run on H20.

Can you share the ONNX model segment about the operation /motion_head/Atan? We can see if anything you can do with your HW/SW setups.