dense model using tensorrt infernece on the A30 got wrong

[zhaohb]

Environment

• TensorRT 8.6.1

• Versions of CUDA(12.1), CUBLAS(12.1.3.1)

• Container used (tensorrt:23.07-py3)

• NVIDIA driver version (510.47.03)

Model：

• https://drive.google.com/file/d/1Sn7vJYf_dUqRyBn-KNSnH851D7atHu0o/view?usp=drive_link

Reproduction Steps

1. A30

   polygraph run gs_concat.onnx --onnxrt --trt --tf32 --atol 1e-4 --pool-imit workspace:10G

   output：

Expected Behavior

• The result of inference using tensorrt is correct

Actual Behavior

• It can be seen that there is a certain gap between the output of trt and onnx, and the inference result is wrong.

Additional Notes

• We also tested it on the A6000 and found it correct,

  polygraph run gs_concat.onnx --onnxrt --trt --tf32 --atol 1e-4 --pool-imit workspace:10G

  output:

So maybe it's hardware related.

bug已经经过导师确认，nv内部的bug id是4259240.