Zero Zeng

I can not reproduce in my test(official docker image nvcr.io/nvidia/tensorrt:22.07-py3 with TRT 8.4): ``` [08/10/2022-15:02:06] [I] === Performance summary === [08/10/2022-15:02:06] [I] Throughput: 44.5772 qps [08/10/2022-15:02:06] [I] Latency: min =...

@pranavm-nvidia looks like polygraphy issue.

Oops, can you try remove `--trt-outputs mark all --onnx-outputs mark all` in your command?

We don't recommend setting all intermediate layers as output, it will break TRT's layers fusion and usually affect the output accuracy. the better solution is to check the verbose log...

The layers name will remain unchanged in the engine, e.g. ``` Layer(CaskConvolution): node_of_gpu_0/conv1_1 + node_of_gpu_0/res_conv1_bn_1 + node_of_gpu_0/res_conv1_bn_2, Tactic: 0xc2a5fc6b5e7cef5e, gpu_0/data_0[Float(1,3,224,224)] -> gpu_0/res_conv1_bn_2[Float(1,64,112,112)] ``` Layer(CaskConvolution): A + B + C, A,B...

@nvpohanh I can reproduce the error, filed internal bug 3707368.

My try on this bug: constant folding won't work, the error happened on the last matmul layer. if I don't dynamic shape then trt can build engine successfully. 

@rajeevsrao ^ ^

Looks like a bug, can you share the onnx model here?

> @zerollzeng Unfortunately no. It's 485Mb and I can't upload it. You can upload it to Google Drive and share the link here.