Zero Zeng comments

Results 581 comments of


                                            Zero Zeng

fp16 onnx -> fp16 tensorrt mismatched outputs

The diff between FP32 TRT and FP32 ONNXRT is very close.

fp16 onnx -> fp16 tensorrt mismatched outputs

I've filed internal bug 3813586 to track this.

fp16 onnx -> fp16 tensorrt mismatched outputs

This is not a bug. Model has LayeNorm subgraph in it, when running it in fp16 the results differ between ORT and TRT as ORT. This happens because this subgraph's...

How use TensorRT API to add conv1d layers?

@nvpohanh ^ ^

Unexpected result on multi-batch gather.

Tried to reproduce the issue with TRT 8.4.1.5 using polygraphy: ``` [I] onnxrt-runner-N0-09/07/22-08:16:25 | Completed 1 iteration(s) in 0.1693 ms | Average inference time: 0.1693 ms. [I] Accuracy Comparison |...

Unexpected result on multi-batch gather.

I can reproduce this with ``` [I] trt-runner-N0-09/09/22-00:16:40: output | Stats: mean=0.35972, std-dev=0.34652, var=0.12008, median=0.27958, min=0 at (1, 0, 0), max=0.96826 at (0, 0, 1), avg-magnitude=0.35972 [I] ---- Histogram ----...

Unexpected result on multi-batch gather.

The issue has been fixed in TRT 8.5, there will be a preview feature to fix this issue, please wait for the 8.5 release coming soon :-)

trtexec is not installed in docker containers - Quickstart samples are broken

@rajeevsrao @kevinch-nv Can you help check it ^ ^

Is there any BERT model I can use to test the performence of TensorRT?

https://huggingface.co/docs/transformers/model_doc/bert or our [demoBert](https://github.com/NVIDIA/TensorRT/tree/main/demo/BERT)?

Tensorrt output NaN while onnxruntime is fine

I can reproduce this in TRT 8.5.0.9. but the issue is gone when I don't use dynamic shape. ``` [I] trt-runner-N0-09/22/22-00:17:14 | Completed 1 iteration(s) in 17.49 ms | Average...