Segmentation fault with multithreaded engine build

AI & Data Science Deep Learning (Training & Inference) TensorRT
3

Will NVIDIA ever push patches to TensorRT 6, or is its development frozen (and all bugfixes put on TRT7 and beyond)?

There are no plans on back-porting bug fixes to older versions of TensorRT.

Some other things that stand out that you may want to try if sticking with TensorRT 6.

  1. Try calling cudaSetDevice(idx) within each thread to assign one GPU per builder. Having multiple builders per GPU will cause issues with the kernel/tactic selection, and probably lead to bad performance. This likely isn’t the cause of the seg faults, but is important to know/use, even in TensorRT 7 as well.

  2. Try using methods/flags on builder only, and not builderConfig. I personally encountered some issues with builderConfig periodically in TensorRT 6, but builderConfig has worked seamlessly for me in TensorRT 7.

However, I strongly recommend keeping up to date with the latest version of TensorRT to capture bug fixes.