Does `torch.profiler` support NVTX markers?

1

Hi, guys,
We plan to use nvidia.dali to accelerate our training, which says:

As for profiling, DALI doesn’t have any built-in profiling capabilities, still it utilizes NVTX ranges and has a dedicated domain (so it is easy to find in the profile) to show its operations. So you can capture the profile using NVIDIA Nsight Systems or any Deep Learning profile that also supports NVTX markers.

(Since I have little knowledge about NVTX,)
I am wondering whether torch.profiler supports NVTX markers?

Your answer and guide will be appreciated!

2

Hi sonyuc, torch.profiler uses Kineto, which itself uses the CUPTI library to caputre NVTX events as described here: 2. Usage — Cupti 12.4 documentation. At least this is my understanding.

3

Hi, this is false. Kineto does not support NVTX markers via cupti. See: support of torch.cuda.nvtx.range_push() & torch.cuda.profiler.start() · Issue #318 · pytorch/kineto · GitHub