How does kineto work actuallly? about kineto HOT 5 CLOSED

1. How does kineto get pytroch functions name?

Kineto uses CUPTI to get the CUDA Runtime and launch names as well the as CUDA kernel names. Kineto interacts with PyTorch Profiler (inside of pytorch/pytorch repo), which helps Kineto get the PyTorch runtime function names. We also have a feature called Python Tracing, enabled by with_stack=True, which turns on the python tracer, and collects all the python runtime stacks too.

2. What is the purpose of KINETO_USE_DAEMON=1 ?

This is a new OSS feature meant to enable on-demand profiling. When the dynolog daemon is enable via this env var, the daemon can directly request Kineto to perform a GPU trace collection. The env variable does not affect whether Kineto is on or not, since Kineto can be used via PyTorch Profiler API calls.

from kineto.

@stricklandye The first issue, as far as I understand, is that pytorch profiler pushes events to kineto actively, not by CUPTI itself.

from kineto.

@shinytang6 Thanks for your reply. Could you please tell me where is the source code related to this? i.e. Pytroch Profiler pushes events to kineto. :D

from kineto.

@shinytang6 Thanks for your reply. Could you please tell me where is the source code related to this? i.e. Pytroch Profiler pushes events to kineto. :D

maybe you can check this:
https://github.com/pytorch/pytorch/blob/e1f159e6b2c0b5d2d169569c490f182771ac643b/torch/csrc/profiler/collection.cpp#L805

from kineto.

I see. Hope someone else could answer the second questions. I will close this issue in few days.

from kineto.