Easy way to get Kineto to use the new profiler?

#1

Hello, I’m trying out the PyTorch Profiler as per this blog post, but I get this error:

AssertionError: Requested Kineto profiling but Kineto is not available,
                  make sure PyTorch is built with USE_KINETO=1

I dug into the code to see if I could profile without Kineto but it seems the profiler requires it. Is there an easy way to get Pytorch with USE_KINETO set, or does this require building from source?

7 Likes
#2

I also hope there is a way to do that other than rebuild.

Though I don’t think that possible reading pytorch/init.cpp at 728d18f97695def22f237ce343a3a9679aca49b0 · pytorch/pytorch · GitHub.

#3

@alannnna what pytorch version and cuda toolkit have you installed?
Profiler is available starting from pytorch 1.8.1 release

#4

Hi,
I also encountered the same issue, I installed from the source of version 1.9.0a0 on my MacBook Air M1.
Thanks

#5

@Elena_Neroslavskaya Pytorch 1.8.1, no CUDA toolkit as I was just trying it out on Mac.

#6

Did installing 1.9.0a0 from source fix it for you, or are you saying you had the same problem when you installed that version? Thanks.

#7

Hey, I also have the same problem (torch 1.8.1 on windows). Does anyone have a solution to this?

#8

I also encounter the same problem on torch1.8.1 with cuda on windows, is any solution?

#9

Got the same problem here. 1.8.1.

Installed using command: pip3 install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio===0.8.1 -f https://download.pytorch.org/whl/torch_stable.html

#10

I had the same issue. As a workaround I’ve switched to the legacy API
torch.autograd.profiler.profile()
and then it worked.

#11

Same issue. Using Windows, pytorch 1.8.1, torchvision 0.9.1, torch-tb-profiler 0.2.1, cuda 11.4