Hello!
I want to use PyTorch profiler as in this example:
But I get error:
AssertionError: Requested Kineto profiling but Kineto is not available, make sure PyTorch is built with USE_KINETO=1
I instaled PyTorch by command:
pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio===0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
I’m unable to reproduce the issue using the 1.8.1 conda binaries and pip wheels and this simple code:
import torch
import torch.nn as nn
x = torch.randn(1, 1).cuda()
lin = nn.Linear(1, 1).cuda()
with torch.profiler.profile(
activities=[
torch.profiler.ProfilerActivity.CPU,
torch.profiler.ProfilerActivity.CUDA]
) as p:
for _ in range(10):
out = lin(x)
print(p.key_averages().table(
sort_by="self_cuda_time_total", row_limit=-1))
Both show a valid output:
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg Self CUDA Self CUDA % CUDA total CUDA time avg # of Calls
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
aten::addmm 0.68% 22.190ms 99.98% 3.239s 323.913ms 166.000us 100.00% 166.000us 16.600us 10
void cutlass::Kernel<cutlass_80_tensorop_s1688gemm_6... 0.00% 0.000us 0.00% 0.000us 0.000us 142.000us 85.54% 142.000us 14.200us 10
Memcpy DtoD (Device -> Device) 0.00% 0.000us 0.00% 0.000us 0.000us 21.000us 12.65% 21.000us 2.100us 10
Memcpy HtoD (Pageable -> Device) 0.00% 0.000us 0.00% 0.000us 0.000us 3.000us 1.81% 3.000us 3.000us 1
aten::linear 0.02% 544.000us 100.00% 3.240s 323.988ms 0.000us 0.00% 166.000us 16.600us 10
aten::t 0.00% 134.000us 0.01% 204.000us 20.400us 0.000us 0.00% 0.000us 0.000us 10
aten::transpose 0.00% 50.000us 0.00% 70.000us 7.000us 0.000us 0.00% 0.000us 0.000us 10
aten::as_strided 0.00% 24.000us 0.00% 24.000us 1.200us 0.000us 0.00% 0.000us 0.000us 20
aten::empty 0.00% 60.000us 0.00% 60.000us 6.000us 0.000us 0.00% 0.000us 0.000us 10
aten::expand 0.00% 23.000us 0.00% 27.000us 2.700us 0.000us 0.00% 0.000us 0.000us 10
aten::resize_ 0.00% 72.000us 0.00% 72.000us 7.200us 0.000us 0.00% 0.000us 0.000us 10
cudaMemcpyAsync 72.69% 2.355s 72.69% 2.355s 235.493ms 0.000us 0.00% 0.000us 0.000us 10
cudaFree 26.57% 860.777ms 26.57% 860.777ms 286.926ms 0.000us 0.00% 0.000us 0.000us 3
cudaDeviceGetAttribute 0.00% 2.000us 0.00% 2.000us 0.167us 0.000us 0.00% 0.000us 0.000us 12
cudaMalloc 0.03% 943.000us 0.03% 943.000us 235.750us 0.000us 0.00% 0.000us 0.000us 4
cudaMemcpy 0.00% 28.000us 0.00% 28.000us 28.000us 0.000us 0.00% 0.000us 0.000us 1
cudaFuncSetAttribute 0.00% 9.000us 0.00% 9.000us 0.005us 0.000us 0.00% 0.000us 0.000us 1660
cudaEventCreateWithFlags 0.00% 3.000us 0.00% 3.000us 0.167us 0.000us 0.00% 0.000us 0.000us 18
cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFla... 0.00% 2.000us 0.00% 2.000us 0.100us 0.000us 0.00% 0.000us 0.000us 20
cudaLaunchKernel 0.00% 91.000us 0.00% 91.000us 9.100us 0.000us 0.00% 0.000us 0.000us 10
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Self CPU time total: 3.240s
Self CUDA time total: 166.000us
I have got same issue in Your code.
My settings are:
python: 3.9.5
OS: Windows 10
cuda: 11.0
GPU: RTX 3060Ti
I’m unsure, if Kineto is available on Windows, as I was using a Linux OS.
3060 Ti? How did you get it? 
I have the same issue, also on the same versions as @kjnm.
Can someone confirm whether Kineto is supposed to be available on Windows?