Proper way to enable and disable autograd profiler

#1

Currently I use the following. Is there a better way to enable it without manually calling __enter__? Is it necessary (I came up with it when it seemed necessary, but now it was maybe refactored?)?

if args.profile_autograd:
    autograd_profiler = torch.autograd.profiler.profile()
    autograd_profiler.__enter__()

# model running

if args.profile_autograd:
    autograd_profiler.__exit__(None, None, None)
    autograd_profiler.export_chrome_trace(args.profile_autograd)

I don’t want to use with construct because I want to keep enabling the profiler under the flag and prefer not to factor out the model code in a separate function. enable()-kind of API exists for autograd itself, so I thought maybe it exists for the profiler as well. It also exists for nvprof: torch.cuda.profiler.start()

#2

Doesn’t profile take an enabled optional argument?

with torch.autograd.profiler.profile(enabled=args.profile_autograd):

1 Like
#3

It would start then without with construct also, right?

autograd_profiler = torch.autograd.profiler.profile(enabled=args.profile_autograd)
# model code
autograd_profiler.export_chrome_trace(args.profile_autograd)

Do you know if I can stop it explicitly? I could not find a method for that in docs: https://pytorch.org/docs/stable/autograd.html#torch.autograd.profiler.profile

#4

No, you still want the context manager.

#5

Given the migration to the Kineto-based profiler, is there a comparable solution to this problem (some way to create a no-op-ing profiler)? Seems like enabled is no longer part of the profile context manager API:

In [9]: torch.__version__
Out[9]: '1.11.0'

In [10]:     with torch.profiler.profile(
    ...:         activities=[
    ...:             torch.profiler.ProfilerActivity.CPU,
    ...:             torch.profiler.ProfilerActivity.CUDA,
    ...:         ], enabled=False
    ...:     ) as prof:
    ...:         print('hi')
    ...: 
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-10-6132ab91e726> in <module>
      3         torch.profiler.ProfilerActivity.CPU,
      4         torch.profiler.ProfilerActivity.CUDA,
----> 5     ], enabled=False
      6 ) as prof:
      7     print('hi')

TypeError: __init__() got an unexpected keyword argument 'enabled'

Thanks!

cc @tom

#6

You are using the Kineto profiler via torch.profiler.profile which does not have the enabled argument.
(Note the missing .autograd namespace)

#7

My (perhaps incorrect) understanding is that the autograd profiler is deprecated. Is the autograd based profiler still fully supported (and recommended for use)?

#8

I’m not sure about its deprecation status, but I would probably use the newer profiler.

#9

Good to know, thanks @ptrblck