Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Simple fix for memory leak on GPU0 #1094

Closed
wants to merge 7 commits into from
Closed

Simple fix for memory leak on GPU0 #1094

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
5c554a1
SA: for #958: set torch cuda device when finding root
shubhamagarwal92 Mar 8, 2020
2f17b2f
SA: for #958: removing root gpu hack in trainer/evaluation_loop
shubhamagarwal92 Mar 8, 2020
6d89505
SA: setting torch cuda device
shubhamagarwal92 Mar 8, 2020
54e9a5e
comment line too long
shubhamagarwal92 Mar 8, 2020
83c291d
check if root gpu exists or available
shubhamagarwal92 Mar 9, 2020
6cba621
Incorporating suggestions on #1094
shubhamagarwal92 Mar 31, 2020
2c7b802
since root gpu returns none instead of -1 for cpu
shubhamagarwal92 Mar 31, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 7 additions & 0 deletions pytorch_lightning/trainer/distrib_parts.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -640,4 +640,11 @@ def determine_root_gpu_device(gpus):
# set root gpu
root_gpu = gpus[0]

# set cuda device to root gpu
# related to https://github.com/PyTorchLightning/pytorch-lightning/issues/958
# Refer solution: https://github.com/pytorch/pytorch/issues/9871#issuecomment-408304190
# root_device = torch.device("cuda", root_gpu)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We can remove this now

root_device = (torch.device("cuda", root_gpu) if root_gpu >= 0 else torch.device("cpu"))
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What would root_device be set if the user wants CPU? None? -1? Maybe we should check for that explicitly

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If the user wants CPU, the function determine_root_gpu_device should not be called (?)

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

in this case it is getting called with gpus=None, and returns None (see first lines of determine_root_gpu_device). So your else torch.device("cpu")) is never relevant.
I don't see anything wrong with just
root_device = torch.device("cuda", root_gpu)

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

but I think the device should be set outside this function anyway

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@awaelchli where do you suggest?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I would search the code base for occurrences of self.root_gpu and check whether it is needed to set the device in each case.
Maybe consider setting directly after here but I am not sure if that's the best place.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

better to ask the core team on this :)

torch.cuda.set_device(root_device)

return root_gpu
13 changes: 12 additions & 1 deletion pytorch_lightning/trainer/evaluation_loop.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -421,9 +421,20 @@ def evaluation_forward(self, model, batch, batch_idx, dataloader_idx, test_mode:
# single GPU data transfer
if self.single_gpu:
# for single GPU put inputs on gpu manually
root_gpu = 0

if isinstance(self.data_parallel_device_ids, list):
root_gpu = self.data_parallel_device_ids[0]

# set cuda device to root gpu
# related to https://github.com/PyTorchLightning/pytorch-lightning/issues/958
# Refer: https://github.com/pytorch/pytorch/issues/9871#issuecomment-408304190
root_device = (torch.device("cuda", root_gpu)
if root_gpu >= 0 else torch.device("cpu"))
torch.cuda.set_device(root_device)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

also need to add tpu device...

else:
raise RuntimeError(
'Expected `data_parallel_device_ids` as a list, cannot determine root gpu.'
)
batch = self.transfer_batch_to_gpu(batch, root_gpu)
args[0] = batch

Expand Down