Fix: handle logical CUDA device IDs for GPUStatsMonitor if CUDA_VISIBLE_DEVICES set

[XuehaiPan]

What does this PR do?

CUDA device IDs may be different from the real GPU IDs if CUDA_VISIBLE_DEVICES set. For example, if I set CUDA_VISIBLE_DEVICES="3,4,5,6", the CUDA device enumeration will be:

• cuda:0 --> GPU 3
• cuda:1 --> GPU 4
• cuda:2 --> GPU 5
• cuda:3 --> GPU 6

The previous version of GPUStatsMonitor callback will log the wrong devices when users have set a non-trivial CUDA_VISIBLE_DEVICES environment variable.

Before submitting

• Was this discussed/approved via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

Anyone in the community is free to review the PR once the tests have passed.
Before you start reviewing make sure you have read Review guidelines. In short, see the following bullet-list:

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified

Did you have fun?

Make sure you had fun coding 🙃

[pep8speaks]

Hello @XuehaiPan! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2021-07-19 11:02:22 UTC

[codecov]

Codecov Report

Merging #8260 (0ff187a) into master (999ef5c) will decrease coverage by 5%.
The diff coverage is 68%.

@@           Coverage Diff           @@
##           master   #8260    +/-   ##
=======================================
- Coverage      93%     88%    -5%     
=======================================
  Files         217     217            
  Lines       14227   14239    +12     
=======================================
- Hits        13201   12547   -654     
- Misses       1026    1692   +666     

[awaelchli]

great observation, thanks for sending the fix!

[tchaton]

Hey @XuehaiPan,

Mind removing this line there: https://github.com/PyTorchLightning/pytorch-lightning/blob/master/pytorch_lightning/trainer/connectors/logger_connector/logger_connector.py#L230

It was filtering the available GPUS. There is an associated test with it.

Best,
T.C

[XuehaiPan]

Mind removing this line there: https://github.com/PyTorchLightning/pytorch-lightning/blob/master/pytorch_lightning/trainer/connectors/logger_connector/logger_connector.py#L230

gpus_metrics contains all GPUs' memory.used fields, but they are enumerated in the real GPU IDs. If CUDA_VISIBLE_DEVICES is set to integer strings, it's easy to map the GPU IDs to CUDA device IDS. But if CUDA_VISIBLE_DEVICES is set to UUID strings, it could be hard to determine the device mapping.

[carmocca]

Pushed a commit to use list comprehensions instead which are simpler to understand