🐛 Fix Multi-GPU Support with torch.compile

[adamshephard]

Change multi-GPU mode from DataParallel to DataDistributedParallel to work with torch.compile. However, this essentially limits the task to using one GPU alone when using torch.compile. It is not a trivial solution to change this to use multiple GPUs that also work with torch.compile. We will release a future fix to fully correct this with the new engine.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.69%. Comparing base (b564590) to head (f1d7cc4).
⚠️ Report is 9 commits behind head on develop.

Additional details and impacted files

@@           Coverage Diff            @@
##           develop     #923   +/-   ##
========================================
  Coverage    99.69%   99.69%           
========================================
  Files           71       71           
  Lines         8939     8947    +8     
  Branches      1170     1170           
========================================
+ Hits          8912     8920    +8     
  Misses          23       23           
  Partials         4        4           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull Request Overview

This PR adapts multi-GPU support to work with torch.compile by introducing a single‐process DDP fallback and updating cleanup, while retaining a DataParallel fallback for non-compile scenarios.

• Switches from DataParallel to a single‐process DistributedDataParallel (DDP) when multiple GPUs are available and compilation is enabled
• Initializes and later destroys the DDP process group in the engine
• Adds a local (skipped-on-CI) test for multi-GPU feature extraction

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 5 comments.

File	Description
tiatoolbox/models/models_abc.py	Added DDP initialization & wrapping logic in model_to and retained DataParallel fallback
tiatoolbox/models/engine/semantic_segmentor.py	Imports and destroys the DDP process group after inference when compile-compatible
tests/models/test_feature_extractor.py	Introduced a skipped local test for multi-GPU feature extraction

Comments suppressed due to low confidence (1)

tests/models/test_feature_extractor.py:128

• This test uses shutil and Path but there are no corresponding imports. Add import shutil and from pathlib import Path at the top of the file.

shutil.rmtree(save_dir, ignore_errors=True)

[Copilot]

Destroying the process group without verifying initialization may error if no group exists. Add if dist.is_initialized(): before calling destroy_process_group().

Suggested change

	dist.destroy_process_group()
	if dist.is_initialized():
	dist.destroy_process_group()

[shaneahmed]

Thanks @adamshephard