[Copilot]

PR Checklist

• Tests pass
• Docs if needed
• pre-commit passes

Changes

The GPU COCO benchmark was forwarding CUDA tensors through a model whose weights were still on CPU. This regressed after RF-DETR moved to lazy device placement, which the benchmark bypasses by calling the raw model directly.

• Benchmark device sync

  • Move the underlying nn.Module onto the requested test device immediately after model construction in the GPU/CPU benchmark path.
  • Keep the fix scoped to the benchmark that uses rfdetr.model.model(...) directly.

• Behavioral intent

  • Align the test with current RF-DETR initialization semantics without changing library code.
  • Preserve the existing benchmark flow and thresholds.

rfdetr = model_cls(device=device)
rfdetr.model.model = rfdetr.model.model.to(device)

Related

[Copilot]

Pull request overview

Fixes the COCO GPU inference benchmark in rfdetr_plus tests when RF-DETR uses lazy device placement, by ensuring the underlying nn.Module weights are moved onto the benchmark’s selected device before directly calling the raw model.

Changes:

• Explicitly moves rfdetr.model.model to the chosen device in the COCO detection inference benchmark.
• Adds inline comments explaining why the benchmark must materialize device placement (since it bypasses public inference APIs).

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94%. Comparing base (5ac42d6) to head (ca45e8e).

Additional details and impacted files

@@        Coverage Diff         @@
##           main   #22   +/-   ##
==================================
  Coverage    94%   94%           
==================================
  Files         6     6           
  Lines        71    71           
==================================
  Hits         67    67           
  Misses        4     4           

🚀 New features to boost your workflow:

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