Feat/mlx inference

[RacineAI-comp]

What does this PR do?

Adds native MLX segmentation inference so RFDETRSegNano/Small/Medium/Large work with optimize_for_inference(backend="mlx") on Apple Silicon.

• New MLXSegInferenceModel with compiled FP16 forward pass (backbone → decoder with intermediates → seg head → masks)
• New SegHead module (depthwise conv blocks + einsum mask generation)
• Seg head weight conversion (convert_seg_weights) with key remapping
• Decoder return_intermediate support to expose spatial features and per-layer hidden states
• Automatic routing: segmentation_head=True in config → MLXSegInferenceModel, otherwise → MLXInferenceModel
• Masks are resized to original image dimensions and returned as sv.Detections(mask=...)

Related Issue(s): None

Type of Change

• New feature (non-breaking change that adds functionality)

Testing

• I have tested this change locally
• I have added/updated tests for this change

Test details:

• 14 new tests in test_mlx_seg_inference.py: weight conversion key remapping, conv transposition, num_blocks auto-detection, decoder return_intermediate shapes, SegHead output shapes, build_seg_head from dict, postprocess output
  keys/shapes/values
• 2 new routing tests in test_mlx_inference.py: verifies seg config routes to MLXSegInferenceModel and det config routes to MLXInferenceModel
• Full suite: 331 pass, 2 skipped, 0 failures (pytest src/ tests/ -n 2 -m "not gpu")
• pre-commit run --all-files clean
• Verified on real image with RFDETRSegLarge: correct person/couch/remote detections with masks via MLX backend

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code where necessary, particularly in hard-to-understand areas
• My changes generate no new warnings or errors
• I have updated the documentation accordingly (if applicable)

Additional Context

Tested on Apple M4 Pro. The segmentation pipeline reuses the same backbone and decoder as the detection MLX backend, adding only the seg head and return_intermediate plumbing. No changes to existing detection inference behavior.

[CLAassistant]

All committers have signed the CLA.

[Copilot]

Pull request overview

Adds native MLX-based segmentation inference so RF-DETR segmentation models can run via optimize_for_inference(backend="mlx") on Apple Silicon, extending the existing MLX detection backend.

Changes:

• Introduces an MLX segmentation inference path (MLXSegInferenceModel) including an MLX SegHead, seg-weight conversion, and decoder intermediate outputs.
• Adds MLX routing + prediction support in RFDETR.optimize_for_inference() / RFDETR.predict().
• Adds MLX-only test coverage and a new mlx optional dependency extra.

Reviewed changes

Copilot reviewed 10 out of 11 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/rfdetr/detr.py	Adds backend="mlx" routing, MLX model caching, and an MLX prediction path.
src/rfdetr/mlx/__init__.py	Exposes MLX builders for detection vs segmentation inference models.
src/rfdetr/mlx/inference.py	Implements compiled MLX detection + segmentation inference and postprocessing.
src/rfdetr/mlx/decoder.py	Adds return_intermediate support for segmentation features.
src/rfdetr/mlx/seg_head.py	Adds MLX segmentation head implementation + builder from converted weights.
src/rfdetr/mlx/convert_weights.py	Adds segmentation-head weight extraction/remapping and conv transposition support.
src/rfdetr/mlx/backbone.py	Adds MLX DINOv2 backbone implementation used by the MLX inference pipeline.
tests/models/test_mlx_inference.py	Adds MLX detection backend tests and routing tests for seg vs det builds.
tests/models/test_mlx_seg_inference.py	Adds MLX segmentation backend tests (seg weights, intermediates, seg head, postprocess).
pyproject.toml	Adds mlx optional extra and registers an mlx pytest marker.
.gitignore	Ignores additional local artifacts (e.g., .pth, scratch/demo outputs).

[codecov]

Codecov Report

❌ Patch coverage is 2.58718% with 866 lines in your changes missing coverage. Please review.
✅ Project coverage is 69%. Comparing base (7450c16) to head (b0a789e).
⚠️ Report is 9 commits behind head on develop.

❌ Your patch check has failed because the patch coverage (3%) is below the target coverage (95%). You can increase the patch coverage or adjust the target coverage.
❌ Your project check has failed because the head coverage (69%) is below the target coverage (95%). You can increase the head coverage or adjust the target coverage.

Additional details and impacted files

@@           Coverage Diff           @@
##           develop   #767    +/-   ##
=======================================
- Coverage       77%    69%    -8%     
=======================================
  Files           97    103     +6     
  Lines         7538   8426   +888     
=======================================
+ Hits          5801   5830    +29     
- Misses        1737   2596   +859     

🚀 New features to boost your workflow:

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