fix(rfdetr): honor --show-boxes instead of ignoring it

[Anai-Guo]

Problem

The RF-DETR CLI declares --show-boxes with help="Show bounding boxes and labels (default: on)", but the parsed value is never used. show_boxes as an identifier appears nowhere in mlx_vlm/models/rfdetr/ outside that one add_argument call:

$ grep -rn "show_boxes" mlx_vlm/models/rfdetr/
$          # (no output — only the "--show-boxes" string literal exists)

Two things are wrong, and the file itself supplies the baseline for both:

1. The value never reaches the annotator. The adjacent --show-fps flag in the very same parser is threaded through (predict_video(..., show_fps=args.show_fps) at generate.py:802). Both sibling backends honor show_boxes too — sam3/generate.py passes args.show_boxes into three call sites, and sam3_1/generate.py threads it down to draw_frame.

2. The flag could not be turned off even once wired up. It is action="store_true" with default=True, so passing --show-boxes is a no-op and there is no way to get boxes off. sam3_1/generate.py:1002-1003 pairs exactly this flag with a --no-show-boxes dual; the RF-DETR parser only has the positive half.

Net effect: a user following the documented (default: on) wording has no way to render mask-only segment output, which is the one thing the option exists for.

Fix

• _get_annotator() takes show_boxes: bool = True. When it is off, the default segment chain keeps MaskAnnotator and drops BoxAnnotator/LabelAnnotator — the same meaning sam3/generate.py documents for draw_frame: "If False, draw only mask overlays + contours (no boxes/labels)."
• A detect chain contains nothing but boxes and labels, so it keeps them regardless. This mirrors sam3's show_boxes=args.show_boxes if args.task == "segment" else True, and avoids rendering an empty overlay.
• An explicit --annotator chain still wins, since it names the annotators outright.
• Added --no-show-boxes (dest="show_boxes", action="store_false"), matching sam3_1/generate.py.

The default path is unchanged. With show_boxes=True the chain is the same MaskAnnotator + BoxAnnotator + LabelAnnotator as before, and the two internal _get_annotator(None, ...) calls in predict_video/predict_realtime keep the default.

Verification

mlx is Apple-silicon only, so I exercised the real module on another platform by stubbing mlx alone and importing mlx_vlm/models/rfdetr/generate.py unmodified — sam3/annotators.py is numpy+cv2 only, so the annotator chain under test is the real one, not a mock.

Before the change (3 of the 4 new tests fail):

FAILED test_rfdetr_show_boxes.py::test_segment_chain_drops_boxes_and_labels_when_show_boxes_is_off
FAILED test_rfdetr_show_boxes.py::test_detect_keeps_boxes_because_nothing_else_is_drawn
FAILED test_rfdetr_show_boxes.py::test_cli_can_turn_boxes_off - SystemExit: 2
   __main__.py: error: unrecognized arguments: --no-show-boxes
3 failed, 1 passed

After: 4 passed.

test_show_boxes_defaults_to_on is the one that passes on both sides — it pins the unchanged default chain so this stays a no-op for existing invocations.

Observed chains:

task	show_boxes	annotator chain
segment	True (default)	MaskAnnotator, BoxAnnotator, LabelAnnotator
segment	False	MaskAnnotator
detect	either	BoxAnnotator, LabelAnnotator

black==26.3.1, isort==5.13.2 and autoflake==2.2.1 (the pinned .pre-commit-config.yaml revisions) leave both files unchanged.

Not in scope

sam3/generate.py declares --show-boxes without default=True, so it defaults to off there and is already honored — I left its semantics alone rather than unifying the two defaults, since that would change existing sam3 output.

🤖 Generated with Claude Code

[Anai-Guo]

The red test job is not caused by this PR. All 4 failures are the MRoPE mx.split() signature change:

tests/test_rope_utils.py::test_sectioned_precomputed_rotary_fast_path_matches_fallback[sectioned_half_split]
tests/test_rope_utils.py::test_sectioned_precomputed_rotary_fast_path_matches_fallback[sectioned_even_odd]
tests/test_models.py::TestMRoPETrainingVJP::test_precomputed_rotary_vjp_matches_pure_mlx  (2 subcases)

TypeError: split(): incompatible function arguments.
  Invoked with types: mlx.core.array, mlx.core.array, kwargs = { axis: int }

models/rope_utils.py:1364 passes split_indices as an mx.array, which recent mlx no longer accepts. Unrelated branches hit the same two test ids in the same window — e.g. runs 32888401433 (feature/model-registry) and 32916522886 (dev-deepseek-v4-speculative-chunked-prefill) — and the fix-mrope-split-mlx-0322 branch is green, so this looks already handled elsewhere.

This PR touches only models/rfdetr/generate.py and adds tests/test_rfdetr_show_boxes.py; the run reports 4 failed, 2581 passed with the 4 new tests among the passing set. I have not re-triggered the job, since re-running would not change the outcome.

[lucasnewman]

Thanks!

[lucasnewman]

@Anai-Guo Need signed commits

[Anai-Guo]

Also re-signed this one while I was at it (single GPG-signed commit, verified: true, tree unchanged). The reopen notification is just the force-push side effect.