args pass

[resolve #2] Review comment by @Copilot (PR #890):
"open(..., \"w\") applies platform-dependent newline translation (e.g. Windows will write \r\n)..."

---
Co-authored-by: Claude Code <noreply@anthropic.com>

[resolve roboflow#2] Review comment by Copilot (PR roboflow#664):
"from_checkpoint is not currently exposed as a public import..."

Adds module-level from_checkpoint(path, **kwargs) wrapper in __init__.py
that delegates to RFDETR.from_checkpoint(). Also adds it to __all__ so
the documented usage `import rfdetr; rfdetr.from_checkpoint(...)` now works.

---
Co-authored-by: Claude Code <noreply@anthropic.com>

Note that the DeprecatedConfig RuntimeError guard is only reachable via
the CLI/PTL path; RFDETR.train() pre-populates model_name before saving,
so the config type-name branch is never reached from that path.

[resolve roboflow#2] /review finding by sw-engineer (report: _outputs/2026/04/output-review-feat-887-checkpoint-model-name-2026-04-01.md):
"[MEDIUM] _resolve_model_name behavior diverges between train() and CLI/PTL paths — src/rfdetr/training/callbacks/best_model.py:145"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

dtype now declared as torch.dtype | str to reflect the string coercion
added in this PR.

[resolve #2] /review finding by sw-engineer (report: _outputs/2026/04/output-review-copilot-fix-cuda-context-leak-2026-04-01.md):
"Missing type annotations on optimize_for_inference signature — dtype now accepts str"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

[resolve #2] /review finding by sw-engineer (report: _outputs/2026/04/output-review-copilot-fix-wrong-number-of-classes-in-onnx-export-2026-04-01.md):
"Guard linear.bias.data access with None check in _resize_linear — accessing .data unconditionally raises AttributeError when bias=False"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

[resolve #2] /review finding by qa-specialist (report: _outputs/2026/04/output-review-fix-stable-2026-04-02.md):
"two tests assert weights.bin.tmp which no longer exists after mkstemp switch — assertions are now no-ops"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

- Add explicit type guard on `counts` in `convert_coco_poly_to_mask`: dicts
  whose `counts` field is not str/bytes/list now raise a descriptive ValueError
  instead of an opaque pycocotools error downstream
- Clarify `_is_rle` docstring: structural-only check, counts=None passes check
  but fails downstream
- Update `convert_coco_poly_to_mask` docstring: note str/bytes/list variants
  for counts and add empty-dict validity note to Args section

[resolve #1,#2,roboflow#7,roboflow#8] /review findings by sw-engineer + doc-scribe
(report: _outputs/2026/04/output-review-feat-416-native-rle-annotation-support-2026-04-02.md)

---
Co-authored-by: Claude Code <noreply@anthropic.com>

…ation

The two YOLO builder functions shared ~80% of their logic. Extract a shared
_build_yolo_samples(img_folder, lb_folder, data_file, *, include_polygons)
helper that handles image enumeration, dimension reading, and label parsing.
Each public builder becomes a one-liner delegation. Reduces the surface area
for future divergence when shared logic (error handling, caching, new fields)
needs to change.

[resolve #2] /review finding by sw-engineer (report: _outputs/2026/04/output-review-copilot-fix-multi-gpu-training-issue-2026-04-02.md):
"Consider extracting shared builder helper to reduce duplication between the two builder functions"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

[resolve #2-3] Review comment by @copilot-pull-request-reviewer (PR #918):
"tests assume get_version() resolves to non-None; patch to a fixed string for determinism"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

- Add TestMSDeformAttnModule: 4 tests for MSDeformAttn.forward covering backward compat (no hw param), hw param path, export mode with hw param, and _export flag guard
- Add test_ms_deform_attn_core_pytorch_export_compatible: end-to-end torch.export.export regression test validating the FakeTensor fix
- Fix gen_encoder_output_proposals backward compat: memory_padding_mask and spatial_shapes now default to None, restoring callers that omit memory_padding_mask

[resolve #5] /review finding by qa-specialist (report: .temp/output-review-copilot-fix-ptl-gradient-accumulation-bug-2026-04-09-2.md):
"[MEDIUM] Missing module-level test for MSDeformAttn.forward"
[resolve #2] @Copilot (gh): "Add torch.export.export smoke test on MSDeformAttn"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

- Add `_has_cuda_device()` and `resolve_augmentation_backend()` to `kornia_transforms.py` as single canonical implementations
- `_has_cuda_device` uses `rfdetr.config.DEVICE` (fork-safe) instead of `torch.cuda.is_available()` — mirrors module_data.py
- `_resolve_runtime_augmentation_backend` in coco.py now delegates to the shared resolver; yolo.py import unchanged
- `build_o365_raw` inline resolution replaced with single `resolve_augmentation_backend()` call, removing duplicate logic and two direct `torch.cuda.is_available()` calls

[resolve #1] /review finding by sw-engineer (report: .temp/output-review-aug-kornia-2026-04-09.md): "Duplicated backend-resolution logic across three modules"
[resolve #2] /review finding by linting-expert (report: .temp/output-review-aug-kornia-2026-04-09.md): "Inconsistent CUDA detection: fork-safe vs direct"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

- Add save_dataset_grids=False to the SimpleNamespace at test_auto_batch.py:171 so the new config.save_dataset_grids attribute access at detr.py:603 doesn't raise AttributeError in CI

[resolve #2] /review finding by sw-engineer (report: .temp/output-review-copilot-fix-save-dataset-grids-kwarg-2026-04-09.md):
"Add save_dataset_grids=False to the SimpleNamespace in test_auto_batch.py:171 — fixes CI AttributeError"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

[resolve #2] /review finding by sw-engineer (report: .temp/output-review-copilot-fix-checkpoint-file-prediction-2026-04-09.md):
"Add logger.debug() when PTL normalization fires (consistent with other branches logging their actions)"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

…cated

- Add RFDETRSegPreviewConfig (res=432, ps=12, PE=36, formula-derived) to the
  positive parametrize list as seg_preview_360 case
- Expand the single-config negative test into a parametrized test covering both
  RFDETRBaseConfig and RFDETRLargeDeprecatedConfig (both have PE=37 ≠ 560÷14=40)

[resolve #2] Review comment by @copilot-pull-request-reviewer (PR #956):
"RFDETRSegPreviewConfig not included in parameterized coverage"

[resolve #3] /review finding by qa-specialist (report: .temp/output-review-copilot-...):
"Missing explicit negative case for RFDETRLargeDeprecatedConfig"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

- Change `bs, c, h, w = src.shape` to `_, c, h, w = src.shape` — `bs` is no longer used after .repeat(bs, ...) was replaced with .expand(memory.shape[0], ...) for ONNX symbolic batching

[resolve roboflow#2] Review comment by @Copilot (PR roboflow#950):
"After replacing .repeat(bs, ...) with .expand(memory.shape[0], ...), bs is no longer used..."

---
Co-authored-by: Claude Code <noreply@anthropic.com>

- Add missing `]` closing the tflite extra array — parse error broke all 20 CI checks
- Add `onnx>=1.16.0,<1.20` to tflite extra — converter imports onnx directly; users installing rfdetr[tflite] alone would hit ImportError at runtime
- Add `tf-keras>=2.16.0` to tflite extra — onnx2tf requires tf-keras at runtime (noted in PR description but omitted from diff)

[resolve roboflow#5] /review finding by sw-engineer (report: .temp/output-review-feat-tflite-export-2026-04-10.md): pyproject.toml missing ] on tflite extra
[resolve roboflow#8] /review finding by sw-engineer (report: .temp/output-review-feat-tflite-export-2026-04-10.md): tflite extra must declare onnx dependency
[resolve roboflow#2] @copilot (gh): tf-keras discrepancy between PR description and diff

---
Co-authored-by: Claude Code <noreply@anthropic.com>

The existing shape assertion would pass even if the loader unnecessarily
re-interpolated the tensor (producing the same shape but different values).
Add `torch.equal` to verify the tensor is truly left untouched.

[resolve #2] Review comment by @Copilot (PR #964):
"Consider asserting value equality (e.g., `torch.equal`/`allclose`) and/or
that the original tensor object is preserved."

---
Co-authored-by: Claude Code <noreply@anthropic.com>

[resolve #1] Review comment by @Copilot (PR #970):
"test_background_class_id_does_not_emit_oob_warning can become order-dependent
because warning_once suppresses repeated warnings globally — clear _warned_once
before predict() so the assertion reliably fails if a background warning is reintroduced"

[resolve #2] Review comment by @Copilot (PR #970):
"test_truly_oob_class_id_still_maps_to_empty_string should also assert the
out-of-range warning IS emitted for truly OOB class IDs"

- Switch both warning tests from caplog to _warned_once inspection; the RF-DETR
  logger sets propagate=False so caplog never receives records via root handler
- Clear _warned_once before each predict() call to guarantee order-independence
- Rename test to test_truly_oob_class_id_still_maps_to_empty_string_and_warns
  to reflect the added warning assertion

---
Co-authored-by: Claude Code <noreply@anthropic.com>

…oco_transforms

- Remove claim that PadIfNeeded pads bottom/right edges only; position is
  determined by Albumentations default (center), not pinned to top_left
- Clarify that the 1333 px long-side cap in make_coco_transforms is applied
  before divisor padding; final dims may exceed 1333 by up to block_size-1 px

[resolve roboflow#2] Review comment by @Copilot (PR roboflow#991):
"The docstring says padding is applied on the bottom/right edges, but the PadIfNeeded config does not pin..."

[resolve roboflow#3] Review comment by @Copilot (PR roboflow#991):
"make_coco_transforms' docstring states the longest side is capped at 1333px for val/test, but..."

---
Co-authored-by: Claude Code <noreply@anthropic.com>

[resolve roboflow#2] Review comment by @Copilot (PR roboflow#992):
"docstrings here say list_of_targets but batch = list(zip(*batch)) makes batch[1] a tuple"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

Replace hardcoded 16 and 384 in TestL1FacadePEInterpolationEndToEnd with
values derived from the model:
- patch_size = default_model.model_config.patch_size
- pe_dim = default_pe.shape[-1]

These feed into the SimpleNamespace checkpoint args and the expected_pe_grid
calculation, making the test robust to future architecture changes.

[resolve roboflow#2] Review comment by @Copilot (also flagged by /review):
"This test hard-codes several architecture-specific constants (patch_size == 16
via new_resolution // 16, and embedding dim == 384)..."

---
Co-authored-by: Claude Code <noreply@anthropic.com>

…ons unit tests

- _pytest_params: locally skip missing checkpoints; in CI (CI=true) run the
  test so artifact download failures cause a visible assertion error instead
  of a silent skip
- _read_versions: accept optional path parameter so callers can inject a
  file directly — removes need for module-level monkeypatching in tests
- TestReadVersions: five parametrized cases covering blank lines, comment
  lines, surrounding whitespace, empty file, and non-existent file

[resolve #2] @Copilot inline (gh): "missing checkpoints silently skipped in CI masks regressions" (test_checkpoint_compat.py:81)
[resolve #8] /review finding by foundry:qa-specialist (report: .temp/output-review-ci-legacy-2026-04-24.md):
"No test for _read_versions() with malformed input"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

- TestGetModelCacheDir (4 tests): default path, custom RF_HOME, tilde
  expansion, return type
- TestMaybeDownloadPretrainWeightsCacheDir (4 tests): bare filename
  resolved via RF_HOME, explicit path used as-is, None skips download,
  cache dir auto-created

[resolve roboflow#2] Review comment by @SkalskiP (PR roboflow#130):
"add automated tests for download utility"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

Model tests live under tests/models/test_model.py, not tests/test_model.py
The wrong path would fail for contributors running a specific test file.

[resolve #2] Review comment by @Copilot (PR #996):
"tests/test_model.py but this repo's model tests live under tests/models/"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

Renames _interpolate_position_embeddings → interpolate_position_embeddings
in rfdetr.models.weights, adds it to __all__, and updates the cross-module
import in training.module_model — eliminating coupling to a private symbol.

[resolve #2] @Copilot (gh): "Consider promoting _interpolate_position_embeddings
to a public utility to avoid coupling training code to a private symbol"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

[resolve #2] Review comment by @Copilot (PR #1011):
"_run_inference() always produces a float32 input tensor, but TFLite mo..."

---
Co-authored-by: Claude Code <noreply@anthropic.com>

…lback

Adds test_three_outputs_with_rank4_masks_resolves_correctly to
TestShapeBasedOutputFallback: verifies that boxes [1,Q,4] / logits [1,Q,82]
/ masks [1,Q,28,28] with generic Identity names resolves without error after
the rank-3 candidate filter fix.

[resolve #2] Review comment by @Copilot (PR #1015):
"The shape-based fallback tests don't cover the realistic segmentation-export
scenario where there are 3 outputs and one is a rank-4 masks tensor…"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

…head

Previously the field was in _PRETRAIN_BREAKING_FIELDS unconditionally,
causing detector-only configs to warn about a field that has no effect
on non-segmentation checkpoints.

Removed from the shared tuple; added explicit conditional check in the
validator guarded by self.segmentation_head being True.

- [resolve #2] Review comment by @Copilot (gh)

---
Co-authored-by: Claude Code <noreply@anthropic.com>

… path

[resolve #1] @Copilot (gh): grammar fix in _labelled_query_tensor docstring
[resolve #2] @Copilot (gh): e2e tests now assert query_feat.weight ordering in all three scenarios
[resolve #5] /review finding by foundry:qa-specialist: add Namespace-style args e2e test covering _ckpt_args_get getattr branch
[resolve #6] /review finding by foundry:qa-specialist: add test for args-present-but-keys-absent legacy fallback path

---
Co-authored-by: Claude Code <noreply@anthropic.com>
Co-authored-by: OpenAI Codex <codex@openai.com>

- Add response.ok check + try/catch with avatar fallback to GitHub API
  fetch in renderCard(); prevents broken markup when API rate-limits
- Add rel="noopener noreferrer" to both target="_blank" author links

[resolve #1] Review comment by @Copilot (PR #1022):
"fetch('https://api.github.com/users/...') is done client-side for each..."

[resolve #2] Review comment by @Copilot (PR #1022):
"Links opened with target=\"_blank\" should also set rel=\"noopener noreferrer\"..."

---
Co-authored-by: Claude Code <noreply@anthropic.com>

- Replace absolute /assets/roboflow-logo.svg with favicon href so logo
  renders correctly under mike versioned docs /<version>/assets/
- Scope .author-name event listener loop to current card element so
  hover listeners don't accumulate across all cards on each renderCard
- Add authorCache object; same GitHub login fetched only once per page

[resolve #1] Review comment by @Copilot (PR #1022):
"src=\"/assets/roboflow-logo.svg\" is an absolute-root path..."

[resolve #2] Review comment by @Copilot (PR #1022):
"Event listeners are attached inside renderCard() via document.querySe..."

[resolve #4] Review comment by @Copilot (PR #1022):
"Each card triggers one unauthenticated GitHub API request per author..."

---
Co-authored-by: Claude Code <noreply@anthropic.com>

[resolve #2] Review comment by @Copilot (PR #1026):
"The comment says this mimics the official pretrain weights filename (rf-detr-seg-nano.pt), but the test saves rf-detr-seg-nano.pth."

---
Co-authored-by: Claude Code <noreply@anthropic.com>

Reinstate the DINOv2 positional-embedding bicubic-resize block in
`RFDETRModelModule.on_load_checkpoint`.  PTL applies
`checkpoint["state_dict"]` immediately after this hook, so when a
`.ckpt` saved at resolution R is resumed at R' ≠ R the PE tensor must
already be reshaped — otherwise `Trainer(ckpt_path=...)` raises
`RuntimeError: size mismatch for backbone.0.encoder.encoder.embeddings.position_embeddings`.

`_load_pretrain_weights` (now `load_pretrain_weights` from canonical) and
`on_load_checkpoint` cover non-overlapping paths: the former fires for
fresh-train bootstrap, the latter for PTL resume.  Both must perform PE
interpolation; removing it from the resume hook in PR #1040 re-opened
regression #998.

Docstring updated to:
- list the PE-interpolation step alongside the existing legacy-format and
  EMA-stash steps;
- cross-reference the divergent `.ckpt` normalisation now handled by
  `load_pretrain_weights` for the fresh-train path (review M4).

[resolve #2] @Copilot (gh): `on_load_checkpoint()` no longer interpolates positional embeddings in `checkpoint["state_dict"]` — breaks regression coverage for issue #998
[resolve #13] /review finding by doc-scribe (report: .reports/review/2026-05-15T07-09-07Z/review-report.md): cross-reference `.ckpt` normalisation in `on_load_checkpoint` docstring

---
Co-authored-by: Claude Code <noreply@anthropic.com>

… type

- Add `if mask_logits.ndim != 3: raise ValueError(...)` guard at top of
  `_decode_masks` so mismatched tensors from the rank-4 heuristic fail
  loudly rather than producing silent shape corruption downstream
- Tighten return type annotation from `NDArray[Any]` to `NDArray[np.bool_]`
  and `dtype=bool` → `dtype=np.bool_` for consistency
- Soften docstring claim from "Mirrors PostProcess.forward" to "Approximates"
  and document that Pillow bilinear and PyTorch F.interpolate may differ at
  border pixels due to distinct half-pixel conventions

Resolve /review finding roboflow#2 (foundry:sw-engineer HIGH) + roboflow#7 (foundry:doc-scribe MEDIUM)
from .reports/review/2026-05-19T10-06-08Z/review-report.md

---
Co-authored-by: Claude Code <noreply@anthropic.com>

Accidental commit of Gemini chat transcript with jupytext CLI snippets.
Not release content — no user-facing value.

[resolve #2] @Copilot (gh): discussion_r3283481989

---
Co-authored-by: Claude Code <noreply@anthropic.com>

- [resolve roboflow#2] /review finding by foundry:qa-specialist (report: .reports/review/2026-05-26T19-40-29Z/review-report.md): _create_interpreter prefers ai_edge_litert first but TestCreateInterpreter had zero coverage for that path; add _mock_ai_edge_litert fixture and three tests mirroring the tflite_runtime suite

---
Co-authored-by: Claude Code <noreply@anthropic.com>

The hook exists to reset _ema_has_updates before test, preventing a stale
True value left by a preceding validation epoch from triggering an
empty-state EMA compute pass that logs sentinel -1 values. Document this
invariant explicitly so future refactors don't remove the hook as apparent
dead construction.

[resolve roboflow#2] /review finding by foundry:challenger (report: .reports/review/2026-06-02T05-43-56Z/review-report.md)

---
Co-authored-by: Claude Code <noreply@anthropic.com>

…s count

Add three regression tests to TestFromCheckpointNumClassesProvenance and
a supporting fixture (args_only_checkpoint):

1. test_checkpoint_num_classes_from_args_not_marked_user_set — verifies
   that num_classes injected via the legacy checkpoint args dict (not via
   model_config) is also cleared from model_fields_set by from_checkpoint.
   Addresses the gap where the two_class_checkpoint fixture only exercised
   the model_config injection path.

2. test_explicit_default_num_classes_does_not_block_alignment — documents
   that passing num_classes=<ModelConfig default> explicitly does not pin
   the head; _align_num_classes_from_dataset treats user_overrode=False
   when value == default, so adaptation still proceeds.  Guards against
   accidental removal of the value!=default guard.

3. test_equal_class_count_does_not_rebuild_head — confirms the fix has
   no side-effect when dataset class count matches checkpoint class count;
   head shape must remain unchanged.

[resolve roboflow#2] foundry:qa-specialist (report: .reports/review/2026-06-13T18-05-46Z/review-report.md)
[resolve roboflow#4] foundry:qa-specialist (report: .reports/review/2026-06-13T18-05-46Z/review-report.md)
[resolve roboflow#5] foundry:qa-specialist (report: .reports/review/2026-06-13T18-05-46Z/review-report.md)

---
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

Adding eps unconditionally breaks the identity IoU(A,A)=1.0 for
non-degenerate boxes (inter==union but union+eps > union). Use
clamp(min=eps) so normal boxes are exact and only zero-area boxes
avoid division-by-zero.

- box_iou: union.clamp(min=eps) instead of union+eps
- generalized_box_iou: area.clamp(min=eps) instead of area+eps

[resolve #2 #3] @Copilot (gh) box_ops.py:64,85

---
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

- fix(roboflow#1): label dir mis-resolves for Ultralytics COCO images/<split> intermediate
  layout; replace name-equality check with segment-replace approach (replace
  "images" component anywhere in path with "labels", mirroring Ultralytics
  img2label_paths convention)
- fix(roboflow#2): narrow exception swallow to OSError (silent); ValueError/TypeError and
  yaml.YAMLError now emit logger.warning with data_file + split context before
  falling back; import yaml + get_logger
- fix(roboflow#3): document intentional divergence between is_valid_yolo_dataset (coarse
  filesystem pre-check) and _resolve_yolo_split_dirs (yaml-first) in docstring
- fix(roboflow#8): remove unused REQUIRED_SPLIT_DIRS constant — superseded by
  _VALID_VAL_DIR_NAMES after PR's validation logic change
- fix(roboflow#11): remove dead no-op ternary yaml_split_key = split if split != "val"
  else "val" — always evaluated to split; inline split directly in data.get(split)

---
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

Resolve /review findings on PR roboflow#1237 (report 2026-07-27T11-48-47Z):
- notebook H1 -> "ExecuTorch Export & Inference", matching the
  cards.yaml / NOTES.md rename (finding roboflow#2)
- notebook warning + section 6: state the contiguity fix lives in
  infer_transforms (_ensure_contiguous) and label the notebook's own
  .contiguous() call defensive / no-op, not load-bearing (finding roboflow#1)
- notebook + export.md: replace "confident-looking but wrong
  predictions ... below threshold" with "plausible-shaped output, but
  every detection's score collapses below threshold" (finding roboflow#6)

export.md's manual-numpy example keeps its np.ascontiguousarray note
unchanged — that path genuinely permutes without copying.

---
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

Add _assert_equal_length guard to elementwise_box_iou and
elementwise_generalized_box_iou so a length-1 operand raises ValueError
instead of silently broadcasting torch.max/min to a wrong "matched"
result. Document that elementwise numerics match the box_iou diagonal
with no added FP16/FP32 dtype dependence.

[resolve #2] @Copilot (gh) + /review MED (doc-scribe/challenger/codex consensus)
[resolve #1] @Borda (gh): dtype question answered via docstring note

---
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

Adds keypoint-positive forwarding coverage for build_coco, the
mirror-direction empty-list case for both builders, the explicit-None
coercion case, square-resize parametrization, and consolidates/renames
the sentinel-forwarding test class. Removes a redundant local re-import.

[resolve #2,#3,#6,#9,#10,#11] PR #1248 — qa-specialist findings

- #1: assert F.grid_sample not called on the manual gather path (as-suggested)
- #2: assert explicit device= is honored for all_gather buffer allocation (self-resolved — avoided mocking torch.tensor directly per challenge)

Challenge log: 1 as-suggested, 1 self-resolved, 0 rejected

---
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

- README methodology note: qualify that Params are deployment (fused) nn.Module
  parameter counts (model.parameters()), not raw checkpoint tensor counts, and
  that † rows are author-reported (items #1 @adhavan18, #2 Copilot)
- README RF100-VL table: SAM 3 † License "SAM" -> "N/A" (author-reported row,
  not a packaged checkpoint) (item #3 Copilot)
- benchmarks.md Params (M) definition: reflow to a single sentence, add the
  nn.Module-vs-checkpoint-tensor clarification and the † author-reported note
  (items #1, #4 Copilot)

Challenge log: 4 as-suggested, 0 self-resolved, 0 rejected

---
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

_as_odd_kernel() gains an optional default_pair: when the incoming
pair matches Albumentations' own Blur default (3, 7), the "pair
collapses" notice demotes to logger.debug (expected CPU/GPU
divergence, not actionable); any other user-supplied pair still
warns. The warning also logged the pre-rounding max(value) rather
than the actual resolved kernel — (3, 6) logged "kernel_size=6" but
returned 7 — now logs the resolved value. Adds the same length
validation _as_range already has: a 1-element sequence is treated
as degenerate, 0- or >2-element raises a clear ValueError instead
of a bare max() crash.

[resolve roboflow#2, roboflow#3] PR roboflow#1277 — @Copilot (gh) + foundry:sw-engineer (report: .reports/review/2026-08-04T17-36-54Z/review-report.md)

---
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

Resolves review findings from /oss:review PR roboflow#1279 (report: .reports/review/2026-08-05T09-41-01Z/review-report.md), items roboflow#2,3,4,5,6,7,8,9,10,14,15,18. All findings cross-validated by foundry:challenger/sw-engineer/qa-specialist before implementation.

- Reset torch.cuda.reset_peak_memory_stats at fit start (gated by _is_cuda), so max_mem reflects this run's footprint rather than a process-lifetime / batch_size="auto"-probe peak (roboflow#2)
- Warn via rank_zero_warn + setdefault instead of silently overwriting a same-named user-logged metric (roboflow#3)
- Reword class docstrings from "GPU" to "CUDA" for accuracy — implementation is CUDA-only (roboflow#4)
- Add ProgressBar base to _GpuMemoryMetricsMixin, dropping the now-unresolvable type:ignore[misc] (roboflow#5, subsumes roboflow#17)
- Test peak-reset semantics across sequential get_metrics() calls (roboflow#6)
- Add missing Rich progress bar coverage for the cuda-present-but-not-initialized branch (roboflow#7)
- Test _is_cuda with None/non-torch.device input (roboflow#8)
- Add rounding-boundary test cases (0 bytes, 0.5MB boundary) (roboflow#9)
- Fold xla/tpu device-type case into the existing non-cuda parametrize (roboflow#10)
- Fix module docstring overstating on_test_end's gating mechanism (roboflow#14)
- Document max_mem as rank-local under DDP (roboflow#15)
- Match legacy "123 MB" spacing (space before unit) for restoration parity (roboflow#18)

34/34 tests pass (tests/training/callbacks/test_gpu_memory_progress_bar_callback.py + tests/training/test_trainer.py).

---
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

- Reset torch.cuda.reset_peak_memory_stats at fit start (gated by _is_cuda), so max_mem reflects this run's footprint rather than a process-lifetime / batch_size="auto"-probe peak (#2)
- Warn via rank_zero_warn + setdefault instead of silently overwriting a same-named user-logged metric (#3)
- Reword class docstrings from "GPU" to "CUDA" for accuracy — implementation is CUDA-only (#4)
- Add ProgressBar base to _GpuMemoryMetricsMixin, dropping the now-unresolvable type:ignore[misc] (#5, subsumes #17)
- Test peak-reset semantics across sequential get_metrics() calls (#6)
- Add missing Rich progress bar coverage for the cuda-present-but-not-initialized branch (#7)
- Test _is_cuda with None/non-torch.device input (#8)
- Add rounding-boundary test cases (0 bytes, 0.5MB boundary) (#9)
- Fold xla/tpu device-type case into the existing non-cuda parametrize (#10)
- Fix module docstring overstating on_test_end's gating mechanism (#14)
- Document max_mem as rank-local under DDP (#15)
- Match legacy "123 MB" spacing (space before unit) for restoration parity (#18)

---------

Co-authored-by: Jesús Royeth <JESUSROYETH@users.noreply.github.com>
Co-authored-by: Jirka Borovec <6035284+Borda@users.noreply.github.com>
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>