Review Legacy Models Registries

Reviewing legacy-models-registries changes

Scope

Trigger this skill when a PR changes any of:

• inference/models/** — concrete legacy model implementations (rfdetr, owlv2, sam/sam2/sam3, clip, gaze, yolo*, easy_ocr, doctr, florence2, paligemma, qwen*, smolvlm, etc.), inference/models/utils.py (ROBOFLOW_MODEL_TYPES + get_roboflow_model), and inference/models/aliases.py.
• inference/core/models/** — base contracts (base.py, roboflow.py), task bases (object_detection_base.py, instance_segmentation_base.py, keypoints_detection_base.py, classification_base.py, semantic_segmentation_base.py), stubs.py, and inference_models_adapters.py (the USE_INFERENCE_MODELS bridge adapters).
• inference/core/registries/** — base.py (ModelRegistry) and roboflow.py (get_model_type, auth checks, metadata cache).

OUT of scope (defer to other skills): Workflows blocks, HTTP API handlers (inference/core/interfaces/**), model managers (inference/core/managers/**), the standalone inference_models package itself, and the tensor-data-representation _tensor sibling effort.

Review checklist

Severity-tagged. Each item points at its canonical rule in Standards below.

• BLOCK — new model missing any leg of the wiring triple: env flag (env.py) + guarded ROBOFLOW_MODEL_TYPES entry (utils.py) + registry recognition (GENERIC_MODELS or API task type). (S3)
• BLOCK — a served (task, variant) has no ROBOFLOW_MODEL_TYPES key → ModelNotRecognisedError. (S3)
• BLOCK — a raw model_id reaches chunking / auth / weight fetch without resolve_roboflow_model_alias first. (S2)
• BLOCK — get_model_type can return None; new branch drops the MODEL_TYPE_DEFAULTS fallback or _ensure_model_supported_on_this_deployment gating. (S4)
• BLOCK — new ort_session.run / run_session_via_iobinding / shared per-image cache mutation without an instance Lock. (S5)
• BLOCK — class-name / environment.json parsing indexes by sorted(keys) instead of numeric str(i). (S6)
• BLOCK — optional/deprecated model import not wrapped in try/except + ModelDependencyMissing warning (breaks slim-image utils.py load). (S3)
• FLAG — response-schema field added as Optional instead of with a default (breaks model_dump(exclude_none=True) callers). (S7)
• FLAG — new metadata/weight fetch path drops countinference / service_secret. (S8)
• FLAG — USE_INFERENCE_MODELS adapter parity missing, or backend= ignores DISABLED_INFERENCE_MODELS_BACKENDS / VALID_INFERENCE_MODELS_BACKENDS. (S9)
• FLAG — postprocess / NMS index math or empty-prediction handling changed without a covering test. (S10)
• FLAG — torch.compile / monkey-patch / background-compilation / download-streaming change without explicit justification + tests (repeatedly reverted). (S11)
• FLAG — model overriding infer_from_request (SAM2/SAM3) drops the @usage_collector("model") decorator. (S12)
• NIT — per-inference logging above debug, or f-strings instead of lazy %-args in hot paths. (S13)
• NIT — HF/transformer artifact download not guarded by FileLock. (S13)

Not blocking

• Do NOT demand an inference/core/version.py bump — inference releases are versioned separately from feature/bugfix PRs.
• Do NOT demand the full wiring triple when the PR only edits an existing model's internals (no new (task, variant)).
• Do NOT demand a new adapter when the touched model has no inference_models backend.
• Do NOT block on Optional-vs-default when the field is genuinely nullable in the response contract and callers do not use exclude_none.
• Do NOT flag missing locks on stateless, read-only compute that never touches a shared ORT session or cache.

Standards

The base contract. Every model is a Model (inference/core/models/base.py). BaseInference.infer() runs preprocess → predict → postprocess under telemetry spans and is wrapped by @usage_collector("model"). make_response(...) is a separate abstract method that builds the InferenceResponse. Model.infer_from_request(request) orchestrates: calls infer(...), then stamps response.time, response.inference_id, and optional visualization. ONNX Roboflow models flow through inference/core/models/roboflow.py (artifact download → cache → environment.json parse → class names/colors/preproc).

Registry / id resolution. inference/core/registries/roboflow.py::get_model_type(model_id, api_key) returns (task_type, model_type) by: resolving aliases, GENERIC_MODELS (whole model_id then dataset_id), _get_local_model_type (local inference_models dirs), stub versions, then the Roboflow API. The tuple keys into ROBOFLOW_MODEL_TYPES (inference/models/utils.py) to pick a class.

• S2 — Alias resolution first. Call resolve_roboflow_model_alias(model_id) (inference/models/aliases.py) before get_model_id_chunks, before obtaining chunks/weights, and before auth checks (missing before chunk-fetch in _check_if_api_key_has_access_to_model broke aliased-model authorization, #1226; pattern established in both model and registry layers, #193).
• S3 — New model wiring is a triple. (a) env enablement flag in inference/core/env.py (e.g. QWEN_3_5_ENABLED, CORE_MODEL_*_ENABLED, #2105); (b) guarded import + ROBOFLOW_MODEL_TYPES entry in inference/models/utils.py; (c) registry recognition via GENERIC_MODELS or API task type. A registry key must exist for every served variant name (missing bare ("object-detection","rfdetr") / ("instance-segmentation","rfdetr") keys yielded ModelNotRecognisedError even though the inference-models backend served the variant, #2128). Optional/deprecated imports (Gaze) must be wrapped in try/except emitting a warnings.warn(category=ModelDependencyMissing), or a crash at import breaks the whole utils.py load in slim images (#2338).
• S4 — get_model_type never returns None. Unknowns fall back via MODEL_TYPE_DEFAULTS (inference/core/roboflow_api.py); _ensure_model_supported_on_this_deployment still gates. A silent None surfaces downstream as ModelArtefactError.
• S5 — Thread-safety for stateful/session models. ONNX sessions and cache-mutating models need an instance-level threading.Lock. self._session_lock (initialized in inference/core/models/roboflow.py) guards run_session_via_iobinding / ort_session.run in classification, clip, gaze, rfdetr, yolo* (#1510). SAM/SAM2 wrap the whole infer_from_request in self._state_lock (renamed from _ort_session_lock) because embedding/cache state is shared (#1549). Any new .run(...) or mutable per-image cache without a lock is a finding.
• S6 — environment.json / CLASS_MAP parsing must be order-stable. Class names come from CLASS_MAP indexed by integer string keys str(i) in numeric order, NOT sorted(keys()) — lexicographic sort put "10" before "2" and scrambled classes (#225). See get_class_names_from_environment_file in inference/core/models/roboflow.py.
• S7 — Response-schema additions ship a default, not Optional. Schemas in inference/core/entities/responses/inference.py are public API; callers use model_dump(exclude_none=True), so a new field needs a concrete default. Classification also keeps top/confidence defaulting to ""/0.0 when the filtered list is empty — watch for results[0] IndexError on empty predictions (#1036; earlier #873 dropped **kwargs on classification infer).
• S8 — countinference / service_secret threading. Weight/metadata fetches accept and forward countinference and service_secret down to get_roboflow_* / get_extra_weights_provider_headers (#1382, #1575). New fetch paths must plumb them through, not drop them.
• S9 — USE_INFERENCE_MODELS bridge parity. When adding a legacy model that also has an inference-models backend, add the matching InferenceModels*Adapter swap in the if USE_INFERENCE_MODELS: block of utils.py and register local-dir task types (#2479, #2105). Adapters must respect DISABLED_INFERENCE_MODELS_BACKENDS / VALID_INFERENCE_MODELS_BACKENDS (inference/core/env.py) when computing backend= (#2096).
• S10 — Postprocess / NMS index math is load-bearing. Padding-undo uses infer_shape[1] for x and [0] for y (x/y swap shifted every box, #495); keypoint slicing is pred[7:] with class id at pred[4 + len(class_names)] (#216); np.argpartition(-logits_flat, max_detections) must guard max_detections < len(logits_flat) or it raises (inference/models/rfdetr/rfdetr.py, #1920); empty per-image predictions append [] and continue rather than indexing an empty array (#387). Treat changes as high-scrutiny.
• S11 — Compile / download-mode changes are high-risk. OWLv2 torch.compile monkey-patching, background compilation, precomputed-embeddings, and streaming-download changes were repeatedly reverted (#2270/#2271/#1046/#976/#943/#1329/#1196/#1787). Also: ModelArtefactError-wrapping retry loops around cache reads hid corrupt-download handling — treat corruption at the root cache layer (#1404); don't gate on "torch" in dir() (#1054). Expect a revert plan and strong test evidence.
• S12 — @usage_collector("model") must survive an override. BaseInference.infer carries the decorator; models overriding infer_from_request (SAM2/SAM3) must re-apply @usage_collector("model") themselves (#1768).
• S13 — Logging is lazy and low-noise; downloads are race-safe. Use logger.debug for per-inference logs (OWLv2 moved INFO→DEBUG, #1793) and %-style lazy args, not f-strings, in hot paths (#1578). Guard HF/transformer artifact downloads with FileLock to avoid worker races (#1578).

Required companions

• inference/core/env.py — env flag for any new model or global toggle (*_ENABLED, DISABLED_INFERENCE_MODELS_BACKENDS).
• inference/models/utils.py — new (task, variant) → class mapping, guarded by the enablement flag and try/except, plus the USE_INFERENCE_MODELS adapter swap when a backend exists.
• GENERIC_MODELS in inference/core/registries/roboflow.py — for foundation/core models keyed by whole model_id or dataset_id.
• Tests. Registry/id changes → tests/inference/unit_tests/core/registries/test_roboflow.py; base/adapter changes → tests/inference/unit_tests/core/models/test_roboflow.py, test_inference_models_adapters.py, test_stubs.py; aliases → tests/inference/unit_tests/models/test_aliases.py; concrete models → tests/inference/unit_tests/models/test_* and tests/inference/models_predictions_tests/.
• Docs. New foundation model → docs/foundation/<model>.md; new fine-tuned family → docs/fine-tuned/<model>.md.

Key files & entry points

• inference/core/registries/roboflow.py — get_model_type, _get_local_model_type, _check_if_api_key_has_access_to_model, _ensure_model_supported_on_this_deployment, GENERIC_MODELS.
• inference/models/utils.py — ROBOFLOW_MODEL_TYPES, get_roboflow_model, USE_INFERENCE_MODELS swap block, per-model enablement wiring.
• inference/core/models/base.py — BaseInference/Model, infer, infer_from_request, make_response, @usage_collector.
• inference/core/models/roboflow.py — ONNX artifact download/cache, get_class_names_from_environment_file, _session_lock, FileLock.
• inference/core/models/inference_models_adapters.py — InferenceModels{ObjectDetection,InstanceSegmentation,KeyPointsDetection,Classification,SemanticSegmentation}Adapter.
• inference/models/aliases.py — resolve_roboflow_model_alias, REGISTERED_ALIASES.
• inference/core/models/{object_detection,instance_segmentation,keypoints_detection,classification,semantic_segmentation}_base.py, stubs.py.
• inference/core/env.py (flags), inference/core/roboflow_api.py (MODEL_TYPE_DEFAULTS).

Reference PRs

• #2105 — redirect versionless model auth/metadata to new registry via USE_INFERENCE_MODELS.
• #2479 — load locally-stored inference_models packages via registry (_get_local_model_type, adapter wiring, tests).
• #2128 — add missing ROBOFLOW_MODEL_TYPES keys so RFDETR/NAS works with inference-models backend.
• #2338 — try/except-guard Gaze import with ModelDependencyMissing warning.
• #1510 — per-instance ONNX _session_lock against GPU concurrency corruption.
• #1549 — wrap SAM/SAM2 handlers in _state_lock.
• #1036 — apply confidence to classification + backward-compatible response defaults.
• #225 — fix CLASS_MAP ordering (numeric str(i), not sorted).
• #193 — introduce model aliases + resolve at model and registry layers.
• #495 — fix x/y swap in padding-undo.
• #1920 — guard argpartition when max_detections >= len(logits_flat).
• #2270 / #1046 — OWLv2 torch.compile monkey-patch / background-compilation reverts.

Related topic skills

When the PR also exhibits these cross-cutting concerns, load the matching topic skill too (see each skill's description for the trigger):

• review-topic-prediction-integrity
• review-topic-concurrency-and-resource-safety
• review-topic-backward-compat-and-versioning
• review-topic-test-hygiene