Fix OpenVINO Matcher Model on XPU and Remove SAM-HQ-Tiny as default

[rajeshgangireddy]

Pull Request

Description

Matcher's openvino model wasn't working as expected on XPU. There were a few small issues and a major one : SAM-HQ-TINY which was our default decoder doesn't work as expected as an OpenVINO model.
The model doesn't produce useful masks and the output is non-deterministic as well - same images gives different masks with non-negligible differences. On CPU, the model works deterministically but without great accuracy.

In this PR :

• Remove SAM-HQ-Tiny as the default model. Now the default is SAM-HQ-Base
• Expose the Large variant of SAM-HQ
• Make fixes for OpenVINO model for Matcher
• Add a script to evalutate performance of different combinations of encoders and decoders here: library/tools/benchmark_variants.py
• Add a notebook with OpenVINO Matcher Model

TODO

• Remove library/tools/benchmark_variants.py if it's not useful
• Make sure application is working fine
• Clean script files which are not needed.

Type of Change

• ✨ feat - New feature
• 🐞 fix - Bug fix
• 📚 docs - Documentation
• ♻️ refactor - Code refactoring
• 🧪 test - Tests
• 🔧 chore - Maintenance

Related Issues

Breaking Changes

---

Examples

Screenshots

[Copilot]

Pull request overview

This pull request updates Instant Learn’s SAM-HQ model defaults and improves the Matcher OpenVINO export/runtime path to address XPU/GPU issues, while updating API/UI defaults, tests, docs, and adding developer benchmarking artifacts.

Changes:

• Switch default SAM-HQ decoder from SAM-HQ-tiny to SAM-HQ-base, add SAM-HQ-large, and deprecate SAM-HQ-tiny.
• Rework Matcher export/inference graph behavior for ONNX/OpenVINO (static shapes, output naming fixes, export-specific prompt/mask paths, XPU runtime fixes).
• Update app/backend + UI defaults/tests/docs and add benchmarking scripts + an OpenVINO export notebook.

Reviewed changes

Copilot reviewed 32 out of 32 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
library/tools/benchmark_variants.py	Adds a dev benchmark script for encoder/decoder combinations across PyTorch/OpenVINO.
library/tools/benchmark_matcher_variants.py	Adds an additional benchmark script (currently overlaps with the above).
library/tools/init.py	Introduces a tools package marker/docstring for dev utilities.
library/tests/unit/scripts/test_benchmark.py	Updates unit tests to use SAM-HQ-base as the benchmark backbone default.
library/tests/integration/models/test_model_integration.py	Updates integration tests to use SAM-HQ-base instead of SAM-HQ-tiny and minor formatting fixes.
library/tests/integration/models/test_matcher_export.py	Updates export/inference integration tests for OpenVINO static shapes and new default SAM.
library/src/instantlearn/utils/constants.py	Adds SAM_HQ_BASE/SAM_HQ_LARGE and updates model map metadata; marks SAM_HQ_TINY deprecated.
library/src/instantlearn/utils/args.py	Changes benchmark CLI default SAM from SAM-HQ-tiny to SAM-HQ-base.
library/src/instantlearn/models/soft_matcher/soft_matcher.py	Changes default SAM for SoftMatcher to SAM-HQ-base.
library/src/instantlearn/models/per_dino/per_dino.py	Changes default SAM for PerDino to SAM-HQ-base.
library/src/instantlearn/models/matcher/prompt_generators.py	Adds an export-optimized prompt selection path to reduce ONNX/OpenVINO graph complexity.
library/src/instantlearn/models/matcher/matcher.py	Adjusts Matcher defaults and improves ONNX/OpenVINO export reliability (static reshape, output renaming, FP16 compression option).
library/src/instantlearn/models/grounded_sam/grounded_sam.py	Changes default SAM for GroundedSAM to SAM-HQ-base.
library/src/instantlearn/components/sam/predictor.py	Adds SAM-HQ-base/large support, deprecation warning for tiny, XPU-safe mask decoder patching, and improved device/dtype syncing.
library/src/instantlearn/components/sam/decoder.py	Adds export-safe prompt/mask paths for fixed-shape ONNX/OpenVINO and adjusts mask tensor handling.
library/pyproject.toml	Adds ruff per-file ignores for tools/notebooks to support dev workflows.
library/examples/matcher_openvino_export.ipynb	Adds a notebook demonstrating Matcher → OpenVINO export and CPU inference/visualization.
library/docs/02-quick-start.md	Updates quick start example to use SAM-HQ-base.
library/docs/01-introduction.md	Updates supported SAM model list to base/large/HQ (removes tiny).
library/README.md	Updates README examples/model tables/memory notes to base/large/HQ and changes defaults.
application/ui/src/test-utils/mocks/mock-model.ts	Updates mocked model config default sam_model to SAM-HQ-base.
application/ui/src/setup-test.ts	Updates UI test fixtures default sam_model to SAM-HQ-base.
application/ui/src/features/prompts/models/model-toolbar/model-configuration/model-configuration-dialog.test.tsx	Updates UI test expectations for decoder selection (Tiny → Base).
application/ui/src/features/prompts/models/model-toolbar/model-configuration/model-configuration-dialog.component.tsx	Updates decoder dropdown options (adds Base/Large, removes Tiny).
application/ui/src/features/prompts/models/api/use-get-models.ts	Updates default UI model configurations to SAM-HQ-base.
application/backend/tests/unit/runtime/core/components/models/test_openvino_model.py	Updates backend unit tests for OpenVINO handler output-port based inference and mask typing expectations.
application/backend/tests/unit/runtime/core/components/factories/test_model.py	Updates backend factory tests to use SAM-HQ-base.
application/backend/tests/unit/api/endpoints/test_models.py	Updates API endpoint tests default sam_model payloads to SAM-HQ-base.
application/backend/app/runtime/core/components/models/openvino_model.py	Improves OpenVINO runtime handler (static reshape, infer request usage, output-port mapping, input resizing).
application/backend/app/domain/services/schemas/processor.py	Updates allowed SAM models and defaults to Base; adds Large to allowed set.
application/backend/app/api/endpoints/models.py	Updates OpenAPI examples/defaults from SAM-HQ-tiny to SAM-HQ-base.
README.md	Updates top-level supported SAM model list to base/large/HQ (removes tiny).