[Ti-Tai Wang (titaiwangms)]

Description

Introduce op-owned contrib source declarations v0, using CUDA Unit Swish -> ScaledSiLU and CausalConvWithState as the first consumers.

This PR targets main; prerequisite #34021 has merged. MoE is separate, dependent #34086. Auto-merge remains disabled.

• Each EP's op implementation owns its executable source pattern, qualification guard, replacement, initializer transformation, and numerical scope. The catalog is derived from those declarations, not a second authored support matrix.
• Reuse Rule / Pattern / Replacement, PatternRewriter, and existing EP capability callbacks. No new matcher, kernel, provider-bridge API, or independent support registry.
• Collect finalized schema registrations under exact (domain, name, since_version) keys. New schemas/versions need an executable declaration or an explicit PendingTodo / Special policy.
• Preserve exact-version legacy baselines: 194 inference keys and 101 training keys. New versions cannot inherit an old-version exemption.
• Derive Available for an EP/variant from its executable declaration; CUDA still performs typed target-kernel probing before offering a rewrite.
• Compile op-associated source companions in the host optimizer, including CPU-only full-contrib builds; exclude them from normal and plugin CUDA provider recursive source lists.

No production kernel changes, expanded pilot matching, numerical-profile changes, priority changes, or default-on enablement are included. Compiled-EP execution and ONNX function expansion are unchanged.

Earlier review follow-ups: the earlier opset-domain report and all four findings in the subsequent Copilot review are addressed:

• Share non-mutating opset and parsed tensor/attribute-binding validation with PatternRewriter, across every source alternative.
• Move executable-body validation into the optimizer. Schema-only builds omit runtime policy recording/registration wrappers and do not acquire common/optimizer link dependencies.
• Reject CUDA recipes unsupported by the current single-target adapter, including multi-node replacements, reversed output mappings, and extra target outputs, before probing or graph mutation.
• Add focused regressions and a standalone consumer that constructs real ScaledSiLU, CausalConvWithState and policy-authored MoEGroupRouter schemas without linking ORT common or optimizer. Full legacy registration is not this consumer's scope: MatMulFpQ4 already has an unrelated MLAS dependency.

Latest review fixes: addressed both inline findings and the two earlier omissions identified in review 5851551387.

• Export schema-only mode to consumers and keep its finalizing helper available with DISABLE_CONTRIB_OPS; both standalone consumer variants enumerate deprecated ONNX schemas.
• Reject parsed replacement/source return-count mismatches before discovery, while preserving match-only placeholder compatibility and runtime unknown-operator tests.
• Include authored ONNX/ATen registrations in policy-bypass coverage using normalized source-file provenance, without claiming upstream ONNX or derived schemas.

Motivation and Context

Make each EP op implementation's source-replacement intent and qualification explicit and executable, rather than equating registered kernels with source declarations. ONNX function expansion runs in the opposite direction and does not replace runtime source-subgraph matching, guards, boundary mappings, ownership, or numerical qualification.

Upstream integration also provides owned pending policies for MoEGroupRouter, GatherPacked, MatMulPacked, and MoEPacked; it does not claim executable source replacements for those schemas.

Current test-only cleanup (2026-11-11)

Parent: 04f2a6e58cb216fc182b8955b48018d448eed66f; combined stack: b776d13387af86c78036e48262ca8b7cdd05f1f9. Normal commits/merge and fast-forward publication only. The final committed test contents match the validated source overlay byte-for-byte; production code, kernels, matching scope, default-off behavior and numerical gates are unchanged.

Remove owner-policy English-reason coupling and isolate malformed/duplicate registered-schema and legacy-key failures. Complete the other declarations so missing-policy rejection cannot mask schema-key validation.

Validation: passed on the combined source overlay: fresh CPU rebuild and 45/45 targeted host tests, zero failures/errors/skips; actual CUDA test translation-unit compilation with the configured CUDA-build C++ flags; canonical CLANGFORMAT and whitespace checks. The compiled old/new generator comparison passed (96 generated cases versus 88 unique cases, equal distinct configurations). Temporary generator source/executable were removed afterward.

Exact commands and results:

# PASS: rebuild the two changed host test objects and relink
cmake --build /mnt/data/titaiwang/runtime-fusion/moe-fusion-integration-20261109/host/Release --target onnxruntime_test_all --parallel 2
# PASS: 45/45, from the build output directory
cd /mnt/data/titaiwang/runtime-fusion/moe-fusion-integration-20261109/host/Release
./onnxruntime_test_all --gtest_filter='MoE*:ContribSourceReplacementPolicyTest.*' --gtest_output=xml:/home/titaiwang/.copilot/session-state/07b6ecde-d659-4b29-aab6-78ba683dbbb3/files/test-cleanup-host.xml
# PASS: actual old/new C++ parameter generators, no GPU use
python /home/titaiwang/.copilot/session-state/07b6ecde-d659-4b29-aab6-78ba683dbbb3/files/validate_test_cleanup_cases.py

Not rerun for this test-only cleanup: GPU numerical execution, full/cross-platform suites, and standalone schema consumers. The older 386-host/98-GPU results below are historical, not reruns on these heads. The 45-test run used the combined stack, not a separate parent-only build. No new-head CI success is claimed.

Previous CI policy validation after syncing main 0b22d27a85 (before test-only cleanup)

Parent: 5362926ba8374fa250943346647c9269212db011; combined stack: e63f8268ea5d7c4113c861ef4d9a1418694ce12a. Normal merges only; no history rewrite.

Upstream #33901 added MatMulNBitsQLoRAMlp-1 and MatMulNBitsQLoRAQkv-1. Both now author PendingTodo beside their schema definitions; neither is added to the frozen legacy baseline or advertised as an executable source. The existing regression covers both exact keys, unqualified variants, CPU/CUDA/WebGPU, rejection of version 2, and rejection after removing either policy. This follow-up does not edit kernels, matching scope or default-off behavior.

Passed: 386/386 CPU-build host tests on the exact clean combined commit, including full inventory and the extended owner-policy regression; 2/2 standalone schema consumers; all 3 disabled-contrib syntax probes (bert_defs.cc, contrib_defs.cc, training_op_defs.cc, with both training macros); canonical CLANGFORMAT and diff whitespace checks. The host/schema builds used the newly pinned Eigen archive (SHA1 18eca78e961ac5e9123ccdff5035ab340dcc5e11), isolated from the earlier Eigen 3.4 cache. At most four compilation jobs. Source-tree identity and artifact hashes were checked before publication.

Exact build/test commands (all passed; tests run from their respective build output directories):

# PASS: schema-build
cmake --build /mnt/data/titaiwang/runtime-fusion/review-fixes-schema-validation-20261110/Release --parallel 4 --target ort_opschema_lib_test ort_opschema_lib_disabled_contrib_test
# PASS: schema-consumers
ctest --test-dir /mnt/data/titaiwang/runtime-fusion/review-fixes-schema-validation-20261110/Release -R '^ort_opschema_lib(_disabled_contrib)?_test$' --output-on-failure --no-tests=error --timeout 60
# PASS: runtime-build
cmake --build /mnt/data/titaiwang/runtime-fusion/moe-fusion-integration-20261109/host/Release --parallel 4 --target onnxruntime_test_all
# PASS: host
/mnt/data/titaiwang/runtime-fusion/moe-fusion-integration-20261109/host/Release/onnxruntime_test_all '--gtest_filter=MoE*:ContribSourceReplacementPolicyTest.*:UnitSwishHostTest.*:CUDARuntimeFusionHostTest.*:PatternRewriterTests.*:PatternComponentTests.*:PatternCapabilityTests.*:ReplacementComponentTests.*:CheckAndBindComponentTests.*:MatchComponentTests.*:FunctionTest.*' --gtest_output=xml:/home/titaiwang/.copilot/session-state/07b6ecde-d659-4b29-aab6-78ba683dbbb3/files/fusion-ci-20261111T1809/validation-3/host.xml

Exclusions for that previous main-sync follow-up: no GPU workload or CUDA kernel rebuild was performed for this metadata-only follow-up. The 98-case CUDA result below belongs to the earlier pre-main-sync tree and is not qualification of the new merged head. Windows/macOS/Android/Web/plugin execution and full upstream integration coverage remain CI responsibilities; new-head CI is not claimed green. The bounded two-file authored diff was reviewed directly; the upstream commit set was not independently re-reviewed.

Previous review-fix validation before the main sync

Previous review-fix validation: ca027638c1e15f596a82081180e73644af999caf

Parent-only overlay: 358/358 host, 2/2 real CUDA EP, 2/2 standalone schema consumers passed. That run preceded only extraction of an unchanged test DSL literal into a constant.

The final combined stack passed 386/386 host, 98/98 real CUDA EP tests (96 MoE cases / 288 sessions plus the two existing pilots), 2/2 standalone schema consumers, and disabled-contrib syntax probes for bert_defs.cc, contrib_defs.cc, and training_op_defs.cc with both training macros. CUDA execution used idle physical GPU 3, H200/sm90, under the fail-closed guard; zero gtest failures/errors/skips. The newly built task-owned CUDA provider was loader-verified and artifact hashes stayed unchanged during execution. Four build jobs, NVCC threads=1.

Commits were created after validation; their merged tree was verified byte-for-byte against the compiled final source overlay before normal atomic publication. These are source-tree-qualified runs, not a claim that embedded build revision strings were regenerated after committing. The validation tree is clean at the published merge head.

Commands below ran from the existing build-output directories with the existing CUDA/cuDNN loader paths. Each completed successfully:

timeout --signal=TERM --kill-after=15s 14400s cmake --build /mnt/data/titaiwang/runtime-fusion/moe-fusion-integration-20261109/cuda/Release --parallel 4 --target onnxruntime_test_all onnxruntime_providers_cuda onnxruntime_provider_test
# Passed.
/mnt/data/titaiwang/runtime-fusion/moe-fusion-integration-20261109/cuda/Release/onnxruntime_test_all '--gtest_filter=MoE*:ContribSourceReplacementPolicyTest.*:UnitSwishHostTest.*:CUDARuntimeFusionHostTest.*:PatternRewriterTests.*:PatternComponentTests.*:PatternCapabilityTests.*:ReplacementComponentTests.*:CheckAndBindComponentTests.*:MatchComponentTests.*:FunctionTest.*' --gtest_output=xml:/home/titaiwang/.copilot/session-state/07b6ecde-d659-4b29-aab6-78ba683dbbb3/files/review-fixes-20261111/validation-combined-2/host.xml
# Passed.
timeout --signal=TERM --kill-after=15s 1800s cmake --build /mnt/data/titaiwang/runtime-fusion/review-fixes-schema-validation-20261110/Release --parallel 4 --target ort_opschema_lib_test ort_opschema_lib_disabled_contrib_test
# Passed.
ctest --test-dir /mnt/data/titaiwang/runtime-fusion/review-fixes-schema-validation-20261110/Release -R '^ort_opschema_lib(_disabled_contrib)?_test$' --output-on-failure --no-tests=error --timeout 60
# Passed.
/mnt/data/titaiwang/runtime-fusion/moe-fusion-integration-20261109/cuda/Release/onnxruntime_provider_test '--gtest_filter=CudaRuntimeFusionTest.UnitSwishScalarFloatBitwise:CudaRuntimeFusionTest.ExtendedOptimizersNumericsAndCudaDispatch:MoE/CudaMoERuntimeFusionTest.SourceFusedDirectTF32OffAndOn/*' --gtest_output=xml:/home/titaiwang/.copilot/session-state/07b6ecde-d659-4b29-aab6-78ba683dbbb3/files/review-fixes-20261111/validation-combined-3/cuda.xml
# Passed.

Canonical lintrunner --take CLANGFORMAT <changed C++ paths> and git diff --check passed in both working trees. Lintrunner used the existing virtual environment plus PYTHONPATH=/usr/lib/python3.12/site-packages; the earlier adapter blockage is resolved without installing dependencies. The task-created driver cache was removed.

Previous exclusions: actual pre-SM80 execution is not qualified on H200; host tests cover metadata validation without materialization/lookup and the actual-device regression was extended for a future pre-SM80 run. Actual Windows/MSVC, full training/plugin suites, other GPU architectures, authentic fixtures and whole-model accuracy/performance were not run locally. No new high-precision kernel-accuracy gate is added. New-head CI is reported separately, not claimed green. The earlier Android comparison below is historical, not a new-head size measurement. No new full-team/model-diversity review coverage is claimed.

Historical validation and CI follow-ups (earlier heads only)

Earlier CUDA validation baseline: 06c0d81720b17bd20ac780bb62a2106799553c7b

• Consistent CUDA build: passed, four build jobs and NVCC threads=1.
• Targeted host regressions: 356/356 passed, zero failures/errors/skips, including both new declaration/output-mapping regressions.
• Targeted provider regressions: 95/95 passed, zero failures/errors/skips, including positive ScaledSiLU/causal CUDA EP placement and typed-kernel checks.
• Actual execution used physical GPU 4, H200/sm90, CUDA 13.3. Loader logs verified the newly built task-owned CUDA provider; source cleanliness and artifact hashes were checked.
• Standalone schema-library consumer: 1/1 passed, linked without ORT common/optimizer. The real schema CMake module was included in an isolated offline CPU configuration; full-training configuration was blocked by uncached unrelated dependencies and was not claimed.
• Changed C++ ranges passed clang-format 18 and git diff --check. Canonical lintrunner was then blocked by missing lintrunner_adapters; no dependencies were installed.

CUDA commands ran from /mnt/data/titaiwang/runtime-fusion/moe-fusion-integration-20261109/cuda/Release, with the virtual environment activated and task-owned CUDA/cuDNN loader paths. Tests used GPU 4 under the fail-closed resource guard (30-second telemetry-query timeout, 30-minute workload budget). Sampling is not an exclusive GPU lease.

cmake --build /mnt/data/titaiwang/runtime-fusion/moe-fusion-integration-20261109/cuda/Release --parallel 4 --target onnxruntime_test_all onnxruntime_providers_cuda onnxruntime_provider_test
# Passed.
./onnxruntime_test_all --gtest_filter='MoE*:ContribSourceReplacementPolicyTest.*:UnitSwishHostTest.*:CUDARuntimeFusionHostTest.*:PatternRewriterTests.*:PatternComponentTests.*:PatternCapabilityTests.*:ReplacementComponentTests.*:CheckAndBindComponentTests.*:MatchComponentTests.*:FunctionTest.*'
# Passed: 356 tests.
./onnxruntime_provider_test --gtest_filter='CudaRuntimeFusionTest.*:CudaRuntimeFusionProbe.ZeroBatch:CausalConvWithStateTest.*-*LocalAuthentic*:CudaRuntimeFusionTest.MoEActualPreSM80DeclinesBeforeTargetLookup'
# Passed: 95 tests.
cmake --build /mnt/data/titaiwang/runtime-fusion/review-fixes-schema-validation-20261110/Release --parallel 4 --target ort_opschema_lib_test
# Passed.
ctest --test-dir /mnt/data/titaiwang/runtime-fusion/review-fixes-schema-validation-20261110/Release -R '^ort_opschema_lib_test$' --output-on-failure
# Passed: 1 test.

Earlier CI portability follow-up: fa608643669a80093616f64d0df34b5767db98ea

Move the four pending-policy builders' preprocessor guards outside schema-macro arguments. The shared runtime helper preserves finalized keys and the exact existing policies; schema-only builds use a translation-unit-local helper that preserves schema finalization while omitting only runtime metadata collection, without runtime linkage dependencies.

• CUDA-enabled build passed; 15/15 focused policy tests and the standalone schema consumer 1/1 passed. Build logs confirm both runtime and schema-only recompilation of the two schema files and the policy implementation.
• The strengthened standalone consumer checks the policy-authored schema's finalized arity and resolved input types. It was reproduced failing with an empty schema-only helper, then passed after preserving finalization.
• 2/2 real CUDA EP tests passed, covering Unit Swish and causal-convolution fused dispatch, on physical GPU 4 under the fail-closed guard. Loader logs verify the task-owned CUDA provider; zero failures/errors/skips.
• Full-source Clang preprocessing with -Werror=embedded-directive reproduced the old failure and accepted the fix in all four file/build-mode combinations (both schema files, runtime and schema-only). This is a standards-diagnostic regression, not an MSVC run.

Commands for that earlier follow-up used the existing CUDA/cuDNN loader paths and build-output cwd:

timeout --signal=TERM --kill-after=15s 14400s cmake --build /mnt/data/titaiwang/runtime-fusion/moe-fusion-integration-20261109/cuda/Release --parallel 4 --target onnxruntime_test_all onnxruntime_providers_cuda onnxruntime_provider_test
# Passed.
/mnt/data/titaiwang/runtime-fusion/moe-fusion-integration-20261109/cuda/Release/onnxruntime_test_all '--gtest_filter=ContribSourceReplacementPolicyTest.*' --gtest_output=xml:/home/titaiwang/.copilot/session-state/07b6ecde-d659-4b29-aab6-78ba683dbbb3/files/fusion-ci-20261110T221446/validation-parent-final/host.xml
# Passed.
timeout --signal=TERM --kill-after=15s 1800s cmake --build /mnt/data/titaiwang/runtime-fusion/review-fixes-schema-validation-20261110/Release --parallel 4 --target ort_opschema_lib_test
# Passed.
ctest --test-dir /mnt/data/titaiwang/runtime-fusion/review-fixes-schema-validation-20261110/Release -R '^ort_opschema_lib_test$' --output-on-failure --no-tests=error --timeout 60
# Passed.
/mnt/data/titaiwang/runtime-fusion/moe-fusion-integration-20261109/cuda/Release/onnxruntime_provider_test --gtest_filter=CudaRuntimeFusionTest.UnitSwishScalarFloatBitwise:CudaRuntimeFusionTest.ExtendedOptimizersNumericsAndCudaDispatch --gtest_output=xml:/home/titaiwang/.copilot/session-state/07b6ecde-d659-4b29-aab6-78ba683dbbb3/files/fusion-ci-20261110T221446/validation-parent-cuda-final/cuda.xml
# Passed.

Changed C++ ranges match clang-format 18 and pass git diff --check. Canonical lintrunner was attempted for the changed files but was then blocked by missing lintrunner_adapters; no dependencies were installed.

Known upstream Android baseline failure: the exact upstream main commit d7ab0d271b0399d889c6b9540cacf04a84b74333 fails the same gate: section total 1,611,530 bytes versus threshold 1,610,778, exceeding it by 752 bytes. #34022's captured earlier CI build reports the same total, on-disk size, and all 27 section sizes. This measured size failure is already present upstream; these fixes do not raise the threshold or claim the gate is green.

Exclusions recorded for that earlier follow-up: new-head Windows/MSVC and broad CI results are pending. The earlier 356-host/95-provider run remains historical evidence, not a rerun of that entire selection for this metadata-only patch. Full ORT/plugin/training suites, other GPU architectures, optional authentic fixtures, whole-model accuracy/performance, and earlier fixed-model review coverage gaps remain unqualified. This follow-up received direct bounded review with no actionable findings, not another whole-team review.

[Copilot]

🟡 Changes recommended

The test-schema exclusion misses Windows path separators, allowing GraphTest schemas to be misreported as missing contrib registrations and causing order-dependent test failures.

1 open issue

• Normalize Windows paths when excluding test schemas · New

What changed in this PR

This PR adds op-owned contrib source declarations on top of the causal-fusion stack. It moves existing CUDA fusion rules into op-associated host companions while preserving kernels, eligibility, numerical guarantees, and default-off behavior.

Changes:

• Add exact-version schema inventory, explicit source policies, and frozen inference/training exemptions.
• Connect existing Swish and causal-convolution declarations to capability selection, keeping host companions outside CUDA provider libraries.
• Add policy and ownership tests, plus documentation for declaration authoring and qualification.

File	Description
onnxruntime/​test/​optimizer/​contrib_source_replacement_policy_test.cc	Tests inventory, metadata validation, ownership, guards, and unsupported recipes.
onnxruntime/​core/​optimizer/​cuda_runtime_fusion.cc	Collects declarations and routes selection through their callbacks.
onnxruntime/​core/​optimizer/​cuda_runtime_fusion_internal.h	Removes discovery classes replaced by companion callbacks.
onnxruntime/​core/​graph/​contrib_ops/​source_replacement.h	Defines executable bodies and numerical/transformation metadata.
onnxruntime/​core/​graph/​contrib_ops/​source_replacement_training_legacy.h	Freezes 101 exact training-schema exemptions.
onnxruntime/​core/​graph/​contrib_ops/​source_replacement_policy.h	Declares schema keys, policy records, and validation interfaces.
onnxruntime/​core/​graph/​contrib_ops/​source_replacement_policy.cc	Implements inventory collection, validation, and status resolution.
onnxruntime/​core/​graph/​contrib_ops/​source_replacement_legacy.h	Freezes 194 inference keys and combines legacy inventories.
onnxruntime/​core/​graph/​contrib_ops/​onnx_deprecated_opset.h	Collects deprecated ONNX schema registrations.
onnxruntime/​core/​graph/​contrib_ops/​ms_schema.h	Includes source-policy declarations for schema registration.
onnxruntime/​core/​graph/​contrib_ops/​ms_opset.h	Collects Microsoft opset schema registrations.
onnxruntime/​core/​graph/​contrib_ops/​contrib_defs.h	Routes contrib registration macros through inventory collection.
onnxruntime/​contrib_ops/​cuda/​fusion/​scaled_silu_sources.cc	Owns the existing Unit Swish rule and qualification metadata.
onnxruntime/​contrib_ops/​cuda/​fusion/​cuda_source_replacements.h	Defines CUDA companions, factories, and catalog interfaces.
onnxruntime/​contrib_ops/​cuda/​fusion/​causal_conv_with_state_sources.cc	Owns the existing causal rule and bias-transformation metadata.
docs/​Pattern_Rewriter.md	Explains declaration ownership and reuse of rewrite primitives.
docs/​design/​Extensible_Runtime_Fusion.md	Separates implemented v0 policy from broader design proposals.
docs/​Cuda_Runtime_Fusion.md	Documents CUDA declaration integration and qualification steps.
cmake/​onnxruntime_providers_cuda.cmake	Excludes host companions from the normal CUDA provider.
cmake/​onnxruntime_providers_cuda_plugin.cmake	Excludes host companions from the CUDA plugin.
cmake/​onnxruntime_optimizer.cmake	Compiles companions into full-contrib host optimizer builds.

_{Review effort: Balanced}

[Chi Lo (chilo-ms)]

Reject declarations missing source-domain opsets — source_replacement_policy.cc:73–79

Keeping only  {kMSDomain, 1, 1}  in the ScaledSiLU declaration’s  rule.opsets  passes both declaration validators. However, discovery constructs a  PatternRewriter , which throws because the Swish pattern’s ONNX-domain requirement is missing. The coverage check therefore accepts a non-executable declaration and defers failure until enabled session initialization.

Validate opset coverage for every pattern/replacement domain, preferably sharing the rewriter’s validation, and add this negative regression case.

This finding is based on static review

[Copilot]

🟡 Changes recommended

The schema-library change leaves consumers using its existing dependencies with unresolved Status symbols, while validation gaps admit inapplicable or incorrectly mapped rewrite recipes.

4 open issues

• Avoid unresolved Status dependencies in opschema builds · New
• Validate rewrite opsets and bindings before discovery · New
• Reject unsupported multi-node recipes in CUDA validation · New
• Reject non-identity or mismatched causal output mappings · New

1 resolved since last review

• Normalize Windows paths when excluding test schemas

_{Review effort: Max}

[Copilot]

🟡 Changes recommended

Critical schema-library compilation and standalone schema-enumeration link failures remain, alongside missing replacement return-count validation and incomplete ONNX/ATen schema policy coverage.

2 open issues

• Export schema enumeration for standalone consumers · New
• Keep schema-only policy helper available when contrib ops are disabled · New

2 previously missed issues

In code that hasn't changed since last review

Reject replacement return counts that differ from source outputs

onnxruntime/​core/​optimizer/​pattern_rewriter.cc:2705

ValidateRule accepts a parsed replacement whose return count differs from the source pattern's output count. Such a declaration passes common validation and can resolve to Available, but every matched rewrite fails in StageReplacement with a return-count error. Check the count for every source alternative here and add a negative regression.

Include ONNX and ATen registrations in policy coverage checks

onnxruntime/​test/​optimizer/​contrib_source_replacement_policy_test.cc:43

This bypass check ignores authored ONNX-domain and ATen schemas, although both are covered by the inventory and frozen baselines. A new schema in either domain registered outside the collection wrappers is therefore absent from validation, allowing the coverage test to pass without its required policy. Use schema-file provenance to include contrib/training registrations while excluding standard ONNX and derived schemas, and add regressions for both domains.

4 resolved since last review

• Avoid unresolved Status dependencies in opschema builds
• Reject non-identity or mismatched causal output mappings
• Reject unsupported multi-node recipes in CUDA validation
• Validate rewrite opsets and bindings before discovery

_{Review effort: Max}

[G. Ramalingam (gramalingam)]

It would be nice if there were a compact "TLDR" summary in the documentation that uses a short example to illustrate the key idea, and mentions the key elements that are brought together.

[Ti-Tai Wang (titaiwangms)]

It would be nice if there were a compact "TLDR" summary in the documentation that uses a short example to illustrate the key idea, and mentions the key elements that are brought together.

Done