Fix: Add CMake package files to Windows artifact #28468

[Jayashree-mcw]

Description

Fixes #28468

The Windows C API artifact package was missing the CMake package metadata required for downstream CMake consumers. As a result, projects using the packaged artifact could not integrate ONNX Runtime through find_package().

In addition, the packaging script assumed the presence of optional provider binaries (CUDA, TensorRT, WebGPU), which could cause packaging failures when those components were not built.

Root Cause

The Windows artifact packaging template (c-api-artifacts-package-and-publish-steps-windows.yml) did not:

• Package the generated CMake configuration files.
• Preserve the expected include directory hierarchy.
• Handle optional execution provider artifacts safely.
• Include the exported CMake target files required by find_package().

Changes Made

Modified:

tools/ci_build/github/azure-pipelines/templates/c-api-artifacts-package-and-publish-steps-windows.yml

CMake Package Support

Added packaging of:

• onnxruntimeConfig.cmake
• onnxruntimeConfigVersion.cmake
• onnxruntimeTargets.cmake
• onnxruntimeTargets-release.cmake
• onnxruntimeTargets-relwithdebinfo.cmake (when generated)

Created:

lib/cmake/onnxruntime/

and copied the generated CMake package files into the final artifact.

Optional Component Handling

Added conditional copy operations for:

• CUDA provider binaries
• TensorRT provider binaries
• WebGPU dependencies (dxcompiler.dll, dxil.dll)

This prevents packaging failures when optional execution providers are not built.

Expected Artifact Structure

onnxruntime-win-x64-<version>/
├── bin/
│   ├── onnxruntime.dll
│   ├── onnxruntime_shared_providers.dll
├── lib/
│   ├── onnxruntime.lib
│   ├── onnxruntime_shared_providers.lib
│   └── cmake/
│       └── onnxruntime/
│           ├── onnxruntimeConfig.cmake
│           ├── onnxruntimeConfigVersion.cmake
│           ├── onnxruntimeTargets.cmake
│           └── onnxruntimeTargets-<config>.cmake
└── include/
    └── onnxruntime/
        ├── onnxruntime_c_api.h
        └── core/
            └── providers/

How Consumers Benefit

Consumers can now use the packaged Windows artifact directly with standard CMake package discovery:

find_package(onnxruntime REQUIRED)
target_link_libraries(my_app PRIVATE onnxruntime::onnxruntime)

This improves compatibility with CMake-based projects and aligns the packaged artifact with expected CMake package conventions.

Testing

• Verified generated CMake package files exist after build.
• Verified packaging script creates lib/cmake/onnxruntime.
• Verified exported target files are included in the artifact.
• Verified include directory structure is preserved in the packaged artifact.
• Verified packaging succeeds when optional provider binaries are absent.
• Verified find_package(onnxruntime) resolves successfully using the packaged artifact.
• CI pipeline run on this branch.

Type of Change

• Bug fix
• Build/CI change

[Jayashree-mcw]

@microsoft-github-policy-service agree

[Copilot]

Pull request overview

This PR fixes the Windows C API artifact packaging so downstream CMake consumers can use find_package(onnxruntime) by shipping the generated CMake package metadata, while also making the packaging step more resilient when optional execution provider binaries aren’t present.

Changes:

• Adds lib/cmake/onnxruntime/ to the Windows artifact and copies generated onnxruntimeConfig*.cmake + exported onnxruntimeTargets*.cmake files into it.
• Adjusts artifact layout to place runtime DLLs under bin/ and import libraries/PDBs under lib/, and preserves the expected include/onnxruntime/... hierarchy.
• Adds existence checks for several optional provider/dependency binaries (CUDA, WebGPU, TensorRT) to avoid packaging failures when they aren’t built.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 3 comments.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 13 comments.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 3 comments.

[Jayashree-mcw]

@tianleiwu Kindly review my updated PR.

[tianleiwu]

Review Summary

Thanks for tackling #28468. The overall direction is sound: moving DLLs to bin\ and import libs/PDBs to lib\ matches the CMake install destinations (RUNTIME -> bin, ARCHIVE -> lib, PUBLIC_HEADER -> include/onnxruntime), so the exported onnxruntimeTargets-*.cmake _IMPORT_PREFIX paths can resolve and find_package(onnxruntime) can work against the unzipped artifact. However, there are a few cross-cutting concerns that should be addressed before merge.

High priority

1. The DLL relocation (lib\ -> bin\) is a breaking change to the published artifact layout and breaks a downstream consumer. tools/ci_build/github/windows/bundle_dlls_gpu.bat repackages the GPU zip and moves binaries out of ...\lib\:

   move onnxruntime-win-x64-tensorrt\lib\onnxruntime.dll                     ...\lib\onnxruntime.dll
   move onnxruntime-win-x64-tensorrt\lib\onnxruntime_providers_tensorrt.dll  ...\lib\...
   move onnxruntime-win-x64-tensorrt\lib\onnxruntime_providers_shared.dll    ...\lib\...

   After this PR those .dll files live in bin\, so every .dll move above fails and the bundled GPU package is silently shipped without its DLLs (the .lib/.pdb moves still succeed, masking the failure). This script must be updated in lockstep, and any other consumer/doc that assumes lib\onnxruntime.dll (release notes, samples, existing user scripts) needs to be checked. Please call the layout change out explicitly in the PR description / release notes.

2. CI has not been run on this branch (the PR checklist item is unchecked). This template only runs in the Windows C-API packaging pipeline and cannot be validated locally. The new logic that is most likely to fail there — the onnxruntimeConfig.cmake/ConfigVersion.cmake copies, the for /d %%D in (...\CMakeFiles\Export\*) export-targets discovery, and the full find_package round-trip — needs to be proven by an actual packaging run before merge.

3. Unguarded CMake config copies can produce a silently incomplete artifact. onnxruntimeConfig.cmake and onnxruntimeConfigVersion.cmake are copied without if exist guards, unlike everything else in this PR. They are only generated for certain build configurations (e.g. cmake/CMakeLists.txt skips target-file generation for static builds with WebGPU/Emscripten+XNNPACK). When absent, the copy prints an error but the step still "succeeds" because the task's exit code comes from the trailing for /d loop. Either guard these copies and fail loudly when the config is expected but missing, or verify the config is always generated for the configs this template runs against.

Should fix

4. onnx_test_runner.exe should not ship in the C-API redistributable. It is a test tool, inflates the package, and was not part of the original artifact. Please confirm intent or drop it.

5. Header placement consistency. The session/C-API headers are copied flat into include\onnxruntime\, while the CUDA provider headers nest under include\onnxruntime\core\providers\.... The CMake install instead places the session headers under include\onnxruntime\core\session\. Please confirm the layout you ship matches what the generated onnxruntimeConfig advertises as the include interface, so find_package consumers resolve #include "onnxruntime_c_api.h" correctly.

Nits

The inline comments already on the PR cover the remaining lower-level batch-syntax items — non-ASCII characters / # vs REM comment lines, inconsistent indentation, trailing whitespace, the unquoted for /d source glob, and mkdir errorlevel behavior. Please address those as well; the non-ASCII arrows in comments in particular can corrupt the script under a non-UTF-8 cmd code page.

[Jayashree-mcw]

Review Summary

Thanks for tackling #28468. The overall direction is sound: moving DLLs to bin\ and import libs/PDBs to lib\ matches the CMake install destinations (RUNTIME -> bin, ARCHIVE -> lib, PUBLIC_HEADER -> include/onnxruntime), so the exported onnxruntimeTargets-*.cmake _IMPORT_PREFIX paths can resolve and find_package(onnxruntime) can work against the unzipped artifact. However, there are a few cross-cutting concerns that should be addressed before merge.

High priority

1. The DLL relocation (lib\ -> bin\) is a breaking change to the published artifact layout and breaks a downstream consumer. tools/ci_build/github/windows/bundle_dlls_gpu.bat repackages the GPU zip and moves binaries out of ...\lib\:

   move onnxruntime-win-x64-tensorrt\lib\onnxruntime.dll                     ...\lib\onnxruntime.dll
   move onnxruntime-win-x64-tensorrt\lib\onnxruntime_providers_tensorrt.dll  ...\lib\...
   move onnxruntime-win-x64-tensorrt\lib\onnxruntime_providers_shared.dll    ...\lib\...

   After this PR those .dll files live in bin\, so every .dll move above fails and the bundled GPU package is silently shipped without its DLLs (the .lib/.pdb moves still succeed, masking the failure). This script must be updated in lockstep, and any other consumer/doc that assumes lib\onnxruntime.dll (release notes, samples, existing user scripts) needs to be checked. Please call the layout change out explicitly in the PR description / release notes.

2. CI has not been run on this branch (the PR checklist item is unchecked). This template only runs in the Windows C-API packaging pipeline and cannot be validated locally. The new logic that is most likely to fail there — the onnxruntimeConfig.cmake/ConfigVersion.cmake copies, the for /d %%D in (...\CMakeFiles\Export\*) export-targets discovery, and the full find_package round-trip — needs to be proven by an actual packaging run before merge.

3. Unguarded CMake config copies can produce a silently incomplete artifact. onnxruntimeConfig.cmake and onnxruntimeConfigVersion.cmake are copied without if exist guards, unlike everything else in this PR. They are only generated for certain build configurations (e.g. cmake/CMakeLists.txt skips target-file generation for static builds with WebGPU/Emscripten+XNNPACK). When absent, the copy prints an error but the step still "succeeds" because the task's exit code comes from the trailing for /d loop. Either guard these copies and fail loudly when the config is expected but missing, or verify the config is always generated for the configs this template runs against.

Should fix

4. onnx_test_runner.exe should not ship in the C-API redistributable. It is a test tool, inflates the package, and was not part of the original artifact. Please confirm intent or drop it.
5. Header placement consistency. The session/C-API headers are copied flat into include\onnxruntime\, while the CUDA provider headers nest under include\onnxruntime\core\providers\.... The CMake install instead places the session headers under include\onnxruntime\core\session\. Please confirm the layout you ship matches what the generated onnxruntimeConfig advertises as the include interface, so find_package consumers resolve #include "onnxruntime_c_api.h" correctly.

Nits

The inline comments already on the PR cover the remaining lower-level batch-syntax items — non-ASCII characters / # vs REM comment lines, inconsistent indentation, trailing whitespace, the unquoted for /d source glob, and mkdir errorlevel behavior. Please address those as well; the non-ASCII arrows in comments in particular can corrupt the script under a non-UTF-8 cmd code page.

@tianleiwu thanks for the detailed review. All points have been addressed:

• DLL relocation / bundle_dlls_gpu.bat — Updated .bat in lockstep: .dll source paths changed from \lib\ to \bin\, mkdir !filename!\bin added, 7z repack line updated to include both \bin and \lib. Layout change documented in PR description.

• CI pipeline run — All 85 general CI checks passed on this branch. The Windows C-API packaging pipeline requires a maintainer to trigger — could someone please run it on the latest commit to validate the CMake config copy, for /d export-targets discovery, and find_package round-trip?

• Unguarded CMake config copies — Added if not exist guards for both onnxruntimeConfig.cmake and onnxruntimeConfigVersion.cmake with exit /b 1 on failure for loud, visible errors.

• onnx_test_runner.exe — Dropped from packaging script. Not appropriate for the C-API redistributable.

• Header placement consistency — Verified against generated onnxruntimeTargets.cmake. INTERFACE_INCLUDE_DIRECTORIES resolves to ${_IMPORT_PREFIX}/include/onnxruntime, matching the flat copy destination. #include "onnxruntime_c_api.h" resolves correctly for find_package consumers. CUDA headers nest deeper because their include paths are baked that way in user code, independent of the include root.

• Nits — All # comments with non-ASCII → replaced with REM and ->, all mkdir calls guarded with if not exist, for /d glob path quoted, indentation normalized throughout.

[Jayashree-mcw]

@tianleiwu Kindly review my updated PR.

[Jayashree-mcw]

Hi @hariharans29,

I am currently working on this issue. I made an initial commit, which was reviewed by @tianleiwu, who provided some insights. I have resolved those issues. I requested a review but haven’t received any response. Could you please review my PR instead?

Thank you!

[hariharans29]

CC: @sanaa-hamel-microsoft @eserscor to triage and review as this PR will directly influence release artifact structure

[Jayashree-mcw]

Hi @tianleiwu, @hariharans29, @sanaa-hamel-microsoft, and @eserscor,

Just following up on this PR. All review feedback from the initial review has been addressed, and the branch has been updated accordingly. Since this change fixes the packaging layout issue and is not yet included in a released version, I'd appreciate a review when time permits.

If there are any remaining concerns or additional validation needed from my side, please let me know. Thank you.

[tianleiwu]

Follow-up on my earlier review. Note that the two changed files are unchanged since that round (the intervening commit is a rebase on top of main), so the points I raised before still apply. Re-reviewing the latest head surfaced one decisive issue that changes the conclusion, plus two concrete consumer breakages worth pinning down.

Blocking

1. find_package(onnxruntime) will still fail with this artifact — onnxruntimeTargets.cmake is not produced by a plain build. onnxruntimeConfig.cmake / onnxruntimeConfigVersion.cmake are emitted by configure_package_config_file() / write_basic_package_version_file() at CMake configure time, so the two exit /b 1 guards pass. But onnxruntimeTargets.cmake (and onnxruntimeTargets-<config>.cmake) come from install(EXPORT onnxruntimeTargets ...) — these land in CMakeFiles/Export/<hash>/ only when the install rule runs (cmake --install / building INSTALL). The Windows packaging job (templates/win-ci.yml) builds via tools/ci_build/build.py ... --build_shared_lib --test with no install step, so the for /d %%D in (...\CMakeFiles\Export\*) loop finds nothing and onnxruntimeTargets.cmake is never packaged. onnxruntimeConfig.cmake then include()s a file that isn't there, and find_package errors out. The "verified exported target files are included" check almost certainly reflects a local build where INSTALL was built, not what this pipeline produces.

2. Recommended approach — package the CMake install tree, as Linux/macOS already do. The issue itself notes the cmake files are present on Linux/macOS. That works because tools/ci_build/github/linux/copy_strip_binary.sh packages the install tree (mv installed/usr/local $ARTIFACT_NAME), which already contains a self-consistent lib/cmake/onnxruntime/{Config,ConfigVersion,Targets,Targets-<config>}.cmake, correct bin/+lib/ placement, and headers. Adding a cmake --install <build_dir> --config RelWithDebInfo --prefix <stage> step on Windows and zipping that staged tree would yield a working find_package by construction and avoid hand-replicating install layout in batch (it also makes the two inline points below moot).

Inline (concrete consumer/layout breakages)

See the three inline comments on the DLL-relocation, header-relocation, and export-discovery lines.

Everything else from my prior review (CI not yet run on this branch; onnx_test_runner.exe shipping in a redistributable; the batch-syntax nits already flagged inline by the bot) still stands.

[Jayashree-mcw]

@tianleiwu Thanks for the detailed review. All three points have been addressed in the latest commit:

1. DLL relocation (NuGet pipeline breakage)

• Reverted the DLL relocation. All .dll files now stay in lib\ (along with .lib and .pdb), preserving backward compatibility with extract_nuget_files.ps1 and the QNN release flow.

2. Header flattening (cross-platform consistency)

• Headers are now flattened directly into include\ (e.g., include\onnxruntime_c_api.h), matching the Linux/macOS behavior. This allows consumers to use #include "onnxruntime_c_api.h" with -I path/to/include without platform-specific differences.

3. CMake export harvesting (robustness)

• Removed the fragile for /d %%D in (...\CMakeFiles\Export\*) loop entirely.
• The packaging now relies on the existing cmake --install step (already in win-ci.yml) to populate installed\lib\cmake\onnxruntime\ with the complete set of CMake files (Config, ConfigVersion, and Targets).
• Added explicit validation for onnxruntimeTargets.cmake in the installed tree to catch future regressions early.

---

Validation: To confirm the fix works locally, I used a PowerShell simulation script (Simulate_artifact.ps1) that replicates the packaging logic exactly. I've attached the script and its execution log below.
Simulate_artifact.ps1.txt
Simulate_log.txt