onnxruntime

b2021624 - Isolate volatile build provenance from shared configuration (#33687)

Isolate volatile build provenance from shared configuration (#33687) ### Description <!-- Describe your changes. --> Move `ORT_BUILD_INFO` from the widely included generated `onnxruntime_config.h` into a private generated `onnxruntime_build_info.h`. Explicitly include it in all four existing consumers: the C API, inference Python bindings, training Python bindings, and CUDA GEMM tactic cache. String construction is unchanged: Git branch/commit, existing capability flags (including INT4/FP8/FP4 settings), and build type retain their existing semantics. No public API, dependency, kernel, or operator changes. ### Motivation and Context <!-- - Why is this change required? What problem does it solve? - If it fixes an open issue, please link to the issue here. --> A real Git commit changes build provenance at the next configure. Putting that value in the shared feature-config header invalidates unrelated objects, even when feature flags and compiler commands are unchanged. Keeping it private limits that invalidation to code that actually reports or consumes build identity. Based on current main `123752d361d2da8ff592b347772f31362e0966bd`; signed-off head `3bb32435737be91ea76805473a2a27ea95a0e352`. ### Validation Fresh isolated Linux x86-64 CPU build; no GPU used. Release configuration with `-O0` to reduce validation compile cost; contrib and ML operators disabled. Build/test artifacts and isolated Python environment are under `R=/tmp/ort-build-info-7144cd59`. Commands below ran from `$R/src`, except the gtest, which ran from `$R/build/Release`. | Command | Result | | --- | --- | | `$R/venv/bin/python tools/ci_build/build.py --build_dir $R/build --config Release --update --cmake_generator Ninja --build_shared_lib --enable_pybind --disable_ml_ops --disable_contrib_ops --skip_tests --compile_no_warning_as_error --cmake_extra_defines CMAKE_CXX_FLAGS_RELEASE=-O0 CMAKE_C_FLAGS_RELEASE=-O0` | Passed after installing initially missing NumPy into the isolated environment. Original and patched generation at the same HEAD reported byte-identical build-info strings. | | `cmake --build $R/build/Release --target onnxruntime onnxruntime_pybind11_state --parallel 4` | Passed: fresh C API/inference Python build and final-commit incremental rebuild. | | `cmake --compile-no-warning-as-error -S $R/src/cmake -B $R/build/Release -Donnxruntime_USE_TELEMETRY=OFF` | Passed before and after the real main-parent-to-patch commit transition. | | `cmp $R/controlled-commands-before.txt $R/controlled-commands-after.txt` | Passed: complete Ninja command lists for both targets were identical. | | `g++ $R/get_build_info.cc -I $R/src/include/onnxruntime/core/session -L $R/build/Release -Wl,-rpath,$R/build/Release -lonnxruntime -o $R/get_build_info` | Passed: C API probe linked against the newly built shared library. | | `$R/venv/bin/python $R/validate_identity.py after` | Passed: exact C API/Python runtime equality and actual HEAD, unchanged feature-config content/mtime, updated private header, exactly two recompiled translation units. | | `$R/venv/bin/python $R/check_build_info_matrix.py` | Passed all 5 generation cases: ordinary, every existing flag enabled, INT4 without CUDA, Git unavailable, and minimal-build generation. Each private-header string matched the original template's string. | | `cmake --build $R/build/Release --target onnxruntime_shared_lib_test --parallel 4` | Passed. | | `./onnxruntime_shared_lib_test --gtest_filter=CApiTest.get_build_info_string --gtest_output=xml:$R/get-build-info-test.xml` | Passed: **1 test ran**, 0 failures; XML test count also asserted. | | `$R/venv/bin/git-clang-format --binary $R/venv/bin/clang-format --diff 9f807084bed8431f7a2734116b4c9322c81a6e6a -- onnxruntime/core/session/onnxruntime_c_api.cc onnxruntime/python/onnxruntime_pybind_module.cc orttraining/orttraining/python/orttraining_python_module.cc onnxruntime/contrib_ops/cuda/llm/gemm_tactic_cache.cc` | Passed with repository-pinned clang-format 20.1.8: no changes needed. | | `git diff --check HEAD^ HEAD` | Passed. | The controlled transition used the actual main parent and actual patch commit above. The validation clone's source tree was checked against the real commit; only Git metadata advanced between measurements, with no fabricated stamps or timestamp manipulation. | Measured invariant | Result | | --- | --- | | Global `onnxruntime_config.h` SHA-256 | `7050ac14554db17722928ba7cd404c733e622f16729e06746b45844e51812baf`, unchanged | | Global header `mtime_ns` | `1793205513366988156`, unchanged | | Private header | Updated from the real parent identity to the real patch identity | | Recompiled translation units | Exactly `onnxruntime_c_api.cc` and `onnxruntime_pybind_module.cc`; no unrelated ORT or dependency compilation | | Complete compiler/linker command-list SHA-256 | `02dbebafbe13f39254fca0890cb20e957ab7333c2e4b3a48d3334d7b504c770c`, unchanged | Both newly built runtimes returned exactly: ```text ORT Build Info: git-branch=justinchuby-isolate-build-info, git-commit-id=3bb3243, fp8-kv-cache=1, build type=Release ``` Telemetry was disabled consistently on both sides of the measured transition because its vendored curl independently rewrites `curl_config.h` on configure, even at unchanged HEAD. The initial telemetry-enabled build also passed runtime/header invariants and recompiled only the two expected ORT files, but additionally rebuilt curl's 192 C files. That unrelated dependency churn is not changed by this PR. Training bindings and CUDA tactic-cache include paths were audited through shared target configuration and CUDA plugin/test targets. Full training, CUDA, and minimal-runtime builds were not run; the minimal-build generation case is a CMake string-equivalence check, not a minimal-runtime execution claim. Signed-off-by: Justin Chu <justinchuby@users.noreply.github.com> Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>