onnxruntime

78c83491 - [CUDA] Follow up INT2 grouped GEMM test coverage and plugin build (PR4a2) (#33045)

[CUDA] Follow up INT2 grouped GEMM test coverage and plugin build (PR4a2) (#33045) ## Summary PR4a2 follows up the five non-blocking comments from the final review of #32963 at `d3d60acf26`. Based on main after #32963 was squash-merged as `a4d55e2b6f`. No production kernel math, alignment contract, or runtime eligibility changes. ## Review Follow-ups - **Coverage** ([comment](https://github.com/microsoft/onnxruntime/pull/32963#discussion_r4158213954)): allow compute capability major 8 in the direct-kernel fixture instead of exactly 8.0. Test parameters continue to select the SM80 implementation. Legacy W4/W8 alignment regressions remain enabled, removing their fixture-level skip on A10/SM86. - **Plugin build footprint** ([comment](https://github.com/microsoft/onnxruntime/pull/32963#discussion_r4158213967)): restrict dense/W4/W8 baselines and disabled benchmarks to legacy tests; add source-specific plugin/adapter definitions. Reduce plugin internal-test contrib implementation sources from 16 to 4, retaining the adaptor needed for INT2 unpacking. Plugin retains INT2 FP16/BF16 numerical, metadata, valid-offset and misalignment coverage. No measured CI wall-time improvement is claimed. - **Diagnostic options** ([comment](https://github.com/microsoft/onnxruntime/pull/32963#discussion_r4158213972)): remove unconditional duplicate suppressions. CUDA 13+ inherits 970/2189 from shared options. CUDA <13 retains target-local suppressions with an explanatory comment because the internal INT2 TU exposes third-party Abseil/Protobuf header diagnostics under warning-as-error. Removing them reproduced a CUDA 12.8 compile failure; the explicit version guard fixes it. - **Rebase** ([comment](https://github.com/microsoft/onnxruntime/pull/32963#discussion_r4158213979)): #32932 is already in the main base. This PR does not replay that fix and has no `moe_quantization.cc` diff. - **Headers** ([comment](https://github.com/microsoft/onnxruntime/pull/32963#discussion_r4158214003)): add the requested Microsoft copyright/MIT license headers to the helper header, implementation CU and test TU. ## Validation All compilation, formatting, CMake checks and GPU tests ran inside Docker `jiafa-dev` on an A100-SXM4-80GB. | Check | Result | | --- | --- | | CUDA 12.8 plugin standalone harness | 16 passed, zero skipped | | CUDA 12.8 legacy standalone harness | 23 enabled passed, zero skipped; 3 benchmarks disabled | | CUDA 13.0 plugin standalone harness | 16 passed, zero skipped | | Fresh legacy/plugin test TU compilation with CUDA 12.8 and 13.0 headers | Passed | | Fresh four plugin implementation CUs with CUDA 12.8 and 13.0, warning-as-error | Passed | | Fresh legacy INT2 implementation CU, CUDA 12.8 | Passed | | Scoped execution of actual CMake source/definition and option blocks | Passed; four contrib sources enabled / zero disabled, exactly one suppression per diagnostic per toolkit | | Symbol audit | Plugin baseline dependencies/benchmarks absent; legacy coverage retained | | Production body comparison | Unchanged apart from requested headers | | clang-format dry run and git diff --check | Passed | ### Limitations - Focused standalone validation, not a full top-level CMake configure/build or provider integration run. Existing Ninja recipes were adapted to compile current sources and new options into fresh objects; actual changed CMake blocks were extracted and executed with capture functions. Full target integration remains for CI. - Plugin harnesses freshly compile the test and all four contrib implementation CUs, reusing unchanged core/static libraries and CUDA support objects. Legacy freshly compiles the test and INT2 CU, reusing unchanged baseline kernel objects and standalone plugin support shims. - Only SM80 hardware was available. SM86/SM89 execution, Windows, full regression and sanitizer runs were not exercised for this follow-up. CUDA 13 legacy test compilation passed; legacy runtime execution used CUDA 12.8. - Exact local commands and compiler/test logs are recorded under `build/pr4a2-review-validation/` on the validation machine; these ignored scratch artifacts are not committed.