onnxruntime

1b64fba3 - Handle empty initializers gracefully in optimizer passes (#27976)

Handle empty initializers gracefully in optimizer passes (#27976) ## Overview Hardens several graph optimizer passes against zero-element (empty) and other unexpected-shape constant initializers by validating initializer element counts before dereferencing `data<T>()`. Prevents out-of-bounds reads / crashes during optimization when a pass that assumes a scalar input encounters a 0-element or multi-element tensor. ## Changes ### Source | File | Description | | ---- | ----------- | | `onnxruntime/core/optimizer/div_mul_fusion.cc` | Require `size() == 1` for the div constant. | | `onnxruntime/core/optimizer/layer_norm_fusion.cc` | Require `epsilon` initializer `size() == 1`; fall back to the default epsilon otherwise (both `LayerNormFusion` and `SimplifiedLayerNormFusion`). | | `onnxruntime/core/optimizer/dropout_elimination.cc` | Require `ratio.size() == 1` before reading it as a scalar. | | `onnxruntime/core/optimizer/double_qdq_pairs_remover.cc` | Require `size() == 1` for zero-point / scale in `FindNewZeroPointAndScale`; `ApplyNewInputValue` now uses `ORT_ENFORCE` to make the pre-validated invariant loud. | | `onnxruntime/core/optimizer/utils.cc` | `GetClipConstantMinMax` now requires `size() == 1` for Clip min/max. | | `onnxruntime/core/optimizer/noop_elimination.cc` | A zero-element initializer is no longer treated as a proven identity value; the node is preserved. | Rationale: these operator inputs are defined by ONNX as scalar / 1-element tensors. Previously the passes only checked for `size() > 0` (or not at all) and would silently read `data<T>()[0]` of a multi-element initializer, changing semantics. Requiring exactly one element makes the optimizer skip fusion / elimination for malformed or unexpected shapes instead. ### Tests - New regression tests covering zero-element initializers: - `GraphTransformationTests.DivMulFusion_ZeroElementInitializer` / `DivMulFusion_MultiElementInitializer` - `GraphTransformationTests.NoopElimination_ZeroElementInitializer` / `NoopElimination_ZeroElementInitializer_InternalNode` - `GraphTransformationTests.LayerNormFusion_ZeroElementEpsilon` / `SimplifiedLayerNormFusion_ZeroElementEpsilon` - The LayerNorm / SimplifiedLayerNorm tests assert that fusion fired and that the fused node carries `epsilon == DEFAULT_LAYERNORM_EPSILON (1e-5f)`. - The DivMulFusion test asserts that both `Div` and `Mul` are preserved after the pass when fusion is correctly skipped. - Tests use `ASSERT_STATUS_OK(Model::Load(...))` so load failures are reported, not silently skipped. - Zero-element initializers are intentionally not declared as graph inputs, so they remain constant initializers and the guarded code paths are actually exercised (ORT treats initializers appearing in graph inputs as overrideable). ### CI infrastructure (test filter) - `onnxruntime/test/testdata/onnx_backend_test_series_filters.jsonc` adds `^test_reduce_max_empty_set_cuda` and `^test_reduce_min_empty_set_cuda` to the CUDA exclusion list. This is **unrelated to the optimizer hardening** but is required to unblock the Windows GPU CUDA CI Pipeline on this PR. Every sibling empty-set Reduce test (`reduce_l1/l2/log_sum/log_sum_exp/prod/sum/sum_square_empty_set_cuda`) is already excluded; ReduceMax/ReduceMin were missed. The CUDA `PrepareForReduce` (`reduction_ops.cc:304`) rejects shapes with a zero-sized reduction axis when `keepdims=false` instead of producing the spec-defined identity (`-inf` / `+inf`); the proper fix is a larger change in the CUDA reduction op and can be split into a follow-up. ## Validation - Built `onnxruntime_test_all` (RelWithDebInfo, Windows). - 11 PR-related tests pass (7 new optimizer tests + 4 existing `DoubleQDQ*` regressions). - No regressions: 278 `GraphTransformationTests` and 134 `QDQTransformerTests` all pass. --------- Co-authored-by: Gopalakrishnan Nallasamy <gopalakrishnan.nallasamy@microsoft.com>