onnxruntime

9c6b7a48 - Merge ConstantOfShape and Cast in clean_graph optimization (#27482)

Merge ConstantOfShape and Cast in clean_graph optimization (#27482) ## Summary - Resolve the TODO in `BertOnnxModel.clean_graph()` by folding the `Cast` node into `ConstantOfShape` - Update the `value` attribute dtype so `ConstantOfShape` directly produces the target type - Add first dedicated test suite for `clean_graph` with 3 tests ## Motivation Fixes #27481 In the attention mask shape optimization path, `clean_graph()` simplifies a `Concat → Unsqueeze → Gather` chain to a direct `Shape → ConstantOfShape` connection, but leaves a redundant `ConstantOfShape(float) → Cast(to=int64)` sequence. The existing TODO (line 271) noted this should be merged. ## Changes - **onnx_model_bert.py**: After the Concat path simplification, update `ConstantOfShape`'s `value` attribute to the Cast target dtype using `helper.tensor_dtype_to_np_dtype`, then redirect consumers from the Cast output to the ConstantOfShape output. The Cast node is collected for removal. (+2 imports: `numpy`, `numpy_helper`) - **test_clean_graph.py** (new): 3 tests covering the merge (Cast removed, ConstantOfShape produces int64), fill value preservation (0 → int64(0)), and Concat path simplification (Concat/Gather pruned). ## Test Plan - [x] 3 new tests pass: `python -m unittest test_clean_graph.TestCleanGraph -v` - [x] Existing `test_attention_fusion` tests pass - [x] `ruff check` clean on both files