onnxruntime

5192b907 - Validate token_id bounds in NGramRepeatBlock to prevent OOB write (#28039)

Validate token_id bounds in NGramRepeatBlock to prevent OOB write (#28039) ### Description In `NGramRepeatBlock` (CPU and CUDA EP), token values from the `input_ids` tensor are used directly as array indices into the `scores` output buffer without adequate bounds checking. The CPU path only checked `token_id < vocab_size` (missing lower bound), and the CUDA kernel had no bounds checks at all. A crafted model with negative token IDs can write at attacker-controlled negative offsets, causing heap corruption or SIGSEGV. Fixes https://portal.microsofticm.com/imp/v5/incidents/details/31000000558069/summary ### Changes - **ngram_repeat_block.h** (CPU): Replace `ORT_ENFORCE(token_id < vocab_size)` with full `[0, vocab_size)` bounds check returning `INVALID_ARGUMENT` Status via atomic error flag (avoids `abort()` under `ORT_NO_EXCEPTIONS`) - **ngram_repeat_block_impl.cu** (CUDA): Add `CUDA_KERNEL_ASSERT` + bounds guard with skip for release safety ### Tests 2 regression tests: negative token_id, token_id >= vocab_size (CPU only, CUDA EP excluded to avoid debug assert context corruption). --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>