Validate whisper beginning_timestamp_token_id bounds

[apsonawane]

This pull request improves validation for the Whisper model's beginning_timestamp_token_id parameter in both runtime and test code. The main focus is to ensure that invalid values for this parameter are caught early, preventing out-of-range errors during inference and processing.

Validation improvements for Whisper timestamp token ID:

• Added a check in BeamSearchParameters::SetSubgraphParameters to enforce that beginning_timestamp_token_id is within the valid range [0, vocab_size) for Whisper models, throwing an error if not.
• Added a similar runtime check in TimestampLogitsProcessor to ensure beginning_timestamp_token_id_ is within bounds before processing logits.

Unit test enhancements:

• Introduced new tests in beam_search_test.cc to verify that SetSubgraphParameters correctly rejects negative or out-of-range beginning_timestamp_token_id values and accepts valid ones.

[Copilot]

Pull request overview

This PR adds early validation for Whisper’s beginning_timestamp_token_id to prevent out-of-range indexing during timestamp logits processing, and extends unit tests to cover invalid/valid values.

Changes:

• Add bounds checks for beginning_timestamp_token_id during subgraph-parameter setup for Whisper generation.
• Add a runtime guard in TimestampLogitsProcessor to fail fast if beginning_timestamp_token_id is outside the logits width.
• Add unit tests to ensure negative/out-of-range values are rejected.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
onnxruntime/test/contrib_ops/beam_search_test.cc	Adds new unit tests covering invalid/valid Whisper beginning_timestamp_token_id values.
onnxruntime/contrib_ops/cpu/transformers/logits_processor.h	Adds runtime enforcement that beginning_timestamp_token_id is within the logits vector bounds.
onnxruntime/contrib_ops/cpu/transformers/beam_search_parameters.cc	Adds Whisper-specific validation of beginning_timestamp_token_id during SetSubgraphParameters.

Suppressed comments (2)

onnxruntime/test/contrib_ops/beam_search_test.cc:67

• Like the previous Whisper validation test, initialize parameters.vocab_size explicitly to avoid using an indeterminate value in SetSubgraphParameters's vocab_size-override logic.

TEST(BeamSearchParametersTest, SetSubgraphParametersRejectsWhisperBeginningTimestampTokenIdEqualToVocabSize) {
  contrib::transformers::BeamSearchParameters parameters;
  parameters.model_type = contrib::transformers::IGenerationParameters::kModelTypeWhisper;
  parameters.logits_processor = contrib::transformers::IGenerationParameters::kLogitsProcessorTypeWhisper;
  parameters.beginning_timestamp_token_id = 128;

  EXPECT_THROW(parameters.SetSubgraphParameters(128, 2, 4, 6), OnnxRuntimeException);

onnxruntime/test/contrib_ops/beam_search_test.cc:78

• beginning_timestamp_token_id = 0 is accepted by the test, but the Whisper timestamp logits processor assumes beginning_timestamp_token_id > 0 (it computes a max over the non-timestamp prefix [0, beginning_timestamp_token_id)). Use a positive in-range id here, and also set parameters.vocab_size explicitly to avoid indeterminate default state.

TEST(BeamSearchParametersTest, SetSubgraphParametersAcceptsValidWhisperBeginningTimestampTokenId) {
  contrib::transformers::BeamSearchParameters parameters;
  parameters.model_type = contrib::transformers::IGenerationParameters::kModelTypeWhisper;
  parameters.logits_processor = contrib::transformers::IGenerationParameters::kLogitsProcessorTypeWhisper;
  parameters.beginning_timestamp_token_id = 0;

  EXPECT_NO_THROW(parameters.SetSubgraphParameters(128, 2, 4, 6));
  EXPECT_EQ(parameters.vocab_size, 128);
}