Fix WhisperProcessor divide-by-zero when single prompt is provided

[Copilot]

WhisperProcessor::Process unconditionally called tokenizer.EncodeBatch(payload.prompts), ignoring payload.prompt. When invoked via the single-string overload of MultiModalProcessor::Process, payload.prompts is an empty span, causing a divide-by-zero in EncodeBatch (encoded.size() / strings.size()).

Changes

• src/models/whisper_processor.cpp: Replace the unconditional EncodeBatch(payload.prompts) call with a priority check:
  1. payload.prompts non-empty → use as-is (batch path)
  2. payload.prompt non-empty → wrap in a single-element const char* span and pass to EncodeBatch (single-prompt path)
  3. Neither set → throw a descriptive std::runtime_error

std::shared_ptr<Tensor> input_ids;
if (!payload.prompts.empty()) {
  input_ids = tokenizer.EncodeBatch(payload.prompts);
} else if (!payload.prompt.empty()) {
  const char* prompt_ptr = payload.prompt.c_str();
  input_ids = tokenizer.EncodeBatch(std::span<const char*>(&prompt_ptr, 1));
} else {
  throw std::runtime_error("No prompt or prompts provided to WhisperProcessor::Process.");
}

• test/python/test_onnxruntime_genai_api.py: Added test_audio_preprocessing_single_prompt to validate the single-prompt code path. The test passes a plain string (not a list) to the processor with a single audio file, which exercises the payload.prompt branch that previously caused the divide-by-zero.

[skottmckay]

@copilot all changes should have tests to validate. please add.

[Copilot]

@copilot all changes should have tests to validate. please add.

Added test_audio_preprocessing_single_prompt in test/python/test_onnxruntime_genai_api.py (commit 43256d3). It uses the same audio-preprocessing model and 1272-141231-0002.mp3 audio file as the existing tests, but passes a single string prompt (not a list) to the processor — which exercises the fixed payload.prompt code path that previously caused a divide-by-zero.

[Copilot]

Pull request overview

Fixes a crash in the audio preprocessing path for Whisper models when MultiModalProcessor is invoked with a single prompt string. The core issue was WhisperProcessor::Process always calling Tokenizer::EncodeBatch with an empty prompt span, triggering a divide-by-zero in the tokenizer’s batching logic.

Changes:

• Update WhisperProcessor::Process to prioritize payload.prompts (batch) and fall back to payload.prompt (single prompt), otherwise throw a clear std::runtime_error.
• Add a Python regression test that calls the multimodal processor with a single prompt string and a single audio input.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/models/whisper_processor.cpp	Avoids passing an empty prompt span into EncodeBatch by correctly handling the single-prompt payload field and emitting a descriptive error for missing prompts.
test/python/test_onnxruntime_genai_api.py	Adds coverage for the single-string prompt call path to prevent regressions of the divide-by-zero crash.