feat: add initial prompt (prompt_ids) support for Whisper generation

[jhlee111]

Closes #923
Closes #1028

Summary

Add prompt_ids support to WhisperForConditionalGeneration.generate(), enabling initial prompt conditioning for Whisper transcription. This is a long-requested feature (both issues are from 2024) that matches the behavior of the Python transformers library.

What this does

When prompt_ids (an array of token IDs, typically starting with <|startofprev|>) is provided via generation config, it is prepended to init_tokens following the Whisper training format:

[<|startofprev|>, ...prompt_text..., <|startoftranscript|>, <|lang|>, <|task|>, ...]

After generation, the prompt tokens are stripped from the output sequences so they don't appear in the transcription results.

Usage

// Encode prompt (e.g., domain-specific terms)
const prompt_ids = tokenizer.encode("<|startofprev|> " + text, { add_special_tokens: false });

// Pass to pipeline or model.generate()
const output = await model.generate({
  inputs: features,
  prompt_ids,
  language: "en",
  return_timestamps: true,
});

Changes

• packages/transformers/src/models/whisper/modeling_whisper.js — 1 file, ~20 lines added
  • Prepend prompt_ids to init_tokens when provided
  • Strip prompt tokens from output sequences after generation

No breaking changes — when prompt_ids is not provided, behavior is identical to before.

Checklist

• Build passes (pnpm build)
• Prettier formatting passes (pnpm format:check)
• No breaking changes to existing functionality
• Tests — no existing Whisper generation tests found in the repo; happy to add if desired