add multimodal preprocessing support

[shx2005]

Purpose

Enable multimodal-aware(currently single image-text pair) preprocessing for the EAGLE offline data-generation pipeline while preserving backward compatibility for existing text-only workflows.

Description

This PR adds multimodal preprocessing support and refactors related preprocessing internals:

• Added --multimodal flag to offline generation entrypoint and passed it through to preprocessing.
• Added helpers for:
  • multimodal content normalization,
  • image reference extraction,
  • singleton batch flattening with clearer error context.
• Generalized assistant-mask/pattern logic to work with either tokenizer or processor callers.
• Kept compatibility for existing callers:
  • is_multimodal=None defaults to text mode.
• Updated tests to new helper/function names and added a regression test for default text fallback behavior.

Related Issue

#290

Tests

• Ran an existing pure-text Qwen example end-to-end to ensure this change does not affect the original text-only path.
• Ran training in my adapted VL code path to verify multimodal preprocessing outputs are consumable and behavior is correct.

Notes:

• The data generation and training workflow in speculators will be updated soon and this PR intentionally focuses on organizing and stabilizing the preprocessing part only.

I have filled in:

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan/results, such as providing test command and pasting the results.
• (Optional) The necessary documentation update.
• I (a human) have written or reviewed the code in this pr to the best of my ability.

[Copilot]

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.

[shanjiaz]

Hi @shx2005 This looks good to me! Can you make sure to rebase off of main once this PR lands? Since this PR refactors preprocessing. Thank you so much for your patience and contribution!

[mergify]

The quality checks have failed. Please run make style and make quality under
the root directory to address the lint failures. You will need to install the
dev optional install to get the required linting packages:
https://github.com/vllm-project/speculators/blob/main/CONTRIBUTING.md

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @shx2005.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[shx2005]

@shanjiaz Thanks so much for your review! I saw that the other PR has already landed, so I’ll update this one based on the latest main.

[shx2005]

I’ve rebased onto the latest PR and also fixed a missing dtype argument in scripts/train.py, which could cause dtype mismatches during matrix multiplication.

I ran tox and verified against the updated train.py in the new PR that this does not affect the current logic.

I haven’t re-tested the image-text pair path yet because the training logic has changed. I’m planning to finish the remaining image-text support over the next few weeks and then test them together.

Thank you again for your patience and for all the guidance throughout this process！