fix(vlm): expose _tokenizer on VLMTokenizerWrapper for guided decoding

[creedx12]

Problem

VLMTokenizerWrapper.__init__ (chat/mlx/model_types.py) sets self.tokenizer but not self._tokenizer. OutlinesLogitsProcessor reads tokenizer._tokenizer (chat/mlx/outlines_logits_processor.py:48,58), so any request that uses json_schema / guided decoding against a VLM model raises:

AttributeError: 'VLMTokenizerWrapper' object has no attribute '_tokenizer'

Fix

Alias self._tokenizer = self.tokenizer in __init__ so the Outlines path resolves the underlying tokenizer for VLM models. One-line change (+ comment).

Present on main (0.5.3). Verified against Gemma-4 with a json_schema response format.

Summary by CodeRabbit

• Bug Fixes
  • Fixed a compatibility error in Vision Language Model support that was preventing advanced decoding features from working properly.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 9e1c836a-fc90-4d1b-a1e9-82b8aada2f16

📥 Commits

Reviewing files that changed from the base of the PR and between 4f8e9ef and 3cb8e0e.

📒 Files selected for processing (1)

• src/mlx_omni_server/chat/mlx/model_types.py

---

📝 Walkthrough

Walkthrough

VLMTokenizerWrapper inside load_mlx_model gains a _tokenizer attribute assigned from self.tokenizer. This exposes the attribute expected by downstream logits processing and guided/json-schema decoding components that access tokenizer._tokenizer on VLM models.

Changes

VLM Tokenizer Compatibility Fix

Layer / File(s)	Summary
Expose _tokenizer on VLMTokenizerWrapper src/mlx_omni_server/chat/mlx/model_types.py	Adds self._tokenizer = self.tokenizer with explanatory comments inside VLMTokenizerWrapper so decoding paths that access tokenizer._tokenizer no longer raise AttributeError on VLM models.

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~2 minutes

Poem

A wrapper once bare, now patched with care,
_tokenizer assigned, no error in the air! 🐇
The VLM decodes without a fright,
Six lines of fix make everything right.
Hop hop hooray, the attribute's there! ✨

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The pull request title clearly and specifically describes the main change: exposing the _tokenizer attribute on VLMTokenizerWrapper to fix guided decoding compatibility.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}