fix: add missing return in load_model_with_fallback

[mikepixelmagic-dev]

Summary

load_model_with_fallback() in utils/tokenizer.py is missing a return statement on the happy path. When mlx_lm.load() succeeds without raising a ValueError, the function falls through the try block and implicitly returns None.

This causes a TypeError in MLXLanguageModel.load():

self.model, self.tokenizer = load_model_with_fallback(...)
TypeError: cannot unpack non-iterable NoneType object

The fix is a single line — adding return model, tokenizer after the successful load() call.

How to reproduce

import asyncio
from vllm_mlx.models.llm import MLXLanguageModel

async def test():
    m = MLXLanguageModel('mlx-community/Qwen3.5-35B-A3B-4bit', trust_remote_code=True)
    m.load()  # TypeError: cannot unpack non-iterable NoneType object

asyncio.run(test())

Note: the bug only manifests when called inside an async context (e.g., asyncio.run()), which is how the server starts the engine via SimpleEngine.start(). Direct calls outside async may work due to Python caching behaviour.

Related issues

Fixes #252, #249

Test

Verified with Qwen3.5-35B-A3B (4-bit and 8-bit) on Apple M5 Max. Server starts and serves requests successfully after the fix.

[mikepixelmagic-dev]

Closing as duplicate of #243 which has the same fix. 👍 on getting it merged.