fix: add missing return in load_model_with_fallback

[sjswerdloff]

When mlx_lm.load() succeeds without raising ValueError, the function falls through without returning the (model, tokenizer) tuple, causing a TypeError in the caller ('cannot unpack non-iterable NoneType').

This affects any model that loads successfully on the first try (the common case). The ValueError catch paths all have returns, but the success path was missing one.

[waybarrios]

This is a duplicate of #237. The same issue applies here, line 54 already has return load(model_name, tokenizer_config=tokenizer_config) and all paths return or raise correctly:

line 54: return load(model_name, tokenizer_config=tokenizer_config)   # success
line 59: return _load_with_tokenizer_fallback(model_name)             # TokenizersBackend fallback
line 68: return _load_strict_false(model_name, tokenizer_config)      # extra weights fallback
line 69: raise                                                        # unhandled ValueError

return load(...) has been on main since the original commit (90d6dc3), so the function already handles all cases correctly. Closing as duplicate.

[Thump604]

Worth noting that #235 and #237 are independent submissions of the same one-line fix. Three people hit this bug and submitted PRs within 3 days of each other because nothing was merged.

If review bandwidth is the bottleneck, I am happy to help triage incoming PRs or act as a community reviewer. I have production experience with the codebase (122B on M2 Ultra, continuous batching, MTP) and already review most new submissions. Let me know if that would be useful.

[waybarrios]

sounds good @Thump604

[Thump604]

One-line fix, correct. Merging.