Fix Qwen35 VLM crash on text-only inference

[dirvine]

Problem

Loading a Qwen3.5 MoE model (e.g. mlx-community/Qwen3.5-35B-A3B-4bit) via VLMModelFactory works fine — the containr loads successfully. But the first text-only inference call crashes with:

Fatal error: SmallVector out of range.
at mlx-c/mlx/c/array.cpp:335

The same model loads and runs without issue through LLMModelFactory.

Root Cause

Qwen35Language.LanguageModel.callAsFunction passes inputs directly to Qwen3VLLanguage.getRopeIndex(), which does:

let (batchSize, seqLength) = (inputIds.dim(0), inputIds.dim(1))

When the VLM container is used for text-only generation (no images), callers like WiredMemoryUtils.tune() and TokenIterator pass 1D [seq] token tensors. dim(1) on a 1D array is out of range → fatal crash.

The same issue affects inputs.dim(1) at lines 946 and 966 in the same function.

Fix

Add an ndim check at the top of callAsFunction to expand 1D inputs to 2D [1, seq] before any dimension-dependent logic runs. This matches what the LLM path does via tokens[text: .newAxis] in prefillOnly.

Testing

Verified locally with mlx-community/Qwen3.5-35B-A3B-4bit on M4 Max 96GB:

• Container loads via VLMModelFactory ✅
• WiredMemoryUtils.tune() completes (was crashing) ✅
• Text inference produces correct output ✅
• Full e2e test suite passes ✅

Fixes #148

[davidkoski]

I wonder if that is incorrect (as in this is the bug). I would expect the shape to always be [B, S].

So:

• are these paths buggy in that they allow 1d arrays?
• if not, perhaps the callers should ensure the correct shape -- otherwise this is a one-off fix for a single model

[davidkoski]

Please look at my comment on the shape -- let me know what you think. I suspect it is a caller issue, but not 100% sure.