Description

Loading mlx-community/Qwen3.5-35B-A3B-4bit via VLMModelFactory.shared.loadContainer() works fine — the container loads and weights map correctly. But the first text-only inference call crashes with:

MLX/ErrorHandler.swift:343: Fatal error: SmallVector out of range.
at mlx-c/mlx/c/array.cpp:335

The same model loads and runs fine through LLMModelFactory.

Root Cause

Qwen3VLLanguage.getRopeIndex() (Qwen3VL.swift L1341) assumes inputIds is always 2D [batch, seq]:

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

When the VLM container is used for text-only generation (no images), the token tensor can be 1D [seq]. inputIds.dim(1) on a 1D array triggers the SmallVector fatal error.

The call path is:

1. Qwen35.prepare(input:cache:windowSize:) — no image/video, so pixelValues is nil
2. languageModel(inputIds, ...) — forwards raw token IDs
3. Qwen35Language.LanguageModel.callAsFunction — positionIds == nil, ropeDeltas == nil, so enters the getRopeIndex branch
4. Qwen3VLLanguage.getRopeIndex(inputIds:...) — crashes on inputIds.dim(1) becasue the tensor is 1D

This shows up concretely when WiredMemoryUtils.tune() runs its prefill measurement — it creates a plain LMInput(tokens:) with 1D tokens, calls model.prepare(), and the VLM path crashes.

Reproduction

import MLXLMCommon
import MLXVLM

let config = ModelConfiguration(id: "mlx-community/Qwen3.5-35B-A3B-4bit")
let container = try await VLMModelFactory.shared.loadContainer(configuration: config)

// This crashes:
try await container.perform { context in
    try await WiredMemoryUtils.tune(
        context: context,
        tokenCount: 512,
        parameters: GenerateParameters(maxTokens: 1)
    )
}

Suggested Fix

The simplest fix would be to ensure inputIds is 2D before calling getRopeIndex. Something like:

// In Qwen35Language.LanguageModel.callAsFunction, before calling getRopeIndex:
var inputsForRope = inputs
if inputsForRope.ndim == 1 {
    inputsForRope = inputsForRope.expandedDimensions(axis: 0)
}

Or alternatively, getRopeIndex itself could handle 1D input gracefully.

Environment

• mlx-swift-lm: main branch (HEAD = bc3c20e)
• mlx-swift: main branch (HEAD = b6e128c)
• Model: mlx-community/Qwen3.5-35B-A3B-4bit (model_type: qwen3_5_moe)
• macOS 26.0, Apple M4 Max 96GB