[Bug] FastVLM-0.5B outputs raw <start_of_*> tokens on macOS (M4) — text and vision both broken

[Melaga]

Environment

• Device: Mac Mini M4
• OS: macOS 26.2 (Apple Silicon / arm64)
• Flutter Gemma version: 0.15.0
• Flutter version: 3.41.6 (stable)
• Dart version: 3.11.4
• Model: FastVLM-0.5B.litertlm

Bug Description

When using FastVLM-0.5B on macOS, both text-only and image+text inputs produce garbled output consisting entirely of raw <start_of_*> tokens instead of readable text.

Steps to Reproduce

1. Download FastVLM-0.5B.litertlm from HuggingFace
2. Run the example app on macOS (M4)
3. Send a text message (e.g. "hi") — broken output
4. Send a message with an image attached — also broken output

Actual Output

<start_of_9!!!<start_of_something!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

Expected Output

A normal text response describing the input.

Key Comparison

✅ Gemma 4 E2B (vision + text) works correctly on the same machine and setup
❌ FastVLM-0.5B produces garbage tokens for both text-only and image+text

This suggests the issue is FastVLM-specific, not a general macOS vision limitation.

Additional Notes

The model downloads and loads successfully (HTTP 200, 100% progress). The failure is purely at inference/decoding time.

[DenisovAV]

Reproduced — known upstream regression: google-ai-edge/LiteRT-LM#1829