Describe the bug
Using the latest commit of mlx-swift-lm and mlx-swift on main branch, MLXVLM generates only 1 token and stops. (EOS)

To Reproduce

Run MLXChatExample from mlx-swift-examples and use qwen2.5VL:3b with an image, you will see a response of only 1 token (The or It or To for example)

Where as with the latest release (2.29.2) of mlx-swift-lm MLXChatExample works fine.