Failure with non-default group sizes on quantized models.

[aaronn]

Running the following to quantize models using a group size of 32 causes a number of error when running the model:

mlx_lm.convert --hf-path microsoft/phi-4 --quantize --q-bits 2 --q-group-size 32

Attempting to run this with MLXLLM.loadModelContainer will cause the following error:

Error downloading model: mismatchedSize(key: "biases", expectedShape: [100352, 80], actualShape: [100352, 160])

Steps to reproduce:

1. Convert a model using a group size of 32 (or alternatively, here: https://huggingface.co/mlx-community/phi-4-2bit).
2. Load it using MLXLLM.loadModelContainer

Expected Result: The model should load.
Actual result: You'll get an error, instead of it loading.

[awni]

I just tried running it and it ran fine for me on the main branch of mlx-swift-examples. Are you using an outdated version?