MLX runtime for Muse Glimmer 30B

# muse-glimmer-mlx

MLX (Apple Silicon) port of [**meta-models/Muse-Glimmer-30B**](https://huggingface.co/meta-models/Muse-Glimmer-30B) —
a 30B image-text-to-text VLM: a 52-layer text stack with interleaved sliding/full attention, and a
50-layer vision tower behind a bilinear-resampled position grid.

## Why this exists

`muse_glimmer` is carried by **neither mlx-lm nor mlx-vlm** — not in released mlx-vlm 0.6.10, and
not on either project's `main`. Fourteen MLX conversions of this model were nonetheless published to
the Hub within hours of its release, and none of them can be loaded:

```python
>>> import importlib; importlib.import_module("mlx_vlm.models.muse_glimmer")
ModuleNotFoundError: No module named 'mlx_vlm.models.muse_glimmer'
```

`load()` resolves `mlx_vlm.models.<model_type>`, so every published build is weights without a
runtime. **This repository is that runtime**, and it reads those checkpoints unmodified — see
[Loading existing conversions](#loading-existing-conversions).

## Architecture

What makes this model non-generic, and what a port has to get right:

| feature | detail |
|---|---|
| attention scale | `qk_scale_factor` **3.87** applied to the *query only*, on top of the usual `head_dim ** -0.5` |
| QK norm | weightless RMS norm on Q and K before the extra scale |
| output gating | attention result multiplied by `sigmoid(gate_proj(x))`, computed from the **layer input** |
| RoPE | θ=500000 on the 39 sliding layers; **θ=0 on the 13 full-attention layers, meaning no rotary at all** |
| layers | 52, interleaved 3 sliding (window 2048) : 1 full |
| GQA | 32 query heads : **2** KV heads |
| norms | three variants, two epsilons — see below |
| logits | `logits * output_multiplier`, then a Gemma-style `T * tanh(x / T)` softcap at **T = 20** |
| vision | 50 layers, 37 windowed (448 px) + 13 full, variable-length by `cu_seqlens` |
| projector | `pixel_shuffle` folds 2x2 blocks into channels, so the projector takes **6144**, not 1536 |

### Three norms, two epsilons

| form | math | used by |
|---|---|---|
| `RMSNorm(with_scale=True)` | `x/rms · w` | the final norm, and nothing else |
| `RMSNorm(with_scale=False)` | `x/rms` | the QK norm and the embedding norm — **no parameter at all** |
| `CenteredRMSNorm` | `x/rms · (1 + w)` | all four per-layer norms |

The pre-norms use `rms_norm_eps` (1e-5) and the post-norms `post_norm_eps` (**1e-8**). Subs