Modelpublic

immanuelpeter/Muse-Glimmer-Vision

sourceHugging Faceapache-2.0updated 23d agoView on Hugging Face

2likes134downloads

Model Card

Muse Glimmer Vision

This repository packages the Tower and Projector from Muse-Glimmer-30B.

File	Tensors	What it holds
`model.safetensors`	806	Muse Glimmer Tower
`projector.safetensors`	3	Two vision-adapter matrices and the language-width projection
`projector_config.json`, `projector.py`		Projector shapes and loader
`config.json`, `preprocessor_config.json`		Vision-only model and image-processing configuration

Architecture

Component	Details
Tower	50 layers, 1536 hidden, 16 heads, 8960 intermediate, patch size 14
Token compression	2x2 pixel shuffle, no learned parameters
Projector	`Linear(6144, 4096)`, GELU, `Linear(4096, 4096)`, GELU, `Linear(4096, 6656)`, scale-free `RMSNorm(6656)`

All three Projector linear layers omit bias. The native MuseGlimmerVisionModel returns the merged 6144-wide tokens; projector.py maps them to the 6656-wide language space.

Usage

See `examples/inference.py` for image feature extraction.

Validation

The release tests compare all 806 Tower tensors and three Projector tensors with the pinned parent checkpoint using torch.equal. Fixed-image merged and projected outputs also match the parent implementation bit-for-bit on CPU and in BF16 on an NVIDIA A100.

Reproduction

The export script reads the Tower and Projector ranges from both meta-models/Muse-Glimmer-30B shards. It removes the parent prefixes, preserves the original BF16 weights, and extracts the image section of the parent processor configuration. The script pins the parent revision.

Credits

Meta released the Muse Glimmer weights and the native Transformers implementation.

License

Apache License 2.0, the same license as the source model. The source usage policy is included.