immanuelpeter's picture
Link the card to its export and parity check in vision-tower-bench
41247dd verified
|
Raw History Blame Contribute Delete
2.11 kB
---
license: other
license_name: openmdw-1.1
license_link: https://openmdw.ai/license/1-1/
pipeline_tag: image-feature-extraction
library_name: pytorch
tags:
- vision
- feature-extraction
- nemotron
base_model:
- nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16
---
# C-RADIOv4-H-3.5-Bolt
This repository packages the C-RADIOv4-H Tower and Projector from [Nemotron 3.5 Bolt](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16).
## Contents
| File | Tensors | What it holds |
|---|---|---|
| `model.safetensors` | 391 | RADIO Tower from Bolt `vision_model.*`, including `summary_idxs` |
| `projector.safetensors` | 3 | Bolt `mlp1` Projector |
| `config.json` | | Source and code pins and widths |
| `vision.py` | | Loader used by `examples/inference.py` |
## Architecture
| Component | Details |
|---|---|
| Tower | C-RADIOv4-H, 32 layers, 1280 hidden, patch size 16, ten prefix tokens |
| Token compression | InternVL v2 2x2 pixel shuffle, scale 0.5, no learned parameters |
| Projector (`mlp1`) | `RMSNorm(5120)`, `Linear(5120, 20480)` no bias, SquaredReLU, `Linear(20480, 4096)` no bias |
At 448 the bench reports 784 Tower tokens of width 1280, 196 `merged` tokens of
width 5120, and 196 `projected` tokens of width 4096.
## Usage
See [`examples/inference.py`](examples/inference.py) for image feature extraction.
## Validation
The [parity check](https://github.com/immanuel-peter/vision-tower-bench/blob/main/vtb/parity/nemotron35.py) compares every exported tensor with the pinned parent using `torch.equal`, then checks all stages against the parent's own `RadioModel` and `VisionProjector` code.
## Reproduction
The [export script](https://github.com/immanuel-peter/vision-tower-bench/blob/main/vtb/adapters/nemotron35.py) reads `vision_model.*` and `mlp1.*` from the pinned parent and writes the original BF16 tensors.
## Credits
NVIDIA released [Nemotron 3.5 Bolt](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16) and [C-RADIOv4-H](https://huggingface.co/nvidia/C-RADIOv4-H).
## License
[OpenMDW-1.1](LICENSE), the license from the source model card.