Link the card to its export and parity check in vision-tower-bench
Browse files
README.md
CHANGED
|
@@ -44,11 +44,11 @@ See [`examples/inference.py`](examples/inference.py) for image feature extractio
|
|
| 44 |
|
| 45 |
## Validation
|
| 46 |
|
| 47 |
-
The [
|
| 48 |
|
| 49 |
## Reproduction
|
| 50 |
|
| 51 |
-
The [export script](https://github.com/immanuel-peter/vision-tower-bench/blob/main/
|
| 52 |
|
| 53 |
## Credits
|
| 54 |
|
|
|
|
| 44 |
|
| 45 |
## Validation
|
| 46 |
|
| 47 |
+
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.
|
| 48 |
|
| 49 |
## Reproduction
|
| 50 |
|
| 51 |
+
The [export script](https://github.com/immanuel-peter/vision-tower-bench/blob/main/vtb/adapters/nemotron35.py) reads `vision_model.*`, `mlp1.*`, and `vision_projector.vision_final_layernorm.*` from the pinned parent and writes the original BF16 tensors.
|
| 52 |
|
| 53 |
## Credits
|
| 54 |
|