File size: 768 Bytes
e5e4742
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
The Vela decision readout in converter.py adapts the mathematical forward pass in
vllm-sr/vela3-fast/vela3_encoder.py (Apache-2.0):
https://huggingface.co/vllm-sr/vela3-fast/blob/fa60e3aa41132c60bf05a84cabfdced39cb6d4f7/vela3_encoder.py

Changes: tensor inputs for marker positions and pooling, fixed-shape export,
and no upstream batching or task dispatch. The upstream license is reproduced
in licenses/vllm-sr-Apache-2.0.txt. Model weights are not included in this package.

The standalone host adapter also adapts rendering from vela3_encoder.py and ub_d3_common.py at the same upstream commit. Changes: tokenizers/NumPy input packing, explicit bounds and unsupported-task rejection, Core AI inference, and Python/Swift examples. No model training or fine-tuning.