| gte-encoder-neuron-kernels -- NKI kernels for NewBERT (gte) encoders on AWS Trainium2 | |
| Licensed under the Apache License, Version 2.0 (see LICENSE). | |
| All kernels in build/torch-neuron/ are new code. They use the following ideas and components: | |
| * Lane-permuted Q/K projection layout, so rotate-half RoPE is elementwise with no transposes: | |
| idea from the V-JEPA2 encoder kernels published at huggingface.co/kernels/jburtoft/vjepa2-neuron-kernels | |
| (Apache-2.0, developed by Antonio Mena), as generalised in | |
| huggingface.co/kernels/jburtoft/vit-encoder-neuron-kernels (Apache-2.0). | |
| * Long-sequence attention (L >= 768) calls attention_cte from the AWS Neuron NKI library | |
| (nkilib, https://github.com/aws-neuron/nki-library), Copyright Amazon.com, Inc. or its affiliates, | |
| Apache-2.0. nkilib is not redistributed here; it is imported at run time from the Neuron environment. | |
| * Target architecture: the NewBERT ("new-impl") modeling code by Alibaba-NLP used by the gte model | |
| family on the Hugging Face Hub (Apache-2.0). No modeling code is redistributed. | |