kyegomez/attn_res

A clean, single-file PyTorch implementation of Attention Residuals (Kimi Team, MoonshotAI, 2026), integrated with Grouped Query Attention (GQA), SwiGLU feed-forward networks, and Rotary Position Embeddings (RoPE).

43
/ 100
Emerging

This project provides a clean, single-file implementation of the Attention Residuals mechanism for Transformer models. It allows researchers and AI practitioners to experiment with different ways that information flows between layers in a language model, moving beyond simple additive connections. You input token sequences, and it outputs logits or loss for training, giving you a powerful building block for advanced language model architectures.

Available on PyPI.

Use this if you are a machine learning researcher or engineer working on large language models and want to explore novel architectural components like Attention Residuals to improve model performance or efficiency.

Not ideal if you are looking for a pre-trained, production-ready language model or a high-level API for everyday NLP tasks.

large-language-models transformer-architecture deep-learning-research neural-networks model-training
Maintenance 13 / 25
Adoption 4 / 25
Maturity 18 / 25
Community 8 / 25

How are scores calculated?

Stars

8

Forks

1

Language

Python

License

Apache-2.0

Last pushed

Mar 16, 2026

Commits (30d)

0

Dependencies

1

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/transformers/kyegomez/attn_res"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.