Skip to content

tuned-org-uk/tauformer-paper

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

28 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Topological Transformer: A Redesign of the Transformer Architecture

DOI

Starting from prior research on topological spaces for vector search (spectral search with taumode), this paper introduces the definition, a concept implementation, and tests for a new class of Transformers focused on improving the attention mechanism.

The Topological Transformer (Tauformer) preserves the standard attention-based Transformer features (with nanoGPT as the baseline) but redesigns attention at its core to:

  • Deliver domain-specific context directly within the attention mechanism for more domain-relevant token generation.
  • Improve time per token by ~20% and reduce KV-cache memory size by ~50%.
  • Provide a pathway to better performance on larger context windows (longer prompts) and high-dimensional embeddings.

These gains are enabled by substituting the standard inner-product attention kernel with taumode’s synthetic index–based distance, aiming for meaningful linear improvements in training and generation compared to current GPT-style models.

Paper: https://github.com/tuned-org-uk/tauformer-paper/blob/main/Topological%20Transformer-uploaded%20version.pdf

Pre-print: https://www.techrxiv.org/users/685780/articles/1375955-topological-transformer-a-redesign-for-domain-memory-and-cheaper-kernel-operations

Code: https://github.com/tuned-org-uk/tauformer

Bench: https://github.com/tuned-org-uk/taugpt-kvcache-bench

Copyright 2025-2026 Lorenzo Moriondo

About

Topological Transformer paper

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages