Skip to content
Proyectos

johnma2006/mamba-minimal

Ver en GitHub

johnma2006/mamba-minimal has 2,969 GitHub stars and 218 forks. GitHot measured 0 stars in 24 hours and 0 in 7 days. Last checked 2026-11-15 00:44 UTC. Last push 2024-03-08.

Cómo se miden estas cifras

2,969 estrellas218 Bifurcaciones

Crecimiento de estrellas medido

in 24h
0
in 7 days
0
in 15 days
Recopilando historial
in 1 month
Recopilando historial

Cambios netos en períodos que se solapan, no totales diarios. Los datos no disponibles se dejan en blanco. Las estrellas indican interés, no calidad.

Acerca de

Simple, minimal implementation of the Mamba SSM in one file of PyTorch.

Aprendizaje automático

README.md

Original

mamba-minimal

Simple, minimal implementation of Mamba in one file of PyTorch.

Featuring:

  • Equivalent numerical output as official implementation for both forward and backward pass
  • Simplified, readable, annotated code

Does NOT include:

  • Speed. The official implementation is heavily optimized, and these optimizations are core contributions of the Mamba paper. I kept most implementations simple for readability.
  • Proper parameter initialization (though this could be added without sacrificing readability)

Demo

See demo.ipynb for examples of prompt completions.

from model import Mamba
from transformers import AutoTokenizer

model = Mamba.from_pretrained('state-spaces/mamba-370m')
tokenizer = AutoTokenizer.from_pretrained('EleutherAI/gpt-neox-20b')

generate(model, tokenizer, 'Mamba is the')

Mamba is the world's longest venomous snake with an estimated length of over 150 m. With such a large size and a venomous bite, Mamba kills by stabbing the victim (which is more painful and less effective than a single stab of the bite)

150 meters... 🫢 scary!

References

The Mamba architecture was introduced in Mamba: Linear-Time Sequence Modeling with Selective State Spaces by Albert Gu and Tri Dao.

The official implementation is here: https://github.com/state-spaces/mamba/tree/main