alxndrTL/mamba.py

Health v2 · maintenance only

Adoption not part of the score

1478 stars · 130 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

A simple, readable pure-PyTorch (plus MLX) implementation of the Mamba state-space model architecture with a parallel scan for efficient training. It also includes Jamba, Vision Mamba, and muP variants, and has been integrated into Hugging Face Transformers.

Use cases

• implement mamba in pytorch
• train a mamba model from scratch
• run mamba inference on a mac with mlx
• understand how the mamba selective scan works
• try the jamba architecture mixing mamba and attention
• transfer hyperparameters from small to large models with muP

When to choose

• you want readable, hackable Mamba code for learning or research
• you train on PyTorch without the official CUDA kernels or on Apple Silicon with MLX
• you need Jamba or Vision Mamba variants in a simple codebase

When to avoid

• you need maximum training throughput on NVIDIA GPUs - the official CUDA implementation is faster
• you want a production-hardened, feature-complete library rather than an educational implementation

Facets

library · maturity active

machine-learning deep-learning llm-training llm-inference deep-learning machine-learning large-language-models python mamba state-space-models pytorch mlx parallel-scan jamba educational natural-language-processing macos gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="alxndrTL/mamba.py")