mambax 0.4.0

Optimized Mamba implementation with chunk processing and ONNX export

MambaX

PyTorch implementation of the Mamba architecture with enhanced production-ready features:

1. ONNX Export - Full model export support for deployment
2. Chunk Processing - Single-forwardpass chunk handling (no token loops)
3. CPU-First - Optimized execution without CUDA dependencies

Installation

Install the package directly from PyPI:

pip install mambax

Key Advantages

• Production Ready: ONNX-compatible for serving
• No Token Loops: Processes entire chunks in single forward pass
• Hardware Agnostic: Runs equally well on CPU/GPU

Acknowledgements

Builds upon reference work from alxndrTL/mamba.py

Usage

1. Standard Forward Pass

import torch
import torch.nn as nn
from mambax import Mamba

# Initialize model
model = Mamba(
    d_model=512,
    d_inner=1024,
    d_conv=4,
    d_state=16,
    dt_rank=64,
    use_cuda=False
)

# Process full sequence
x = torch.rand(1, 128, 512)  # (batch, seq_len, dim)
output = model(x)  # single forward pass

2. Single Token Processing

x_token = torch.rand(1, 1, 512)  # (batch, 1, dim)
output, new_state, new_conv = model(x_token, state_cache, conv_cache)

3. Chunk-Based Processing

# Initialize with empty caches
state_cache = torch.zeros(1, 1024, 16)  # (batch, d_inner, d_state)
conv_cache = torch.zeros(1, 1024, 3)     # (batch, d_inner, d_conv-1)

# Process chunks (e.g. 8 tokens at once)
x_chunk = torch.rand(1, 8, 512)
output, new_state, new_conv = model(x_chunk, state_cache, conv_cache)