haloblocks 0.1.0

A modular neural network component library

HaloBlocks

A modern, high-performance Python library for modular neural network components. In HaloBlocks, every model component is treated as a first-class, composable "block"—from specialized attention layers and heads to advanced Mixture-of-Experts (MoE) modules.

Features

• Modular Architecture: Build complex models by composing small, reusable blocks.
• Mixture of Experts: Built-in support for Deepseek-style MoE with routed and shared experts.
• VLA Integration: Includes specialized blocks for Vision-Language-Action models, such as Flow Matching action decoders.
• Aesthetic First: Designed for clarity, performance, and ease of use.

Installation

pip install haloblocks

Quick Start

from haloblocks.core.factory import BlockFactory

# Create a Multi-Head Attention block
config = {
    'type': 'multi_head_attn',
    'emb_dim': 256,
    'num_heads': 8
}
attn_block = BlockFactory.create(config)