haloblocks 0.1.3

Modern, modular, composability-first neural network building blocks for PyTorch.

HaloBlocks

Modern, Modular, and Composability-First Neural Network Components.

---

HaloBlocks is a high-performance Python library for assembling complex neural network architectures from simple, composable building blocks. Whether you are prototyping a new Transformer variant, experimenting with Mixture-of-Experts (MoE) scaling, or building a Vision-Language-Action (VLA) model from scratch, HaloBlocks provides the foundational "bricks" you need — without getting in your way.

[!TIP] New to HaloBlocks? Jump straight into the interactive Tutorial Notebook for a hands-on tour of the library.

---

Features

Feature	Description
First-Class Composability	Every component is a Block that can be freely nested and combined via CompositeBlock
Rich Attention Zoo	MHA, MQA, GQA, Cross, Gated, Sliding-Window, Linear, Trinity Attention & more
MoE Ready	DeepSeek-style Mixture-of-Experts with routed + shared experts and noisy Top-K routing
Positional Embeddings	Sinusoidal, Learned, RoPE (Rotary), and ALiBi — all plug-and-play
VLA Integration	Flow Matching decoder blocks for Vision-Language-Action pipelines
Config-Driven	Build entire model graphs from plain Python dicts or YAML/JSON configs via BlockFactory
Keras-like API	Access any block directly from haloblocks.blocks (registry keys) without touching a config
PyTorch Native	Zero magic — pure nn.Module subclasses built for speed and debuggability

---

Installation

# Using pip
pip install haloblocks

# Using uv (recommended — faster)
uv add haloblocks

Requirements: Python ≥ 3.10, PyTorch ≥ 2.0, NumPy ≥ 2.0

Running tests

From a clone of this repository (at the repo root):

./scripts/run_all_unit_tests.sh
./scripts/run_all_unit_tests.sh -v
./scripts/run_all_unit_tests.sh tests/test_core.py::TestCore::test_composite_block

The script runs uv sync --dev and then pytest. With no arguments it runs the full suite under tests/; any extra arguments are forwarded to pytest.

Releasing a new version

Requires GitHub CLI (gh) authenticated (gh auth login).

./scripts/open_release_pr.sh 0.2.0

This updates version in pyproject.toml (and uv.lock when uv is available), commits on branch release/v0.2.0, pushes it, and opens a PR to main. After the PR is merged, GitHub Actions creates tag v0.2.0 and pushes it; the existing Publish to PyPI workflow runs on that tag.

---

Quick Start

HaloBlocks supports three usage styles — choose whichever fits your workflow.

Style 1 — Keras-like direct access

No config dictionaries needed. Import blocks from haloblocks and call any block by its class name, or import attention classes from the attention package:

from haloblocks import blocks
from haloblocks.blocks import attention

# Multi-Head Attention (registry key = class name)
attn = blocks.MultiHeadAttention(emb_dim=512, num_heads=8)

# Same family, explicit class (namespace style)
mqa = attention.MultiQueryAttention(emb_dim=512, num_heads=8)

# Grouped-Query Attention
gqa = blocks.GroupedQueryAttention(emb_dim=512, num_heads=8, num_kv_heads=2)

# Rotary Positional Embedding (RoPE is per-head; pass head_dim, not full emb_dim)
rope = blocks.RotaryPositionalEmbedding(head_dim=64, max_len=8192)

# DeepSeek Mixture-of-Experts
moe = blocks.DeepseekMoE(
    emb_dim=512,
    hid_dim=2048,
    num_router_exprts=8,
    best_k=2,
    num_shared_exprts=2,
)

Style 2 — hb.create convenience function

import haloblocks as hb

attn = hb.create('MultiHeadAttention', emb_dim=512, num_heads=8)
moe = hb.create(
    'DeepseekMoE',
    emb_dim=512,
    hid_dim=2048,
    num_router_exprts=8,
    best_k=2,
    num_shared_exprts=2,
)

Style 3 — Config-driven (YAML / JSON friendly)

Perfect for experiment configs or hyperparameter sweeps:

import haloblocks as hb

config = {
    'type': 'TransformerBlock',
    'emb_dim': 768,
    'num_heads': 12,
    'mlp_dim': 3072,
    'use_moe': False,
}
block = hb.create(config)

Building a Composite Model

import haloblocks as hb

model = hb.CompositeBlock([
    hb.create('SinusoidalPositionalEmbedding', emb_dim=512, max_len=1024),
    hb.create('TransformerBlock', emb_dim=512, num_heads=8),
    hb.create('TransformerBlock', emb_dim=512, num_heads=8),
])

---

Block Catalogue

Registry keys now match class names exactly. Use them with blocks.<ClassName>, hb.create('<ClassName>', ...), and hb.create({'type': '<ClassName>', ...}).

Every block supports three equivalent ways to instantiate. For example, Multi-Head Attention:

# 1. blocks.<ClassName>
from haloblocks import blocks
attn = blocks.MultiHeadAttention(emb_dim=256, num_heads=8)

# 2. hb.create with keyword args
import haloblocks as hb
attn = hb.create('MultiHeadAttention', emb_dim=256, num_heads=8)

# 3. hb.create with a config dict
attn = hb.create({'type': 'MultiHeadAttention', 'emb_dim': 256, 'num_heads': 8})

Attention

Block	Registry Key	Description
Scaled Dot-Product	ScaledDotProductAttention	Bare-metal scaled dot-product attention
Self-Attention	SelfAttention	Single-head self-attention (QKV from same input)
Attention head (sub-block)	HeadAttention	Single head in a reduced subspace (building block inside MHA)
Multi-Head Attention	MultiHeadAttention	Classic MHA (Vaswani et al.)
Multi-Query Attention	MultiQueryAttention	MQA — single shared KV head
Grouped-Query Attention	GroupedQueryAttention	GQA — configurable KV head groups
Cross-Attention (single-head)	CrossAttention	Single-head encoder–decoder cross-attention
Cross-Attention (multi-head)	MultiHeadCrossAttention	Multi-head cross-attention
Gated Attention	GatedAttention	Attention with gating mechanism
Sliding Window Attention	SlidingWindowAttention	Local context window (Longformer-style)
Linear Attention	LinearAttention	Sub-quadratic linear attention
Multi-Head Latent Attention	MultiHeadLatentAttention	Latent-space compressed attention
Trinity Attention	TrinityAttention	Combined local + global + linear attention

Positional Embeddings

Block	Registry Key	Description
Sinusoidal	SinusoidalPositionalEmbedding	Fixed sinusoidal PE (original Transformer)
Learned	LearnedPositionalEmbedding	Trainable position embeddings
Rotary (RoPE)	RotaryPositionalEmbedding	Rotary position encoding
ALiBi	AlibiPositionalBias	Attention with Linear Biases

MLP

Block	Registry Key	Description
MLP	MLP	Configurable feed-forward block (activations, bias, last-layer options)

Mixture-of-Experts

Block	Registry Key	Description
DeepSeek MoE	DeepseekMoE	Routed + shared experts with noisy Top-K routing

Transformer

Block	Registry Key	Description
Transformer Block Builder	TransformerBlockBuilder	Highly composable layer supporting arbitrary Attn & FFN combinations
Stacked Transformer (config)	StackedTransformerBlock	Same as TransformerBlockBuilder with num_layers; use inside CompositeBlock configs
Transformer Block	TransformerBlock	Pre-norm Transformer layer (attn + MLP)
Decoder	DecoderTransformer	Stacked Transformer decoder

Vision-Language-Action

Block	Registry Key	Description
Flow Decoder	FlowActionDecoder	Flow-matching decoder for VLA action prediction

---

Project Structure

HaloBlocks/
├── src/haloblocks/
│  ├── __init__.py       # Top-level API: Block, BlockFactory, create
│  ├── layers.py        # Alias of ``blocks`` (same module; backward compatible)
│  ├── core/
│  │  ├── block.py      # Base Block (nn.Module subclass)
│  │  ├── registry.py     # BlockRegistry — central name → class map
│  │  ├── factory.py     # BlockFactory.create() dispatcher
│  │  ├── composite.py    # CompositeBlock for sequential pipelines
│  │  └── builder.py     # TransformerBlockBuilder, stacked blocks
│  └── blocks/
│    ├── attention/     # MHA, MQA, GQA, Cross, Gated, Sliding-Window, Linear, Trinity
│    ├── norm/          # RMSNorm (shared by attention and builder)
│    ├── positional_embedding/ # Sinusoidal, Learned, RoPE, ALiBi
│    ├── mlp/        # Configurable MLP block
│    ├── moe/        # DeepSeek Mixture-of-Experts
│    ├── transformer/    # Transformer Block & Decoder
│    └── vla/        # Flow Matching Decoder
├── scripts/
│  ├── run_all_unit_tests.sh  # uv sync --dev && pytest (optional args)
│  ├── open_release_pr.sh     # bump version + PR; merge triggers tag + PyPI
│  └── format.sh             # black, isort, pyflakes
├── notebooks/
│  └── tutorial.ipynb     # Interactive getting-started guide
├── tests/           # Pytest test suite
└── pyproject.toml

---

Contributing

Contributions are welcome! Please read our Contributing Guide before opening a pull request. When adding a new block:

1. Create your module under src/haloblocks/blocks/<category>/.
2. Subclass Block and decorate with @BlockRegistry.register() (auto-uses the class name).
3. Re-export from the category's __init__.py.
4. Add a test in tests/ and a usage example in the tutorial notebook.

---

License

HaloBlocks is released under the MIT License.

---

Built with by Naveen