zarvan 0.2.2

Zarvan: A Hybrid MoE Architecture for Advanced Sequence Modeling.

Zarvan: A Hybrid MoE Architecture for Advanced Sequence Modeling

Zarvan is an advanced neural network architecture designed to overcome the fundamental limitations of Transformers and RNNs. By unifying the strengths of parallel processing and stateful reasoning, Zarvan provides a powerful, scalable solution for the next generation of sequence modeling challenges.

This library is built on pure PyTorch, offering a lightweight, independent, and high-performance implementation of the Zarvan architecture.

---

🚀 Key Features

• Hybrid Mixture-of-Experts (MoE) Architecture: Employs an intelligent MoE system that dynamically chooses between three "experts" to process a sequence: two for global pattern recognition and a dedicated state machine for step-by-step reasoning.
• Linear Time Complexity ($O(S)$): By replacing the quadratic ($O(S^2)$) self-attention mechanism, Zarvan is significantly more efficient and ideal for processing ultra-long sequences.
• 🧠 Stateful Sequential Reasoning: Features the Sequential Extractor, a deterministic state machine that maintains a perfect, non-decaying memory of the sequence history, enabling it to solve complex, path-dependent tasks where Transformers fail.
• ⚡ Lightweight & Independent: Built on pure PyTorch with zero external dependencies beyond torch, ensuring easy integration, maximum flexibility, and no version conflicts.

---

🏛️ Architecture Overview

The core of Zarvan is a stack of identical blocks. Each block is a Mixture-of-Experts model that dynamically combines the outputs of three specialist modules via a learned gating network.

1. Holistic Extractor: Captures the "gist" or overall summary of the sequence.
2. Associative Extractor: Acts as a "focused memory" retriever for salient, sparse information.
3. Sequential Extractor (The State Machine): Functions as a parallelized state machine that tracks the sequence history losslessly using gated accumulation and phase representation.

An Expert Gate then learns to weigh the outputs of these three modules for each token, allowing the model to adapt its strategy based on the input.

---

🚀 Installation

Install the package directly from PyPI:

pip install zarvan

Or after cloning the repository locally:

git clone [https://github.com/systbs/zarvan-torch.git](https://github.com/systbs/zarvan-torch.git)
cd zarvan-torch
pip install .

✨ Quick Start

Using the independent zarvan library is clean and simple.

import torch
from zarvan import Config, Zarvan

# --- Step 1: Create and Save Each Component Independently ---

# 1a. The Backbone
backbone_config = Config.Backbone(
    embed_dim=256,
    hidden_dim=1024,
    num_layers=6,
    num_heads=4,
)
backbone = Zarvan.Backbone(backbone_config)
backbone_save_dir = "./saved-zarvan-backbone"
backbone.save(backbone_save_dir)
print(f"Backbone saved to '{backbone_save_dir}'.")

# 1b. The Text Modality Head
text_modality_config = Config.Text(
    embed_dim=256, # Must match backbone's embed_dim
    vocab_size=30522,
    max_len=128
)
text_modality_head = Zarvan.Modality.Text(text_modality_config)
text_head_save_dir = "./saved-zarvan-text"
text_modality_head.save(text_head_save_dir)
print(f"Text Modality Head saved to '{text_head_save_dir}'.")

# 1c. The Classification Task Head (Finisher)
classification_config = Config.Classification(
    embed_dim=256, # Must match backbone's embed_dim
    num_classes=10
)
classification_task_head = Zarvan.Task.Classification(classification_config)
task_head_save_dir = "./saved-zarvan-classification"
classification_task_head.save(task_head_save_dir)
print(f"Classification Task Head saved to '{task_head_save_dir}'.")
print("-" * 50)


# --- Step 2: Load and Combine Components for Inference ---

# Load each independent component
loaded_backbone = Zarvan.Backbone.load(backbone_save_dir)
loaded_text_head = Zarvan.Modality.Text.load(text_head_save_dir)
loaded_task_head = Zarvan.Task.Classification.load(task_head_save_dir)

# Set all components to evaluation mode
loaded_backbone.eval()
loaded_text_head.eval()
loaded_task_head.eval()

print("All components loaded successfully.")
print("-" * 50)


# --- Step 3: Perform a Forward Pass ---

# Create dummy input data
input_ids = torch.randint(0, text_modality_config.vocab_size, (2, 50)) # (Batch, SeqLen)

# The forward pass is a chain of the three components
with torch.no_grad():
    # 1. Modality head converts tokens to embeddings
    embeddings = loaded_text_head(input_ids)
    # 2. Backbone processes the embeddings
    hidden_states = loaded_backbone(embeddings)
    # 3. Task head produces the final output (logits)
    logits = loaded_task_head(hidden_states, task='classification')

print("--- I/O Shapes ---")
print("Input IDs Shape:     ", input_ids.shape)
print("Embeddings Shape:    ", embeddings.shape)
print("Hidden States Shape: ", hidden_states.shape)
print("Logits Shape:        ", logits.shape)
# Expected Logits shape: torch.Size([2, 10])
print("-" * 50)

print("Model composition and forward pass successful. ✅")