nef2 0.2.3

NEF2: Unified Multi-Backend AI Runtime. Use --extra-index-url for GPU-accelerated wheels.

NEF2: The AI Operating Substrate

A hardware-native, framework-independent intelligence stack for the next generation of autonomous systems.

---

NEF2: The AI Operating Substrate

A hardware-native, framework-independent intelligence stack for the next generation of autonomous systems.

---

The Vision

NEF2 is not a library; it is a Substrate. It eliminates the "framework tax" by bypassing heavy abstractions like PyTorch and JAX, communicating directly with the silicon through a custom-built hardware-native stack.

It is designed for a world where AI is not just a model, but a distributed, agentic system requiring zero-copy memory movement, hardware-peak performance, and intelligent memory virtualization.

Key Pillars

• Zero-Dependency Core: Pure Python/C++/Rust. No external ML frameworks.
• Hardware-Native Stack (NEF-HNS): Direct NVIDIA Driver API integration using raw PTX assembly.
• HyperCache Memory: Transparent virtualization of VRAM across System RAM and NVMe for trillion-parameter scale.
• Agent-Native Primitives: Built-in support for model-chaining, shared tensor buses, and streaming inference.

---

Feature Matrix

Feature	Status	Technology
NEFCore Runtime	✅ Production	Hybrid C++/Rust/Python execution
CUDA Driver Backend	✅ Production	Raw PTX Kernel execution
HyperCache (VRAM/RAM)	🚧 Beta	Intelligent memory paging
TurboQuant	🚧 Beta	Adaptive precision (FP8, INT4, NF4)
Multi-GPU Fabric	✅ Active	Unified logical accelerator
NEF Compiler	✅ Active	Graph capture & kernel fusion

---

Documentation Suite

For deep dives into specific areas of the NEF2 ecosystem:

• Getting Started: Installation, your first model, and basic training.
• Architecture Deep-Dive: Understanding NEFCore, HyperCache, and the Compiler stack.
• Hardware Support: Details on CUDA, HIP, Metal, and NPU integration.
• Developer Guide: Contributing, coding standards, and kernel development.

---

Quick Start

Installation

pip install nef2

Hardware-Native Tensors

from nef2 import Tensor
import nef2.gpu as gpu

# NEF2 automatically handles device placement
x = Tensor([[1, 2], [3, 4]], requires_grad=True)

if gpu.cuda_available():
    # Direct hardware-native matmul
    a, b = gpu.tensor([[1.0, 2.0]]), gpu.tensor([[3.0], [4.0]])
    c = a.matmul(b)
    print(f"Result on {gpu.device_name()}: {c.tolist()}")

---

Roadmap

1. Phase 1 (Active): Establish the Foundation with NEFCore and custom CUDA kernels.
2. Phase 2 (Active): Implement the Rust-based safe concurrency layer and distributed networking.
3. Phase 3 (Active): Launch the HyperScale Multi-GPU Fabric for unified cluster execution.
4. Phase 4 (Active): Realize Agent-Native Infrastructure for autonomous, model-agnostic intelligence.

---

Built for the future of Distributed Intelligence.

Join the revolution at github.com/Hexa08/NEF2