A hardware-native, unified multi-backend intelligence stack for distributed AI execution.
A hardware-native, framework-independent intelligence stack for the next generation of autonomous systems.
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.
- 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 | 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 |
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.
# Standard install (CPU & Core logic)
pip install nef2
# For GPU-accelerated wheels (CUDA 12.1+):
pip install nef2 --extra-index-url https://Hexa08.github.io/whl/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()}")- Phase 1 (Active): Establish the Foundation with NEFCore and custom CUDA kernels.
- Phase 2 (Active): Implement the Rust-based safe concurrency layer and distributed networking.
- Phase 3 (Active): Launch the HyperScale Multi-GPU Fabric for unified cluster execution.
- 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