corepy-ai 0.2.4

A unified, high-performance core runtime for data, computation, and AI workflows.

Corepy

High-Performance Tensor Runtime for Python. Rust-Powered Backend. Correctness First. Hardware Aware.

📖 What is Corepy?

Corepy is a high-performance tensor library that bridges Python's ease of use with the raw speed and safety of Rust and C++.

Unlike untyped array libraries, Corepy is built on a strict runtime that ensures:

1. Safety: Rust ownership guarantees prevent common segmentation faults.
2. Performance: C++ kernels (AVX2, NEON) and Metal shaders for heavy number crunching.
3. Observability: Zero-overhead built-in profiler to visualize bottlenecks.

It is designed for developers building AI foundations, scientific simulations, and high-performance systems.

Key Features

• NumPy-Compatible API: Familiar cp.array, cp.zeros, cp.add interface.
• 🚀 Metal GPU: Native acceleration on macOS (Apple Silicon) with device="metal".
• 📊 Profiler Export: Visualise traces in Chrome/Perfetto with cp.profiler.export_chrome_trace().
• ⚡ Hybrid Runtime: Rust dispatcher + C++ kernels + Objective-C++ (Metal).
• 🛡️ Cross-Platform Build: CMake-based build system that works on Linux, macOS, and Windows.
• 📈 Efficient Stats: Compute multiple statistics in one pass with cp.compute_stats.

---

💻 Supported Platforms

Platform	Architecture	Accelerators	Status
Linux	x86_64	AVX2, OpenBLAS	✅ Production
macOS	Apple Silicon	Metal, NEON	✅ Beta (0.2.4+)
Windows	x86_64	AVX2	✅ Experimental

---

🛠️ Installation

Preferred Method (uv)

We recommend uv for fast, correct cross-platform installation.

uv pip install corepy

Fallback (pip)

pip install corepy

👨‍💻 Development

For detailed instructions on setting up a development environment, building from source, and running tests, please refer to DEVELOPMENT.md.

Quick Build

git clone https://github.com/ai-foundation-software/corepy.git
cd corepy
make install

---

⚡ Quick Start

1. Metal Acceleration (macOS)

import corepy as cp

# Automatically uses Metal if available on macOS
t = cp.array([1.0, 2.0, 3.0], device="metal")
result = t.sum()
print(f"Result (GPU): {result}")

2. Performance Profiling

Stop guessing where your code is slow. Corepy has a built-in profiler.

import corepy as cp

# 1. Enable profiling
cp.enable_profiling()

# 2. Run your heavy workload
x = cp.ones(1_000_000)
y = x * 3.14159
result = y.mean()

# 3. Export to Chrome Tracing format
cp.profiler.export_chrome_trace("trace.json")

---

📚 Documentation

• Getting Started: First steps and installation details.
• Metal GPU Guide: using Apple Silicon acceleration.
• Architecture: Deep dive into the Rust/C++ hybrid runtime.
• Examples: Runnable scripts for common patterns.
• Contributing: Build and test guide.

---

🤝 Stability & Roadmap

Corepy is currently Alpha (v0.2.4).

• v0.2.4: Local CI Simulation, Metal Framework linking, Pinned Deps.
• v0.3.0: CUDA Support and Tiled Matmul Optimization.
• v1.0: Stable API promise.

See Roadmap for details.