---

OmniNumPy (Unfinished Experiment)

Stop fighting NumPy version hell. OmniNumPy is an attempt at a universal compatibility layer — run legacy NumPy code on NumPy 2.x and even swap backends (Torch, CuPy, JAX) with zero refactoring. It mostly works. Sometimes it explodes. That’s research.

---

⚠️ Status: Prototype, not Production

• Unfinished — several backends half-wired; some ops crash outright.
• Partial coverage — about 20 functions dispatch correctly; thousands don’t.
• Fragile — error handling is explicit, but not graceful.
• Performance — benchmarks are interesting, not reliable.

If you can make it stable, please do. PRs welcome.

---

🚀 Why This Exists

Scientific computing shouldn’t be a walled garden. You should be able to:

• Write once, run on CPU or GPU.
• Keep ancient NumPy 1.x code alive under NumPy 2.x.
• Experiment with Torch, CuPy, or JAX without rewriting everything.

OmniNumPy proves that idea mostly works — for now.

---

---

📖 Usage (When It Does Work)

import omninumpy as np

np.set_backend("torch")
a = np.array([1, 2, 3])
b = np.dot(a, a)

You can also emulate old NumPy behavior:

np.emulate("1.21")
print(np.int)

---

🧠 Architecture Sketch

User Code → Compatibility Layer → Backend Abstraction → Version Emulator → NumPy ≥ 2.x

The idea: same API, different engines underneath.

---

✅ Currently Functional

Function	NumPy	Torch	CuPy	JAX
array	✅	✅	✅	✅
dot	✅	✅	✅	⚠️ (JIT)
matmul	✅	✅	✅	⚠️ (JIT)
mean, sum	✅	✅	✅	✅
linalg.inv	✅	✅	✅	⚠️ (JIT)
linalg.svd	✅	✅	✅	⚠️ (JIT)

Everything else → NumPy fallback → hope.

---

🗺️ Roadmap

• Wire up more functions across all backends
• Add isolation and smarter fallbacks
• Expand linear algebra suite
• Harden tests and CI

---

📜 License & Credits

MIT License. Built on NumPy (© NumPy Developers, BSD 3-Clause). Torch, CuPy, and JAX belong to their respective developers.

---

TL;DR: OmniNumPy doesn’t work reliably — yet. But the architecture’s there. If you enjoy fighting the laws of tensor physics, fork it and evolve it.

---