Shixin Zhang

Posted on Feb 10

🚀 TensorCircuit-NG: The Universal, Differentiable Quantum Infrastructure

#quantum #ai #opensource #differentiable

👋 Hello World!

If you are a developer exploring quantum machine learning, or a physicist tired of rewriting code to make it run on GPUs, you have likely faced the "Framework Dilemma."

Do you write in PyTorch because you need the dataloaders?
Do you switch to JAX for that sweet JIT compilation speed?
Do you stick to TensorFlow because of legacy production pipelines?

What if your quantum simulator didn't care?

Meet TensorCircuit-NG (Next Generation)—the open-source, tensor-native platform that unifies quantum physics, AI, and High-Performance Computing.

🌟 What is TensorCircuit-NG?

TensorCircuit-NG is not just another circuit simulator. It is a backend-agnostic computational infrastructure.

It is designed to let you define your physics logic once and execute it anywhere. It wraps industry-standard ML frameworks (JAX, TensorFlow, PyTorch) into a unified engine, making quantum simulation end-to-end differentiable and hardware-accelerated.

🛠️ The "Write Once, Run Anywhere" Philosophy

The killer feature of TensorCircuit-NG is Infrastructure Unification.

You don't need to learn a new dialect for every backend. You simply switch the engine with one line of code:

Python

import tensorcircuit as tc

# Want JAX for JIT speed?
tc.set_backend("jax")

# Want PyTorch for easy integration with your existing DL models?
tc.set_backend("pytorch")

# Legacy TensorFlow project?
tc.set_backend("tensorflow")

This flexibility enables radical interoperability. You can train a hybrid model where the data pipeline lives in PyTorch, but the heavy-duty quantum circuit simulation is JIT-compiled via JAX/XLA for massive speedups—all handling zero-copy tensor transfers (DLPack) under the hood.

⚡ Why You Should Try It

1. Native Machine Learning Integration

We treat quantum circuits as first-class citizens in the computational graph.

Plug-and-Play Layers: Use tc.TorchLayer or tc.KerasLayer to insert parameterized quantum circuits directly into classical ResNets or Transformers.
Automatic Differentiation (AD): Forget parameter-shift rules. We compute gradients via backpropagation through the tensor network, making VQE and QML training exponentially faster.

2. HPC-Ready Scalability

Stop simulating on your CPU. TensorCircuit-NG supports:

GPU/TPU Acceleration: Move simulations to NVIDIA GPUs or Google TPUs without changing your physics code.
Distributed Computing: We support automated data parallelism (scaling to multiple devices) and model parallelism (tensor network slicing across GPU clusters).
Benchmark: We've demonstrated near-linear speedups on 8x NVIDIA H200 GPU clusters, simulating end-to-end variational quantum algorithms with 40+ qubits.

3. Advanced Physics Engines

It’s not just for qubits. TCNG comes with batteries included for:

Fermions Gaussian States: Efficiently simulate thousands of fermions.
Qudits: Native support for high-dimensional systems (d≥3).
Noise Modeling: Customizable noise channels for realistic hardware simulation.

💻 Show Me The Code

Here is how simple it is to build a differentiable variational circuit (VQE) that runs on any backend:

Python

import tensorcircuit as tc

# 1. Select your fighter (Backend)
tc.set_backend("jax") # or "pytorch", "tensorflow"

def vqe_loss(params, n=6):
    c = tc.Circuit(n)

    # 2. Build circuit (Hardware efficient ansatz)
    for i in range(n):
        c.rx(i, theta=params[i])
    for i in range(n-1):
        c.cnot(i, i+1)

    # 3. Calculate Expectation
    # This entire process is differentiable!
    e = c.expectation_ps(z=[0, 1]) 
    return tc.backend.real(e)

# 4. Get Gradients (Backend Agnostic API)
# This works regardless of whether you chose JAX, TF, or Torch
val_and_grad = tc.backend.jit(tc.backend.value_and_grad(vqe_loss))

# Run it!
print(val_and_grad(tc.backend.ones(6)))

🤝 Join the Community

TensorCircuit-NG is Open Source (Apache 2.0) and ready for you to hack on.

GitHub: Check out the Repository
Install: pip install tensorcircuit-ng
Docs: Read the Documentation

Whether you are building the next QML image classifier or simulating many-body physics, we'd love to see what you build.

Happy Coding! ⚛️

DEV Community