Quick Example

Using QQN as an optimizer

PyTorch — qqn-torch on GitHub

from qqn_torch import QQN

model = MyModel()
optimizer = QQN(model.parameters())  # no extra hyperparameters to tune

def closure():
    optimizer.zero_grad()
    loss = loss_fn(model(x), y)
    loss.backward()
    return loss

for step in range(num_steps):
    loss = optimizer.step(closure)  # line search needs a closure

JAX / Optax — qqn-jax on GitHub

import optax
from qqn_jax import qqn

solver = qqn()                     # drop-in GradientTransformation
opt_state = solver.init(params)

value, grad = jax.value_and_grad(loss_fn)(params)
updates, opt_state = solver.update(grad, opt_state, params, value=value,
                                   value_fn=loss_fn)
params = optax.apply_updates(params, updates)

Running the benchmark tournament (Rust)

qqn-optimizer on GitHub

# Quick benchmark (subset of problems)
cargo run --release -- --quick

# Full suite: 62 problems x 25 optimizers x 50 runs
cargo run --release -- --full

# Generate reports in a specific format
cargo run --release -- --format latex
cargo run --release -- --format markdown

Read the theory

• Plain-language essay + interactive tutorial → — start here
• Paper (DOI: 10.13140/RG.2.2.15200.19206) — derivation, proofs, full results
• Optimization Mechanics Visualizer — see QQN race GD, Adam, and L-BFGS in your browser

👉 Read the essay · Install for PyTorch · Install for JAX · Reproduce the benchmarks