mobiu-q 3.2.0

Soft Algebra Optimizer + O(N) Linear Attention for Long Context LLMs

Mobiu-Q v3.2.0

Soft Algebra for Optimization & Attention

---

Overview

Mobiu-Q is a framework built on Soft Algebra (nilpotent ε²=0) that provides:

1. MobiuOptimizer - Stable optimization in noisy environments
2. MobiuAttention 🧪 - O(N) linear attention for long sequences

Both share the same mathematical foundation but serve different purposes.

---

Installation

pip install mobiu-q

---

Quick Start

MobiuOptimizer (Stable API)

from mobiu_q import MobiuOptimizer
import torch

# Your license key (get one at https://app.mobiu.ai)
LICENSE_KEY = "your-license-key-here"

# Wrap any PyTorch optimizer
model = MyModel()
base_opt = torch.optim.Adam(model.parameters(), lr=0.0003)
opt = MobiuOptimizer(
    base_opt,
    license_key=LICENSE_KEY,
    method="adaptive",
    use_soft_algebra=True
)

for batch in dataloader:
    loss = criterion(model(batch))
    loss.backward()
    opt.step(loss.item())  # Pass loss for Soft Algebra

opt.end()  # Important: release resources

🆕 AUTO Mode - Don't Know Which Method to Use?

from mobiu_q import MobiuOptimizer

# Let Mobiu-Q choose automatically!
opt = MobiuOptimizer(base_opt, license_key=LICENSE_KEY, method="auto")

for batch in dataloader:
    loss = criterion(model(batch))
    loss.backward()
    opt.step(loss.item())

print(f"Selected mode: {opt.current_mode}")  # boost, dampen, or off
opt.end()

Results: ~35% average improvement, 0% hurts across 7 domains!

---

License Key

MobiuOptimizer requires a license key to access the cloud API:

from mobiu_q import MobiuOptimizer

LICENSE_KEY = "your-license-key-here"

# PyTorch mode (pass optimizer)
opt = MobiuOptimizer(base_opt, license_key=LICENSE_KEY, method="adaptive")

# Quantum/NumPy mode (pass params array)
opt = MobiuOptimizer(params, license_key=LICENSE_KEY, method="standard")

Get your key: https://app.mobiu.ai

Tier	API Calls	Price
Free	20/month	$0
Pro	Unlimited	$19/month

Note: MobiuAttention and AUTO mode run locally and do NOT require API calls for each step.

---

Methods

Method	Use Case	Default LR
auto 🆕	Don't know? Use this! Auto-selects best	0.01
auto-safe 🆕	Safer auto with DecayBoost	0.01
standard	Smooth landscapes, chemistry, physics	0.01
deep	Deep circuits, noisy hardware, complex opt	0.1
adaptive	RL, LLM fine-tuning, high-variance problems	0.0003

🆕 AUTO Mode Details

AUTO mode runs 3 virtual optimizers in parallel and picks the winner:

Mode	Description	When It Wins
boost	Higher LR when making progress	Biased gradients (Federated, Imbalanced)
dampen	Lower LR when gradients noisy	High variance, oscillations
off	Standard Adam	Clean problems

# Maximum improvement (~35% avg)
opt = MobiuOptimizer(base_opt, method="auto")

# Safer, more conservative (~20% avg, fewer hurts)
opt = MobiuOptimizer(base_opt, method="auto-safe")

AUTO with Maximize (RL / Trading)

AUTO mode supports both minimization and maximization:

# Loss minimization (default) - VQE, supervised learning
opt = MobiuOptimizer(base_opt, method="auto")
opt.step(loss.item())

# Reward maximization - RL, trading
opt = MobiuOptimizer(base_opt, method="auto", maximize=True)
opt.step(episode_return)

How AUTO Differs from Other Methods

Feature	standard/deep/adaptive	auto
Uses Frustration Engine	✅ Yes	❌ No
Uses Cloud API	✅ Yes	❌ No (local)
Adapts LR	Fixed strategy	Picks best strategy
Extra latency	~1ms per sync	Zero

Why no Frustration Engine in AUTO?

They solve the same problem differently:

• Frustration Engine: Reactive - detects stagnation, then boosts LR ×3
• AUTO: Proactive - continuously picks the best LR strategy

Using both together could cause conflicts (both trying to adjust LR).

Benchmark (7 domains, 30 seeds each):

Domain	AUTO	AUTO-Safe
Federated Learning	+40.8%	+22.4%
Noisy Labels	+22.6%	+14.3%
Sim-to-Real	+28.0%	+16.9%
Imbalanced Data	+27.8%	+17.0%
Meta-Learning	+42.7%	+22.8%
GAN Training	+42.6%	+23.0%
Clean	+43.9%	+23.2%
Hurts	0/210	0/210

---

Benchmarks

Reinforcement Learning & Trading

Domain	Improvement	Win Rate	p-value
Crypto Trading	+56% profit	100%	<0.001
LunarLander-v3	+128%	97%	<0.001
MuJoCo InvertedPendulum	+111%	100%	<0.001

Quantum Computing

Domain	Improvement	Win Rate	p-value
VQE H₂ (FakeFez)	+52%	100%	<0.001
QAOA MaxCut	+45%	95%	<0.001

Noisy & Distributed Learning

Domain	Improvement	Win Rate	p-value	Bias Source
Federated Learning	+67%	100%	<0.001	Non-IID client data
Imbalanced Data	+52%	100%	<0.001	Majority class dominates
Sim-to-Real	+47%	100%	<0.001	Simulator ≠ reality
Noisy Labels	+40%	100%	<0.001	Systematic mislabeling

---

Monitoring Training

opt = MobiuOptimizer(base_opt, license_key=LICENSE_KEY, method="adaptive")
# ... training ...

# Track metrics
print(opt.lr_history)      # Learning rates over time
print(opt.warp_history)    # Gradient warp factors
print(opt.current_mode)    # AUTO mode selection (if using auto)

---

Maximize vs Minimize

By default, Mobiu-Q assumes you're minimizing (loss, energy). For RL/Trading where you maximize (reward, profit), set maximize=True:

LICENSE_KEY = "your-license-key-here"

# Loss minimization (default) - for supervised learning, VQE
opt = MobiuOptimizer(base_opt, license_key=LICENSE_KEY, method="adaptive")
opt.step(loss.item())

# Reward maximization - for RL, trading
opt = MobiuOptimizer(base_opt, license_key=LICENSE_KEY, method="adaptive", maximize=True)
opt.step(episode_return)

---

A/B Testing

LICENSE_KEY = "your-license-key-here"

# Test with Soft Algebra
opt_on = MobiuOptimizer(base_opt, license_key=LICENSE_KEY, use_soft_algebra=True)

# Test without (baseline)
opt_off = MobiuOptimizer(base_opt, license_key=LICENSE_KEY, use_soft_algebra=False)

---

Examples by Domain

🆕 AUTO Mode (Recommended for New Users)

from mobiu_q import MobiuQCore
import numpy as np

LICENSE_KEY = "your-license-key-here"

# Don't know which method? Use AUTO!
params = np.random.randn(20)
opt = MobiuQCore(license_key=LICENSE_KEY, method="auto")

def energy_fn(p):
    return np.sum(p ** 2)  # Your loss function

for step in range(100):
    params = opt.step(params, energy_fn)
    
    if step % 20 == 0:
        print(f"Step {step}: energy={energy_fn(params):.4f}, mode={opt.current_mode}")

opt.end()

Federated Learning

import numpy as np
from mobiu_q import MobiuOptimizer

LICENSE_KEY = "your-license-key-here"

# Simulate federated aggregation with non-IID clients
class FederatedTrainer:
    def __init__(self, n_clients=10, non_iid_strength=0.5):
        self.n_clients = n_clients
        self.non_iid = non_iid_strength
        self.client_biases = [np.random.randn(dim) * non_iid_strength 
                             for _ in range(n_clients)]
    
    def aggregate_gradients(self, params, sampled_clients):
        grads = []
        for c in sampled_clients:
            local_grad = compute_gradient(params) + self.client_biases[c]
            grads.append(local_grad)
        return np.mean(grads, axis=0)

params = np.random.randn(100)
opt = MobiuOptimizer(
    params,
    license_key=LICENSE_KEY,
    method="standard",  # or "auto"!
    base_lr=0.01
)

for round in range(100):
    clients = np.random.choice(n_clients, size=5, replace=False)
    gradient = trainer.aggregate_gradients(params, clients)
    loss = compute_global_loss(params)
    params = opt.step(params, gradient, loss)

opt.end()

Reinforcement Learning (REINFORCE)

import torch
import gymnasium as gym
from mobiu_q import MobiuOptimizer

LICENSE_KEY = "your-license-key-here"

policy = torch.nn.Sequential(
    torch.nn.Linear(8, 64), torch.nn.Tanh(),
    torch.nn.Linear(64, 64), torch.nn.Tanh(),
    torch.nn.Linear(64, 4)
)

base_opt = torch.optim.Adam(policy.parameters(), lr=3e-4)
opt = MobiuOptimizer(
    base_opt,
    license_key=LICENSE_KEY,
    method="adaptive",
    maximize=True,
    sync_interval=50,
    verbose=True
)

env = gym.make("LunarLander-v3")

for episode in range(1000):
    state, _ = env.reset()
    log_probs, rewards = [], []
    
    done = False
    while not done:
        logits = policy(torch.FloatTensor(state))
        dist = torch.distributions.Categorical(logits=logits)
        action = dist.sample()
        log_probs.append(dist.log_prob(action))
        state, reward, terminated, truncated, _ = env.step(action.item())
        rewards.append(reward)
        done = terminated or truncated
    
    # REINFORCE update
    returns = []
    G = 0
    for r in reversed(rewards):
        G = r + 0.99 * G
        returns.insert(0, G)
    returns = torch.tensor(returns)
    returns = (returns - returns.mean()) / (returns.std() + 1e-8)
    
    loss = sum(-lp * G for lp, G in zip(log_probs, returns))
    
    opt.zero_grad()
    loss.backward()
    opt.step(sum(rewards))

opt.end()

Quantum Chemistry (VQE with Qiskit)

import numpy as np
from qiskit.circuit.library import EfficientSU2
from qiskit.quantum_info import SparsePauliOp
from qiskit_aer import AerSimulator
from qiskit.primitives import BackendEstimatorV2
from mobiu_q import MobiuOptimizer

LICENSE_KEY = "your-license-key-here"

hamiltonian = SparsePauliOp.from_list([
    ("II", -0.4804), ("ZZ", 0.3435), ("ZI", -0.4347),
    ("IZ", 0.5716), ("XX", 0.0910), ("YY", 0.0910)
])

backend = AerSimulator()
estimator = BackendEstimatorV2(backend=backend)
estimator.options.default_shots = 4096

ansatz = EfficientSU2(2, reps=2, entanglement="linear")
params = np.random.uniform(-0.3, 0.3, ansatz.num_parameters)

opt = MobiuOptimizer(
    params,
    license_key=LICENSE_KEY,
    method="standard",  # or "auto"!
    mode="hardware",
    use_soft_algebra=True
)

for step in range(100):
    delta = np.random.choice([-1, 1], size=len(params))
    shift = 0.1
    
    job = estimator.run([
        (ansatz, hamiltonian, params),
        (ansatz, hamiltonian, params + shift * delta),
        (ansatz, hamiltonian, params - shift * delta)
    ])
    results = job.result()
    
    energy = float(results[0].data.evs)
    grad = (float(results[1].data.evs) - float(results[2].data.evs)) / (2 * shift) * delta
    
    params = opt.step(params, grad, energy)

opt.end()
print(f"Final energy: {energy:.4f}")

---

Base Optimizers

Optimizer	Description	Best For
Adam	Adaptive moments	Default, most cases
AdamW	Adam with weight decay	LLM, Transformers
NAdam	Adam with Nesterov	Alternative to Adam
AMSGrad	Adam with stability	Drug discovery
SGD	Simple gradient descent	QAOA, convex
Momentum	SGD with momentum	RL, LLM fine-tuning
LAMB	Layer-wise adaptive	Large batch

---

🛠️ Troubleshooting

1. Don't Know Which Method to Use?

# Just use AUTO!
opt = MobiuOptimizer(base_opt, method="auto")

2. Switch Base Optimizer

Problem Type	Recommended
LoRA / LLM	Momentum or AdamW
VQE / Chemistry	Adam
QAOA	NAdam
RL / Trading	Momentum
Drug Discovery	AMSGrad

3. Adjust Learning Rate

# Try lower LR if diverging
base_opt = torch.optim.Adam(model.parameters(), lr=0.0001)

# Try higher LR if stuck
base_opt = torch.optim.Adam(model.parameters(), lr=0.001)

---

MobiuAttention 🧪

Why?

Standard Transformer attention is O(N²). MobiuAttention is O(N).

Seq Length	Transformer	MobiuAttention	Speedup
2,048	21s	9s	2.3x
4,096	39s	10s	3.9x
8,192	42s	7s	6.0x
16,384	OOM 💥	5s ✅	∞

Usage

from mobiu_q.experimental import MobiuBlock

# No license key needed - runs locally!
class LongContextLM(nn.Module):
    def __init__(self, vocab, d=512, h=8, layers=6):
        super().__init__()
        self.embed = nn.Embedding(vocab, d)
        self.blocks = nn.Sequential(*[MobiuBlock(d, h) for _ in range(layers)])
        self.head = nn.Linear(d, vocab)
    
    def forward(self, x):
        return self.head(self.blocks(self.embed(x)))

# Works with 16K+ tokens!
model = LongContextLM(50000)
x = torch.randint(0, 50000, (1, 16384))
out = model(x)  # No OOM!

---

Full Examples

File	Domain	Description
test_lunarlander_hybrid.py	RL	LunarLander with REINFORCE
test_mujoco_maximize.py	RL	MuJoCo continuous control
ppo_mobiu_test.py	RL	PPO from scratch
crypto_trading_benchmark.py	Trading	Crypto with regime switching
test_fakefez_h2.py	VQE	H₂ molecule on FakeFez
test_fakefez_qaoa.py	QAOA	MaxCut optimization
test_federated_fair.py	FL	Federated learning
test_noisy_labels_fair.py	Noisy	Noisy labels
test_sim_to_real_fair.py	Robotics	Sim-to-real
test_imbalanced_fair.py	Classification	Imbalanced data

---

License

Tier	API Calls	Price	Get Started
Free	20/month	$0	Sign up
Pro	Unlimited	$19/month	Get one

Note: MobiuAttention and AUTO mode run locally, no API calls required.

---

Links

• PyPI
• GitHub

---

Citation

@software{mobiu_q,
  title={Mobiu-Q: Soft Algebra for Optimization and Attention},
  author={Mobiu Technologies},
  year={2026},
  url={https://mobiu.ai}
}