photokan 0.3.0

Photonic KAN: Bridging PyTorch, KAN, and Q.ANT photonic hardware

PhotoKAN

Photonic Kolmogorov-Arnold Networks — bridging PyTorch · KAN · Q.ANT photonic hardware.

---

Why PhotoKAN?

Standard MLPs use fixed activations on nodes and linear weights on edges.
KANs invert this: learnable nonlinear functions sit on the edges, summed at nodes.

Photonic hardware is physically structured around edge-level nonlinear transforms — light through a waveguide produces exactly the kind of parametric nonlinear function a KAN edge needs. This is not an analogy; it is a direct structural match.

Published benchmarks show:

• 43% fewer parameters vs equivalent MLPs
• 46% fewer operations vs equivalent MLPs
• 30x energy efficiency gain on Q.ANT NPU vs CMOS

---

Installation

pip install photokan          # CPU simulation (no hardware required)
pip install photokan[qpal]    # + Q.ANT NPU support
pip install photokan[llm]     # + HuggingFace / PEFT integration
pip install photokan[onnx]    # + ONNX export
pip install photokan[dev]     # + development tools

---

Quick Start

import torch
import photokan as pk

# Works on CPU sim if no NPU — no code changes needed
model = pk.PhotoKAN(
    layer_sizes=[4, 16, 16, 1],
    activation='sine',   # 'sine' | 'fourier' | 'spline' | 'relu'
    backend='auto',      # auto-detects NPU, falls back to CPU
)

x = torch.randn(32, 4)
y = model(x)

# Standard PyTorch training
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
loss = torch.nn.MSELoss()(y, torch.randn(32, 1))
loss.backward()
optimizer.step()

# Check what hardware is available
print(pk.available_backends())
# -> {'cpu': True, 'cuda': True, 'qpal': False}

---

Activation Variants

Name	Formula	Best for	Photonic native
sine	sum w*sin(f*x + p)	Periodic targets, photonic deployment	Yes
fourier	a0 + sum [a*cos + b*sin]	Multi-frequency signals	Yes
spline	B-spline basis	Non-periodic, high precision	Via LUT
relu	sum w*ReLU(a*x + b)	Edge inference, speed	Yes

---

Photonic Noise Simulation

Test accuracy against realistic hardware impairments before deploying to NPU:

sim = pk.PhotonicSimulator()
sim.set_hardware_profile('npu2')   # SNR=16dB, 8-bit

results = sim.sweep_snr(model, x_test, y_test,
                         snr_range=[8, 10, 12, 14, 16, 20])
sim.plot_snr_accuracy(results)

---

Energy Estimation

Estimate energy consumption with Q.ANT's published efficiency numbers:

from photokan.utils import estimate_model_energy

reports = estimate_model_energy(model, batch_size=64)
# Layer 0: 8.214 uJ (CMOS) -> 0.267 uJ (Photonic), 30.8x better
# Layer 1: ...

---

Convolutional KAN

Use PhotoConvKAN for image and spatial workloads:

model = pk.PhotoConvKAN(
    in_channels=1, out_channels=16,
    kernel_size=3, activation='sine',
)
x = torch.randn(8, 1, 28, 28)
y = model(x)  # [8, 16, 28, 28]

---

LLM Integration

Replace MLP layers in HuggingFace transformers with PhotoKAN using LoRA-style adapters:

from photokan.llm import add_photo_lora, PhotoKANAttention

# Wrap a transformer's MLP layers
model = add_photo_lora(model, target_modules=["mlp"], n_basis=4)

---

AOT Compilation

Compile models to photonic deployment bundles (.npu):

compiler = pk.PhotonicCompiler()
program = compiler.compile(model, './my_model.npu')

# CPU validation (LUT interpreter)
y = program.run(x, backend='cpu')

# NPU inference (requires Q.PAL)
y = program.run(x, backend='qpal')

# Benchmark latency
stats = program.benchmark(x)
# -> {'mean_ms': 0.42, 'throughput_samples_per_sec': 76190}

---

Architecture

User PyTorch model
  |- PhotoKAN / PhotoKANLayer / PhotoConvKAN  (nn.Module, drop-in)
       |- EdgeActivations (sine / fourier / spline / relu)
            |- SimBackend -> CPU physics simulation
            |- QPALBackend -> Q.ANT NPU via Q.PAL (Phase 3)
  |- PhotonicCompiler
       |- LUTCompiler -> int8 quantised lookup tables
       |- QPALGraphBuilder -> NPU execution graph
       |- PhotonicProgram -> run on NPU or CPU LUT interpreter
  |- PhotonicSimulator -> SNR sweeps, transfer functions
  |- LLM integration (LoRA adapters, attention)

---

Project Status

Phase	Features	Status
Phase 1	Activations, SimBackend, Layers, Energy, Profiler	Done
Phase 2	LUT compiler, execution graph, ONNX export, ConvKAN	Done
Phase 3	Q.PAL hardware dispatch, LLM fine-tuning, arXiv paper	Upcoming

---

Running Tests

pip install -e ".[dev]"
pytest                          # 130+ tests
pytest -k "not mnist_real"      # skip slow MNIST training
pytest --cov=photokan           # with coverage

---

References

• Liu et al. (2024) -- KAN: Kolmogorov-Arnold Networks (arXiv 2404.19756)
• Peng et al. (2024) -- Photonic KAN via RAMZI (98% MNIST, 65x energy-area reduction)
• Reinhardt et al. (2024) -- SineKAN
• Q.ANT NPU -- https://qant.com/photonic-computing/

---

PhotoKAN v0.3 -- Build the bridge. Light does the math.