ATOM

Angular-Multiplexed Transformer Optical Model

Transformer attention scores from optical interference — with an exact match to scaled dot-product attention.

Encode query and key as amplitudes with binary phase (0 / π). Interference then recovers the real inner product term by term:

Re(q_wave · conj(k_wave))  =  q · k

so the score matrix is identical to QKᵀ / √d. That is the mechanism claim: a physical two-level phase encoding is not an approximation to attention scores; under that encoding it is attention scores. The identity is unit-tested; the hybrid stack runs it on real model weights.

Architecture: optical (or optical-model) path computes scores; digital path handles softmax, values, MLP, norms, and the lm head — a deliberate hybrid split, not a half-finished pure-optical fantasy.

Results

SmolLM2-135M-Instruct (Hugging Face safetensors, all 30 layers):

• Optical-score greedy decode matches digital-score decode on fixed and natural-language prompts
• Full-forward logits: 100% top-1 agreement on the logged short sequence

That run shows the mechanism plugged into a full transformer, not only a toy matmul.

Details: evidence · result log

Stack

Piece	Role
atom/attention.py	Interference score algebra
atom/hybrid_model.py	Safetensors / GGUF hybrid generate
atom/capacity.py	M#-aware capacity model
atom/refresh.py	Readout-erase / rewrite planning stub
atom/rack.py	Multi-crystal placement (datacenter-style scale-out)
fpga/	Digital hybrid path scaffold for measured energy/latency later

Install

pip install -e .
pip install torch safetensors transformers pytest

Do not commit model weight files.

python examples/16_safetensors_text_eval.py \
  --model /path/to/SmolLM2-135M-Instruct \
  --prompt "The capital of France is" \
  --max-new 4

pytest tests/test_certainty.py -q

Scope

This repo is the software and systems foundation: exact score mechanism, hybrid inference, capacity/refresh/rack planning, FPGA hooks.

A physical photorefractive write/read loop is future work — not claimed as done. Geometric storage ceilings and material defaults are models with stated assumptions; see validation audit when citing capacity or hardware numbers.

Contributing

CONTRIBUTING.md