winnex-nano

Native Winnex inference engine — deterministic spectral tokenizer, multimodel weight balancing, and chunk streaming via the Madhava fast build.

Model-agnostic (Qwen / BERT / DeepSeek / GPT), C++20/OpenCL, no CUDA. Reuses the winnex-madhava engine kernels (QKᵀ matmul, selective-top-K with Cauchy-Schwarz bounds) — no code duplication.

Components

Installation

pip install winnex-nano

Requires winnex-madhava>=1.8.1 (the engine), numpy. Optional: winnex-nano[server] for the SSE server (fastapi, uvicorn, httpx).

Quick start

import winnex_nano as wn

# 1. Deterministic spectral tokenizer (no BPE, no vocabulary)
tok = wn.SpectralTokenizer(embed_dim=64)
states = tok.encode("Winnex AI")       # list of quaternions
text = tok.decode(states)              # round-trip: "Winnex AI"

# 2. Multimodel weight balancing: W' = Σ αᵢ·R(qᵢ)·Wᵢ
bal = wn.WeightBalancer()
model_a = {"layer.0.weight": [1, 2, 3, 4]}
model_b = {"layer.0.weight": [10, 20, 30, 40]}
a = wn.BlendWeight(); a.alpha = 0.5
b = wn.BlendWeight(); b.alpha = 0.5
fused = bal.blend([model_a, model_b], [a, b])   # [5.5, 11, 16.5, 22]

# 3. Streaming via the Madhava fast build
engine = wn.StreamEngine()
chunks = []
engine.stream(
    "O Madhava e deterministico.",
    lambda ctx, top: " chunk" if len(chunks) < 3 else "",
    lambda c: chunks.append(c.text) if not c.done else None,
)

SSE server (OpenAI-compatible)

pip install "winnex-nano[server]"
python -m winnex_nano.server_sse 30002

curl -N http://localhost:30002/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages":[{"role":"user","content":"O Madhava e deterministico"}],"stream":true}'

Honest benchmark (vs BPE)

The spectral tokenizer is 100% deterministic with no vocabulary dependency, but 50–70× slower and ~1000× less compact than BPE. It is an autonomous encoding for the native multimodel engine, NOT a drop-in BPE replacement for a BPE-trained model.

License

Business Source License 1.1 (BSL 1.1) — pay@winnex.ai.

Pre-patent: https://zenodo.org/records/21861809