wiola13m 1.0.0

Wiola13M: A 12.9M parameter decoder-only language model featuring Gated Spiral Attention, Spiral RoPE, and Butterfly MLP.

Wiola

Gated Spiral Attention — a small language model built for the 10–100M parameter regime
Spiral RoPE · content-adaptive attention gating · Butterfly MLP

---

Wiola is a decoder-only small language model whose novelty lives entirely in two sub-components of every layer. It is designed to run on a laptop, train on a single consumer GPU in hours, and publish to the Hugging Face Hub — yet be architecturally distinct enough to serve as a real experimental baseline.

Variant	d	L	H	d_inner	Params
Nano	256	6	8	512	~12.9M
Micro	384	8	12	768	~40M
Small	512	12	16	1024	~90M

What's novel

1. Spiral Rotary Positional Encoding. Standard RoPE frequencies are perturbed by a sqrt-growing, per-dimension-pair factor so phase trajectories fan outward instead of staying collinear — improving long-range discrimination at zero added parameters. Setting spiral_alpha=0.0 recovers standard RoPE exactly.

2. Gated Spiral Attention (GSA). A per-head, content-adaptive scalar gate, derived causally from a cumulative mean of the query projections, modulates attention scores before softmax. Heads that don't help self-suppress — implicit soft head pruning with no sparsity loss. The gate adds 2·H·d_h + H² params (a few hundred for Nano) and is fully KV-cache compatible.

3. Butterfly MLP. A multiplicative feed-forward block, SiLU(a) ⊙ b, plus an intra-block bypass W_bypass·x. With d_inner = 2d it matches a GeLU 4× FFN in parameter count while providing SwiGLU-class gating and steadier gradients in shallow stacks.

See docs/ARCHITECTURE.md for the full math.

Install

# from source (recommended while pre-1.0)
git clone https://github.com/wiola-project/wiola.git
cd wiola
pip install -e .

# with training / hub extras
pip install -e ".[train,hub]"

From PyPI once published:

pip install wiola

Version note: the model uses the modern transformers Cache API. Pinned to transformers>=4.40,<4.46, the range this release is tested against.

Quickstart

import torch
from wiola13m import WiolaConfig, WiolaForCausalLM

model = WiolaForCausalLM(WiolaConfig())          # Wiola Nano, random init
ids = torch.randint(0, 32000, (1, 16))

out = model(input_ids=ids, labels=ids)           # forward + LM loss
out.loss.backward()                              # gradients flow

model.eval()
gen = model.generate(ids[:, :4], max_new_tokens=20, do_sample=False)

Or run the bundled smoke test:

python scripts/quickstart.py

Train on TinyStories

# 1) get a 32k tokenizer (reuse a LLaMA tokenizer, or train your own)
python examples/create_tokenizer.py reuse --source meta-llama/Llama-2-7b-hf --out ./wiola-tokenizer

# 2) pre-train Nano (~2h/epoch on an RTX 3090)
python examples/train_tinystories.py \
    --tokenizer ./wiola-tokenizer \
    --output-dir ./wiola-nano-tinystories \
    --max-steps 20000

# 3) generate
python examples/generate.py --model ./wiola-nano-tinystories --prompt "Once upon a time"

Publish to the Hugging Face Hub

Wiola ships with auto_map support, so anyone can load your model without installing this package:

huggingface-cli login
python examples/push_to_hub.py --model-dir ./wiola-nano-tinystories --repo-id your-name/wiola-nano

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("your-name/wiola-nano", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("your-name/wiola-nano")

If the wiola package is installed, the "wiola" architecture is auto-registered and you don't even need trust_remote_code=True.

Design decision: gate input

The design doc's figure feeds the gate from post-RoPE queries, while the prose describes it as content-adaptive. Wiola defaults to computing the gate from the pre-RoPE query projections (gate_pre_rope=True) — position-independent and numerically stable — and exposes gate_pre_rope=False to match the figure. Both are causally correct and KV-cache safe.

Tests

pip install -e ".[dev]"
pytest

The suite verifies output shapes, weight tying, strict causality (no future-token leakage), exact equivalence between cached step-by-step decoding and a single full-sequence forward (with and without the gate), save/reload round-trips, and greedy/sampling/batched/beam generation.

Project layout

wiola/
├── src/wiola/
│   ├── configuration_wiola.py   # WiolaConfig
│   ├── modeling_wiola.py        # Spiral RoPE, GSA, Butterfly MLP, decoder, CausalLM
│   └── __init__.py              # Auto* registration
├── examples/                    # train / generate / tokenizer / push_to_hub
├── scripts/quickstart.py
├── tests/
└── docs/ARCHITECTURE.md

Citation

If you use Wiola, please cite it (see CITATION.cff).

License

Apache 2.0 — see LICENSE.