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
-
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. Settingspiral_alpha=0.0recovers standard RoPE exactly. -
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 + H²params (a few hundred for Nano) and is fully KV-cache compatible. -
Butterfly MLP. A multiplicative feed-forward block,
SiLU(a) ⊙ b, plus an intra-block bypassW_bypass·x. Withd_inner = 2dit 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)
gh repo clone Wiola-OSCOWL-ai/Wiola13M
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
transformersCacheAPI. Pinned totransformers>=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.
Metadata
Release files for wiola13m 1.0.1
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| wiola13m-1.0.1.tar.gz | 24.5 kB | Without provenance | Details for wiola13m-1.0.1.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| wiola13m-1.0.1-py3-none-any.whl | Python 3 | none | any | Details for wiola13m-1.0.1-py3-none-any.whl |
Total release size: 44.6 kB
Release files / wiola13m-1.0.1.tar.gz
| Download URL | wiola13m-1.0.1.tar.gz |
|---|---|
| Size | 24.5 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
663e8b553c6b96de307703cb9b7920e1b59449c73d3023b218d1ef5331ee02ba
|
|
BLAKE2b-256 checksum How to use checksums |
1f6cf441917404aacb635a7ac1cd8dbeb32b427f4ad30a70de25b2dd4f2ec394
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.9.13
|
Release files / wiola13m-1.0.1-py3-none-any.whl
| Download URL | wiola13m-1.0.1-py3-none-any.whl |
|---|---|
| Size | 20.0 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
bc75c576593b2ec0f80f4a20df798352ed34419c6d5f55e74d75a2e89de5b692
|
|
BLAKE2b-256 checksum How to use checksums |
4da905ea82b1ed950e8ed940e05f0d7aa897bab2e0c14c34c3d1ca63eb6cb920
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.9.13
|
Security for wiola13m 1.0.1
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.