eelbaz/mLattice

选用指南

AI 根据 README 整理 · 2026/10/14 20:16:04

使用方式

通过命令行运行测试和训练脚本，也可启动演示服务器在浏览器中交互使用。

查看 README 依据

**[Live demo on Hugging Face →](https://huggingface.co/spaces/Ex0bit/mLattice)** ## Quickstart (no training required) ```bash # unit tests: kernel invariants, scan equivalence (runs on CPU, ~30s) python3 tests/test_kernel.py # acceptance tests (require CUDA GPU + checkpoints in checkpoints/) python3 tests/test_capacity.py cuda # ~5s python3 tests/test_forgetting.py cuda # ~7s python3 tests/test_realtext.py cuda # ~30s (requires data/wikitext103_slice.txt) ``` ## Results All scores from single runs on released checkpoints. | Protocol | Score | Test | |---|---|---|

Approximate GPU-hours: capacity ~26h, forgetting ~12h, real-text ~6h (single NVIDIA GB10). ## Training ```bash python3 experiments/train_capacity.py # curriculum to 8,192 pairs python3 experiments/train_forgetting.py # overwrite curriculum python3 experiments/train_realtext.py # real-text retrieval ``` One protocol module (`src/mulattice/protocol.py`) generates all training and evaluation data. ## Demo ```bash python3 experiments/demo_server.py # open http://localhost:8772 ```

Interactive: paste your own text, define named facts, mark overwrites, query the model. Shows the learned address map, measured cost curve, and cell-state inspection. ## Full write-up See [`PAPER.md`](PAPER.md) for the complete research note (architecture, measured design constraints, limitations). ## Cite ```bibtex @misc{mulattice2026, title = {μLattice: A Dynamically Allocated Memory Lattice for Sequence Models}, author = {Eric E. and Tensorbend Research}, year = {2026}, url = {github.com/eelbaz/mLattice} } ```

部署要求

需要 Python 3.10 或更高版本、PyTorch 2.4 或更高版本，以及 CUDA GPU（用于验收测试，CPU 可用于单元测试）。

查看 README 依据

# μLattice **Breaking the quadratic memory barrier in AI.** Today's models pay quadratically more to remember more. μLattice grows memory linearly: each token reads and writes exactly 9 cells on a 2D lattice, so per-token cost is flat by architecture. At 131K tokens, this is 7,280× fewer operations than attention (measured head-to-head on the same GPU). **Tensorbend Research · Eric E. · October 2026** ## Installation ```bash # requirements python >= 3.10 torch >= 2.4 CUDA GPU (recommended for acceptance tests; CPU works for unit tests) pip install -e . ```

已知限制

作者在论文中说明了架构的局限性，但具体内容未在提供的 README 片段中详述。

查看 README 依据

Interactive: paste your own text, define named facts, mark overwrites, query the model. Shows the learned address map, measured cost curve, and cell-state inspection. ## Full write-up See [`PAPER.md`](PAPER.md) for the complete research note (architecture, measured design constraints, limitations). ## Cite ```bibtex @misc{mulattice2026, title = {μLattice: A Dynamically Allocated Memory Lattice for Sequence Models}, author = {Eric E. and Tensorbend Research}, year = {2026}, url = {github.com/eelbaz/mLattice} } ```