PSSA: a plastic state space architecture
PSSA is a recurrent sequence model written from scratch in Rust. No PyTorch and no autograd framework: the forward pass, the backward pass and the optimizer are all hand written, and every batched or GPU path is checked against a scalar CPU reference twin (max gradient difference 2.98e-8).
Code, training chain and reproduction scripts: https://github.com/Sparticle62ops/pssa
Results
PSSA was trained against a parameter matched transformer baseline on cleaned WikiText-103, with the same corpus, the same token budget and the same schedule. The scores below are on a held out slice of 198,939 unseen tokens.
| model | cross entropy | perplexity | next token accuracy |
|---|---|---|---|
| PSSA | 3.9971 | 54.44 | 24.08% |
| transformer baseline | 4.4285 | 83.81 | 18.02% |
Generating 200 tokens on CPU takes 226 ms with PSSA against 2735 ms with the transformer, about 12x faster. Training runs at roughly 900 tokens per second on a single GPU through the CUDA path, against about 210 on CPU.
What is in it
- an affine state space scan as the sequence mixer, no attention
- a plastic memory bank with ordered retrieval and protected slots
- refractory gating that suppresses repeated writes
- a scalar CPU reference implementation kept as a correctness twin for every optimized path
Weights and checkpoints
- NSC64/pssa-100m: a 109M-parameter PyTorch port of this architecture, trained from scratch by NSC64 on 4B tokens of TinyStories and FineWeb-Edu. Ships bf16 safetensors weights, a 16k byte-level BPE tokenizer and a self-contained loader (plain scan on CPU, fused Triton kernel on CUDA).
- sparticle62/pssa-checkpoints-run2: training checkpoints from the 105M run in progress (raw .pssa resume files, not a packaged model).
Status
Research code under active development. The links above hold the published 109M PyTorch port and the in-training 105M checkpoints; the GitHub repository holds the training chain, the evaluation harness and the scripts to reproduce the numbers above.