pararnn-torch documentation

Train nonlinear RNNs in parallel over the sequence; decode one step at a time.

Package pararnn-torch · import pararnn · paper ParaRNN (arXiv:2510.21450).

Start here

Page	Purpose
Architecture fidelity	Which cells match which paper equations, and where we diverge
Spec template	Contract for a cell write-up
Cell catalog	Call-site snippets
Numerics contract	Agreement τ vs Newton residual

Install

pip install pararnn-torch
# or
uv add pararnn-torch

Fused Triton kernels: Linux + NVIDIA CUDA. Elsewhere NewtonConfig(scan_backend="auto") selects the eager Newton+scan path.

Quick mental model

1. A cell defines \(h_t = f(h_{t-1}, x_t)\) and a Jacobian structure.
2. Train: a few Newton iterations + parallel scan over \(T\) (paper Alg. 1).
3. Decode: sequential step; CUDA \(T{=}1\) uses decode_step.

Architecture pages stay at the algorithm level. Kernel / Triton notes live in the repo and maintainer lab.