ParaRNN

Train nonlinear RNNs in parallel over the sequence, decode one step at a time. Package pararnn-torch, import pararnn.

GRU/LSTM/sLSTM (and cousins) walk token-by-token at decode. ParaRNN trains the same cell with a few Newton updates + a parallel scan over time, so train wall-clock scales gently with sequence length T. .eval() keeps the one-step recurrence.

Alpha. Fused kernels need Linux + NVIDIA CUDA (fp32/fp16; lab includes RTX 2080 Ti). bf16 fused needs Ampere+ (CC ≥ 8.0). Elsewhere scan_backend="auto" selects the eager Newton+scan path.

• News
• What you get
• Glossary
• Models
• Install
• Quickstart
• Feel it (60s)
• Benchmarks
• Usage
• Training
• Evaluation
• Examples
• API overview
• Compatibility
• Method
• Citation

News

• [2026-09] Public launch: PyPI pararnn-torch, GitHub public, ParaSLSTM preprint v2 on Zenodo (doi:10.5281/zenodo.22558086; concept 10.5281/zenodo.22302586).
• [2026-09] Long-context campaign through T=131072: several cells need only 2 Newton iterations to match sequential unroll; RWKV-7 needs 0 (exact parallel scan). NewtonConfig(max_iters=None) picks those schedules.
• [2026-09] Cell zoo: ParaTitans, ParaRWKV7, ParaHopfield, ParaCfC, ParaNLRU (v0.13–0.17).
• [2026-09] Product entry: ParaSLSTMBlock, ParaSLSTMForCausalLM (labels CE, safetensors), continuous batch + vLLM plugin hooks.
• [2026-09] Factorized Newton for Dreamer-style ParaGRU(mix='head'); matrix-state M²RNN with measured iteration growth.

Diag-sLSTM fused Newton vs sequential (lab, see Benchmarks).

What you get

Nonlinear recurrent cells with one train / decode API (paper Alg. 1):

• Long-T train — Newton iterations + parallel scan; measured 100–1000× vs sequential unroll of the same cell
• Decode — .eval() sequential step; CUDA T=1 → fused decode_step (CUDA graphs with pinned out= buffers)
• Cell zoo — GRU, LSTM, sLSTM, Liquid CfC, Hopfield, RWKV-7, Titans-style memory, …
• Stack — ParaSLSTMBlock, CausalLM, paged state, continuous batch, speculative verify, optional vLLM plugin
• K*(T) — measured Newton budgets vs T (Glossary); max_iters=None reads those tables
• Train path — packed VJP, compile_safe_config(), DDP / FSDP2

Swap paths (docs/adoption.md):

• Attention trunk → ParaSLSTMBlock
• Dreamer RSSM → ParaGRU(mix='head', n_heads=8)
• Liquid / CfC → ParaCfC (Δt in the last channel of x)

Glossary

More on iterations: FAQs.md.

Models

Full snippets: docs/cells.md. Core algorithm: Danieli et al., ICLR 2026. HF wrappers and serve hooks are this repo’s extras.

Install

# when published on PyPI:
pip install pararnn-torch

# from git (until / beside PyPI):
pip install "pararnn-torch @ git+https://github.com/bugkira/pararnn-torch"

# editable:
git clone https://github.com/bugkira/pararnn-torch && cd pararnn-torch
uv sync --group dev
uv run pytest -q -m "not cuda"

PyPI Trusted Publishing: .github/workflows/release.yml. Hardware / Triton: INSTALL.md · FAQs.md.

Place modules with .to(device). Data parallel: docs/distributed.md. Lab benches (source tree only): scripts/.

Quickstart

1. Trunk block (drop-in residual mixer)

import torch
from pararnn import NewtonConfig, ParaSLSTMBlock

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
block = ParaSLSTMBlock(64, mlp_ratio=4.0, config=NewtonConfig(max_iters=3)).to(device)
x = torch.randn(2, 128, 64, device=device)

block.train()
y = block(x)          # Newton + scan inside the recurrent branch
y.sum().backward()

block.eval()
y_eval = block(x)     # sequential step (CUDA T=1: decode_step)

2. CausalLM

from pararnn import ParaSLSTMConfig, ParaSLSTMForCausalLM

cfg = ParaSLSTMConfig(vocab_size=256, hidden_size=64, num_hidden_layers=2, mlp_ratio=2.0)
model = ParaSLSTMForCausalLM(cfg).to(device)
ids = torch.randint(0, 256, (2, 32), device=device)

logits, loss = model(ids, labels=ids)
loss.backward()
out = model.generate(ids[:1, :8], max_new_tokens=16)
model.save_pretrained("./ckpt")   # config.json + model.safetensors

Smoke: examples/causal_lm_smoke.py · continuous batch: examples/continuous_batch.py.

3. Cell + ParaRNN

from pararnn import NewtonConfig, ParaGRU, ParaRNN

cell = ParaGRU(32, 64, device=device)
model = ParaRNN(cell, config=NewtonConfig(max_iters=3))
y = model(torch.randn(4, 128, 32, device=device))  # .train() → Newton

.train() → parallel Newton · .eval() → sequential step · CUDA T=1 → decode_step. Force either path with solver='newton'|'sequential'.

Feel it (60s)

Toy task: running XOR on a bit string (prefix parity). A stacked ParaSLSTM trained with Newton reaches last-token accuracy 1.0 at length 16 and still 1.0 at held-out length 32 — the nonlinear recurrence carries the bit.

uv run python examples/parity.py

Expect a JSON summary with "newton" arm accuracy 1.0 at T=16 and T=32. Interactive tour: notebooks/paraslstm_demo.ipynb (Colab).

Benchmarks

Two comparisons, both on the same cell:

1. Fused Newton (this library’s parallel train path) vs sequential (torch.compile of the usual time loop).
2. At very long T, how many Newton iterations you actually need (K*) and what wall time that costs.

Lab GPUs (RTX 2080 Ti / 3060) — consumer cards; regenerate plots with uv run python scripts/plot_readme_assets.py.

Diag-sLSTM forward median (ms), B=8, d_h=256, float32, RTX 2080 Ti (scripts/slstm_vs_flashrnn.py):

At T=131072, agreement tolerance τ=1e-4, B=1, RTX 3060 (scripts/bench_k_star.py). Labels under each cell are the measured Newton budget K* (how many iterations match sequential). Blue = parallel path; brown = sequential for-loop:

ParaNLRU smoke (3060, B=8, T=2048, d_h=256, K=3): fused ~2.7 ms vs sequential ~549 ms.

uv run python scripts/bench_k_star.py --cell all --time
uv run python scripts/slstm_vs_flashrnn.py --config configs/bench/newton_slstm_flashrnn.yaml

Usage

Cell zoo: docs/cells.md.

sLSTM / xLSTM-style

from pararnn import ParaRNN, ParaSLSTM

slstm = ParaRNN(ParaSLSTM(64, 64, mix="diag"), device=device)
y = slstm(torch.randn(4, 128, 64, device=device))

Dreamer-style block GRU

from pararnn import ParaGRU, ParaRNN

# Block-diagonal A_*; CUDA factorized Newton. LN stays outside the cell.
rssm_h = ParaRNN(ParaGRU(512, 512, mix="head", n_heads=8), device=device)
y = rssm_h(torch.randn(4, 64, 512, device=device))

Smoke: examples/rssm_recurrent.py.

Research cells

ParaM2RNN, ParaNLRU, ParaCfC, ParaHopfield, ParaRWKV7, ParaTitans — see docs/cells.md. Pin Newton depth with max_iters=int, use max_iters=None for auto schedules from measured K*(T), or pass newton_iters_by_t={…}.

Training

BabyLM / diag-sLSTM stack: scripts/train_babylm.py with configs under configs/train/. Optional extras: pip install "pararnn-torch[train,lm]". Log runs with MLflow (mlflow group). Distributed wrap: docs/distributed.md · examples/ddp_fsdp.py.

Evaluation

• Numerics — parallel ↔ sequential (pytest -m cuda); residual history on NewtonDivergenceError
• Expressivity — running-XOR (examples/parity.py)
• Wall-clock — scripts/bench_k_star.py, scripts/slstm_vs_flashrnn.py
• LM smoke — BabyLM notes under results/

Examples

More: scripts/README.md. Distributed demos: archive/distributed-demos.

API overview

• Cells — docs/cells.md; wrap with ParaRNN or call newton_apply / sequential_apply (ParaM2RNN, ParaRWKV7).
• Trunk — ParaSLSTMBlock(d_model, mlp_ratio=4) (docs/adoption.md).
• CausalLM / serve — ParaSLSTMForCausalLM, BlockStackPool, vllm.general_plugins (docs/inference.md, docs/vllm.md).
• Solver — NewtonConfig(scan_backend="auto", max_iters=None|int); picard_iters warms the first guess for sLSTM / M²RNN; verify_first_step=True for a one-shot agreement smoke on ParaRNN; compile_safe_config() for torch.compile(..., fullgraph=True).
• Numerics check — verify_agreement(module, x) → AgreementReport; full contract: docs/numerics-contract.md.
• Speculative — verify_linear_draft.
• Paged — PagedStatePool / paged_apply.
• Decode — decode_step, decode_wx, can_decode_step.

from pararnn.solvers import newton_apply, sequential_apply

h = newton_apply(cell, x)
h = sequential_apply(cell, x)

Docs: adoption · cells · xlstm.md · distributed.md · vllm.md · numerics-contract.md · oom-cookbook.md · compile-amp.md · shapes-layout.md · inference.md · structure.md · INSTALL.md · FAQs.md.

Compatibility

• torch.compile / AMP: docs/compile-amp.md — compile_safe_config() for fullgraph=True; Newton opts out of outer autocast (explicit .to(dtype) for half). Tests: tests/numerics/test_{compile,autocast}.py.
• DDP / FSDP / checkpoint: docs/distributed.md; ultra-long train VRAM → NewtonConfig(recompute=True) and the OOM cookbook (Hopfield d_h cap, RWKV slim heads).
• Shapes / packing: docs/shapes-layout.md — contiguous copies on fused paths; cu_seqlens support matrix.
• Inference / carry: docs/inference.md — decode_step + out=, generate(), vLLM via Mamba1 pages.
• Determinism: packed VJP uses tile tl.sum then .sum (no tl.atomic*); set CUBLAS_WORKSPACE_CONFIG=:4096:8 under torch.use_deterministic_algorithms(True).

Method

We solve for the whole hidden trajectory at once:

$F(H)_t = h_t - f(h_{t-1}, x_t) = 0$

with Newton updates; each linear solve is an associative scan over time (Danieli et al., Alg. 1). Layout: src/pararnn/layout.py.

K*(T) is the smallest Newton iteration count that still matches a sequential unroll within τ≈1e-4. We measure it per cell on a T grid and ship the envelopes in pararnn.solvers.newton.k_star; NewtonConfig(max_iters=None) reads them. Campaign script: scripts/bench_k_star.py.

GRU/LSTM warm-start follows paper App. A ($h_l^{(0)}=f(0,x_l)$); sLSTM uses zero-hidden + Picard. Backward follows paper eq. 2.6. M²RNN uses a factorized Jacobian map $J[\Delta]=f\Delta+(1-f)(1-Z^{\odot 2})\odot(\Delta W)$.

Citation

If you use this library, please cite the software and the ParaRNN framework.

@software{sereda2026pararnn,
  author       = {Sereda, Daniil},
  title        = {{pararnn-torch}: Hardware-efficient parallel training for nonlinear {RNNs}},
  year         = {2026},
  url          = {https://github.com/bugkira/pararnn-torch},
  version      = {0.17.4}
}

@misc{sereda2026paraslstm,
  author       = {Sereda, Daniil},
  title        = {{ParaSLSTM}: Work-Efficient Parallel Training of Nonlinear {sLSTM} via Tropical Warm-Starts},
  month        = sep,
  year         = 2026,
  note         = {Version 2},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.22558086},
  url          = {https://doi.org/10.5281/zenodo.22558086}
}

@inproceedings{danieli2026pararnn,
  title        = {{ParaRNN}: Unlocking Parallel Training of Nonlinear {RNNs} for Large Language Models},
  author       = {Danieli, Federico and Rodr{\'i}guez, Pau and Sarabia, Miguel and Suau, Xavier and Zappella, Luca},
  booktitle    = {International Conference on Learning Representations},
  year         = {2026},
  note         = {Oral. arXiv:2510.21450},
  url          = {https://arxiv.org/abs/2510.21450}
}

References

• Danieli et al. ParaRNN. ICLR 2026 (Oral). arXiv:2510.21450.
• Mishra et al. M²RNN. arXiv:2603.14360.
• Sereda. ParaSLSTM (v2). doi:10.5281/zenodo.22558086 (concept 10.5281/zenodo.22302586).
• Beck et al. xLSTM. arXiv:2405.04517.
• Lim et al. DEER. ICLR 2024. arXiv:2309.12252.
• Merrill et al. The Illusion of State in State-Space Models. arXiv:2404.08819.

License

MIT, LICENSE.