log-depth-recurrent-modeling 0.1.4

Log Depth Recurrent Modeling - Pytorch

Explorations into the Log Depth Recurrent Modeling proposed by Yiqin Wang of Imperial College London.

Install

$ pip install log-depth-recurrent-modeling

Usage

import torch
from log_depth_recurrent_modeling import ARGRC

model = ARGRC(
    num_tokens = 256,
    dim = 128,
    depth = 2,
    max_seq_len = 65536,
    shift_tokens = True
)

tokens = torch.randint(0, 256, (1, 65535))

# autoregressive loss

loss = model(tokens, return_loss = True)
loss.backward()

# forward for logits

logits = model(tokens) # (1, 65535, 256)

Standalone ARGRCLayer, which automatically pads the sequence to a multiple of max_seq_len (or the next power of two if unset) and strips the padding from the output:

import torch
from log_depth_recurrent_modeling import ARGRCLayer

layer = ARGRCLayer(
    dim = 128,
    max_seq_len = 65536,
    prenorm = True,
    shift_tokens = True,
    separate_grc = False # shares up and down gated recursive cell
)

x = torch.randn(1, 65535, 128)

out = layer(x) + x # (1, 65535, 128)

Feed the sequence in chunks of varying length, carrying memory across calls (equivalent to one pass over the full sequence):

memory = None

for start, length in ((0, 3), (3, 5), (8, 2)):
    chunk = x[:, start : start + length]
    out, memory = layer(chunk, memory = memory, return_memory = True)

Tasks

Run parity task with length generalization:

$ python train_parity_extrapolation.py

Run character language modeling on enwik8 with memory caching during generation:

$ python train_enwik8.py

Citations

@misc{wang2026logdepthrecurrentlanguagemodeling,
    title    = {Log-Depth Recurrent Language Modeling},
    author   = {Yiqin Wang and Nuri Cingillioglu and Charles Pert},
    year     = {2026},
    eprint   = {2609.28212},
    archivePrefix = {arXiv},
    primaryClass = {cs.LG},
    url      = {https://arxiv.org/abs/2609.28212},
}

@misc{shen2019orderedmemory,
    title   = {Ordered Memory},
    author  = {Yikang Shen and Shawn Tan and Arian Hosseini and Zhouhan Lin and Alessandro Sordoni and Aaron Courville},
    year    = {2019},
    eprint  = {1910.13466},
    archivePrefix = {arXiv},
    primaryClass = {cs.LG},
    url     = {https://arxiv.org/abs/1910.13466},
}

@misc{pert2026lengthgeneralizationlogdepthrecurrent,
    title     = {Length Generalization with Log-Depth Recurrent Units},
    author    = {Charles Pert and Dalal Alrajeh and Alessandra Russo},
    year      = {2026},
    eprint    = {2605.26035},
    archivePrefix = {arXiv},
    primaryClass = {cs.LG},
    url       = {https://arxiv.org/abs/2605.26035},
}

@software{peng_bo_2021_5196578,
    author    = {PENG Bo},
    title     = {BlinkDL/RWKV-LM: 0.01},
    month     = {aug},
    year      = {2021},
    publisher = {Zenodo},
    version   = {0.01},
    doi       = {10.5281/zenodo.5196578},
    url       = {https://doi.org/10.5281/zenodo.5196578}
}