minGRU-pytorch 0.2.1

minGRU

minGRU

Implementation of the proposed minGRU in Pytorch, only the log-space numerically stable version.

Yannic's paper review

Appreciation

• Adam Sorenti for contributing minLSTM!

Install

$ pip install minGRU-pytorch

Usage

import torch
from minGRU_pytorch import minGRU

min_gru = minGRU(512)

x = torch.randn(2, 1024, 512)

out = min_gru(x)

assert x.shape == out.shape

Sanity check

import torch
from minGRU_pytorch import minGRU

min_gru = minGRU(dim = 512, expansion_factor = 1.5)

x = torch.randn(1, 2048, 512)

# parallel

parallel_out = min_gru(x)[:, -1:]

# sequential

prev_hidden = None
for token in x.unbind(dim = 1):
    sequential_out, prev_hidden = min_gru(token[:, None, :], prev_hidden, return_next_prev_hidden = True)

assert torch.allclose(parallel_out, sequential_out, atol = 1e-4)

Test

enwik8

$ python train.py

Citations

@inproceedings{Feng2024WereRA,
    title   = {Were RNNs All We Needed?},
    author  = {Leo Feng and Frederick Tung and Mohamed Osama Ahmed and Yoshua Bengio and Hossein Hajimirsadegh},
    year    = {2024},
    url     = {https://api.semanticscholar.org/CorpusID:273025630}
}

@inproceedings{anonymous2024hymba,
    title   = {Hymba: A Hybrid-head Architecture for Small Language Models},
    author  = {Anonymous},
    booktitle = {Submitted to The Thirteenth International Conference on Learning Representations},
    year    = {2024},
    url     = {https://openreview.net/forum?id=A1ztozypga},
    note    = {under review}
}