GPT - gMLP
This repository will attempt to crack long context autoregressive language modeling (GPT) using variations of gMLPs. Specifically, it will contain a variant that does gMLP for local sliding windows. The hope is to be able to stretch a single GPU to be able to train context lengths of 4096 and above efficiently and well.
GPT is technically a misnomer now, since there will be no attention (transformer) at all contained in the architecture.
Install
$ pip install g-mlp-gpt
Usage
import torch
from g_mlp_gpt import gMLPGPT
model = gMLPGPT(
num_tokens = 20000,
dim = 512,
depth = 4,
seq_len = 1024,
window_size = (128, 256, 512, 1024) # window sizes for each depth
)
x = torch.randint(0, 20000, (1, 1000))
logits = model(x) # (1, 1000, 20000)
16k context length
import torch
from g_mlp_gpt import gMLPGPT
model = gMLPGPT(
num_tokens = 20000,
dim = 512,
seq_len = 16384,
depth = 8,
reversible = True,
window = (128, 128, 256, 512, 1024, 1024, 2048, 2048, 4096, 4096, 8192, 8192),
axial = (1, 1, 1, 1, 1, 1, 2, 2, 4, 4, 8, 8)
).cuda()
x = torch.randint(0, 20000, (1, 16384)).cuda()
logits = model(x) # (1, 16384, 20000)
Citations
@misc{liu2021pay,
title = {Pay Attention to MLPs},
author = {Hanxiao Liu and Zihang Dai and David R. So and Quoc V. Le},
year = {2021},
eprint = {2105.08050},
archivePrefix = {arXiv},
primaryClass = {cs.LG}
}
9 Nov 16, 2022
24 Oct 26, 2022
331 Jan 3, 2023
74 Dec 5, 2022
28 Feb 20, 2022
81 Nov 26, 2022
39 Nov 30, 2022
49 Dec 22, 2022
120 Dec 12, 2022
0 Sep 06, 2022
8 Oct 21, 2022
37 Dec 15, 2022
32 Mar 25, 2022
1 Jul 14, 2022
20 Dec 16, 2022
16 Oct 29, 2022
80 Dec 25, 2022
42 Dec 02, 2022
41 Dec 13, 2022
28 Jan 03, 2023
3 Nov 27, 2022
102 Dec 14, 2022
3 Jan 11, 2022
1 Dec 12, 2021
1 Jan 02, 2022
30 Dec 19, 2022
29 Dec 03, 2022
54 Nov 21, 2022
5 Nov 23, 2022