Linear_Tiny_87M

Linear_Tiny_87M

Introduction

Linear transformers have emerged as a subquadratic-time alternative to softmax attention and have garnered significant interest due to their fixed-size recurrent state that lowers inference cost. However, their original formulation suffers from poor scaling and underperforms compute-matched transformers. Recent linear models such as RWKV and Mamba have attempted to address these shortcomings by proposing novel time-mixing and gating architectures, but pre-training large language models requires significant data and compute investments. Thus, the search for subquadratic architectures is limited by the availability of compute and quality pre-training datasets. As a cost-effective alternative to pre-training linear transformers, we propose Scalable UPtraining for Recurrent Attention (SUPRA). For more detail, refer to the paper

Linear_Tiny_87M is a linear model that has been trained on a subset of redpajama dataset for 1 epoch on 1x A4000. It took almost 4 hours for training to be completed.

Usage

Just download the checkpoint and afterwards run the following code snippet:

cd scripts

python generate.py \
--model open_lm_87m \
--checkpoint /path/to/checkpoint.pt \
--positional-embedding-type head_rotary \
--input-text "Machine Learning is a"