Just read twice: closing the recall gap for recurrent language models.

Overview

This is part of a line of research in understanding the key quality-efficiency tradeoffs in how subquadratic architectures perform in-context learning. A key axis that differentiates LM architectures is their recall abilities, i.e., being able to ground next token predictions in information provided in-context beyond merely information the model memorized during training.

• Zoology explores the tradeoffs around input-dependent sequence mixing and the model dimension (efficiency) needed to perform recall.
• Based explores the tradeoffs around varying the recurrent state size for causal linear attention models. Increasing state size increases the GPU memory requirements during inference, but boosts recall quality.
• Just read twice explores going beyond causal language modeling to help recurrent models better decide what to store in the fixed-memory state.

Attention outperforms prior subquadratic LMs by large margins at associative recall. We use our understanding of these tradeoffs to design new methods. The Based architecture extends the Pareto frontier of the throughput-recall tradeoff space beyond prior autoregressive causl subquadratic LMs. This work presents two methods, a new prompting strategy JRT-Prompt and a new architecture JRT-RNN (included in the figure below), to further close the recall gap to attention!

Table of contents

• Installation
• Just read twice prompting
• Just read twice recurrent model evaluation
• Just read twice recurrent model training
• Synthetics
• Benchmarking
• Citations and acknowledgements

Installation

We recommend using a clean environment.

conda create -n dev python=3.9

# clone the repository
git clone git@github.com:HazyResearch/prefix-linear-attention.git
cd prefix-linear-attention

# install torch (matching your CUDA installation)
pip3 install torch torchvision torchaudio

# install package
pip install packaging
pip install -e .

# install flash linear attention
git clone https://github.com/sustcsonglin/flash-linear-attention.git
pip install -U git+https://github.com/sustcsonglin/flash-linear-attention

# pretraining kernel installs
git clone https://github.com/Dao-AILab/flash-attention.git
cd flash-attention/
cd csrc/layer_norm && pip install -e . && cd ../../
cd csrc/rotary && pip install -e . && cd ../../
cd csrc/fused_dense_lib && pip install -e . && cd ../../

Just read twice prompting

The following instructions are for reproducing the JRT-Prompting results.

To get setup:

cd lm-eval-harness
pip install -e .

To reproduce Table 1 in our ArXiv, you can use the following scripts. In the scripts, enter an output_directory name. The code will automatically download datasets and pre-trained models from Huggingface for evaluation.

bash prompt_scripts/run_jrt_prompt_hf.sh
bash prompt_scripts/run_jrt_prompt_hazy.sh

To view the results and save a consolidated dataframe with all the model predictions, enter your output_directory name and prefix_path in this file and run:

python prompt_scripts/collect_results.py

Just read twice recurrent model evaluation

We release the following JRT-RNN checkpoints. These models are trained on the same tokens in the same data order:

• 360M parameters, 30Bn Pile tokens: hazyresearch/JRT-360M-30B, hazyresearch/based-360M-30B, hazyresearch/mamba-360M-30B, hazyresearch/attn-360M-30B

• 1.3B parameters, 50Bn Pile tokens: hazyresearch/JRT-1B-50B, hazyresearch/based-1b-50b, hazyresearch/mamba-1b-50b, hazyresearch/attn-1b-50bn

To get setup:

cd lm-eval-harness
pip install -e .

To reproduce Table 2 in our ArXiv, you can use the following script. In the script, enter an output_directory name.

bash architecture_scripts/run_jrt_rnn.sh

To view the results and save a consolidated dataframe with all the model predictions, enter your output_directory name and prefix_path in this file and run:

python architecture_scripts/collect_results.py

To explore generation independently of these evaluations, we include the following code snippets:

import torch
from transformers import AutoTokenizer
from train.src.models.gpt import GPTLMHeadModel

# JRT and Based
tokenizer = AutoTokenizer.from_pretrained("gpt2")
model = GPTLMHeadModel.from_pretrained_hf("hazyresearch/JRT-1B-50B")
model = GPTLMHeadModel.from_pretrained_hf("hazyresearch/based-1b-50b")

# attention 
from based.models.transformer.gpt import GPTLMHeadModel
tokenizer = AutoTokenizer.from_pretrained("gpt2")
model = GPTLMHeadModel.from_pretrained_hf("hazyresearch/attn-1b-50bn").to("cuda")

# Mamba. Ensure you have run ```pip install mamba_ssm```.
from based.models.mamba import MambaLMHeadModel
tokenizer = AutoTokenizer.from_pretrained("gpt2")
model = MambaLMHeadModel.from_pretrained_hf("hazyresearch/mamba-1b-50b").to("cuda")

Just read twice recurrent model training

Follow the instructions at train/README.md to explore training. Prefix linear attention (PLA) implementation.

Synthetics

Follow the instructions at synthetics/README.md to explore the set disjointness synthetic.

Benchmarking

The following instructions are for comparing the JRT prefill speed to competitive baselines. The Based and JRT-RNN kernels will run on an NVidia H100 GPU since they exploit new hardware features (e.g., TMA, warpgroup ops).

To get setup:

# general
git submodule init
git submodule update
cd ThunderKittens/
source env.src

If you want to exclude any of the basaelines, you can modify the imports and the methods dictionary in the python file, to filter to the subset you care about. You can then choose which kernels to install below:

# 1. based tk kernel setup
cd examples/based/prefix_lin_attn/H100/
python prefill_setup.py install

# 2. jrt-rnn tk kernel setup
cd examples/based/linear_attn_forward/H100/
python lin_attn_setup.py install

# 3. fast transformers kernel setup
cd examples/jrt/csrc/causal_dot_prod/
python setup.py install
cd ../../

# 4. flash linear attention kernel setup
pip install -U git+https://github.com/sustcsonglin/flash-linear-attention

# 5. flash attention kernel setup
pip install flash-attn

To profile and evaluate the correctness of these kernels against PyTorch:

cd ThunderKittens/examples/based/prefix_lin_attn/
python lin_attn_profile.py

Citations and acknowledgements

Please reach out with feedback and questions! If you use or build on this work useful, please consider citing:

@article{arora2024just,
  title={Just read twice: closing the recall gap for recurrent language models},
  author={Arora, Simran and Timalsina, Aman and Singhal, Aaryan and Eyuboglu, Sabri and Zhao, Xinyi and Rao, Ashish and Rudra, Atri and Ré, Christopher},
  year={2024}
}

We also use or reference several prior works including: FlashAttention training code and kernels, LM-Eval Harness evaluation code, benchmark tasks (NQ, SQUAD, TriviaQA, etc.), Flash linear attention kernels, Conv1D kernel, Fast transformers kernel. We adapt the Based repository and use the Zoology synthetics infrastructure for this code release. Please cite them if you use any of their resources.

This work was made possible by compute from Together AI.