xLSTM: Extended Long Short-Term Memory

Paper: https://arxiv.org/abs/2405.04517

About

xLSTM is a new Recurrent Neural Network architecture based on ideas of the original LSTM. Through Exponential Gating with appropriate normalization and stabilization techniques and a new Matrix Memory it overcomes the limitations of the original LSTM and shows promising performance on Language Modeling when compared to Transformers or State Space Models.

xLSTM Large 7B

🚨 We trained a 7B parameter xLSTM Language Model 🚨

We have optimized the xLSTM architecture in terms of training throughput and stability. The code for the updated architecture is located in xlstm/xlstm_large.

The model weights are available on Huggingface at https://huggingface.co/NX-AI/xLSTM-7b.

Minimal Installation

Create a conda environment from the file environment_pt220cu121.yaml. Install the model code only (i.e. the module xlstm) as package:

Install via pip:

pip install xlstm

Clone from github:

git clone https://github.com/NX-AI/xlstm.git
cd xlstm
pip install -e .

For using the 7B xLSTM model install mlstm_kernels via:

pip install mlstm_kernels

Requirements

This package is based on PyTorch and was tested for versions >=1.8. For the CUDA version of sLSTM, you need Compute Capability >= 8.0, see https://developer.nvidia.com/cuda-gpus. For a well-tested environment, install the environment_pt220cu121.yaml as:

conda env create -n xlstm -f environment_pt220cu121.yaml
conda activate xlstm

For the xLSTM Large 7B model we require our mlstm_kernels (TODO add github link) package, which provides fast kernels for the xLSTM.

Models from the xLSTM Paper

This section explains how to use the models from the xLSTM paper.

Usage

For non language applications or for integrating in other architectures you can use the xLSTMBlockStack and for language modeling or other token-based applications you can use the xLSTMLMModel.

xLSTM Block Stack

The xLSTMBLockStack is meant for use as alternative backbone in existing projects. It is similar to a stack of Transformer blocks, but uses xLSTM blocks:

import torch

from xlstm import (
    xLSTMBlockStack,
    xLSTMBlockStackConfig,
    mLSTMBlockConfig,
    mLSTMLayerConfig,
    sLSTMBlockConfig,
    sLSTMLayerConfig,
    FeedForwardConfig,
)

cfg = xLSTMBlockStackConfig(
    mlstm_block=mLSTMBlockConfig(
        mlstm=mLSTMLayerConfig(
            conv1d_kernel_size=4, qkv_proj_blocksize=4, num_heads=4
        )
    ),
    slstm_block=sLSTMBlockConfig(
        slstm=sLSTMLayerConfig(
            backend="cuda",
            num_heads=4,
            conv1d_kernel_size=4,
            bias_init="powerlaw_blockdependent",
        ),
        feedforward=FeedForwardConfig(proj_factor=1.3, act_fn="gelu"),
    ),
    context_length=256,
    num_blocks=7,
    embedding_dim=128,
    slstm_at=[1],

)

xlstm_stack = xLSTMBlockStack(cfg)

x = torch.randn(4, 256, 128).to("cuda")
xlstm_stack = xlstm_stack.to("cuda")
y = xlstm_stack(x)
y.shape == x.shape

If you are working with yaml strings / files for configuration you can also use dacite to create the config dataclasses. This is the same as the snippet above:

from omegaconf import OmegaConf
from dacite import from_dict
from dacite import Config as DaciteConfig
from xlstm import xLSTMBlockStack, xLSTMBlockStackConfig

xlstm_cfg = """ 
mlstm_block:
  mlstm:
    conv1d_kernel_size: 4
    qkv_proj_blocksize: 4
    num_heads: 4
slstm_block:
  slstm:
    backend: cuda
    num_heads: 4
    conv1d_kernel_size: 4
    bias_init: powerlaw_blockdependent
  feedforward:
    proj_factor: 1.3
    act_fn: gelu
context_length: 256
num_blocks: 7
embedding_dim: 128
slstm_at: [1]
"""
cfg = OmegaConf.create(xlstm_cfg)
cfg = from_dict(data_class=xLSTMBlockStackConfig, data=OmegaConf.to_container(cfg), config=DaciteConfig(strict=True))
xlstm_stack = xLSTMBlockStack(cfg)

x = torch.randn(4, 256, 128).to("cuda")
xlstm_stack = xlstm_stack.to("cuda")
y = xlstm_stack(x)
y.shape == x.shape

xLSTM Language Model

The xLSTMLMModel is a wrapper around the xLSTMBlockStack that adds the token embedding and lm head.

from omegaconf import OmegaConf
from dacite import from_dict
from dacite import Config as DaciteConfig
from xlstm import xLSTMLMModel, xLSTMLMModelConfig

xlstm_cfg = """ 
vocab_size: 50304
mlstm_block:
  mlstm:
    conv1d_kernel_size: 4
    qkv_proj_blocksize: 4
    num_heads: 4
slstm_block:
  slstm:
    backend: cuda
    num_heads: 4
    conv1d_kernel_size: 4
    bias_init: powerlaw_blockdependent
  feedforward:
    proj_factor: 1.3
    act_fn: gelu
context_length: 256
num_blocks: 7
embedding_dim: 128
slstm_at: [1]
"""
cfg = OmegaConf.create(xlstm_cfg)
cfg = from_dict(data_class=xLSTMLMModelConfig, data=OmegaConf.to_container(cfg), config=DaciteConfig(strict=True))
xlstm_stack = xLSTMLMModel(cfg)

x = torch.randint(0, 50304, size=(4, 256)).to("cuda")
xlstm_stack = xlstm_stack.to("cuda")
y = xlstm_stack(x)
y.shape[1:] == (256, 50304)

Experiments

The synthetic experiments show-casing the benefits of sLSTM over mLSTM and vice versa best are the Parity task and the Multi-Query Associative Recall task. The Parity task can only be solved with state-tracking capabilities provided by the memory-mixing of sLSTM. The Multi-Query Associative Recall task measures memorization capabilities, where the matrix-memory and state expansion of mLSTM is very beneficial. In combination they do well on both tasks.

To run each, run the main.py in the experiments folder like:

python experiments/main.py --config experiments/parity_xLSTM01.yaml   # xLSTM[0:1], sLSTM only
python experiments/main.py --config experiments/parity_xLSTM10.yaml   # xLSTM[1:0], mLSTM only
python experiments/main.py --config experiments/parity_xLSTM11.yaml   # xLSTM[1:1], mLSTM and sLSTM

Note that the training loop does not contain early stopping or test evaluation.

Citation

If you use this codebase, or otherwise find our work valuable, please cite the xLSTM paper:

@inproceedings{beck:24xlstm,
      title={xLSTM: Extended Long Short-Term Memory}, 
      author={Maximilian Beck and Korbinian Pöppel and Markus Spanring and Andreas Auer and Oleksandra Prudnikova and Michael Kopp and Günter Klambauer and Johannes Brandstetter and Sepp Hochreiter},
      booktitle = {Thirty-eighth Conference on Neural Information Processing Systems},
      year={2024},
      url={https://arxiv.org/abs/2405.04517}, 
}