Sparsity

Introduction

ModelOpt’s Sparsity module (modelopt.torch.sparsity) enables you to sparsify the weights of your model. This can be useful for reducing the memory footprint of your model and can also be used to speed up inference.

Follow the steps described below to obtain a model with sparse weights using ModelOpt’s Sparsity module modelopt.torch.sparsity:

Training:You can either train your model using the existing training pipeline or load a pre-trained checkpoint for your model.
Sparsification: Sparsify the model using the provided mts.sparsify API.
Checkpoint and re-load: Save the model via mto.save and restore via mto.restore. See saving and loading of ModelOpt-modified model to learn more.

To find out more about Sparsity and related concepts, please refer to the section on Sparsity Concepts.

Post-Training Sparsification

Post-training sparsification is the process of converting a dense model to a sparse model without retraining. The simplest way to sparsify a model is to use the mts.sparsify API.

The mts.sparsify API takes a sparsity config and a sparsity format as input and returns a sparse model. The sparsity config is a dictionary specifying the layers to sparsify and the optional dataloader for calibration in data-driven sparsity, e.g., SparseGPT.

mts.sparsify() supports NVIDIA ASP and SparseGPT methods for magnitude-based and data-driven sparsity, respectively.

Example usage:

import torch
from transformers import AutoModelForCausalLM
import modelopt.torch.sparsity as mts

# User-defined model
model = AutoModelForCausalLM.from_pretrained("EleutherAI/gpt-j-6b")

# Configure and convert for sparsity
sparsity_config = {
    # data_loader is required for sparsity calibration
    "data_loader": calib_dataloader,
    "collect_func": lambda x: x,
}
sparse_model = mts.sparsify(
    model,
    "sparsegpt",  # or "sparse_magnitude"
    config=sparsity_config,
)

Note

data_loader is only required in case of data-driven sparsity, e.g., for calibration in sparsegpt. sparse_magnitude does not require data_loader as it uses magnitude-based method for thresholding.

Save and restore the sparse model

To store the sparse model for future usage, call mto.save():

mto.save(sparse_model, "modelopt_sparse_model.pth")

Note

mto.save() will save the model state_dict, along with the sparse masks and metadata to correctly re-create the sparse model later.

To restore the saved sparse model you can use mto.restore():

import modelopt.torch.opt as mto

# Re-initialize the original, unmodified model
model = AutoModelForCausalLM.from_pretrained("EleutherAI/gpt-j-6b")

# Restore the sparse model and metadata.
sparse_model = mto.restore(model, "modelopt_sparse_model.pth")

Note

mto.restore() will restore the model state_dict, along with the sparse masks and metadata of each sparse module. The plain pytorch module will be converted to a sparse module. The sparsity mask will be automatically enforced when the model weight is accessed.

Note

mts.export() will export the sparse model to a plain pytorch model. The sparse masks will be applied to model weights and all the sparse metadata will be removed. After exporting, sparsity will no longer be enforced during subsequent fine-tuning. If you want to continue fine-tuning, do not export the model.

Note

Please see saving and restoring of ModelOpt-modified models to learn about all the available options for saving and restoring.

Sparsity Concepts

Below, we will provide an overview of ModelOpt’s sparsity feature as well as its basic concepts and terminology.

Structured and Unstructured Sparsity

Weight sparsity is a model optimization technique where a fraction of the weights in a model are set to zero. Model sparsity can be broadly categorized as structured and unstructured sparsity. Unstructured sparsity refers to the case where the zero weights are randomly distributed across the weight matrix. Unstructured sparsity is more flexible but can lead to poor utilization on highly-parallelized hardware architectures like GPUs. Structured sparsity, on the other hand, is more efficient in terms of memory access and can be exploited to achieve higher math throughput. Structured sparsity can usually be achieved by enforcing a specific sparsity pattern on the weights.

N:M Sparsity

N:M sparsity refers to special type of fine-grained structured pattern, where in each block of M contiguous elements, at most N are nonzeros. Due to its regularity N:M sparsity can be efficiently implemented on GPU architecture and provides the following benefits:

Reduced memory bandwidth requirement: N:M Sparsity pattern have a smaller memory bandwidth requirement than both dense weights and weights with unstructured sparsity pattern.

Higher math throughput: Sparse Tensor Cores deliver higher math throughput for matrix-multiply operations when the first argument is a compressed N:M sparse matrix. For example, 2:4 sparsity pattern allows for 2x higher math throughput on sparse Tensor Cores.

On current Nvidia architectures (Ampere or later), 2:4 Sparsity, where in each block of four contiguous elements two are nonzeros, is supported for accelerated inference on sparse Tensor Cores.

Sparsification algorithm

There are many ways to achieve weight sparsity. A commonly-used approach is magnitude-based sparsity where in block of M elements, the N largest elements are retained and the rest are set to zero. Magnitude-based sparsity is simple and easy to implement, but may not retain the accuracy of the original model as well. Other methods such as data-driven sparsity, e.g., Optimal Brain Surgeon, usually delivers better accuracy. ModelOpt supports both magnitude-based (NVIDIA ASP) and data-driven sparsity (SparseGPT).