dmx-compressor 0.1.11

d-Matrix Compressor

This project contains tools for deep neural net co-design for custom hardware accelerators.

• Overview
• Getting started
• Usage
  • Basic API
  • Hugging Face pipeline API
• Next steps

---

Overview

In essence this is an extension of the PyTorch framework that implements hardware-efficient features, including

• custom low-precision numerical formats and arithmetic,
• fine-grain structured weight sparsity, and
• custom operator approximation logic.

In addition, the project provides a set of optimization tools for co-design using the above features, and is extensible through a plugin architecture.

Getting started

pip install dmx-compressor

Usage

Basic API

Given a PyTorch model, e.g. Net(), wrap it in a DmxModel container:

from dmx.compressor.modeling import DmxModel

model = DmxModel.from_torch(Net())

Here model is functionally equivalent to Net(), and all torch functionalities are still available, but model is equipped with d-Matrix specific features, making it ready for co-design configuration and/or optimization, at training time or post-training. See advanced topics for further details.

model.dmx_config is a dictionary that contains all, and only those, configurations that affect the functional behavior of the model, different from the behavior of the original Net(). Use method model.transform() to set these configurations, through application of configuration rules. See advanced topics for engineering of configuration rules.

There are two predefined special rule sets config_rules.BASELINE and config_rules.BASIC; the former is a dummy that does not change the original model's functional behavior, whereas the latter brings the model to a functional state that is equivalent to basic-mode execution on d-Matrix's hardware, e.g.

from dmx.compressor import config_rules
model = model.transform(
    model.dmx_config,
    *config_rules.BASIC,
)

Hugging Face pipeline API

To leverage the popularity of Hugging Face's pipeline API for inference, we extend transformers.pipeline() to dmx.compressor.modeling.hf.pipeline(), which retains all existing functionality of pipelines while enabling model transformation and configuration for deployment on d-Matrix hardware.

from dmx.compressor.modeling.hf import pipeline

pipe = pipeline(
    task="text-generation",
    model="facebook/opt-125m",
    dmx_config="BASIC",  # make the model deployable on d-Matrix backend
    ...
)

# Deploy pipe the same way as Hugging Face provides.

Next steps

For more detailed information, go over the following documents on specific topics. Find more usage examples here.

• Configurations
• Numerics
• Weight sparsity
• Custom approximation logic