furiosa-registry 0.6.0

Furiosa Registry

Furiosa Registry

Furiosa Registry defines Model and Artifact which allows client to communicate with several types of registries.

Overview

Client

Client using Furiosa Registry library. It can be a model server to serve models provided by FuriosaAI.

Furiosa Registry

This library

Registry A, B, C..

Registry sources. As long as there is a descriptor file called artifact that defines where to list the model binary and model descriptions, it does not matter what form the registry should be. Github repository is a typical source.

artifacts.yaml

Model descriptor file. You can find a complete schema in artfiact_schema.json or furiosa/registry/artifact.py

For example,

artifacts:
  - name: mlcommons_resnet50_v1.5_int8
    family: ResNet
    location: models/MNISTnet_uint8_quant_without_softmax.tflite
    format: tflite
    metadata:
      description: ResNet50 v1.5 model for MLCommons v1.1
      publication:
        url: https://arxiv.org/abs/1512.03385.pdf
  - name: mlcommons_ssd_mobilenet_v1_int8
    family: MobileNetV1
    location: https://github.com/furiosa-ai/furiosa-models/raw/main/models/mlcommons/mlcommons_ssd_mobilenet_v1_int8.onnx
    format: onnx
    description: MobileNet v1 model for MLCommons
    metadata:
      description: MobileNet v1 model for MLCommons v1.1
      publication:
        url: https://arxiv.org/abs/1704.04861.pdf
  - name: mlcommons_ssd_resnet34_int8
    family: ResNet
    location: models/model.py
    format: code
    metadata:
      description: ResNet34 model for MLCommons v1.1
      publication:
        url: https://github.com/mlcommons/inference/tree/master/vision/classification_and_detection

S3, Disk, HTTP

Data binary transport types. Each repository can store their model binary in several types of transports. location field in artifacts.yaml will be used to decide the transport.

Getting started

Install

pip install furiosa-registry

List up available models via artifact

import asyncio

import furiosa.registry as registry


# Repository where to load artifacts from.
repository = "https://github.com/furiosa-ai/furiosa-models"

# List up the available artifacts.
artifacts: List[registry.Artifact] = registry.list(repository)

# Access models from the artifacts
for artifact in arifacts:
    print(artifact.name)

Load models

import asyncio

import furiosa.registry as registry


# Repository where to load model from.
repository = "https://github.com/furiosa-ai/furiosa-models"

# Model name described in 'artifacts.yaml' at the repository.
model = "mlcommons_resnet50"

# Model version described in 'artifacts.yaml' at the repository.
version = "v1.1

# Load available model from the repository.
# Note that request() is an async function so we have to run the function in eventloop.
models: registry.Model = asyncio.run(registry.load(uri=repository, name=model, version=version))

# Access the model
print(model.name)
print(model.version)
print(model.description)

Print documentation about a model

import asyncio

import furiosa.registry as registry


# Repository where to load artifacts from.
repository = "https://github.com/furiosa-ai/furiosa-models"

# Model name described in 'artifacts.yaml' at the repository.
model = "mlcommons_resnet50"

# Render documentation provided by the repository for the  model.
print(registry.help(repository, model))

Development

Generate artfiact JSON schema from pydantic model definition.

When you changed artifact schema, you can generate modified schema via

python -c 'from furiosa.registry import Artifact;\
            print(Artifact.schema_json(indent=2), file=open("artifact_schema.json", "w"))'

See artifact_schema.json