ModelX

A simple, high-performance, scalable model storage service.

Modelx contains three components

• modelx

  Cli tools for the user side. You can use it to initialized、push、pull models，or management model repository locally.

• modelxd

  modelx repository server. It follows the OCI protocol and provides an http server to receive the modelx cli push models.

• modelxdl

  model deployment tools. It is integrated into the kubegems model deployment flows.

More infomation see: How Modelx born

Deploy Modelxd Server

with CLI

installation

Download modelxd binary from latest release to your PATH.

Determine your version with modelxd --version.

Configuration

Show all CLI options with modelxd --help. Common configurations can be seen below.

Usage:
  modelxd [flags]

Flags:
      --ca-file string               tls ca file
      --cert-file string             tls cert file
      --enable-redirect              enable blob storage redirect
  -h, --help                         help for modelxd
      --key-file string              tls key file
      --listen string                listen address (default ":8080")
      --oidc-issuer string           oidc issuer
      --s3-access-key string         s3 access key
      --s3-bucket string             s3 bucket (default "registry")
      --s3-presign-expire duration   s3 presign expire (default 1h0m0s)
      --s3-region string             s3 region
      --s3-secret-key string         s3 secret key
      --s3-url string                s3 url (default "https://s3.amazonaws.com")
  -v, --version                      version for modelxd

Using with Amazon S3 or Compatible services like Minio or DigitalOcean.

Make sure your environment is properly setup to access my-s3-bucket

modelxd --listen=:8080 \
  --s3-url=http://<Minio_URL>:<Port> \
  --s3-access-key=<AccessKey> \
  --s3-secret-key=<SecretKey> \
  --s3-bucket=<Bucket> \
  --enable-redirect=true

Using OIDC with KubeGems

Make sure you KubeGems API is properly access

modelxd --listen=:8080 \
  --s3-url=http(s)://<Minio_URL>:<Port> \
  --s3-access-key=<AccessKey> \
  --s3-secret-key=<SecretKey> \
  --s3-bucket=<Bucket> \
  --enable-redirect=true \
  --oidc-issuer=http(s)://<Kubegems_URL>:<Port>

with docker compose

Clone this repository, and run this command below.

export ADVERTISED_IP=<Host_IP> //set your host ip
sed -i "s/__ADVERTISED_IP__/${ADVERTISED_IP}/g" docker-compose.yaml
docker compose up -d

with Helm v3

Setup a temp S3 server using minio:

helm install --namespace minio --create-namespace  --repo https://charts.min.io \
--set rootUser=root,rootPassword=password \
--set 'mode=standalone,replicas=1,persistence.enabled=false,buckets[0].name=modelx,buckets[0].policy=none' \
--set service.type=NodePort \
minio minio

Make sure we can access S3 url out of cluster, modelx client pull/push from the address directly.

Setup modelx server

Setup modelx from helm:

export S3_URL="http://$(kubectl get node -o jsonpath='{.items[0].status.addresses[0].address}'):$(kubectl -n minio get svc minio -o jsonpath='{.spec.ports[0].nodePort}')"
echo ${S3_URL} # minio service node port address
helm install --namespace modelx --create-namespace --repo https://charts.kubegems.io/kubegems \
--set "storage.s3.url=${S3_URL},storage.s3.accessKey=root,storage.s3.secretKey=password,storage.s3.bucket=modelx" \
--set service.type=NodePort \
modelx modelx

Access modelx server fom node port:

export MODELX_URL="http://$(kubectl get node -o jsonpath='{.items[0].status.addresses[0].address}'):$(kubectl -n modelx get svc modelx -ojsonpath='{.spec.ports[0].nodePort}')"
echo ${MODELX_URL}  # modelx service node port address
curl ${MODELX_URL}
# {"schemaVersion":0,"manifests":null} # OK, if see this output

For more infomations see setup.

Modelx CLI

installation

Download binary from latest release to your PATH.

Completions provided via modelx completions zsh|bash|fish|powershell.

Quick Start

First, add and login a model repository

# Add model repository
$ modelx repo add modelx http://<your_modelxd_url>

# Login repository, if you don't set oidc iusername, press "enter" to skip token authentication.
$ modelx login modelx
Token:
Login successful for modelx

Second, Init a model locally

$ modelx init class

Modelx model initialized in class

$ tree class
class
├── modelx.yaml
└── README.md

$ cd class
# add model files

$ echo "some script" > scripy.sh
$ echo -n "some binary" > binary.dat

Finally, push your models ! 💪🏻

# add modelx registry

$ modelx push modelx/library/class@v1

Pushing to http://modelx.kubegems.io/library/class@v1
17e682f0 [++++++++++++++++++++++++++++++++++++++++] done
17e682f0 [++++++++++++++++++++++++++++++++++++++++] done
17e682f0 [++++++++++++++++++++++++++++++++++++++++] done
b6f9dd31 [++++++++++++++++++++++++++++++++++++++++] done
test.img [++++++++++++++++++++++++++++++++++++++++] done
4c513e54 [++++++++++++++++++++++++++++++++++++++++] done

Other Commands

list repository models

$ modelx list modelx

+---------+-------+------------------------------------------+
| PROJECT | NAME  | URL                                      |
+---------+-------+------------------------------------------+
| library | class | http://modelx.kubegems.io/library/class  |
+---------+-------+------------------------------------------+

list model versions

$ modelx list test/class

+---------+--------------------------------------------+--------+
| VERSION | URL                                        | SIZE   |
+---------+--------------------------------------------+--------+
| v1      | http://modelx.kubegems.io/library/class@v1 | 4.29GB |
| v2      | http://modelx.kubegems.io/library/class@v2 | 4.29GB |
| v3      | http://modelx.kubegems.io/library/class@v3 | 4.29GB |
+---------+--------------------------------------------+--------+

get model infomation

$ modelx info modelx/library/class@v1

config:
  inputs: {}
  outputs: {}
description: This is a modelx model
framework: pytorch
maintainers:
- [email protected]
modelFiles: []
tags:
- modelx
task: ""

About modelx.yaml

modelx.yaml contains model's metadata, a full example is:

config:
  inputs: {}
  outputs: {}
description: This is a modelx model
framework: <some framework>
maintainers:
  - maintainer
modelFiles: []
tags:
  - modelx
  - <other>
task: ""