The octoml
CLI helps you create deployable containers for your ML models using the OctoML Platform, the terminal or within your CI/CD pipeline.
You can run inferences on the container locally for development and testing, then deploy the same container to the cloud.
- Download the latest version of the OctoML CLI:
- Mac: Due to MacOS's security setup, you'll need to confirm in your browser that you'd like to retain the downloaded file, go to Finder to unzip the file, right-click on the
octoml
executable, and then click Open With -> Terminal -> Open. - Linux
- Windows
-
Check out the list of available commands by running
./octoml
in your current directory. -
Move the binary to an appropriate location like
/usr/local/bin
by runningmv octoml /usr/local/bin
in your current directory. -
Verify the
octoml
command works and accept OctoML's Terms of Use and Privacy Policy:
$ octoml -V
By use of this CLI application, you agree to OctoML’s terms of use and privacy policy.
https://octoml.ai/legals/terms-of-service/
https://octoml.ai/legals/privacy-policy/
? Do you wish to continue? · yes
octoml 0.4.2 (8f3bfbd 2022-06-09 23:11:41)
$ octoml -V
octoml 0.4.2 (8f3bfbd 2022-06-09 23:11:41)
-
Ensure you have a Docker daemon installed, by running
docker ps
. If the command is not found, install Docker for Mac OS, Linux, or Windows. -
Begin deploying your own model by jumping to the Core Commands section below, or follow one of our demos in Vision, Question Answering, or Text Generation to see an end-to-end user journey first.
Model framework coverage: We support TensorFlow SavedModel, TensorFlow GraphDef, Torchcript (PyTorch), and ONNX models.
Hardware coverage:
You may run our CLI on x86, CUDA machines, and ARM64 machines (including M1 Macs). In unauthenticated usage, the containers built via the CLI may only be deployed on hardware architectures matching the local machine on which you ran
octoml package
. In authenticated usage, you may create accelerated model containers deployable to any x86, CUDA, and ARM64 remote cloud instance, even if that hardware architecture does not match that of your local machine on which you ran octoml package
.
OS coverage: We support MacOS, Linux (Ubuntu 18.04+), and Windows. If you wish to use the CLI on Ubuntu 18.04, please ensure your protobuf version is on 3.19.4.
Note: You do not need to provide an OctoML API access token at this stage. If you run these commands without setting an OctoML API access token beforehand, we do not upload your model to the OctoML platform.
-
octoml init
: This is the first command we recommend that you run. It helps you set up an input configuration file by prompting you for the information required for the CLI to generate a container for deployment. -
octoml package
: Builds a deployment-ready Docker container for the models specified in the input configuration file. The input configuration file is required for this command to complete successfully. Downloading the 12+ GB base image will take approx. 20 minutes; future runs will be nearly instantaneous because the base image will be cached. -
octoml deploy
: Deploys a Docker container to a locally hosted endpoint. After completion of this command, you may rundocker ps
to confirm the container has been successfully generated for you. -
To run inferences against the container, follow our demos to configure an appropriate
run.py
file for your specific model.run.py
provides a standard schema on how to run inferences against the OctoML-deployed model, while requiring you to customize the pre- and post-processing code for your own model use case.
OctoML combines state-of-the-art compiler technologies (TVM, ONNX-RT, and others) to give you the best-performing package for any model. To access OctoML's acceleration and benchmarking services, you will need to sign up for an OctoML account after which you can generate an API access token.
-
octoml setup acceleration
: Prompts you for information required for acceleration, including an OctoML API access token, hardware, and dynamic shape disambiguation. Populates the information into your input configuration file for downstream use inoctoml package
andoctoml deploy
. If you choose not to runoctoml setup acceleration
, make sure to configure your input configuration file manually with the requisite fields for acceleration before callingoctoml package -e
oroctoml package -a
. -
octoml package -e
: We recommend running acceleration in express mode, which completes within 20 minutes. If you are willing to wait for several hours for potentially better latency via a fuller exploration of the optimization space, runoctoml package -a
. Both modes return the best-performing package with minimal latency for each hardware you've selected, after exploring multiple acceleration strategies including TVM, ONNX-RT, and the native training framework. After this step we recommend that you send this built container to a remote cloud repository for downstream remote usage. -
(Optional)
octoml deploy -a
oroctoml deploy -e
: Only run this command if you're on hardware with an architecture matching the cloud instance for which you have accelerated the model. If you previously ranpackage
on more than one hardware target, make sure to calldeploy
with a specific hardware input.
See demos for how to deploy OctoML model containers to downstream cloud registries like ECR and Kubernetes services like EKS.
OctoML by default collects telemetry on your usage of the CLI. However, we do not collect your model unless you submit an OctoML API access token.
The first time you run the CLI's octoml init
command, you will get prompted on whether you'd like to opt out of telemetry. You can opt out of telemetry anytime by setting the following environment variable: export OCTOML_TELEMETRY=false
.