-
Notifications
You must be signed in to change notification settings - Fork 127
Asset naming convention
Louie Larson edited this page May 3, 2023
·
21 revisions
These conventions apply to all assets, regardless of type. Names must:
- Start with a letter.
- Use only lowercase letters, numbers, dashes, and periods.
- Be up to 255 characters long.
- Use dashes to separate words. Underscores are currently used for components due to an SDK limitation.
- Not include prefixes like "azureml" or "microsoft". Registries containing these assets will already provide this information.
- Not include the type of asset as a suffix (examples: "component" or "environment"). The asset type is already clearly represented via CLI, SDK, and UI.
- Not include a version number of the asset. Versions of included packages and frameworks are okay though.
Examples of valid asset names:
Name | Notes |
---|---|
tremendous-transform | Starts with a letter, uses lowercase letters and a dash |
pytorch-1.9-cuda11.6-gpu | Version numbers refer to PyTorch and CUDA, not the asset itself |
Examples of invalid asset names:
Name | Issue |
---|---|
microsoft-master-transform | Stars with "microsoft" prefix |
awesome_component | Ends with "component" |
fastsupertransform | Words aren't separated by dashes |
1-amazing-sort | Doesn't start with a letter |
pattern-detect-2.0 | Asset version should be in its version, not its name |
If you have a lot of related components, consider using a prefix to help group them. For example: forecast-preprocessing, forecast-threat-model, etc.
Your environment will need a meaningful name that will make it easy to find. Follow the naming convention below for framework-based environments:
<framework>-<framework-version>[-<os><os-version>][-py<python-version>][{-<gpu-driver><gpu-driver-version>|-gpu}][-inference]
Field | Description | Examples |
---|---|---|
framework |
Machine learning framework supported by the environment | pytorch, sklearn, tensorflow |
framework-version |
Framework version, major and minor values only | 1.1, 11.1, 2.8 |
os |
Operating system, specified only when not Ubuntu. | centos, debian, win |
os-version |
Operating system version | 9, 11, 2022 |
python-version |
Python version, major and minor values only, no period. Use only when absolutely required. | 38 (for Python 3.8) |
gpu-driver |
GPU driver | cuda |
gpu-driver-version |
GPU driver version, major and minor values only | 11.6 |
- Environments are assumed to be based on Ubuntu unless
-os<os-version>
is included - For a GPU environment, use either
-<gpu-driver><gpu-driver-version>
or-gpu
, not both-
-gpu
is only used when the GPU driver name and version aren't important
-
- Environments without a GPU driver or
-gpu
suffix are CPU-only - Environments with an
-inference
suffix provide inferencing support - In general, use fewer fields in the name to minimize maintenance costs. Otherwise, you'll spend time deprecating environments that are locked to a specific OS or Python version that have reached end-of-life, or other situations where a required update would invalidate the environment's name.
- If your environment isn't focused on a particular framework, choose a simple name and use the fields above as needed. Consider using a project-specific prefix to group multiple environments.
- pytorch-1.9 (CPU version)
- pytorch-1.9-cuda11.6
- pytorch-1.9-gpu
- pytorch-1.9-py38-gpu-inference
- pytorch-1.9-win2022-py38
- myproject-indexer
To keep them easy to maintain, Docker image names should match the name of their associated environment.
Examples:
Environment Name | Docker Image Name |
---|---|
pytorch-1.9 | azureml/curated/pytorch-1.9 |
pytorch-1.9-cuda11.6 | azureml/curated/pytorch-1.9-cuda11.6 |