Features

Model registry

The Model Registry is a centralized hub for managing ML models' versioning and metadata. It provides a systematic way to track different versions of models along with details like their parameters, performance metrics, and approval status. You can for example:

• View all models and their versions in one place
• Manage model states (approved, pending, rejected) for seamless workflow management.
• Utilize single or multiple files to represent a “model”.
• Provide comprehensive descriptions, use cases, explanations, and deployment usage instructions.
• Trace the model creation process for transparency and reproducibility.

You can read more about the Model Registry in the Valohai documentation. 

• Introduction to the Model Registry
• Get started with the Model Registry 

Webhook triggers

In addition to scheduled pipelines and executions, the Valohai triggers can now be used to launch pipelines on demand. Webhooks are web requests that many services and applications can send in response to events happening within. In other words, you can have your webhook-sending service sending a POST request to start a pipeline in Valohai. There are several methods for authenticating and verifying the requests.

To learn more about the topic, see the following articles in the Valohai documentation:

• Introduction to webhook triggers 
• Launch pipelines with webhooks 
• Trigger example with Slack 
• Trigger example with GitHub 
• Trigger example with V7 
• Webhook triggers reference 

Compare images side-by-side

You can now compare image file outputs in the Valohai UI. To do this, choose two files under the outputs tab of a single job or under the Data tab of your project and click on the Compare button.

Multiple metadata charts

You can now add tabs for multiple plots under the Metadata section of your executions.

Opt-in for new pipeline experience

We have update the pipeline graph UI. You can choose to view either the classic graph or the new one by toggling the “Try the new pipeline graph” button.

Execution underutilization alerts

Valohai records information about the resource usage for both CPU and GPU during your executions. You can now view alerts for resource underutilization under the execution Alerts tab. If the peak usage is under 50% of the maximum available capacity, you will get an alert. This can be used to optimize the compute resources for your executions.

Noteworthy changes

Pipeline edge behavior

If the step defining a pipeline node has a default value for an input, the files coming via an edge are appended to the default input(s) instead of replacing them. To overcome this, it has been possible to set a pipeline override and set the default input to be empty. In the future this behavior will change and the default input from the step definition will be replaced instead. You can still set the edge inputs to be appended to the default inputs if needed.

You can toggle this with the edge-merge-mode pipeline property

• edge-merge-mode: replace - replace input's default value when edge is overriding the input. If you do not define edge-merge-mode this will be used by default.
• edge-merge-mode: append - combine the input's default value with edge's output

Environment list API

The environment list API will changed very slightly in a future release so unfinished_job_count is no longer included by default; it can be included with ?include=unfinished_job_count.

Inputs read-only

Valohai inputs have always been designed to be in a read-only directory. This will be further enforced in an upcoming peon change. Environments can be customized on a per-request basis to allow read/write access, due to some edge scenarios.