How VaikerAI works?
VaikerAI simplifies running machine learning models by providing a cloud-based API, eliminating the need to master complex machine learning concepts or handle infrastructure management. You can run open-source models shared by others or package and publish your own models, with the option to make them public or private.
Models
In machine learning, the term "model" can have various meanings depending on the context. It might refer to the source code, trained weights, architecture, or a combination of these elements. At VaikerAI, when we use the term "model," we're typically referring to a trained, packaged, and published software program that takes inputs and returns outputs.
Versions
Just like traditional software, machine learning models evolve and improve over time, with updates released as new versions. Whenever a model author retrains a model with new data, fixes a bug, or updates a dependency, these changes can impact the model's behavior. By publishing these updates as new versions, authors can enhance their models without disrupting the experience for users of older versions. Versioning is crucial for ensuring the reproducibility of machine learning, as it helps maintain consistent model behavior regardless of when or where it's being used.
Predictions
Each time you run a model, you generate a prediction. A prediction is an object that captures a single result from running the model, including the inputs you provided, the outputs returned, and other metadata like the model version, user information, prediction status, and timestamps.
Running Models in the Browser
You can run models on VaikerAI via the cloud API or the web interface. The web interface is ideal for trying out a model for the first time, offering a visual view of all model inputs and generating a form to run the model directly from your browser.
Running Models with the API
While the web interface is perfect for familiarizing yourself with a model, the API is essential when you're ready to integrate a model into applications like chatbots, websites, or mobile apps. VaikerAI's HTTP API can be used with any programming language, and client libraries are available for Python, JavaScript, and other languages to simplify API usage.
Python Example:
How Predictions Work
When you run a model, you create a prediction. Some models are fast, returning results in milliseconds, while others, especially generative models, may take longer. For models that require more processing time, you may need to poll the API to check the prediction's status. Predictions can have various statuses:
starting: The prediction is initializing.
processing: The model is running.
succeeded: The prediction completed successfully.
failed: The prediction encountered an error.
canceled: The prediction was canceled by the user.
Predictions timeout after 30 minutes. For longer timeouts, please contact us. Logged-in users can view their predictions on the dashboard, with details like status and run time.
Sharing Predictions
Every prediction you create is linked to your account and is private by default. If you're using the web interface, you can make a prediction public by clicking the "Share" button, allowing others to view it.
Deleting Predictions
Predictions created through the API are automatically deleted after an hour, including all inputs and outputs. If you need to retain any files, make sure to save them. Predictions created through the web interface are stored indefinitely unless deleted manually. To delete a prediction, go to your dashboard and click the "Delete" button on the prediction page.
Which Models Can You Run?
You can run any public model on VaikerAI using the API or web interface. These models may be open-source models published by others, such as meta/meta-llama-3-70b-instruct or stability-ai/sdxl, or you can publish and run your own models.
Finding Models
Explore popular and featured models or search for specific ones on VaikerAI. Search results include models that are public, have at least one published version, and include an example prediction. To make your own models discoverable, ensure they meet these criteria.
Pricing
For details on pricing, refer to the billing page.
Commercial Use
The models on VaikerAI are contributed by various individuals and organizations, each with its own licensing terms. For instance, Stable Diffusion has no ownership claims on the output, while other models like Pixray may have restrictions on commercial use. Check the model's license on its page for specifics.
Safety
Image generation models like Stability AI's SDXL include a safety checker to prevent the creation of unsafe content, such as images depicting nudity or violence. The safety checker is enabled by default for web predictions but can be disabled via the API for greater flexibility.
Cold Boots
VaikerAI hosts a vast catalog of models, but only actively used models remain "warm" (running). If a model hasn't been used recently, it may require a "cold boot," which can take a few minutes. Popular models are less likely to experience cold boots, while less frequently used models may take longer to start.
Rate Limits
API requests to VaikerAI are rate-limited. For more details, see the HTTP API reference documentation.
Pushing Your Own Models
In addition to running models created by others, you can push your own models to VaikerAI. You have the option to make your model public or private. For more information, see the guide on pushing a model to VaikerAI.
Private Models
When creating a model on VaikerAI, you can choose to make it public or private. Private models are only visible to you, or to members of your team if created within an organization. To create a private model, visit vaikerai.com/create and select the "Private" option. For more details, refer to the guide on deploying a custom model.
Last updated