KServe's python server libraries implement a standardized library that is extended by model serving frameworks such as Scikit Learn, XGBoost and PyTorch. It encapsulates data plane API definitions and storage retrieval for models.

It provides many functionalities, including among others:

Registering a model and starting the server
Prediction Handler
Pre/Post Processing Handler
Liveness Handler
Readiness Handlers

It supports the following storage providers:

Google Cloud Storage with a prefix: "gs://"
- By default, it uses GOOGLE_APPLICATION_CREDENTIALS environment variable for user authentication.
- If GOOGLE_APPLICATION_CREDENTIALS is not provided, anonymous client will be used to download the artifacts.
S3 Compatible Object Storage with a prefix "s3://"
- By default, it uses S3_ENDPOINT, AWS_ACCESS_KEY_ID, and AWS_SECRET_ACCESS_KEY environment variables for user authentication.
Azure Blob Storage with the format: "https://{$STORAGE_ACCOUNT_NAME}.blob.core.windows.net/{$CONTAINER}/{$PATH}"
- By default, it uses anonymous client to download the artifacts.
- For e.g. https://kfserving.blob.core.windows.net/triton/simple_string/
Local filesystem either without any prefix or with a prefix "file://". For example:
- Absolute path: /absolute/path or file:///absolute/path
- Relative path: relative/path or file://relative/path
- For local filesystem, we recommended to use relative path without any prefix.
Persistent Volume Claim (PVC) with the format "pvc://{$pvcname}/[path]".
- The pvcname is the name of the PVC that contains the model.
- The [path] is the relative path to the model on the PVC.
- For e.g. pvc://mypvcname/model/path/on/pvc
Generic URI, over either HTTP, prefixed with http:// or HTTPS, prefixed with https://. For example:
- https://<some_url>.com/model.joblib
- http://<some_url>.com/model.joblib

Metrics

For latency metrics, send a request to /metrics. Prometheus latency histograms are emitted for each of the steps (pre/postprocessing, explain, predict). Additionally, the latencies of each step are logged per request.

Metric Name	Description	Type
request_preprocess_seconds	pre-processing request latency	Histogram
request_explain_seconds	explain request latency	Histogram
request_predict_seconds	prediction request latency	Histogram
request_postprocess_seconds	pre-processing request latency	Histogram

KServe Client

Getting Started

KServe's python client interacts with KServe control plane APIs for executing operations on a remote KServe cluster, such as creating, patching and deleting of a InferenceService instance. See the Sample for Python SDK Client to get started.

Documentation for Client API

Please review KServe Client API docs.

Documentation For Models

FAQs

What is kserve?

Is kserve well maintained?

Did you know?

Socket for GitHub automatically highlights issues in each pull request and monitors the health of all your open source dependencies. Discover the contents of your packages and block harmful activity before you install or update your dependencies.

Install