Model Serving Runtimes¶

KServe provides a simple Kubernetes CRD to enable deploying single or multiple trained models onto model serving runtimes such as TFServing, TorchServe, Triton Inference Server. In addition ModelServer is the Python model serving runtime implemented in KServe itself with prediction v1 protocol, MLServer implements the prediction v2 protocol with both REST and gRPC. These model serving runtimes are able to provide out-of-the-box model serving, but you could also choose to build your own model server for more complex use case. KServe provides basic API primitives to allow you easily build custom model serving runtime, you can use other tools like BentoML to build your custom model serving image.

After models are deployed with InferenceService, you get all the following serverless features provided by KServe.

• Scale to and from Zero
• Request based Autoscaling on CPU/GPU
• Revision Management
• Optimized Container
• Batching
• Request/Response logging
• Traffic management
• Security with AuthN/AuthZ
• Distributed Tracing
• Out-of-the-box metrics
• Ingress/Egress control

The table below identifies each of the model serving runtimes supported by KServe. The HTTP and gRPC columns indicate the prediction protocol version that the serving runtime supports. The KServe prediction protocol is noted as either "v1" or "v2". Some serving runtimes also support their own prediction protocol, these are noted with an *. The default serving runtime version column defines the source and version of the serving runtime - MLServer, KServe or its own. These versions can also be found in the runtime kustomization YAML. All KServe native model serving runtimes use the current KServe release version (v0.12). The supported framework version column lists the major version of the model that is supported. These can also be found in the respective runtime YAML under the supportedModelFormats field. For model frameworks using the KServe serving runtime, the specific default version can be found in kserve/python. In a given serving runtime directory the pyproject.toml file contains the exact model framework version used. For example, in kserve/python/lgbserver the pyproject.toml file sets the model framework version to 3.3.2, lightgbm ~= 3.3.2.

*tensorflow - Tensorflow implements its own prediction protocol in addition to KServe's. See: Tensorflow Serving Prediction API documentation

*torchserve - PyTorch implements its own prediction protocol in addition to KServe's. See: Torchserve gRPC API documentation

apiVersion: "serving.kserve.io/v1beta1"
kind: "InferenceService"
metadata:
  name: "torchscript-cifar"
spec:
  predictor:
    triton:
      storageUri: "gs://kfserving-examples/models/torchscript"
      runtimeVersion: 21.08-py3