Skip to content

Latest commit

 

History

History
261 lines (195 loc) · 8.42 KB

deploy-remove.md

File metadata and controls

261 lines (195 loc) · 8.42 KB
Raw

Deploying an LLM model with the Caikit+TGIS Serving runtime

There are two options for deploying and removing an LLM model:

In this procedure, you deploy an example Large Learning Model (LLM) model, flan-t5-small, with the Caikit+TGIS Serving runtime.

Note: The flan-t5-small LLM model has been containerized into an S3 MinIO bucket. For information on how to containerize your own LLM model into a MinIO bucket for testing purposes, see Create your own MinIO image.

Prerequisites

  • You installed the Caikit-TGIS-Serving stack as described in the Caikit-TGIS-Serving README file.

  • Your current working directory is the /caikit-tgis-serving/demo/kserve/ directory.

  • If your LLM is already in an S3-like object storage (for example, AWS S3), change the connection data in the minio-secret.yaml and serviceaccount-minio.yaml as shown here. Please note that the model should be converted to a caikit-compatible format, see [here][/demo/kserve/built-tip.md] for instructions.

Procedure

  1. Deploy the MinIO image that contains the LLM model.

    Note: If your model is already in an S3-like object storage (for example, AWS S3), you can skip this step.

    Create a new namespace for MinIO and deploy it together with the service account and data connection (a secret with generated access key ID and secret access key).

    ACCESS_KEY_ID=admin
SECRET_ACCESS_KEY=password
MINIO_NS=minio
    oc new-project ${MINIO_NS}
oc apply -f ./custom-manifests/minio/minio.yaml -n ${MINIO_NS}
sed "s/<minio_ns>/$MINIO_NS/g" ./custom-manifests/minio/minio-secret.yaml | tee ./minio-secret-current.yaml | oc -n ${MINIO_NS} apply -f -
sed "s/<minio_ns>/$MINIO_NS/g" ./custom-manifests/minio/serviceaccount-minio.yaml | tee ./serviceaccount-minio-current.yaml | oc -n ${MINIO_NS} apply -f -

  2. Deploy the LLM model with Caikit+TGIS Serving runtime

    a. Choose protocol to be used to invoke inferences: Default protocol is HTTP (e.g., curl commands). If you want to use gRPC set INF_PROTO to "-grpc" value, either skip the following command lines.

    INF_PROTO="-grpc"

    b. Create a new namespace.

    export TEST_NS="kserve-demo"
oc new-project ${TEST_NS}

    c. Create a caikit ServingRuntime.

    By default, it requests 4CPU and 8Gi of memory. You can adjust these values as needed.

    oc apply -f ./custom-manifests/caikit/caikit-tgis/caikit-tgis-servingruntime"$INF_PROTO".yaml -n ${TEST_NS}

    d. Deploy the MinIO data connection and service account.

    oc apply -f ./minio-secret-current.yaml -n ${TEST_NS}
oc create -f ./serviceaccount-minio-current.yaml -n ${TEST_NS}

    e. Deploy the inference service.

    The ISVC template file shown below contains all that is needed to set up the Inference Service (or gRPC ISVC template file for gRPC)

    Before using it, the following details have to be added:

    • <caikit-tgis-isvc-name> should be replaced by the name of the inference service
    • <NameOfAServiceAccount> should be replaced by the actual name of the Service Account
    • proto://path/to/model should be replaced by the actual path to the model that will run the inferences
    • <NameOfTheServingRuntime should be replaced by the name of the ServingRuntime

    Note: If you followed all the steps to this point, the following code will create the needed Inference Service using the Minio with the flan-t5-small model and the service account that have been created in the previous steps.

    ISVC_NAME=caikit-tgis-isvc$INF_PROTO
oc apply -f ./custom-manifests/caikit/caikit-tgis/"$ISVC_NAME".yaml -n ${TEST_NS}

    f. Verify that the inference service's READY state is True.

    oc get isvc/$ISVC_NAME -n ${TEST_NS}

  3. Perform inference using either using HTTP or gRPC

    Compute ISVC_HOSTNAME:

    export ISVC_URL=$(oc get isvc "$ISVC_NAME" -n ${TEST_NS} -o jsonpath='{.status.components.predictor.url}')

    • http only. Perform inference with HTTP. This example uses cURL.

      a. Run the following curl command for all tokens in a single call:

      curl -kL -H 'Content-Type: application/json' -d '{"model_id": "flan-t5-small-caikit", "inputs": "At what temperature does Nitrogen boil?"}' ${ISVC_URL}/api/v1/task/text-generation

      The response should be similar to the following:

      {
  "generated_token_count": "5",
  "text": "74 degrees F",
  "stop_reason": "EOS_TOKEN",
  "producer_id": {
    "name": "Text Generation",
    "version": "0.1.0"
  }
}

      b. Run curl to generate a token stream.

      curl -kL -H 'Content-Type: application/json' -d '{"model_id": "flan-t5-small-caikit", "inputs": "At what temperature does Nitrogen boil?"}' ${ISVC_URL}/api/v1/task/server-streaming-text-generation

      The response should be similar to the following:

      {
  "details": {

  }
},
{
  "tokens": [
    {
      "text": "▁",
      "logprob": -1.599083423614502
    }
  ],
  "details": {
    "generated_tokens": 1
  }
}
{
  "generated_text": "74",
  "tokens": [
    {
      "text": "74",
      "logprob": -3.3622500896453857
    }
  ],
  "details": {
    "generated_tokens": 2
  }
}
...

    • gRPC only. Perform inference with Remote Procedure Call (gPRC) commands. This example uses the grpcurl command-line utility.

      a. Determine whether the HTTP2 protocol is enabled in the cluster.

      oc get ingresses.config/cluster -ojson | grep ingress.operator.openshift.io/default-enable-http2

      If the annotation is set to true, skip to Step 3c.

      b. If the annotation is set to either false or not present, enable it.

      oc annotate ingresses.config/cluster ingress.operator.openshift.io/default-enable-http2=true

      c. Run the following grpcurl command for all tokens in a single call:

      export ISVC_HOSTNAME=$(oc get isvc "$ISVC_NAME" -n ${TEST_NS} -o jsonpath='{.status.components.predictor.url}' | cut -d'/' -f 3-)
grpcurl -insecure -d '{"text": "At what temperature does liquid Nitrogen boil?"}' -H "mm-model-id: flan-t5-small-caikit" ${ISVC_HOSTNAME}:443 caikit.runtime.Nlp.NlpService/TextGenerationTaskPredict

      The response should be similar to the following:

      {
  "generated_token_count": "5",
  "text": "74 degrees F",
  "stop_reason": "EOS_TOKEN",
  "producer_id": {
    "name": "Text Generation",
    "version": "0.1.0"
  }
}

      d. Run grpcurl to generate a token stream.

      grpcurl -insecure -d '{"text": "At what temperature does liquid Nitrogen boil?"}' -H "mm-model-id: flan-t5-small-caikit" ${ISVC_HOSTNAME}:443 caikit.runtime.Nlp.NlpService/ServerStreamingTextGenerationTaskPredict

      The response should be similar to the following:

      {
  "details": {

  }
},
{
  "tokens": [
    {
      "text": "▁",
      "logprob": -1.599083423614502
    }
  ],
  "details": {
    "generated_tokens": 1
  }
}
{
  "generated_text": "74",
  "tokens": [
    {
      "text": "74",
      "logprob": -3.3622500896453857
    }
  ],
  "details": {
    "generated_tokens": 2
  }
}
...

  4. Remove the LLM model

    a. To remove (undeploy) the LLM model, delete the Inference Service and its containing namespace:

    oc delete isvc --all -n ${TEST_NS} --force --grace-period=0
oc delete ns ${TEST_NS}

    b. Delete the MinIO resources by deleting the MinIO namespace.

    oc delete ns ${MINIO_NS}