WASI-NN Inference

This runs the WASI-NN inference module through the Propeller Kubernetes operator. It uses an OpenVINO backend to classify images.

Prerequisites

The operator must be deployed. Follow the end-to-end example first.

Proplet Configuration

Unlike other examples, WASI-NN needs a different proplet image compiled with neural-network host functions:

apiVersion: propeller.propeller.absmach.eu/v1
kind: Proplet
metadata:
  name: wasi-nn-proplet
spec:
  type: k8s
  k8s:
    image: "ghcr.io/absmach/propeller/proplet-wasi-nn:latest"
  connectionConfig:
    mqttAddress: "tcp://host.k3d.internal:1883"
    tenantId: "<your-tenant-id>"
    channelId: "<your-channel-id>"
    entityId: "<your-entity-id>"
    apiKey: "<your-api-key>"
    mqttQos: 2
    mqttTimeout: 30s

Apply it:

kubectl apply -n propeller-workloads -f wasi-nn-proplet.yaml

Prepare Model Files

The module expects OpenVIR model files (model.xml, model.bin) and a BGR image tensor (tensor.bgr). See the full WASI-NN example for instructions.

Build the WASM Module

cd propeller
make wasi-nn

Apply the Task

WASM_B64=$(base64 -w0 propeller/build/wasi-nn.wasm)
kubectl apply -n propeller-workloads -f - <<EOF
apiVersion: propeller.propeller.absmach.eu/v1
kind: Task
metadata:
  name: wasi-nn-example
spec:
  name: wasi-nn-example
  functionName: _start
  file: "${WASM_B64}"
  cliArgs:
    - "-S"
    - "nn"
    - "--dir=/home/proplet/fixture::fixture"
  propletSelector:
    propletId: "wasi-nn-proplet"
EOF

Watch and Verify

kubectl get task wasi-nn-example -n propeller-workloads -w

Expected result: top 5 ImageNet predictions with confidence scores.

Reference

Field	Value
`functionName`	`_start`
`cliArgs`	`["-S", "nn", "--dir=/home/proplet/fixture::fixture"]`
`daemon`	`false`
Proplet image	`ghcr.io/absmach/propeller/proplet-wasi-nn:latest`