def test_model_spec():
ms = ModelSpec(
model_details=ModelDetails(
name="test",
local_folder="",
uri="",
platform=ModelFramework.XGBoost,
inputs=ModelDataArgs(args=[ModelDataArg(ty=str)]),
outputs=ModelDataArgs(args=[]),
),
protocol=V2Protocol(),
runtime_options=KFServingOptions().local_options,
)
s = ms.json()
j = json.loads(s)
ms2 = ModelSpec(**j)
assert isinstance(ms2.protocol, V2Protocol)
assert ms2.model_details.inputs.args[0].ty == str
def test_tensorflow_spec():
md = ModelDetails(
name="test",
local_folder="",
uri="",
platform=ModelFramework.Tensorflow,
inputs=ModelDataArgs(args=[]),
outputs=ModelDataArgs(args=[]),
)
protocol = SeldonProtocol()
options = KubernetesOptions(namespace="production", replicas=1)
runtime_options = RuntimeOptions(k8s_options=options)
model_spec = ModelSpec(model_details=md,
protocol=protocol,
runtime_options=runtime_options)
spec = get_container_spec(model_spec)
assert "image" in spec
assert "command" in spec
def deserialize(d: dict, client_details: ClientDetails = None) -> ClientModel:
ms = ModelSpec(**d)
return ClientModel(ms, client_details=client_details)
def _get_spec(self, model_spec: ModelSpec) -> dict:
if model_spec.model_details.platform == ModelFramework.TempoPipeline:
serviceAccountName = model_spec.runtime_options.k8s_options.serviceAccountName
if serviceAccountName is None:
serviceAccountName = DefaultServiceAccountName
return {
"apiVersion": "serving.kubeflow.org/v1beta1",
"kind": "InferenceService",
"metadata": {
"name": model_spec.model_details.name,
"namespace": model_spec.runtime_options.k8s_options.namespace,
"labels": {
TempoK8sLabel: "true",
},
"annotations": {
TempoK8sDescriptionAnnotation: model_spec.model_details.description,
TempoK8sModelSpecAnnotation: model_spec.json(),
},
},
"spec": {
"predictor": {
"serviceAccountName": serviceAccountName,
"containers": [
{
"image": MLSERVER_IMAGE,
"name": "mlserver",
"env": [
{
"name": "STORAGE_URI",
"value": model_spec.model_details.uri,
},
{
"name": "MLSERVER_HTTP_PORT",
"value": DefaultHTTPPort,
},
{
"name": "MLSERVER_GRPC_PORT",
"value": DefaultGRPCPort,
},
{
"name": "MLSERVER_MODEL_IMPLEMENTATION",
"value": "tempo.mlserver.InferenceRuntime",
},
{
"name": "MLSERVER_MODEL_NAME",
"value": model_spec.model_details.name,
},
{
"name": "MLSERVER_MODEL_URI",
"value": DefaultModelsPath,
},
{
"name": ENV_TEMPO_RUNTIME_OPTIONS,
"value": json.dumps(model_spec.runtime_options.dict()),
},
],
},
],
},
},
}
elif model_spec.model_details.platform in Implementations:
model_implementation = Implementations[model_spec.model_details.platform]
spec: Dict = {
"apiVersion": "serving.kubeflow.org/v1beta1",
"kind": "InferenceService",
"metadata": {
"name": model_spec.model_details.name,
"namespace": model_spec.runtime_options.k8s_options.namespace,
"labels": {
TempoK8sLabel: "true",
},
"annotations": {
TempoK8sDescriptionAnnotation: model_spec.model_details.description,
TempoK8sModelSpecAnnotation: model_spec.json(),
},
},
"spec": {
"predictor": {
model_implementation: {"storageUri": model_spec.model_details.uri},
},
},
}
if model_spec.runtime_options.k8s_options.serviceAccountName is not None:
spec["spec"]["predictor"][
"serviceAccountName"
] = model_spec.runtime_options.k8s_options.serviceAccountName
if isinstance(model_spec.protocol, KFServingV2Protocol):
spec["spec"]["predictor"][model_implementation]["protocolVersion"] = "v2"
return spec
else:
raise ValueError(
"Can't create spec for implementation ",
model_spec.model_details.platform,
)
def deserialize(d: dict) -> DeployedModel:
ms = ModelSpec(**d)
return DeployedModel(ms)