aioli-sdk 0.1.2.dev0__tar.gz → 0.2.0__tar.gz

{aioli-sdk-0.1.2.dev0 → aioli-sdk-0.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: aioli-sdk
-Version: 0.1.2.dev0
+Version: 0.2.0
 Summary: Aioli (AI OnLine Inference), a platform for deploying AI models at scale.
 Home-page: https://github.com/determined-ai/aioli
 Author: HPE AI Solutions

{aioli-sdk-0.1.2.dev0 → aioli-sdk-0.2.0}/aioli/__version__.py

@@ -1,2 +1,2 @@
 # © Copyright 2024 Hewlett Packard Enterprise Development LP
-__version__ = "0.1.2-dev0"
+__version__ = "0.2.0"

{aioli-sdk-0.1.2.dev0 → aioli-sdk-0.2.0}/aioli/cli/cli.py

@@ -20,7 +20,6 @@ from aioli.cli.deployment import args_description as deployment_args_description
 from aioli.cli.model import args_description as model_args_description
 from aioli.cli.registry import args_description as registry_args_description
 from aioli.cli.role import args_description as role_args_description
-from aioli.cli.service import args_description as service_args_description
 from aioli.cli.sso import args_description as sso_args_description
 from aioli.cli.user import args_description as user_args_description
 from aioli.cli.version import args_description as version_args_description
@@ -55,7 +54,6 @@ all_args_description: ArgsDescription = (
     args_description
     + registry_args_description
     + model_args_description
-    + service_args_description
     + deployment_args_description
     + version_args_description
     + user_args_description

{aioli-sdk-0.1.2.dev0 → aioli-sdk-0.2.0}/aioli/cli/deployment.py

@@ -5,9 +5,11 @@ from typing import Any, List
 import aiolirest
 from aioli import cli
 from aioli.cli import render
+from aioli.common import api
 from aioli.common.api import authentication
 from aioli.common.api.errors import NotFoundException
-from aioli.common.declarative_argparse import Arg, ArgsDescription, Cmd
+from aioli.common.declarative_argparse import Arg, ArgsDescription, Cmd, Group
+from aiolirest.models.autoscaling import Autoscaling
 from aiolirest.models.deployment import Deployment, DeploymentState
 from aiolirest.models.deployment_request import DeploymentRequest
 from aiolirest.models.security import Security
@@ -20,14 +22,17 @@ def show_deployment(args: Namespace) -> None:
 
     deployment = lookup_deployment(args.name, api_instance)
 
-    # For a more useful display, replace the service ID with its name
-    services_api = aiolirest.ServicesApi(session)
-    deployment.service = services_api.services_id_get(deployment.service).name
+    # For a more useful display, replace the model ID with its name
+    packaged_models_api = aiolirest.PackagedModelsApi(session)
+    deployment.model = packaged_models_api.models_id_get(deployment.model).name
 
     d = deployment.to_dict()
     # Remove clusterName for now - INF-243
     d.pop("clusterName")
-    print(render.format_object_as_yaml(d))
+    if args.json:
+        render.print_json(d)
+    else:
+        print(render.format_object_as_yaml(d))
 
 
 @authentication.required
@@ -36,9 +41,39 @@ def list_deployments(args: Namespace) -> None:
         api_instance = aiolirest.DeploymentsApi(session)
         response = api_instance.deployments_get()
 
-    def format_deployment(e: Deployment, reg_api: aiolirest.ServicesApi) -> List[Any]:
-        service = reg_api.services_id_get(e.service)
-
+    model_api = aiolirest.PackagedModelsApi(session)
+
+    if args.json:
+        format_json(response, model_api)
+    else:
+        format_deployments(response, args, model_api)
+
+
+def format_json(response: List[Deployment], model_api: aiolirest.PackagedModelsApi) -> None:
+    deps = []
+    for d in response:
+        # Don't use the d.to_json() method as it adds backslash escapes for double quote
+        m_dict = d.to_dict()
+        m_dict.pop("id")
+        m_dict.pop("modifiedAt")
+        # Use model name instead of id
+        model = model_api.models_id_get(d.model)
+        m_dict["model"] = model.name
+        # Remove pop as part of INF-503
+        m_dict.pop("service", None)
+        m_dict.pop("clusterName", None)
+        deps.append(m_dict)
+
+    render.print_json(deps)
+
+
+def format_deployments(
+    response: List[Deployment],
+    args: Namespace,
+    packaged_models_api: aiolirest.PackagedModelsApi,
+) -> None:
+    def format_deployment(e: Deployment, models_api: aiolirest.PackagedModelsApi) -> List[Any]:
+        model = models_api.models_id_get(e.model)
         state = e.state
         if state is None:
             state = DeploymentState()
@@ -49,33 +84,32 @@ def list_deployments(args: Namespace) -> None:
 
         assert e.security is not None
 
+        auto_scaling = e.auto_scaling
+        if auto_scaling is None:
+            auto_scaling = Autoscaling()
+
         result = [
             e.name,
-            service.name,
+            model.name,
             e.namespace,
             e.status,
             e.security.authentication_required,
             state.status,
             state.traffic_percentage,
-            secondary_state.status,
-            secondary_state.traffic_percentage,
         ]
 
         return result
 
     headers = [
         "Name",
-        "Service",
+        "Model",
         "Namespace",
         "Status",
         "Auth Required",
         "State",
         "Traffic %",
-        "Secondary State",
-        "Traffic %",
     ]
-    services_api = aiolirest.ServicesApi(session)
-    values = [format_deployment(r, services_api) for r in response]
+    values = [format_deployment(r, packaged_models_api) for r in response]
     render.tabulate_or_csv(headers, values, args.csv)
 
 
@@ -88,13 +122,27 @@ def create(args: Namespace) -> None:
             val = args.authentication_required.lower() == "true"
             sec.authentication_required = val
 
+        auto = Autoscaling(
+            metric=args.autoscaling_metric,
+        )
+
+        if args.autoscaling_target is not None:
+            auto.target = args.autoscaling_target
+
+        if args.autoscaling_max_replicas is not None:
+            auto.max_replicas = args.autoscaling_max_replicas
+
+        if args.autoscaling_min_replicas is not None:
+            auto.min_replicas = args.autoscaling_min_replicas
+
         r = DeploymentRequest(
             name=args.name,
-            service=args.service,
+            model=args.model,
             security=sec,
             namespace=args.namespace,
+            autoScaling=auto,
+            canaryTrafficPercent=args.canary_traffic_percent,
         )
-
         api_instance.deployments_post(r)
 
 
@@ -114,18 +162,39 @@ def update(args: Namespace) -> None:
             name=found.name,
             namespace=found.namespace,
             security=found.security,
-            service=found.service,
+            model=found.model,
+            autoScaling=found.auto_scaling,
+            canaryTrafficPercent=found.canary_traffic_percent,
         )
 
+        if request.auto_scaling is None:
+            # Not likely, but testing these prevents complaints from mypy
+            raise api.errors.BadResponseException("Unexpected null result")
+
         if args.name is not None:
             request.name = args.name
 
-        if args.service is not None:
-            request.service = args.service
+        if args.model is not None:
+            request.model = args.model
 
         if args.namespace is not None:
             request.namespace = args.namespace
 
+        if args.autoscaling_min_replicas is not None:
+            request.auto_scaling.min_replicas = args.autoscaling_min_replicas
+
+        if args.autoscaling_max_replicas is not None:
+            request.auto_scaling.max_replicas = args.autoscaling_max_replicas
+
+        if args.autoscaling_metric is not None:
+            request.auto_scaling.metric = args.autoscaling_metric
+
+        if args.autoscaling_target is not None:
+            request.auto_scaling.target = args.autoscaling_target
+
+        if args.canary_traffic_percent is not None:
+            request.canary_traffic_percent = args.canary_traffic_percent
+
         assert request.security is not None
 
         if args.authentication_required is not None:
@@ -149,9 +218,23 @@ def delete_deployment(args: Namespace) -> None:
 common_deployment_args: ArgsDescription = [
     Arg(
         "--authentication-required",
-        help="Deployed service requires callers to provide authentication",
+        help="Deployed model requires callers to provide authentication",
     ),
     Arg("--namespace", help="The Kubernetes namespace to be used for the deployment"),
+    Arg("--autoscaling-min-replicas", help="Minimum number of replicas", type=int),
+    Arg(
+        "--autoscaling-max-replicas",
+        help="Maximum number of replicas created based upon demand",
+        type=int,
+    ),
+    Arg("--autoscaling-metric", help="Metric name which controls autoscaling"),
+    Arg("--autoscaling-target", help="Metric target value", type=int),
+    Arg(
+        "--canary-traffic-percent",
+        help="Percent traffic to pass to new model version",
+        type=int,
+        default=100,
+    ),
 ]
 
 main_cmd = Cmd(
@@ -166,6 +249,7 @@ main_cmd = Cmd(
             "list deployments",
             [
                 Arg("--csv", action="store_true", help="print as CSV"),
+                Arg("--json", action="store_true", help="print as JSON"),
             ],
             is_default=True,
         ),
@@ -180,7 +264,7 @@ main_cmd = Cmd(
                     help="The name of the deployment. Must begin with a letter, but may contain "
                     "letters, numbers, underscore, and hyphen",
                 ),
-                Arg("--service", help="Service name or ID to be deployed", required="true"),
+                Arg("--model", help="Model to be deployed", required="true"),
             ]
             + common_deployment_args,
         ),
@@ -194,6 +278,10 @@ main_cmd = Cmd(
                     "name",
                     help="The name of the deployment.",
                 ),
+                Group(
+                    Arg("--yaml", action="store_true", help="print as YAML", default=True),
+                    Arg("--json", action="store_true", help="print as JSON"),
+                ),
             ],
         ),
         # Update command
@@ -208,7 +296,7 @@ main_cmd = Cmd(
                     help="The new name of the deployment. Must begin with a letter, but may "
                     "contain letters, numbers, underscore, and hyphen",
                 ),
-                Arg("--service", help="Service name or ID to be deployed"),
+                Arg("--model", help="Model to be deployed", required="true"),
             ]
             + common_deployment_args,
         ),