Update inferenece SDK examples

Author: zhaoqizqwang

examples/inference/SDK/inference-fsx-model-e2e.ipynb

@@ -7,10 +7,19 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "from sagemaker.hyperpod.hyperpod_manager import HyperPodManager\n",
-    "\n",
-    "HyperPodManager.list_clusters(region='us-east-2')\n",
-    "HyperPodManager.set_context('<hyperpod-cluster-name>', region='us-east-2')"
+    "from sagemaker.hyperpod import list_clusters, set_cluster_context\n",
+    "list_clusters(region='us-east-2')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "765ef3fd",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# choose the HP cluster\n",
+    "set_cluster_context('<my-cluster>', region='us-east-2')"
    ]
   },
   {
@@ -20,7 +29,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "from sagemaker.hyperpod.inference.config.hp_endpoint_config import CloudWatchTrigger, PrometheusTrigger, AutoScalingSpec, ModelMetrics, Metrics, FsxStorage, ModelSourceConfig, Tags, TlsConfig, ConfigMapKeyRef, FieldRef, ResourceFieldRef, SecretKeyRef, ValueFrom, EnvironmentVariables, ModelInvocationPort, ModelVolumeMount, Claims, Resources, Worker\n",
+    "from sagemaker.hyperpod.inference.config.hp_endpoint_config import FsxStorage, ModelSourceConfig, TlsConfig, EnvironmentVariables, ModelInvocationPort, ModelVolumeMount, Resources, Worker\n",
     "from sagemaker.hyperpod.inference.hp_endpoint import HPEndpoint\n",
     "import yaml\n",
     "import time"
@@ -33,13 +42,13 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "tls_config=TlsConfig(tls_certificate_output_s3_uri='s3://<your-tls-bucket-name>')\n",
+    "tls_config=TlsConfig(tls_certificate_output_s3_uri='s3://<my-tls-bucket-name>')\n",
     "\n",
     "model_source_config = ModelSourceConfig(\n",
     "    model_source_type='fsx',\n",
-    "    model_location=\"<your-model-folder-in-fsx>\",\n",
+    "    model_location=\"<my-model-folder-in-fsx>\",\n",
     "    fsx_storage=FsxStorage(\n",
-    "        file_system_id='<your-fs-id>'\n",
+    "        file_system_id='<my-fs-id>'\n",
     "    ),\n",
     ")\n",
     "\n",
@@ -73,7 +82,7 @@
    "outputs": [],
    "source": [
     "fsx_endpoint = HPEndpoint(\n",
-    "    endpoint_name='test-endpoint-name-fsx-pysdk',\n",
+    "    endpoint_name='<my-endpoint-name>',\n",
     "    instance_type='ml.g5.8xlarge',\n",
     "    model_name='deepseek15b-fsx-test-pysdk',\n",
     "    tls_config=tls_config,\n",
@@ -165,7 +174,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "endpoint = HPEndpoint.get(name='<your-endpoint-name>')"
+    "endpoint = HPEndpoint.get(name='<my-endpoint-name>')"
    ]
   },
   {

examples/inference/SDK/inference-jumpstart-e2e.ipynb

@@ -8,22 +8,14 @@
     "## Inference Operator PySDK E2E Expereience (JumpStart model)"
    ]
   },
-  {
-   "cell_type": "markdown",
-   "id": "1b3ce5c1-3c3d-4139-b7ae-042f360f3032",
-   "metadata": {},
-   "source": [
-    "<b>Prerequisite:</b> Data scientists should list clusters and set cluster context"
-   ]
-  },
   {
    "cell_type": "code",
    "execution_count": null,
    "id": "e22c86d6-0d3d-4c51-bef0-3f4c59ce111c",
    "metadata": {},
    "outputs": [],
    "source": [
-    "from sagemaker.hyperpod.hyperpod_manager import HyperPodManager"
+    "from sagemaker.hyperpod import list_clusters, set_cluster_context"
    ]
   },
   {
@@ -33,8 +25,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "#Set region \n",
-    "region = \"us-west-2\""
+    "list_clusters(region='us-east-2')"
    ]
   },
   {
@@ -44,8 +35,8 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# choose the HP cluster user works on\n",
-    "HyperPodManager.set_context('sagemaker-hyperpod-eks-cluster-demo-05-01', region=region)"
+    "# choose the HP cluster\n",
+    "set_cluster_context('<my-cluster>', region='us-east-2')"
    ]
   },
   {
@@ -67,7 +58,7 @@
     "from jumpstart_public_hub_visualization_utils import get_all_public_hub_model_data\n",
     "\n",
     "# Load and display SageMaker public hub models\n",
-    "get_all_public_hub_model_data(region=\"us-west-2\")"
+    "get_all_public_hub_model_data(region=\"us-east-2\")"
    ]
   },
   {
@@ -122,8 +113,8 @@
     "server=Server(\n",
     "    instance_type='ml.g5.8xlarge',\n",
     ")\n",
-    "endpoint_name=SageMakerEndpoint(name='deepsek7bsme-testing-jumpstart-7-1')\n",
-    "tls_config=TlsConfig(tls_certificate_output_s3_uri='s3://tls-bucket-inf1-beta2')\n",
+    "endpoint_name=SageMakerEndpoint(name='<my-endpoint-name>')\n",
+    "tls_config=TlsConfig(tls_certificate_output_s3_uri='s3://<my-tls-bucket>')\n",
     "\n",
     "# create spec\n",
     "js_endpoint=HPJumpStartEndpoint(\n",
@@ -230,7 +221,7 @@
    "outputs": [],
    "source": [
     "# output is similar to kubectl describe jumpstartmodel\n",
-    "endpoint = HPJumpStartEndpoint.get(name='deepseek-llm-r1-distill-qwen-1-5b')\n",
+    "endpoint = HPJumpStartEndpoint.get(name='<my-endpoint-name>')\n",
     "print_yaml(endpoint)"
    ]
   },
@@ -265,10 +256,7 @@
    "outputs": [],
    "source": [
     "# get operator logs\n",
-    "print(js_endpoint.get_operator_logs(since_hours=1))\n",
-    "\n",
-    "# get specific pod log\n",
-    "# js_endpoint.get_logs(pod='pod-name')"
+    "print(js_endpoint.get_operator_logs(since_hours=0.1))"
    ]
   },
   {

examples/inference/SDK/inference-s3-model-e2e.ipynb

@@ -7,10 +7,19 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "from sagemaker.hyperpod.hyperpod_manager import HyperPodManager\n",
-    "\n",
-    "HyperPodManager.list_clusters(region='us-east-2')\n",
-    "HyperPodManager.set_context('<hyperpod-cluster-name>', region='us-east-2')"
+    "from sagemaker.hyperpod import list_clusters, set_cluster_context\n",
+    "list_clusters(region='us-east-2')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "14cd61ab",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# choose the HP cluster\n",
+    "set_cluster_context('<my-cluster>', region='us-east-2')"
    ]
   },
   {
@@ -33,13 +42,13 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "tls_config=TlsConfig(tls_certificate_output_s3_uri='s3://<your-tls-bucket-name>')\n",
+    "tls_config=TlsConfig(tls_certificate_output_s3_uri='s3://<my-tls-bucket-name>')\n",
     "\n",
     "model_source_config = ModelSourceConfig(\n",
     "    model_source_type='s3',\n",
-    "    model_location=\"<your-model-folder-in-s3>\",\n",
+    "    model_location=\"<my-model-folder-in-s3>\",\n",
     "    s3_storage=S3Storage(\n",
-    "        bucket_name='<your-model-artifacts-bucket>',\n",
+    "        bucket_name='<my-model-artifacts-bucket>',\n",
     "        region='us-east-2',\n",
     "    ),\n",
     ")\n",
@@ -67,7 +76,7 @@
     "\n",
     "# Create dimensions\n",
     "dimensions = [\n",
-    "    Dimensions(name=\"EndpointName\", value=\"<your-endpoint-name>\"),\n",
+    "    Dimensions(name=\"EndpointName\", value=\"<my-endpoint-name>\"),\n",
     "    Dimensions(name=\"VariantName\", value=\"AllTraffic\")\n",
     "]\n",
     "\n",
@@ -102,7 +111,7 @@
    "outputs": [],
    "source": [
     "s3_endpoint = HPEndpoint(\n",
-    "    endpoint_name='s3-test-endpoint-name',\n",
+    "    endpoint_name='<my-endpoint-name>',\n",
     "    instance_type='ml.g5.8xlarge',\n",
     "    model_name='deepseek15b-test-model-name',  \n",
     "    tls_config=tls_config,\n",
@@ -120,7 +129,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "s3_endpoint.create(debug=True)"
+    "s3_endpoint.create()"
    ]
   },
   {
@@ -193,7 +202,17 @@
    "outputs": [],
    "source": [
     "endpoint_list = HPEndpoint.list()\n",
-    "print_yaml(endpoint_list[1])"
+    "print_yaml(endpoint_list[0])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "660e8d47",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "s3_endpoint = HPEndpoint.get(name='<my-endpoint-name>')"
    ]
   },
   {
@@ -206,10 +225,7 @@
    "outputs": [],
    "source": [
     "# get operator logs\n",
-    "print(s3_endpoint.get_operator_logs(since_hours=0.5))\n",
-    "\n",
-    "# get specific pod log\n",
-    "# js_endpoint.get_logs(pod='pod-name')"
+    "print(s3_endpoint.get_operator_logs(since_hours=0.1))"
    ]
   },
   {