Deploy Oss Sagemaker

1# Cell 0 [markdown]: Model Deployment — SageMaker
2
3# Cell 1: Setup
4
5# %pip install --upgrade 'sagemaker>=3.17.0,<4.0' --quiet  # NOTEBOOK_ONLY
6
7# Cell 2: Configuration
8
9import json
10import os
11
12os.environ["AWS_DEFAULT_REGION"] = "[REGION]"
13
14from sagemaker.core import Attribution, set_attribution
15from sagemaker.core.resources import TrainingJob
16from sagemaker.serve import ModelBuilder
17
18set_attribution(Attribution.SAGEMAKER_AGENT_PLUGIN)
19
20TRAINING_JOB_NAME = "[TRAINING_JOB_NAME]"
21ROLE_ARN = "[ROLE_ARN]"
22INSTANCE_TYPE = "[INSTANCE_TYPE]"
23ENDPOINT_NAME = "[ENDPOINT_NAME]"
24ADAPTER_IC_NAME = f"{ENDPOINT_NAME}-adapter"
25# NOTE FOR AGENT: Only include the ACCEPT_EULA line below for Meta/Llama models.
26# For all other models (Apache 2.0, MIT, Qwen License, etc.) remove the ACCEPT_EULA
27# variable and the model_builder.accept_eula line entirely — they do not apply.
28ACCEPT_EULA = [ACCEPT_EULA]  # Meta/Llama only — remove this line for non-Meta models
29
30# Cell 3: Build Model
31
32training_job = TrainingJob.get(training_job_name=TRAINING_JOB_NAME)
33print(f"Training job: {training_job.training_job_name}")
34print(f"Model package: {training_job.output_model_package_arn}")
35
36model_builder = ModelBuilder(
37    model=training_job,
38    role_arn=ROLE_ARN,
39)
40# Apply the hosting config selected in Step 2 (image, env, and compute for this instance).
41model_builder.set_deployment_config(instance_type=INSTANCE_TYPE)
42# NOTE FOR AGENT: Only include model_builder.accept_eula for Meta/Llama models.
43# Remove this line for all other models.
44model_builder.accept_eula = ACCEPT_EULA  # Meta/Llama only — remove for non-Meta models
45model = model_builder.build(model_name=ENDPOINT_NAME)
46print(f"Model: {model.model_arn}")
47
48# Cell 4: Deploy Endpoint
49
50endpoint = model_builder.deploy(
51    endpoint_name=ENDPOINT_NAME,
52    inference_component_name=ADAPTER_IC_NAME,
53)
54print(f"Endpoint: {endpoint.endpoint_name}")
55
56# Cell 5: Test Inference
57
58output = endpoint.invoke(
59    body=json.dumps(
60        {
61            "inputs": "What is the capital of France?",
62            "parameters": {"max_new_tokens": 50},
63        }
64    ),
65    inference_component_name=ADAPTER_IC_NAME,
66)
67print(f"Response: {output.body.read()}")
68
69# Cell 6: Save Manifest
70# Save manifest - record output of workflow step for future reference
71from pathlib import Path
72
73manifest_dir = Path("[PROJECT_DIR]") / "manifests"
74manifest_dir.mkdir(parents=True, exist_ok=True)
75manifest_path = manifest_dir / f"deploy-{ENDPOINT_NAME}.json"
76manifest_path.write_text(
77    json.dumps(
78        {
79            "endpoint_name": ENDPOINT_NAME,
80        },
81        indent=2,
82    )
83)
84print(f"Manifest saved: {manifest_path}")