aqua-deployment

name: aqua-deployment description: Deploy LLM models on OCI using AI Quick Actions (AQUA) - single model, multi-model, stacked (LoRA), with GPU shape selection, vLLM configuration, streaming, and tool calling. Triggered when user wants to deploy, update, or manage model deployments. user-invocable: true disable-model-invocation: false

AQUA Model Deployment

Use this skill when the user wants to deploy, manage, or configure LLM model deployments on OCI Data Science using AI Quick Actions.

Deployment Types

Python SDK Usage

Import

from ads.aqua.modeldeployment import AquaDeploymentApp
deployment_app = AquaDeploymentApp()

Create Single Model Deployment

from ads.aqua.modeldeployment.entities import CreateModelDeploymentDetails

details = CreateModelDeploymentDetails(
    model_id="ocid1.datasciencemodel.oc1.iad.xxx",
    instance_shape="VM.GPU.A10.2",
    display_name="llama-3.1-8b-deployment",
    compartment_id="ocid1.compartment.oc1..xxx",
    project_id="ocid1.datascienceproject.oc1.iad.xxx",
    log_group_id="ocid1.loggroup.oc1.iad.xxx",
    log_id="ocid1.log.oc1.iad.xxx",
    env_var={
        "MODEL_DEPLOY_PREDICT_ENDPOINT": "/v1/completions",
        "PARAMS": "--max-model-len 4096",
    },
)
deployment = deployment_app.create(create_deployment_details=details)
print(f"Deployment: {deployment.id} | State: {deployment.state}")

Create with Chat Completions Endpoint

details = CreateModelDeploymentDetails(
    model_id="ocid1.datasciencemodel.oc1.iad.xxx",
    instance_shape="VM.GPU.A10.2",
    display_name="llama-3.1-8b-chat",
    env_var={
        "MODEL_DEPLOY_PREDICT_ENDPOINT": "/v1/chat/completions",
        "PARAMS": "--max-model-len 4096",
    },
)

Create Multi-Model Deployment

from ads.aqua.common.entities import AquaMultiModelRef

details = CreateModelDeploymentDetails(
    models=[
        AquaMultiModelRef(
            model_id="ocid1.datasciencemodel.oc1.iad.model1",
            model_name="llama-3.1-8b",
            gpu_count=1,
        ),
        AquaMultiModelRef(
            model_id="ocid1.datasciencemodel.oc1.iad.model2",
            model_name="mistral-7b",
            gpu_count=1,
        ),
    ],
    instance_shape="VM.GPU.A10.2",
    display_name="multi-model-deployment",
    compartment_id="ocid1.compartment.oc1..xxx",
    project_id="ocid1.datascienceproject.oc1.iad.xxx",
)
deployment = deployment_app.create(create_deployment_details=details)

Create Stacked Deployment (Base + LoRA Fine-Tunes)

from ads.aqua.common.entities import AquaMultiModelRef, LoraModuleSpec

details = CreateModelDeploymentDetails(
    models=[
        AquaMultiModelRef(
            model_id="ocid1.datasciencemodel.oc1.iad.base_model",
            model_name="llama-3.1-8b",
            fine_tune_weights=[
                LoraModuleSpec(
                    model_id="ocid1.datasciencemodel.oc1.iad.ft1",
                    model_name="llama-3.1-8b-customer-support",
                ),
                LoraModuleSpec(
                    model_id="ocid1.datasciencemodel.oc1.iad.ft2",
                    model_name="llama-3.1-8b-summarization",
                ),
            ],
        ),
    ],
    instance_shape="VM.GPU.A10.2",
    display_name="stacked-llama-deployment",
    deployment_type="STACKED",
)
deployment = deployment_app.create(create_deployment_details=details)

List Deployments

deployments = deployment_app.list(compartment_id="ocid1.compartment.oc1..xxx")
for d in deployments:
    print(f"{d.display_name} | {d.state} | {d.endpoint}")

Get Deployment Details

deployment = deployment_app.get(model_deployment_id="ocid1.datasciencemodeldeployment.oc1.iad.xxx")

Get Deployment Config (Recommended Shapes)

config = deployment_app.get_deployment_config(model_id="ocid1.datasciencemodel.oc1.iad.xxx")

List Available Shapes

shapes = deployment_app.list_shapes(compartment_id="ocid1.compartment.oc1..xxx")

Shape Recommendation

recommendation = deployment_app.recommend_shape(model_id="ocid1.datasciencemodel.oc1.iad.xxx")

CLI Usage

Create Deployment

ads aqua deployment create \
  --model_id "ocid1.datasciencemodel.oc1.iad.xxx" \
  --instance_shape "VM.GPU.A10.2" \
  --display_name "llama-3.1-8b-deployment" \
  --compartment_id "ocid1.compartment.oc1..xxx" \
  --project_id "ocid1.datascienceproject.oc1.iad.xxx" \
  --log_group_id "ocid1.loggroup.oc1.iad.xxx" \
  --log_id "ocid1.log.oc1.iad.xxx"

Create Multi-Model Deployment

ads aqua deployment create \
  --models '[{"model_id":"ocid1...model1","model_name":"llama-8b","gpu_count":1},{"model_id":"ocid1...model2","model_name":"mistral-7b","gpu_count":1}]' \
  --instance_shape "VM.GPU.A10.2" \
  --display_name "multi-model"

Create Stacked Deployment

ads aqua deployment create \
  --models '[{"model_id":"ocid1...base","model_name":"llama-8b","fine_tune_weights":[{"model_id":"ocid1...ft1","model_name":"ft-support"}]}]' \
  --instance_shape "VM.GPU.A10.2" \
  --display_name "stacked-deployment" \
  --deployment_type "STACKED"

List / Get

ads aqua deployment list --compartment_id "ocid1.compartment.oc1..xxx"
ads aqua deployment get --model_deployment_id "ocid1.datasciencemodeldeployment.oc1.iad.xxx"

Invoking a Deployed Model

Python SDK (Streaming)

import ads
import oci
import requests

ads.set_auth("resource_principal")
endpoint = "https://modeldeployment.us-ashburn-1.oci.customer-oci.com/ocid1.datasciencemodeldeployment.oc1.iad.xxx"

# Non-streaming
response = requests.post(
    f"{endpoint}/predict",
    json={
        "model": "odsc-llm",
        "prompt": "Write a haiku about clouds",
        "max_tokens": 256,
        "temperature": 0.7,
    },
    auth=oci.auth.signers.get_resource_principals_signer(),
)
print(response.json())

OpenAI-Compatible Client (ADS)

from ads.aqua.client.openai_client import OpenAI

client = OpenAI(
    model_deployment_url="https://modeldeployment.us-ashburn-1.oci.customer-oci.com/ocid1.datasciencemodeldeployment.oc1.iad.xxx",
    auth={"signer": oci.auth.signers.get_resource_principals_signer()},
)
response = client.chat.completions.create(
    model="odsc-llm",
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=500,
)
print(response.choices[0].message.content)

GPU Shape Reference

Quick sizing rule: GPU_memory_GB = num_params_billions × 2 for FP16/BF16, plus ~20% for KV cache.

For the full shape table, per-model recommendations, multi-model GPU count constraints, and quantization options, see references/shapes.md.

vLLM Configuration Parameters

Set via PARAMS environment variable or --params CLI flag:

Tool Calling / Function Calling

Enable during deployment:

env_var={
    "MODEL_DEPLOY_PREDICT_ENDPOINT": "/v1/chat/completions",
    "PARAMS": "--enable-auto-tool-choice --tool-call-parser llama3_json --max-model-len 4096",
}

Supported parsers: llama3_json, llama4_json, granite, hermes, mistral, jamba, pythonic, internlm.

Advanced Topics

Key Source Files

• ads/aqua/modeldeployment/deployment.py — AquaDeploymentApp (create, list, get, update)
• ads/aqua/modeldeployment/entities.py — CreateModelDeploymentDetails, AquaDeployment
• ads/aqua/common/entities.py — AquaMultiModelRef, LoraModuleSpec
• ads/aqua/client/openai_client.py — OpenAI-compatible client
• ads/aqua/shaperecommend/recommend.py — GPU shape recommendation engine