Run this notebook online:

vLLM LLAMA-13B rollingbatch deployment guide¶

In this tutorial, you will use LMI container from DLC to SageMaker and run inference with it.

Please make sure the following permission granted before running the notebook:

SageMaker access

Step 1: Let's bump up SageMaker and import stuff¶

%pip install sagemaker --upgrade  --quiet

import sagemaker
from sagemaker.djl_inference.model import DJLModel

role = sagemaker.get_execution_role()  # execution role for the endpoint
session = sagemaker.session.Session()  # sagemaker session for interacting with different AWS APIs

Step 2: Start building SageMaker endpoint¶

In this step, we will build SageMaker endpoint from scratch

Getting the container image URI (optional)¶

Check out available images: Large Model Inference available DLC

# Choose a specific version of LMI image directly:
# image_uri = "763104351884.dkr.ecr.us-west-2.amazonaws.com/djl-inference:0.28.0-lmi10.0.0-cu124"

Create SageMaker model¶

Here we are using LMI PySDK to create the model.

Checkout more configuration options.

model_id = "OpenAssistant/llama2-13b-orca-8k-3319" # model will be download form Huggingface hub

env = {
    "TENSOR_PARALLEL_DEGREE": "4",            # use 4 GPUs
    "OPTION_ROLLING_BATCH": "vllm",           # use vllm for rolling batching
    "OPTION_TRUST_REMOTE_CODE": "true",
}

model = DJLModel(
    model_id=model_id,
    env=env,
    role=role)

Create SageMaker endpoint¶

You need to specify the instance to use and endpoint names

instance_type = "ml.g5.12xlarge"
endpoint_name = sagemaker.utils.name_from_base("lmi-model")

predictor = model.deploy(initial_instance_count=1,
             instance_type=instance_type,
             endpoint_name=endpoint_name,
             # container_startup_health_check_timeout=3600
            )

Step 3: Test and benchmark the inference¶

predictor.predict(
    {"inputs": "tell me a story of the little red riding hood", "parameters": {}}
)

%%timeit -n3 -r1
predictor.predict(
    {"inputs": "tell me a story of the little red riding hood", "parameters": {}}
)

Clean up the environment¶

session.delete_endpoint(endpoint_name)
session.delete_endpoint_config(endpoint_name)
model.delete_model()