|
| 1 | +import argparse |
| 2 | +import json |
| 3 | +import os |
| 4 | +import sagemaker |
| 5 | +from sagemaker.model import Model |
| 6 | +from sagemaker import serializers |
| 7 | +from sagemaker.predictor import Predictor |
| 8 | + |
| 9 | + |
| 10 | +def deploy_endpoint( |
| 11 | + endpoint_name, container_uri, iam_role, instance_type, model_id, hf_token |
| 12 | +): |
| 13 | + """Deploy vLLM model to SageMaker endpoint""" |
| 14 | + try: |
| 15 | + print(f"Starting deployment of endpoint: {endpoint_name}") |
| 16 | + print(f"Using image: {container_uri}") |
| 17 | + print(f"Instance type: {instance_type}") |
| 18 | + |
| 19 | + print("Creating SageMaker model...") |
| 20 | + model = Model( |
| 21 | + name=endpoint_name, |
| 22 | + image_uri=container_uri, |
| 23 | + role=iam_role, |
| 24 | + env={ |
| 25 | + "SM_VLLM_MODEL": model_id, # Model to load |
| 26 | + "SM_VLLM_HF_TOKEN": hf_token, # HuggingFace token for model access |
| 27 | + }, |
| 28 | + ) |
| 29 | + print("Model created successfully") |
| 30 | + print("Starting endpoint deployment (this may take 10-15 minutes)...") |
| 31 | + |
| 32 | + model.deploy( |
| 33 | + instance_type=instance_type, |
| 34 | + initial_instance_count=1, |
| 35 | + endpoint_name=endpoint_name, |
| 36 | + wait=True, # Wait for deployment to complete |
| 37 | + ) |
| 38 | + print(f"Endpoint {endpoint_name} deployed successfully") |
| 39 | + return True |
| 40 | + except Exception as e: |
| 41 | + print(f"Deployment failed: {str(e)}") |
| 42 | + return False |
| 43 | + |
| 44 | + |
| 45 | +def cleanup_endpoint(endpoint_name): |
| 46 | + """Delete SageMaker endpoint and model""" |
| 47 | + try: |
| 48 | + import boto3 |
| 49 | + |
| 50 | + sagemaker_client = boto3.client("sagemaker") |
| 51 | + |
| 52 | + print(f"Cleaning up endpoint: {endpoint_name}") |
| 53 | + sagemaker_client.delete_endpoint(EndpointName=endpoint_name) |
| 54 | + sagemaker_client.delete_endpoint_config(EndpointConfigName=endpoint_name) |
| 55 | + sagemaker_client.delete_model(ModelName=endpoint_name) |
| 56 | + print(f"Endpoint {endpoint_name} cleaned up successfully") |
| 57 | + return True |
| 58 | + except Exception as e: |
| 59 | + print(f"Cleanup failed: {str(e)}") |
| 60 | + return False |
| 61 | + |
| 62 | + |
| 63 | +def invoke_endpoint(endpoint_name, prompt, max_tokens=2400, temperature=0.01): |
| 64 | + """Invoke SageMaker endpoint with vLLM model for text generation""" |
| 65 | + try: |
| 66 | + predictor = Predictor( |
| 67 | + endpoint_name=endpoint_name, |
| 68 | + serializer=serializers.JSONSerializer(), |
| 69 | + ) |
| 70 | + |
| 71 | + payload = { |
| 72 | + "messages": [{"role": "user", "content": prompt}], # Chat format |
| 73 | + "max_tokens": max_tokens, # Response length limit |
| 74 | + "temperature": temperature, # Randomness (0=deterministic, 1=creative) |
| 75 | + "top_p": 0.9, # Nucleus sampling |
| 76 | + "top_k": 50, # Top-k sampling |
| 77 | + } |
| 78 | + |
| 79 | + response = predictor.predict(payload) |
| 80 | + |
| 81 | + # Handle different response formats |
| 82 | + if isinstance(response, bytes): |
| 83 | + response = response.decode("utf-8") |
| 84 | + |
| 85 | + if isinstance(response, str): |
| 86 | + try: |
| 87 | + response = json.loads(response) |
| 88 | + except json.JSONDecodeError: |
| 89 | + print("Warning: Response is not valid JSON. Returning as string.") |
| 90 | + |
| 91 | + return response |
| 92 | + |
| 93 | + except Exception as e: |
| 94 | + print(f"Inference failed: {str(e)}") |
| 95 | + return None |
| 96 | + |
| 97 | + |
| 98 | +def main(): |
| 99 | + parser = argparse.ArgumentParser(description="SageMaker vLLM Inference") |
| 100 | + parser.add_argument( |
| 101 | + "--endpoint-name", required=True, help="SageMaker endpoint name" |
| 102 | + ) |
| 103 | + parser.add_argument( |
| 104 | + "--container-uri", |
| 105 | + help="DLC image URI", |
| 106 | + default=os.getenv( |
| 107 | + "CONTAINER_URI", |
| 108 | + "public.ecr.aws/deep-learning-containers/vllm:0.11.0-gpu-py312", |
| 109 | + ), |
| 110 | + ) |
| 111 | + parser.add_argument( |
| 112 | + "--iam-role", help="IAM role ARN", default=os.getenv("IAM_ROLE") |
| 113 | + ) |
| 114 | + parser.add_argument( |
| 115 | + "--instance-type", default="ml.g5.12xlarge", help="Instance type" |
| 116 | + ) |
| 117 | + parser.add_argument( |
| 118 | + "--model-id", |
| 119 | + default="deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B", |
| 120 | + help="HuggingFace model ID", |
| 121 | + ) |
| 122 | + parser.add_argument( |
| 123 | + "--hf-token", help="HuggingFace token", default=os.getenv("HF_TOKEN", "") |
| 124 | + ) |
| 125 | + parser.add_argument( |
| 126 | + "--prompt", |
| 127 | + default="Write a python code to generate n prime numbers", |
| 128 | + help="Inference prompt", |
| 129 | + ) |
| 130 | + parser.add_argument("--max-tokens", type=int, default=2400, help="Maximum tokens") |
| 131 | + parser.add_argument( |
| 132 | + "--temperature", type=float, default=0.01, help="Sampling temperature" |
| 133 | + ) |
| 134 | + |
| 135 | + args = parser.parse_args() |
| 136 | + |
| 137 | + if not args.iam_role: |
| 138 | + print("Error: IAM role required") |
| 139 | + return |
| 140 | + |
| 141 | + # Deploy endpoint |
| 142 | + if not deploy_endpoint( |
| 143 | + args.endpoint_name, |
| 144 | + args.container_uri, |
| 145 | + args.iam_role, |
| 146 | + args.instance_type, |
| 147 | + args.model_id, |
| 148 | + args.hf_token, |
| 149 | + ): |
| 150 | + return |
| 151 | + |
| 152 | + # Run inference |
| 153 | + print("\nSending request to endpoint...") |
| 154 | + response = invoke_endpoint( |
| 155 | + endpoint_name=args.endpoint_name, |
| 156 | + prompt=args.prompt, |
| 157 | + max_tokens=args.max_tokens, |
| 158 | + temperature=args.temperature, |
| 159 | + ) |
| 160 | + |
| 161 | + if response: |
| 162 | + print("\nResponse from endpoint:") |
| 163 | + if isinstance(response, (dict, list)): |
| 164 | + print(json.dumps(response, indent=2)) |
| 165 | + else: |
| 166 | + print(response) |
| 167 | + else: |
| 168 | + print("No response received from the endpoint.") |
| 169 | + |
| 170 | + # Cleanup |
| 171 | + print("\nCleaning up resources...") |
| 172 | + cleanup_endpoint(args.endpoint_name) |
| 173 | + |
| 174 | + |
| 175 | +if __name__ == "__main__": |
| 176 | + main() |
0 commit comments