LLM API Gateway on
AWS Lambda

import json import boto3 from transformers import AutoModelForCausalLM, AutoTokenizer # 全局模型缓存（容器复用） model = None tokenizer = None def load_model(): global model, tokenizer if model is None: # 从S3加载模型到EFS model_path = "/mnt/efs/models/llama-7b" tokenizer = AutoTokenizer.from_pretrained(model_path) model = AutoModelForCausalLM.from_pretrained( model_path, device_map="auto", torch_dtype="float16" ) return model, tokenizer def lambda_handler(event, context): # 加载模型（容器复用时跳过） model, tokenizer = load_model() # 解析请求 body = json.loads(event['body']) prompt = body['prompt'] max_tokens = body.get('max_tokens', 100) # 生成响应 inputs = tokenizer(prompt, return_tensors="pt") outputs = model.generate( **inputs, max_new_tokens=max_tokens, temperature=0.7 ) response = tokenizer.decode(outputs[0]) return { 'statusCode': 200, 'body': json.dumps({'response': response}) }