Modal vs AWS Lambda: Which serverless platform for AI workloads?

import modal
from modal import Image, asgi_app
import os

image = Image.debian_slim()\
    .pip_install("vllm", "pydantic")

app = modal.App(name="llm-api", image=image)

@app.cls(gpu="a100", container_idle_timeout=600)
class LLMModel:
    def __init__(self):
        from vllm import LLM
        # Modal's GPU selection abstracts CUDA/ROCm setup
        self.llm = LLM("meta-llama/Llama-2-7b-hf", tensor_parallel_size=1)
    
    def generate(self, prompt: str) -> str:
        output = self.llm.generate(prompt, max_tokens=100)
        return output[0].outputs[0].text

@app.function()
@asgi_app()
def api():
    from fastapi import FastAPI
    app = FastAPI()
    model = LLMModel()
    
    @app.post("/generate")
    async def generate(prompt: str):
        return {"response": model.generate(prompt)}
    
    return app

if __name__ == "__main__":
    modal.serve(api.web_endpoint)

import json
import os
import boto3

# Lambda requires manual model management or external service calls
def lambda_handler(event, context):
    # For GPU inference, you'd call SageMaker instead: Lambda itself can't host 7B models
    # This example uses smaller CPU model or external inference API
    
    import torch
    from transformers import pipeline
    
    # Model loading happens on every cold start (slow for GPU)
    try:
        generator = pipeline(
            "text-generation",
            model="gpt2",  # Small model only: 7B won't fit Lambda's memory/storage
            device=-1  # CPU only: GPU requires provisioned concurrency + g4dn
        )
    except Exception as e:
        return {
            "statusCode": 500,
            "body": json.dumps({"error": f"Model load failed: {str(e)}"})
        }
    
    prompt = json.loads(event.get("body", "{}")).get("prompt", "")
    
    try:
        # Lambda: model reloading on every invocation + CPU inference = 2-5 seconds latency minimum
        output = generator(prompt, max_length=100, do_sample=True)[0]["generated_text"]
        return {
            "statusCode": 200,
            "body": json.dumps({"response": output})
        }
    except Exception as e:
        return {
            "statusCode": 500,
            "body": json.dumps({"error": str(e)})
        }