# NutriPaw AI — Modal Serverless GPU Deployment # Skill: modal-serverless-gpu-cloud | CULTIVA IA import modal from pydantic import BaseModel app = modal.App("nutripaw-ai") # ── Imagen con vLLM + dependencias ────────────────── image = ( modal.Image.debian_slim(python_version="3.11") .uv_pip_install( "vllm==0.6.4", "fastapi==0.115.0", "pydantic==2.8.0", ) ) # ── Volumen persistente para pesos del modelo ──────── weights_vol = modal.Volume.from_name( "nutripaw-model-weights", create_if_missing=True, ) # Ruta: /weights/nutripaw-llama3-8b-q4.gguf (16 GB) # ── Secret HuggingFace ─────────────────────────────── hf_secret = modal.Secret.from_name("huggingface-token") # ── Clase principal con lifecycle hooks ────────────── @app.cls( gpu="L40S", # 48 GB VRAM, mejor precio/perf image=image, volumes={"/weights": weights_vol}, secrets=[hf_secret], min_containers=1, # Warm: latencia baja max_containers=10, scaledown_window=600, ) class NutriPawAI: @modal.enter() def load_model(self): from vllm import LLM, SamplingParams self.llm = LLM( model="/weights/nutripaw-llama3-8b-q4", dtype="float16", max_model_len=4096, ) self.sampling = SamplingParams( temperature=0.3, max_tokens=512, ) # ── Web endpoint (POST /generate) ──────────────── @modal.method() @modal.fastapi_endpoint(method="POST", path="/generate") def generate_plan(self, request: "PetProfileRequest"): prompt = build_prompt(request) outputs = self.llm.generate([prompt], self.sampling) return { "plan": outputs[0].outputs[0].text, "tokens": len(outputs[0].outputs[0].token_ids), } @modal.exit() def cleanup(self): del self.llm # ── Batch: regenerar 5.000 planes en paralelo ──────── @app.function( gpu="L40S", image=image, cpu=4.0, memory=16384, volumes={"/weights": weights_vol}, max_containers=20, timeout=1800, ) def batch_generate_plan(pet_profile: dict) -> dict: # Cada invocación = 1 mascota; .map() paralela from vllm import LLM, SamplingParams llm = LLM(model="/weights/nutripaw-llama3-8b-q4") out = llm.generate([build_prompt(pet_profile)]) return {"pet_id": pet_profile["id"], "plan": out[0].outputs[0].text} # ── Job nocturno programado 02:00 UTC ──────────────── @app.function( schedule=modal.Cron("0 2 * * *"), secrets=[modal.Secret.from_name("nutripaw-db")], timeout=3600, ) def nightly_batch(): import os, json profiles = fetch_all_pets_from_db(os.environ["DATABASE_URL"]) results = list(batch_generate_plan.map(profiles)) save_plans_to_db(results) print(f"✓ {len(results)} planes regenerados")