|
import os |
|
import json |
|
import boto3 |
|
import uvicorn |
|
from fastapi import FastAPI, HTTPException |
|
from pydantic import BaseModel |
|
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline |
|
from huggingface_hub import hf_hub_download |
|
from io import BytesIO |
|
import torch |
|
import safetensors |
|
from dotenv import load_dotenv |
|
import tqdm |
|
import re |
|
|
|
|
|
load_dotenv() |
|
|
|
|
|
AWS_ACCESS_KEY_ID = os.getenv("AWS_ACCESS_KEY_ID") |
|
AWS_SECRET_ACCESS_KEY = os.getenv("AWS_SECRET_ACCESS_KEY") |
|
AWS_REGION = os.getenv("AWS_REGION") |
|
S3_BUCKET_NAME = os.getenv("S3_BUCKET_NAME") |
|
HUGGINGFACE_TOKEN = os.getenv("HUGGINGFACE_TOKEN") |
|
|
|
|
|
s3_client = boto3.client( |
|
's3', |
|
aws_access_key_id=AWS_ACCESS_KEY_ID, |
|
aws_secret_access_key=AWS_SECRET_ACCESS_KEY, |
|
region_name=AWS_REGION |
|
) |
|
|
|
app = FastAPI() |
|
|
|
|
|
class DownloadModelRequest(BaseModel): |
|
model_name: str |
|
pipeline_task: str |
|
input_text: str |
|
|
|
class S3DirectStream: |
|
def __init__(self, bucket_name): |
|
self.s3_client = boto3.client( |
|
's3', |
|
aws_access_key_id=AWS_ACCESS_KEY_ID, |
|
aws_secret_access_key=AWS_SECRET_ACCESS_KEY, |
|
region_name=AWS_REGION |
|
) |
|
self.bucket_name = bucket_name |
|
|
|
def stream_from_s3(self, key): |
|
try: |
|
print(f"[INFO] Descargando archivo {key} desde S3...") |
|
response = self.s3_client.get_object(Bucket=self.bucket_name, Key=key) |
|
return BytesIO(response['Body'].read()) |
|
except self.s3_client.exceptions.NoSuchKey: |
|
raise HTTPException(status_code=404, detail=f"El archivo {key} no existe en el bucket S3.") |
|
except Exception as e: |
|
print(f"[ERROR] Error al descargar {key}: {str(e)}") |
|
raise HTTPException(status_code=500, detail=f"Error al descargar archivo {key} desde S3.") |
|
|
|
def file_exists_in_s3(self, key): |
|
try: |
|
self.s3_client.head_object(Bucket=self.bucket_name, Key=key) |
|
return True |
|
except self.s3_client.exceptions.ClientError: |
|
return False |
|
|
|
def load_model_from_s3(self, model_name): |
|
try: |
|
print(f"[INFO] Cargando el modelo {model_name} desde S3...") |
|
|
|
|
|
model_prefix = model_name.lower() |
|
model_files = self.get_model_file_parts(model_prefix) |
|
|
|
if not model_files: |
|
print(f"[INFO] El modelo {model_name} no est谩 en S3. Procediendo a descargar desde Hugging Face...") |
|
self.download_and_upload_from_huggingface(model_name) |
|
model_files = self.get_model_file_parts(model_prefix) |
|
|
|
if not model_files: |
|
raise HTTPException(status_code=404, detail=f"Archivos del modelo {model_name} no encontrados en S3.") |
|
|
|
|
|
model_streams = [] |
|
for model_file in tqdm.tqdm(model_files, desc="Cargando archivos del modelo", unit="archivo"): |
|
model_streams.append(self.stream_from_s3(model_file)) |
|
|
|
config_stream = self.stream_from_s3(f"{model_prefix}/config.json") |
|
config_data = json.loads(config_stream.read().decode("utf-8")) |
|
|
|
|
|
if any(file.endswith("model.safetensors") for file in model_files): |
|
print("[INFO] Cargando el modelo como safetensor...") |
|
model = AutoModelForCausalLM.from_config(config_data) |
|
model.load_state_dict(safetensors.torch.load_stream(model_streams[0])) |
|
else: |
|
print("[INFO] Cargando el modelo como archivo binario de PyTorch...") |
|
model = AutoModelForCausalLM.from_config(config_data) |
|
model.load_state_dict(torch.load(model_streams[0], map_location="cpu")) |
|
|
|
print("[INFO] Modelo cargado con 茅xito.") |
|
return model |
|
|
|
except Exception as e: |
|
print(f"[ERROR] Error al cargar el modelo desde S3: {e}") |
|
raise HTTPException(status_code=500, detail="Error al cargar el modelo desde S3.") |
|
|
|
def load_tokenizer_from_s3(self, model_name): |
|
try: |
|
print(f"[INFO] Cargando el tokenizer {model_name} desde S3...") |
|
tokenizer_stream = self.stream_from_s3(f"{model_name}/tokenizer.json") |
|
tokenizer = AutoTokenizer.from_pretrained(tokenizer_stream) |
|
return tokenizer |
|
except Exception as e: |
|
print(f"[ERROR] Error al cargar el tokenizer desde S3: {e}") |
|
raise HTTPException(status_code=500, detail="Error al cargar el tokenizer desde S3.") |
|
|
|
def get_model_file_parts(self, model_name): |
|
print(f"[INFO] Listando archivos del modelo en S3 con prefijo {model_name}...") |
|
files = self.s3_client.list_objects_v2(Bucket=self.bucket_name, Prefix=model_name) |
|
model_files = [] |
|
for obj in tqdm.tqdm(files.get('Contents', []), desc="Verificando archivos", unit="archivo"): |
|
key = obj['Key'] |
|
|
|
if re.match(rf"{model_name}/.*", key): |
|
model_files.append(key) |
|
if not model_files: |
|
print(f"[WARNING] No se encontraron archivos para el modelo {model_name}.") |
|
return model_files |
|
|
|
def download_and_upload_from_huggingface(self, model_name): |
|
try: |
|
print(f"[INFO] Descargando {model_name} desde Hugging Face...") |
|
|
|
files_to_download = hf_hub_download(repo_id=model_name, use_auth_token=HUGGINGFACE_TOKEN) |
|
|
|
|
|
for file in files_to_download: |
|
file_name = os.path.basename(file) |
|
s3_key = f"{model_name}/{file_name}" |
|
if not self.file_exists_in_s3(s3_key): |
|
self.upload_file_to_s3(file, s3_key) |
|
|
|
except Exception as e: |
|
print(f"[ERROR] Error al descargar y subir modelo desde Hugging Face: {e}") |
|
raise HTTPException(status_code=500, detail="Error al descargar y subir modelo desde Hugging Face.") |
|
|
|
def upload_file_to_s3(self, file_path, s3_key): |
|
try: |
|
print(f"[INFO] Subiendo archivo {file_path} a S3 con key {s3_key}...") |
|
with open(file_path, 'rb') as data: |
|
self.s3_client.put_object(Bucket=self.bucket_name, Key=s3_key, Body=data) |
|
|
|
os.remove(file_path) |
|
except Exception as e: |
|
print(f"[ERROR] Error al subir archivo a S3: {e}") |
|
raise HTTPException(status_code=500, detail="Error al subir archivo a S3.") |
|
|
|
|
|
@app.post("/predict/") |
|
async def predict(model_request: DownloadModelRequest): |
|
try: |
|
print(f"[INFO] Recibiendo solicitud para predecir con el modelo {model_request.model_name}...") |
|
streamer = S3DirectStream(S3_BUCKET_NAME) |
|
model = streamer.load_model_from_s3(model_request.model_name) |
|
tokenizer = streamer.load_tokenizer_from_s3(model_request.model_name) |
|
|
|
task = model_request.pipeline_task |
|
if task not in ["text-generation", "sentiment-analysis", "translation", "fill-mask", "question-answering", |
|
"text-to-speech", "text-to-video", "text-to-image"]: |
|
raise HTTPException(status_code=400, detail="Pipeline task no soportado") |
|
|
|
|
|
nlp_pipeline = None |
|
if task == "text-generation": |
|
nlp_pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer) |
|
elif task == "sentiment-analysis": |
|
nlp_pipeline = pipeline("sentiment-analysis", model=model, tokenizer=tokenizer) |
|
elif task == "translation": |
|
nlp_pipeline = pipeline("translation", model=model, tokenizer=tokenizer) |
|
elif task == "fill-mask": |
|
nlp_pipeline = pipeline("fill-mask", model=model, tokenizer=tokenizer) |
|
elif task == "question-answering": |
|
nlp_pipeline = pipeline("question-answering", model=model, tokenizer=tokenizer) |
|
elif task == "text-to-speech": |
|
nlp_pipeline = pipeline("text-to-speech", model=model, tokenizer=tokenizer) |
|
elif task == "text-to-video": |
|
nlp_pipeline = pipeline("text-to-video", model=model, tokenizer=tokenizer) |
|
elif task == "text-to-image": |
|
nlp_pipeline = pipeline("text-to-image", model=model, tokenizer=tokenizer) |
|
|
|
result = nlp_pipeline(model_request.input_text) |
|
return {"result": result} |
|
|
|
except Exception as e: |
|
print(f"[ERROR] Error en el proceso de predicci贸n: {str(e)}") |
|
raise HTTPException(status_code=500, detail="Error en el proceso de predicci贸n") |
|
|
|
|
|
|
|
if __name__ == "__main__": |
|
uvicorn.run(app, host="0.0.0.0", port=8000) |
|
|