app.py

import os
import json
import logging
import boto3
from fastapi import FastAPI, HTTPException
from fastapi.responses import JSONResponse
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import asyncio
import concurrent.futures

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

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")

MAX_TOKENS = 1024  # Limite de tokens por fragmento

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()

PIPELINE_MAP = {
    "text-generation": "text-generation",
    "sentiment-analysis": "sentiment-analysis",
    "translation": "translation",
    "fill-mask": "fill-mask",
    "question-answering": "question-answering",
    "text-to-speech": "text-to-speech",
    "text-to-video": "text-to-video",
    "text-to-image": "text-to-image"
}

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

    async def stream_from_s3(self, key):
        loop = asyncio.get_event_loop()
        return await loop.run_in_executor(None, self._stream_from_s3, key)

    def _stream_from_s3(self, key):
        try:
            response = self.s3_client.get_object(Bucket=self.bucket_name, Key=key)
            return response['Body']
        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:
            raise HTTPException(status_code=500, detail=f"Error al descargar {key} desde S3: {str(e)}")

    async def get_model_file_parts(self, model_name):
        loop = asyncio.get_event_loop()
        return await loop.run_in_executor(None, self._get_model_file_parts, model_name)

    def _get_model_file_parts(self, model_name):
        try:
            model_prefix = model_name.lower()
            files = self.s3_client.list_objects_v2(Bucket=self.bucket_name, Prefix=model_prefix)
            model_files = [obj['Key'] for obj in files.get('Contents', []) if model_prefix in obj['Key']]
            return model_files
        except Exception as e:
            raise HTTPException(status_code=500, detail=f"Error al obtener archivos del modelo {model_name} desde S3: {e}")

    async def load_model_from_s3(self, model_name):
        try:
            profile, model = model_name.split("/", 1) if "/" in model_name else ("", model_name)

            model_prefix = f"{profile}/{model}".lower()
            model_files = await 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.")

            config_stream = await self.stream_from_s3(f"{model_prefix}/config.json")
            config_data = config_stream.read()

            if not config_data:
                raise HTTPException(status_code=500, detail=f"El archivo de configuración {model_prefix}/config.json está vacío.")
            
            config_text = config_data.decode("utf-8")
            config_json = json.loads(config_text)

            model = AutoModelForCausalLM.from_pretrained(f"s3://{self.bucket_name}/{model_prefix}", config=config_json, from_tf=False)
            return model

        except HTTPException as e:
            raise e
        except Exception as e:
            raise HTTPException(status_code=500, detail=f"Error al cargar el modelo desde S3: {e}")

    async def load_tokenizer_from_s3(self, model_name):
        try:
            profile, model = model_name.split("/", 1) if "/" in model_name else ("", model_name)

            tokenizer_stream = await self.stream_from_s3(f"{profile}/{model}/tokenizer.json")
            tokenizer_data = tokenizer_stream.read().decode("utf-8")

            tokenizer = AutoTokenizer.from_pretrained(f"{profile}/{model}")
            return tokenizer
        except Exception as e:
            raise HTTPException(status_code=500, detail=f"Error al cargar el tokenizer desde S3: {e}")

    async def create_s3_folders(self, s3_key):
        try:
            folder_keys = s3_key.split('/')
            for i in range(1, len(folder_keys)):
                folder_key = '/'.join(folder_keys[:i]) + '/'
                if not await self.file_exists_in_s3(folder_key):
                    self.s3_client.put_object(Bucket=self.bucket_name, Key=folder_key, Body='')

        except Exception as e:
            raise HTTPException(status_code=500, detail=f"Error al crear carpetas en S3: {e}")

    async def file_exists_in_s3(self, s3_key):
        try:
            self.s3_client.head_object(Bucket=self.bucket_name, Key=s3_key)
            return True
        except self.s3_client.exceptions.ClientError:
            return False

def split_text_by_tokens(text, tokenizer, max_tokens=MAX_TOKENS):
    tokens = tokenizer.encode(text)
    chunks = []
    for i in range(0, len(tokens), max_tokens):
        chunk = tokens[i:i+max_tokens]
        chunks.append(tokenizer.decode(chunk))
    return chunks

def continue_generation(input_text, model, tokenizer, max_tokens=MAX_TOKENS):
    generated_text = ""
    while len(input_text) > 0:
        tokens = tokenizer.encode(input_text)
        input_text = tokenizer.decode(tokens[:max_tokens])
        output = model.generate(input_ids=tokenizer.encode(input_text, return_tensors="pt").input_ids)
        generated_text += tokenizer.decode(output[0], skip_special_tokens=True)
        input_text = input_text[len(input_text):]  # Si la entrada se agotó, ya no hay más que procesar
    return generated_text

@app.post("/predict/")
async def predict(model_request: dict):
    try:
        model_name = model_request.get("model_name")
        task = model_request.get("pipeline_task")
        input_text = model_request.get("input_text")

        if not model_name or not task or not input_text:
            raise HTTPException(status_code=400, detail="Faltan parámetros en la solicitud.")

        streamer = S3DirectStream(S3_BUCKET_NAME)
        model = await streamer.load_model_from_s3(model_name)
        tokenizer = await streamer.load_tokenizer_from_s3(model_name)

        if task not in PIPELINE_MAP:
            raise HTTPException(status_code=400, detail="Pipeline task no soportado")

        nlp_pipeline = pipeline(PIPELINE_MAP[task], model=model, tokenizer=tokenizer)

        result = await asyncio.to_thread(nlp_pipeline, input_text)

        chunks = split_text_by_tokens(result, tokenizer)

        if len(chunks) > 1:
            full_result = ""
            for chunk in chunks:
                full_result += continue_generation(chunk, model, tokenizer)
            return JSONResponse(content={"result": full_result})
        else:
            return JSONResponse(content={"result": result})

    except Exception as e:
        raise HTTPException(status_code=500, detail=f"Error al realizar la predicción: {e}")

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=7860)