Spaces:
Sleeping
Sleeping
Hjgugugjhuhjggg
commited on
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,191 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from huggingface_hub import HfApi
|
2 |
+
from fastapi import FastAPI, HTTPException
|
3 |
+
from pydantic import BaseModel
|
4 |
+
import requests
|
5 |
+
import boto3
|
6 |
+
from dotenv import load_dotenv
|
7 |
+
import os
|
8 |
+
import uvicorn
|
9 |
+
from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer
|
10 |
+
import safetensors.torch
|
11 |
+
from fastapi.responses import StreamingResponse
|
12 |
+
from tqdm import tqdm
|
13 |
+
|
14 |
+
# Cargar las variables de entorno desde el archivo .env
|
15 |
+
load_dotenv()
|
16 |
+
|
17 |
+
# Cargar las credenciales de AWS y el token de Hugging Face desde las variables de entorno
|
18 |
+
AWS_ACCESS_KEY_ID = os.getenv("AWS_ACCESS_KEY_ID")
|
19 |
+
AWS_SECRET_ACCESS_KEY = os.getenv("AWS_SECRET_ACCESS_KEY")
|
20 |
+
AWS_REGION = os.getenv("AWS_REGION")
|
21 |
+
S3_BUCKET_NAME = os.getenv("S3_BUCKET_NAME") # Nombre del bucket de S3
|
22 |
+
HUGGINGFACE_TOKEN = os.getenv("HUGGINGFACE_TOKEN") # Token de Hugging Face
|
23 |
+
|
24 |
+
# Cliente S3 de Amazon
|
25 |
+
s3_client = boto3.client(
|
26 |
+
's3',
|
27 |
+
aws_access_key_id=AWS_ACCESS_KEY_ID,
|
28 |
+
aws_secret_access_key=AWS_SECRET_ACCESS_KEY,
|
29 |
+
region_name=AWS_REGION
|
30 |
+
)
|
31 |
+
|
32 |
+
app = FastAPI()
|
33 |
+
|
34 |
+
# Pydantic Model para el cuerpo de la solicitud del endpoint /download_model/
|
35 |
+
class DownloadModelRequest(BaseModel):
|
36 |
+
model_name: str
|
37 |
+
pipeline_task: str
|
38 |
+
input_text: str
|
39 |
+
revision: str = "main" # Revisi贸n por defecto
|
40 |
+
|
41 |
+
class S3DirectStream:
|
42 |
+
def __init__(self, bucket_name):
|
43 |
+
self.s3_client = boto3.client(
|
44 |
+
's3',
|
45 |
+
aws_access_key_id=AWS_ACCESS_KEY_ID,
|
46 |
+
aws_secret_access_key=AWS_SECRET_ACCESS_KEY,
|
47 |
+
region_name=AWS_REGION
|
48 |
+
)
|
49 |
+
self.bucket_name = bucket_name
|
50 |
+
|
51 |
+
def stream_from_s3(self, key):
|
52 |
+
try:
|
53 |
+
print(f"Descargando archivo {key} desde S3...")
|
54 |
+
response = self.s3_client.get_object(Bucket=self.bucket_name, Key=key)
|
55 |
+
return response['Body'] # Devolver el cuerpo directamente para el StreamingResponse
|
56 |
+
except self.s3_client.exceptions.NoSuchKey:
|
57 |
+
raise HTTPException(status_code=404, detail=f"El archivo {key} no existe en el bucket S3.")
|
58 |
+
|
59 |
+
def file_exists_in_s3(self, key):
|
60 |
+
try:
|
61 |
+
self.s3_client.head_object(Bucket=self.bucket_name, Key=key)
|
62 |
+
return True
|
63 |
+
except self.s3_client.exceptions.ClientError:
|
64 |
+
return False
|
65 |
+
|
66 |
+
def load_model_from_stream(self, model_prefix, revision):
|
67 |
+
try:
|
68 |
+
print(f"Cargando el modelo {model_prefix} desde S3...")
|
69 |
+
if self.file_exists_in_s3(f"{model_prefix}/config.json") and \
|
70 |
+
(self.file_exists_in_s3(f"{model_prefix}/pytorch_model.bin") or self.file_exists_in_s3(f"{model_prefix}/model.safetensors")):
|
71 |
+
print(f"Modelo {model_prefix} ya existe en S3. No es necesario descargarlo.")
|
72 |
+
return self.load_model_from_existing_s3(model_prefix)
|
73 |
+
|
74 |
+
print(f"Modelo {model_prefix} no encontrado. Procediendo a descargar...")
|
75 |
+
self.download_and_upload_to_s3(model_prefix, revision) # Pasamos 'revision' aqu铆
|
76 |
+
return self.load_model_from_stream(model_prefix, revision)
|
77 |
+
except HTTPException as e:
|
78 |
+
print(f"Error al cargar el modelo: {e}")
|
79 |
+
return None
|
80 |
+
|
81 |
+
def load_model_from_existing_s3(self, model_prefix):
|
82 |
+
# Cargar el modelo y los archivos necesarios desde S3
|
83 |
+
print(f"Cargando los archivos {model_prefix} desde S3...")
|
84 |
+
config_stream = self.stream_from_s3(f"{model_prefix}/config.json")
|
85 |
+
config_data = config_stream.read().decode("utf-8")
|
86 |
+
|
87 |
+
print(f"Cargando el modelo de lenguaje {model_prefix}...")
|
88 |
+
|
89 |
+
# Verificar si el archivo es un safetensor o un archivo binario
|
90 |
+
if self.file_exists_in_s3(f"{model_prefix}/model.safetensors"):
|
91 |
+
# Usar safetensors si el archivo es de tipo safetensors
|
92 |
+
model_stream = self.stream_from_s3(f"{model_prefix}/model.safetensors")
|
93 |
+
model = AutoModelForCausalLM.from_config(config_data)
|
94 |
+
model.load_state_dict(safetensors.torch.load_stream(model_stream)) # Cargar el modelo utilizando safetensors
|
95 |
+
else:
|
96 |
+
# Cargar el modelo utilizando pytorch si el archivo es .bin
|
97 |
+
model_stream = self.stream_from_s3(f"{model_prefix}/pytorch_model.bin")
|
98 |
+
model = AutoModelForCausalLM.from_config(config_data)
|
99 |
+
model.load_state_dict(torch.load(model_stream, map_location="cpu"))
|
100 |
+
|
101 |
+
return model
|
102 |
+
|
103 |
+
def load_tokenizer_from_stream(self, model_prefix):
|
104 |
+
try:
|
105 |
+
if self.file_exists_in_s3(f"{model_prefix}/tokenizer.json"):
|
106 |
+
print(f"Tokenizer para {model_prefix} ya existe en S3. No es necesario descargarlo.")
|
107 |
+
return self.load_tokenizer_from_existing_s3(model_prefix)
|
108 |
+
|
109 |
+
print(f"Tokenizer para {model_prefix} no encontrado. Procediendo a descargar...")
|
110 |
+
self.download_and_upload_to_s3(model_prefix) # Pasamos 'revision' aqu铆 tambi茅n
|
111 |
+
return self.load_tokenizer_from_stream(model_prefix)
|
112 |
+
except HTTPException as e:
|
113 |
+
print(f"Error al cargar el tokenizer: {e}")
|
114 |
+
return None
|
115 |
+
|
116 |
+
def load_tokenizer_from_existing_s3(self, model_prefix):
|
117 |
+
print(f"Cargando el tokenizer para {model_prefix} desde S3...")
|
118 |
+
tokenizer_stream = self.stream_from_s3(f"{model_prefix}/tokenizer.json")
|
119 |
+
tokenizer = AutoTokenizer.from_pretrained(tokenizer_stream)
|
120 |
+
return tokenizer
|
121 |
+
|
122 |
+
def download_and_upload_to_s3(self, model_prefix, revision):
|
123 |
+
# URLs de los archivos de Hugging Face
|
124 |
+
model_url = f"https://huggingface.co/{model_prefix}/resolve/{revision}/pytorch_model.bin"
|
125 |
+
safetensors_url = f"https://huggingface.co/{model_prefix}/resolve/{revision}/model.safetensors"
|
126 |
+
tokenizer_url = f"https://huggingface.co/{model_prefix}/resolve/{revision}/tokenizer.json"
|
127 |
+
config_url = f"https://huggingface.co/{model_prefix}/resolve/{revision}/config.json"
|
128 |
+
|
129 |
+
print(f"Descargando y subiendo archivos para el modelo {model_prefix} a S3...")
|
130 |
+
self.download_and_upload_to_s3_url(model_url, f"{model_prefix}/pytorch_model.bin")
|
131 |
+
self.download_and_upload_to_s3_url(safetensors_url, f"{model_prefix}/model.safetensors")
|
132 |
+
self.download_and_upload_to_s3_url(tokenizer_url, f"{model_prefix}/tokenizer.json")
|
133 |
+
self.download_and_upload_to_s3_url(config_url, f"{model_prefix}/config.json")
|
134 |
+
|
135 |
+
def download_and_upload_to_s3_url(self, url: str, s3_key: str):
|
136 |
+
print(f"Descargando archivo desde {url}...")
|
137 |
+
response = requests.get(url)
|
138 |
+
if response.status_code == 200:
|
139 |
+
# Subir archivo a S3
|
140 |
+
print(f"Subiendo archivo a S3 con key {s3_key}...")
|
141 |
+
self.s3_client.put_object(Bucket=self.bucket_name, Key=s3_key, Body=response.content)
|
142 |
+
else:
|
143 |
+
raise HTTPException(status_code=500, detail=f"Error al descargar el archivo desde {url}")
|
144 |
+
|
145 |
+
|
146 |
+
@app.post("/predict/")
|
147 |
+
async def predict(model_request: DownloadModelRequest):
|
148 |
+
try:
|
149 |
+
print(f"Recibiendo solicitud para predecir con el modelo {model_request.model_name}...")
|
150 |
+
|
151 |
+
model_name = model_request.model_name
|
152 |
+
revision = model_request.revision
|
153 |
+
|
154 |
+
# Cargar el modelo y tokenizer desde S3
|
155 |
+
streamer = S3DirectStream(S3_BUCKET_NAME)
|
156 |
+
model = streamer.load_model_from_stream(model_name, revision)
|
157 |
+
tokenizer = streamer.load_tokenizer_from_stream(model_name)
|
158 |
+
|
159 |
+
# Obtener el pipeline adecuado seg煤n la solicitud
|
160 |
+
task = model_request.pipeline_task
|
161 |
+
if task not in ["text-generation", "sentiment-analysis", "translation", "fill-mask", "question-answering", "text-to-speech", "text-to-image", "text-to-audio", "text-to-video"]:
|
162 |
+
raise HTTPException(status_code=400, detail="Pipeline task no soportado")
|
163 |
+
|
164 |
+
# Crear el pipeline din谩micamente basado en el tipo de tarea
|
165 |
+
nlp_pipeline = pipeline(task, model=model, tokenizer=tokenizer, use_auth_token=HUGGINGFACE_TOKEN, revision=revision)
|
166 |
+
|
167 |
+
# Ejecutar el pipeline con el input_text
|
168 |
+
outputs = nlp_pipeline(model_request.input_text)
|
169 |
+
|
170 |
+
# Almacenar el resultado en S3 dependiendo del tipo de tarea
|
171 |
+
if task == "text-to-image":
|
172 |
+
s3_key = f"{model_request.model_name}/generated_image.png"
|
173 |
+
return StreamingResponse(streamer.stream_from_s3(s3_key), media_type="image/png")
|
174 |
+
|
175 |
+
elif task == "text-to-speech":
|
176 |
+
s3_key = f"{model_request.model_name}/generated_audio.wav"
|
177 |
+
return StreamingResponse(streamer.stream_from_s3(s3_key), media_type="audio/wav")
|
178 |
+
|
179 |
+
elif task == "text-to-video":
|
180 |
+
s3_key = f"{model_request.model_name}/generated_video.mp4"
|
181 |
+
return StreamingResponse(streamer.stream_from_s3(s3_key), media_type="video/mp4")
|
182 |
+
|
183 |
+
# Devolver resultados de texto u otros tipos de tarea
|
184 |
+
return {"result": outputs}
|
185 |
+
|
186 |
+
except Exception as e:
|
187 |
+
raise HTTPException(status_code=500, detail=f"Error al procesar la solicitud: {str(e)}")
|
188 |
+
|
189 |
+
|
190 |
+
if __name__ == "__main__":
|
191 |
+
uvicorn.run(app, host="0.0.0.0", port=8000)
|