Spaces:
Running
on
Zero
Running
on
Zero
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,143 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from fastapi import FastAPI, HTTPException
|
2 |
+
from pydantic import BaseModel
|
3 |
+
from llama_cpp import Llama
|
4 |
+
from concurrent.futures import ThreadPoolExecutor, as_completed
|
5 |
+
from tqdm import tqdm
|
6 |
+
import uvicorn
|
7 |
+
from dotenv import load_dotenv
|
8 |
+
from difflib import SequenceMatcher
|
9 |
+
import re
|
10 |
+
from spaces import GPU
|
11 |
+
import httpx
|
12 |
+
|
13 |
+
# Cargar variables de entorno
|
14 |
+
load_dotenv()
|
15 |
+
|
16 |
+
# Inicializar aplicación FastAPI
|
17 |
+
app = FastAPI()
|
18 |
+
|
19 |
+
# Diccionario global para almacenar los modelos
|
20 |
+
global_data = {
|
21 |
+
'models': []
|
22 |
+
}
|
23 |
+
|
24 |
+
# Configuración de los modelos (incluyendo los nuevos)
|
25 |
+
model_configs = [
|
26 |
+
{"repo_id": "Ffftdtd5dtft/gpt2-xl-Q2_K-GGUF", "filename": "gpt2-xl-q2_k.gguf", "name": "GPT-2 XL"},
|
27 |
+
{"repo_id": "Ffftdtd5dtft/Meta-Llama-3.1-8B-Instruct-Q2_K-GGUF", "filename": "meta-llama-3.1-8b-instruct-q2_k.gguf", "name": "Meta Llama 3.1-8B Instruct"},
|
28 |
+
# Otros modelos omitidos por espacio
|
29 |
+
{"repo_id": "Ffftdtd5dtft/Meta-Llama-3.1-70B-Instruct-Q2_K-GGUF", "filename": "meta-llama-3.1-70b-instruct-q2_k.gguf", "name": "Meta Llama 3.1-70B Instruct"},
|
30 |
+
{"repo_id": "Ffftdtd5dtft/codegemma-2b-IQ1_S-GGUF", "filename": "codegemma-2b-iq1_s-imat.gguf", "name": "Codegemma 2B"},
|
31 |
+
{"repo_id": "Ffftdtd5dtft/Mistral-Nemo-Instruct-2407-Q2_K-GGUF", "filename": "mistral-nemo-instruct-2407-q2_k.gguf", "name": "Mistral Nemo Instruct 2407"}
|
32 |
+
]
|
33 |
+
|
34 |
+
# Clase para gestionar modelos
|
35 |
+
class ModelManager:
|
36 |
+
def __init__(self):
|
37 |
+
self.models = []
|
38 |
+
|
39 |
+
def load_model(self, model_config):
|
40 |
+
print(f"Cargando modelo: {model_config['name']}...")
|
41 |
+
return {"model": Llama.from_pretrained(repo_id=model_config['repo_id'], filename=model_config['filename']), "name": model_config['name']}
|
42 |
+
|
43 |
+
@GPU(duration=0)
|
44 |
+
def load_all_models(self):
|
45 |
+
print("Iniciando carga de modelos...")
|
46 |
+
with ThreadPoolExecutor(max_workers=len(model_configs)) as executor:
|
47 |
+
futures = [executor.submit(self.load_model, config) for config in model_configs]
|
48 |
+
models = []
|
49 |
+
for future in tqdm(as_completed(futures), total=len(model_configs), desc="Cargando modelos", unit="modelo"):
|
50 |
+
try:
|
51 |
+
model = future.result()
|
52 |
+
models.append(model)
|
53 |
+
print(f"Modelo cargado exitosamente: {model['name']}")
|
54 |
+
except Exception as e:
|
55 |
+
print(f"Error al cargar el modelo: {e}")
|
56 |
+
print("Todos los modelos han sido cargados.")
|
57 |
+
return models
|
58 |
+
|
59 |
+
# Instanciar ModelManager y cargar modelos una sola vez
|
60 |
+
model_manager = ModelManager()
|
61 |
+
global_data['models'] = model_manager.load_all_models()
|
62 |
+
|
63 |
+
# Modelo global para la solicitud de chat
|
64 |
+
class ChatRequest(BaseModel):
|
65 |
+
message: str
|
66 |
+
top_k: int = 50
|
67 |
+
top_p: float = 0.95
|
68 |
+
temperature: float = 0.7
|
69 |
+
|
70 |
+
# Función para generar respuestas de chat
|
71 |
+
def generate_chat_response(request, model_data):
|
72 |
+
try:
|
73 |
+
user_input = normalize_input(request.message)
|
74 |
+
llm = model_data['model']
|
75 |
+
response = llm.create_chat_completion(
|
76 |
+
messages=[{"role": "user", "content": user_input}],
|
77 |
+
top_k=request.top_k,
|
78 |
+
top_p=request.top_p,
|
79 |
+
temperature=request.temperature
|
80 |
+
)
|
81 |
+
reply = response['choices'][0]['message']['content']
|
82 |
+
return {"response": reply, "literal": user_input, "model_name": model_data['name']}
|
83 |
+
except Exception as e:
|
84 |
+
return {"response": f"Error: {str(e)}", "literal": user_input, "model_name": model_data['name']}
|
85 |
+
|
86 |
+
def normalize_input(input_text):
|
87 |
+
return input_text.strip()
|
88 |
+
|
89 |
+
def remove_duplicates(text):
|
90 |
+
text = re.sub(r'(Hello there, how are you\? \[/INST\]){2,}', 'Hello there, how are you? [/INST]', text)
|
91 |
+
text = re.sub(r'(How are you\? \[/INST\]){2,}', 'How are you? [/INST]', text)
|
92 |
+
text = text.replace('[/INST]', '')
|
93 |
+
lines = text.split('\n')
|
94 |
+
unique_lines = list(dict.fromkeys(lines))
|
95 |
+
return '\n'.join(unique_lines).strip()
|
96 |
+
|
97 |
+
def remove_repetitive_responses(responses):
|
98 |
+
seen = set()
|
99 |
+
unique_responses = []
|
100 |
+
for response in responses:
|
101 |
+
normalized_response = remove_duplicates(response['response'])
|
102 |
+
if normalized_response not in seen:
|
103 |
+
seen.add(normalized_response)
|
104 |
+
unique_responses.append(response)
|
105 |
+
return unique_responses
|
106 |
+
|
107 |
+
# Manejo de errores en la inicialización de modelos (traza mencionada en el error)
|
108 |
+
def handle_initialization_error(allow_token):
|
109 |
+
try:
|
110 |
+
client = httpx.Client()
|
111 |
+
pid = 0 # Variable que simula el proceso actual
|
112 |
+
assert client.allow(allow_token=allow_token, pid=pid) == httpx.codes.OK
|
113 |
+
except AssertionError:
|
114 |
+
raise HTTPException(status_code=500, detail="Error en la inicialización del cliente Spaces")
|
115 |
+
|
116 |
+
# Ruta para generar chat en múltiples modelos
|
117 |
+
@app.post("/chat/")
|
118 |
+
async def chat(request: ChatRequest):
|
119 |
+
try:
|
120 |
+
# Simulación del error `AssertionError` durante la inicialización
|
121 |
+
allow_token = "test_token"
|
122 |
+
handle_initialization_error(allow_token)
|
123 |
+
|
124 |
+
with ThreadPoolExecutor() as executor:
|
125 |
+
futures = [executor.submit(generate_chat_response, request, model) for model in global_data['models']]
|
126 |
+
responses = [future.result() for future in as_completed(futures)]
|
127 |
+
unique_responses = remove_repetitive_responses(responses)
|
128 |
+
return {"responses": unique_responses}
|
129 |
+
except Exception as e:
|
130 |
+
raise HTTPException(status_code=500, detail=f"Error procesando la solicitud: {str(e)}")
|
131 |
+
|
132 |
+
# Uso de template `chat_template.default`
|
133 |
+
chat_template = """
|
134 |
+
User: {message}
|
135 |
+
Bot: {response}
|
136 |
+
"""
|
137 |
+
|
138 |
+
# Plantilla de respuesta de chat
|
139 |
+
def render_chat_template(message, response):
|
140 |
+
return chat_template.format(message=message, response=response)
|
141 |
+
|
142 |
+
if __name__ == "__main__":
|
143 |
+
uvicorn.run(app, host="0.0.0.0", port=8000)
|