Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,11 +1,8 @@
|
|
| 1 |
import os
|
| 2 |
-
import sys
|
| 3 |
import torch
|
| 4 |
import json
|
| 5 |
-
import time
|
| 6 |
-
import gc
|
| 7 |
import re
|
| 8 |
-
from fastapi import FastAPI
|
| 9 |
from fastapi.responses import HTMLResponse
|
| 10 |
from fastapi.middleware.cors import CORSMiddleware
|
| 11 |
from pydantic import BaseModel, Field
|
|
@@ -16,26 +13,13 @@ import torch.nn as nn
|
|
| 16 |
import torch.nn.functional as F
|
| 17 |
import sentencepiece as spm
|
| 18 |
|
| 19 |
-
# ======================
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
if torch.cuda.is_available():
|
| 23 |
-
DEVICE = "cuda"
|
| 24 |
-
print("✅ GPU NVIDIA detectada. Usando CUDA.")
|
| 25 |
-
else:
|
| 26 |
-
DEVICE = "cpu"
|
| 27 |
-
print("⚠️ GPU no detectada. Usando CPU (puede ser más lento).")
|
| 28 |
-
|
| 29 |
-
if DEVICE == "cpu":
|
| 30 |
-
torch.set_num_threads(max(1, os.cpu_count() // 2))
|
| 31 |
-
|
| 32 |
-
torch.set_grad_enabled(False)
|
| 33 |
|
| 34 |
MODEL_REPO = "TeszenAI/MTP-3"
|
| 35 |
|
| 36 |
-
# ======================
|
| 37 |
-
# ARQUITECTURA DEL MODELO MEJORADA
|
| 38 |
-
# ======================
|
| 39 |
class LayerNorm(nn.Module):
|
| 40 |
def __init__(self, d_model: int, eps: float = 1e-5):
|
| 41 |
super().__init__()
|
|
@@ -114,8 +98,8 @@ class PositionalEncoding(nn.Module):
|
|
| 114 |
return x + self.pe[:, :x.size(1), :]
|
| 115 |
|
| 116 |
class MTPModel(nn.Module):
|
| 117 |
-
def __init__(self, vocab_size: int, d_model: int =
|
| 118 |
-
n_layers: int =
|
| 119 |
super().__init__()
|
| 120 |
self.vocab_size = vocab_size
|
| 121 |
self.d_model = d_model
|
|
@@ -125,7 +109,6 @@ class MTPModel(nn.Module):
|
|
| 125 |
self.blocks = nn.ModuleList([TransformerBlock(d_model, n_heads, d_ff, dropout) for _ in range(n_layers)])
|
| 126 |
self.norm = LayerNorm(d_model)
|
| 127 |
self.lm_head = nn.Linear(d_model, vocab_size)
|
| 128 |
-
|
| 129 |
def forward(self, x, mask=None):
|
| 130 |
if mask is None:
|
| 131 |
mask = torch.tril(torch.ones(x.size(1), x.size(1))).unsqueeze(0).unsqueeze(0).to(x.device)
|
|
@@ -136,891 +119,214 @@ class MTPModel(nn.Module):
|
|
| 136 |
x = self.norm(x)
|
| 137 |
return self.lm_head(x)
|
| 138 |
|
| 139 |
-
# ======================
|
| 140 |
-
|
| 141 |
-
# ======================
|
| 142 |
-
class NLPProcessor:
|
| 143 |
-
"""Procesador de lenguaje natural para entender mejor las intenciones"""
|
| 144 |
-
|
| 145 |
-
@staticmethod
|
| 146 |
-
def detect_intent(text):
|
| 147 |
-
"""Detecta la intención del usuario"""
|
| 148 |
-
text_lower = text.lower()
|
| 149 |
-
|
| 150 |
-
intents = {
|
| 151 |
-
'saludo': ['hola', 'buenas', 'que tal', 'cómo estás', 'hey', 'saludos'],
|
| 152 |
-
'despedida': ['adiós', 'chao', 'hasta luego', 'nos vemos', 'bye'],
|
| 153 |
-
'agradecimiento': ['gracias', 'gracias por', 'te agradezco', 'muchas gracias'],
|
| 154 |
-
'pregunta': ['qué es', 'cómo funciona', 'por qué', 'cuándo', 'dónde', 'quién'],
|
| 155 |
-
'ayuda': ['ayuda', 'necesito ayuda', 'puedes ayudarme', 'me ayudas'],
|
| 156 |
-
'presentacion': ['quién eres', 'qué eres', 'presentate', 'eres'],
|
| 157 |
-
'capacidad': ['qué puedes hacer', 'funciones', 'capacidades', 'que sabes hacer'],
|
| 158 |
-
'sentimiento': ['estoy triste', 'estoy feliz', 'me siento', 'emocionado']
|
| 159 |
-
}
|
| 160 |
-
|
| 161 |
-
for intent, keywords in intents.items():
|
| 162 |
-
for keyword in keywords:
|
| 163 |
-
if keyword in text_lower:
|
| 164 |
-
return intent
|
| 165 |
-
return 'general'
|
| 166 |
-
|
| 167 |
-
@staticmethod
|
| 168 |
-
def should_stop(response, min_length=30, max_length=200):
|
| 169 |
-
"""Determina si la respuesta debe terminar"""
|
| 170 |
-
|
| 171 |
-
# Palabras que indican final de respuesta
|
| 172 |
-
stop_phrases = [
|
| 173 |
-
'¿alguna otra pregunta?', '¿en qué más puedo ayudarte?',
|
| 174 |
-
'¿necesitas ayuda con algo más?', '¿tienes alguna otra duda?',
|
| 175 |
-
'espero haberte ayudado', 'que tengas un buen día',
|
| 176 |
-
'hasta luego', 'adiós', 'saludos', 'gracias por consultar'
|
| 177 |
-
]
|
| 178 |
-
|
| 179 |
-
# Si es demasiado corta, continuar
|
| 180 |
-
if len(response) < min_length:
|
| 181 |
-
return False
|
| 182 |
-
|
| 183 |
-
# Si excede el máximo, cortar
|
| 184 |
-
if len(response) > max_length:
|
| 185 |
-
return True
|
| 186 |
-
|
| 187 |
-
# Verificar frases de parada
|
| 188 |
-
for phrase in stop_phrases:
|
| 189 |
-
if phrase in response.lower():
|
| 190 |
-
return True
|
| 191 |
-
|
| 192 |
-
# Verificar si termina con puntuación adecuada
|
| 193 |
-
if len(response) > 50:
|
| 194 |
-
last_chars = response[-10:]
|
| 195 |
-
# Termina con punto, signo de interrogación o exclamación
|
| 196 |
-
if any(last_chars.rstrip().endswith(p) for p in ['.', '?', '!', '…']):
|
| 197 |
-
# Contar oraciones completas
|
| 198 |
-
sentences = re.split(r'[.!?]+', response)
|
| 199 |
-
if len(sentences) >= 2: # Al menos 2 oraciones completas
|
| 200 |
-
return True
|
| 201 |
-
|
| 202 |
-
return False
|
| 203 |
-
|
| 204 |
-
@staticmethod
|
| 205 |
-
def clean_response(text):
|
| 206 |
-
"""Limpia y mejora la respuesta"""
|
| 207 |
-
# Eliminar repeticiones excesivas
|
| 208 |
-
text = re.sub(r'(\b\w+\b)(?:\s+\1\b)+', r'\1', text)
|
| 209 |
-
|
| 210 |
-
# Corregir espaciado
|
| 211 |
-
text = re.sub(r'\s+([.,!?;:])', r'\1', text)
|
| 212 |
-
|
| 213 |
-
# Asegurar mayúscula al inicio
|
| 214 |
-
if text and text[0].islower():
|
| 215 |
-
text = text[0].upper() + text[1:]
|
| 216 |
-
|
| 217 |
-
# Agregar punto final si no tiene
|
| 218 |
-
if text and not text[-1] in '.!?':
|
| 219 |
-
text += '.'
|
| 220 |
-
|
| 221 |
-
return text.strip()
|
| 222 |
-
|
| 223 |
-
@staticmethod
|
| 224 |
-
def extract_key_info(text):
|
| 225 |
-
"""Extrae información clave del texto"""
|
| 226 |
-
# Detectar números
|
| 227 |
-
numbers = re.findall(r'\d+(?:\.\d+)?', text)
|
| 228 |
-
|
| 229 |
-
# Detectar emails
|
| 230 |
-
emails = re.findall(r'[\w\.-]+@[\w\.-]+\.\w+', text)
|
| 231 |
-
|
| 232 |
-
# Detectar URLs
|
| 233 |
-
urls = re.findall(r'https?://(?:[-\w.]|(?:%[\da-fA-F]{2}))+', text)
|
| 234 |
-
|
| 235 |
-
return {
|
| 236 |
-
'has_numbers': bool(numbers),
|
| 237 |
-
'has_emails': bool(emails),
|
| 238 |
-
'has_urls': bool(urls),
|
| 239 |
-
'numbers': numbers,
|
| 240 |
-
'emails': emails,
|
| 241 |
-
'urls': urls
|
| 242 |
-
}
|
| 243 |
-
|
| 244 |
-
# ======================
|
| 245 |
-
# DESCARGA Y CARGA DEL MODELO
|
| 246 |
-
# ======================
|
| 247 |
-
def download_with_retry(repo_id, local_dir, max_retries=3):
|
| 248 |
-
for attempt in range(max_retries):
|
| 249 |
-
try:
|
| 250 |
-
print(f"📦 Intento {attempt + 1}/{max_retries} - Descargando modelo...")
|
| 251 |
-
repo_path = snapshot_download(
|
| 252 |
-
repo_id=repo_id,
|
| 253 |
-
repo_type="model",
|
| 254 |
-
local_dir=local_dir,
|
| 255 |
-
resume_download=True,
|
| 256 |
-
local_files_only=False
|
| 257 |
-
)
|
| 258 |
-
print(f"✅ Modelo descargado")
|
| 259 |
-
return repo_path
|
| 260 |
-
except Exception as e:
|
| 261 |
-
print(f"⚠️ Error: {str(e)[:100]}")
|
| 262 |
-
if attempt < max_retries - 1:
|
| 263 |
-
time.sleep(3)
|
| 264 |
-
else:
|
| 265 |
-
raise
|
| 266 |
-
return local_dir
|
| 267 |
-
|
| 268 |
-
print(f"🚀 Cargando modelo...")
|
| 269 |
|
| 270 |
if os.path.exists("mtp_repo") and os.path.exists("mtp_repo/mtp_model.pt"):
|
| 271 |
-
print("📁 Modelo en caché")
|
| 272 |
repo_path = "mtp_repo"
|
| 273 |
else:
|
| 274 |
try:
|
| 275 |
-
repo_path =
|
| 276 |
except:
|
| 277 |
repo_path = "mtp_repo"
|
| 278 |
|
| 279 |
-
#
|
| 280 |
config_path = os.path.join(repo_path, "config.json")
|
| 281 |
if os.path.exists(config_path):
|
| 282 |
with open(config_path, "r") as f:
|
| 283 |
config = json.load(f)
|
| 284 |
else:
|
| 285 |
-
config = {
|
| 286 |
-
"vocab_size": 2000,
|
| 287 |
-
"d_model": 256,
|
| 288 |
-
"n_heads": 8,
|
| 289 |
-
"n_layers": 6,
|
| 290 |
-
"d_ff": 1024,
|
| 291 |
-
"dropout": 0.1,
|
| 292 |
-
"max_len": 512
|
| 293 |
-
}
|
| 294 |
|
| 295 |
-
#
|
| 296 |
tokenizer_path = os.path.join(repo_path, "mtp_tokenizer.model")
|
| 297 |
if os.path.exists(tokenizer_path):
|
| 298 |
sp = spm.SentencePieceProcessor()
|
| 299 |
sp.load(tokenizer_path)
|
| 300 |
VOCAB_SIZE = sp.get_piece_size()
|
| 301 |
config["vocab_size"] = VOCAB_SIZE
|
| 302 |
-
print(f"✅ Tokenizador: {VOCAB_SIZE} tokens")
|
| 303 |
else:
|
| 304 |
sp = None
|
| 305 |
-
VOCAB_SIZE =
|
| 306 |
-
|
| 307 |
-
print(f"🧠 Inicializando modelo...")
|
| 308 |
-
print(f" → Vocabulario: {VOCAB_SIZE}")
|
| 309 |
-
print(f" → Dimensión: {config['d_model']}")
|
| 310 |
-
print(f" → Capas: {config['n_layers']}")
|
| 311 |
|
| 312 |
model = MTPModel(**config)
|
| 313 |
model.to(DEVICE)
|
| 314 |
|
| 315 |
-
# Cargar pesos
|
| 316 |
model_path = os.path.join(repo_path, "mtp_model.pt")
|
| 317 |
if os.path.exists(model_path):
|
| 318 |
try:
|
| 319 |
state_dict = torch.load(model_path, map_location=DEVICE)
|
| 320 |
model.load_state_dict(state_dict)
|
| 321 |
-
print("✅
|
| 322 |
-
except
|
| 323 |
-
print(
|
| 324 |
-
|
| 325 |
model.eval()
|
| 326 |
|
| 327 |
-
|
| 328 |
-
|
| 329 |
-
|
| 330 |
-
# ======================
|
| 331 |
-
# API CONFIG
|
| 332 |
-
# ======================
|
| 333 |
-
app = FastAPI(title="MTP API - Versión Mejorada", description="API con NLP integrado", version="2.0")
|
| 334 |
-
|
| 335 |
-
app.add_middleware(
|
| 336 |
-
CORSMiddleware,
|
| 337 |
-
allow_origins=["*"],
|
| 338 |
-
allow_methods=["*"],
|
| 339 |
-
allow_headers=["*"],
|
| 340 |
-
)
|
| 341 |
|
| 342 |
class PromptRequest(BaseModel):
|
| 343 |
text: str = Field(..., max_length=2000)
|
| 344 |
-
max_tokens: int = Field(default=150, ge=10, le=300)
|
| 345 |
-
temperature: float = Field(default=0.7, ge=0.1, le=2.0)
|
| 346 |
-
top_k: int = Field(default=50, ge=1, le=100)
|
| 347 |
-
top_p: float = Field(default=0.9, ge=0.1, le=1.0)
|
| 348 |
|
| 349 |
-
|
| 350 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 351 |
|
| 352 |
-
|
| 353 |
-
|
| 354 |
-
#
|
| 355 |
-
|
| 356 |
-
|
| 357 |
-
|
| 358 |
-
|
| 359 |
-
|
| 360 |
-
|
| 361 |
-
|
| 362 |
-
|
| 363 |
-
|
| 364 |
-
|
| 365 |
-
|
| 366 |
-
|
| 367 |
-
|
| 368 |
-
|
| 369 |
-
|
| 370 |
-
|
| 371 |
-
|
| 372 |
-
|
|
|
|
|
|
|
| 373 |
generated = input_ids.copy()
|
| 374 |
eos_id = tokenizer.eos_id()
|
| 375 |
|
| 376 |
-
|
| 377 |
-
|
| 378 |
-
last_chars = []
|
| 379 |
-
|
| 380 |
-
for step in range(max_length):
|
| 381 |
-
input_tensor = torch.tensor([generated[-model.max_len:]], dtype=torch.long).to(device)
|
| 382 |
with torch.no_grad():
|
| 383 |
logits = model(input_tensor)
|
| 384 |
next_logits = logits[0, -1, :] / temperature
|
| 385 |
|
| 386 |
-
# Top-k filtering
|
| 387 |
-
if top_k > 0:
|
| 388 |
-
indices_to_remove = next_logits < torch.topk(next_logits, top_k)[0][..., -1, None]
|
| 389 |
-
next_logits[indices_to_remove] = float('-inf')
|
| 390 |
-
|
| 391 |
-
# Top-p filtering
|
| 392 |
-
if top_p < 1.0:
|
| 393 |
-
sorted_logits, sorted_indices = torch.sort(next_logits, descending=True)
|
| 394 |
-
cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
|
| 395 |
-
sorted_indices_to_remove = cumulative_probs > top_p
|
| 396 |
-
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
|
| 397 |
-
sorted_indices_to_remove[..., 0] = 0
|
| 398 |
-
indices_to_remove = sorted_indices[sorted_indices_to_remove]
|
| 399 |
-
next_logits[indices_to_remove] = float('-inf')
|
| 400 |
-
|
| 401 |
probs = F.softmax(next_logits, dim=-1)
|
| 402 |
next_token = torch.multinomial(probs, 1).item()
|
| 403 |
|
| 404 |
-
# Detener en EOS
|
| 405 |
if next_token == eos_id:
|
| 406 |
break
|
| 407 |
-
|
| 408 |
-
# Detener si hay demasiados signos de puntuación seguidos
|
| 409 |
-
token_str = tokenizer.decode([next_token]) if hasattr(tokenizer, 'decode') else str(next_token)
|
| 410 |
-
if token_str in '.!?':
|
| 411 |
-
consecutive_punctuation += 1
|
| 412 |
-
if consecutive_punctuation >= 3:
|
| 413 |
-
break
|
| 414 |
-
else:
|
| 415 |
-
consecutive_punctuation = 0
|
| 416 |
-
|
| 417 |
-
# Guardar últimos caracteres para análisis
|
| 418 |
-
last_chars.append(token_str)
|
| 419 |
-
if len(last_chars) > 20:
|
| 420 |
-
last_chars.pop(0)
|
| 421 |
-
|
| 422 |
-
# Detectar bucles de repetición
|
| 423 |
-
if len(last_chars) >= 10:
|
| 424 |
-
last_str = ''.join(last_chars[-5:])
|
| 425 |
-
if last_str in ''.join(last_chars[:-5]):
|
| 426 |
-
break
|
| 427 |
-
|
| 428 |
generated.append(next_token)
|
| 429 |
-
|
| 430 |
-
# Verificar si ya es suficiente (para respuestas cortas)
|
| 431 |
-
current_response = tokenizer.decode(generated)
|
| 432 |
-
if "### Respuesta:" in current_response:
|
| 433 |
-
response_part = current_response.split("### Respuesta:")[-1].strip()
|
| 434 |
-
if nlp.should_stop(response_part, min_length=20, max_length=max_length):
|
| 435 |
-
break
|
| 436 |
|
| 437 |
-
# Decodificar respuesta
|
| 438 |
response = tokenizer.decode(generated)
|
| 439 |
-
|
| 440 |
-
# Extraer la parte de la respuesta
|
| 441 |
if "### Respuesta:" in response:
|
| 442 |
response = response.split("### Respuesta:")[-1].strip()
|
| 443 |
-
elif "Respuesta:" in response:
|
| 444 |
-
response = response.split("Respuesta:")[-1].strip()
|
| 445 |
-
elif "[/INST]" in response:
|
| 446 |
-
response = response.split("[/INST]")[-1].strip()
|
| 447 |
|
| 448 |
-
# Limpiar
|
| 449 |
-
|
| 450 |
-
for word in garbage_words:
|
| 451 |
-
response = response.replace(word, '')
|
| 452 |
-
|
| 453 |
-
# Limpiar caracteres especiales
|
| 454 |
-
response = re.sub(r'[^\w\s\u00C0-\u00FF\u0100-\u017F.,!?¿¡()\-:;"\']+', ' ', response)
|
| 455 |
response = re.sub(r'\s+', ' ', response).strip()
|
| 456 |
|
| 457 |
-
#
|
| 458 |
-
|
| 459 |
-
|
| 460 |
-
|
| 461 |
-
|
| 462 |
-
|
| 463 |
-
|
| 464 |
-
'despedida': "¡Hasta luego! Que tengas un excelente día.",
|
| 465 |
-
'agradecimiento': "¡De nada! Estoy aquí para ayudarte cuando lo necesites.",
|
| 466 |
-
'ayuda': "Claro, estoy aquí para ayudarte. ¿Qué necesitas saber?",
|
| 467 |
-
'presentacion': "Soy MTP, un asistente virtual creado para responder preguntas y ayudarte con información.",
|
| 468 |
-
'general': "Entendido. ¿Hay algo específico en lo que pueda ayudarte?"
|
| 469 |
-
}
|
| 470 |
-
response = default_responses.get(intent, default_responses['general'])
|
| 471 |
|
| 472 |
return response
|
| 473 |
|
| 474 |
-
# ======================
|
| 475 |
-
# ENDPOINTS
|
| 476 |
-
# ======================
|
| 477 |
-
ACTIVE_REQUESTS = 0
|
| 478 |
-
|
| 479 |
-
class TokenizerWrapper:
|
| 480 |
-
def __init__(self, sp_model):
|
| 481 |
-
self.sp = sp_model
|
| 482 |
-
def encode(self, text):
|
| 483 |
-
if self.sp is None:
|
| 484 |
-
return [ord(c) % 1000 for c in text[:200]]
|
| 485 |
-
return self.sp.encode(text)
|
| 486 |
-
def decode(self, tokens):
|
| 487 |
-
if self.sp is None:
|
| 488 |
-
return ''.join([chr(t % 128) if 32 <= t % 128 < 127 else ' ' for t in tokens])
|
| 489 |
-
return self.sp.decode(tokens)
|
| 490 |
-
def eos_id(self):
|
| 491 |
-
return self.sp.eos_id() if self.sp else 3
|
| 492 |
-
def bos_id(self):
|
| 493 |
-
return self.sp.bos_id() if self.sp else 2
|
| 494 |
-
def pad_id(self):
|
| 495 |
-
return self.sp.pad_id() if self.sp else 0
|
| 496 |
-
|
| 497 |
-
tokenizer_wrapper = TokenizerWrapper(sp)
|
| 498 |
-
|
| 499 |
@app.post("/generate")
|
| 500 |
async def generate(req: PromptRequest):
|
| 501 |
-
global ACTIVE_REQUESTS
|
| 502 |
-
ACTIVE_REQUESTS += 1
|
| 503 |
-
|
| 504 |
user_input = req.text.strip()
|
| 505 |
if not user_input:
|
| 506 |
-
|
| 507 |
-
return {"reply": "", "tokens_generated": 0, "intent": None}
|
| 508 |
-
|
| 509 |
-
# Detectar intención
|
| 510 |
-
intent = nlp.detect_intent(user_input)
|
| 511 |
|
| 512 |
try:
|
| 513 |
-
response =
|
| 514 |
-
|
| 515 |
-
max_length=req.max_tokens,
|
| 516 |
-
temperature=req.temperature,
|
| 517 |
-
top_k=req.top_k,
|
| 518 |
-
top_p=req.top_p,
|
| 519 |
-
device=DEVICE
|
| 520 |
-
)
|
| 521 |
-
|
| 522 |
-
# Extraer información clave
|
| 523 |
-
key_info = nlp.extract_key_info(response)
|
| 524 |
-
|
| 525 |
-
return {
|
| 526 |
-
"reply": response,
|
| 527 |
-
"tokens_generated": len(response.split()),
|
| 528 |
-
"model": "MTP-Intelligent",
|
| 529 |
-
"intent": intent,
|
| 530 |
-
"has_numbers": key_info['has_numbers'],
|
| 531 |
-
"has_emails": key_info['has_emails']
|
| 532 |
-
}
|
| 533 |
except Exception as e:
|
| 534 |
-
print(f"
|
| 535 |
-
|
| 536 |
-
|
| 537 |
-
ACTIVE_REQUESTS -= 1
|
| 538 |
-
if DEVICE == "cuda":
|
| 539 |
-
torch.cuda.empty_cache()
|
| 540 |
-
gc.collect()
|
| 541 |
|
| 542 |
-
@app.get("/
|
| 543 |
-
def
|
| 544 |
-
return
|
| 545 |
-
"status": "healthy",
|
| 546 |
-
"model": "MTP-Intelligent",
|
| 547 |
-
"device": DEVICE,
|
| 548 |
-
"active_requests": ACTIVE_REQUESTS,
|
| 549 |
-
"vocab_size": VOCAB_SIZE
|
| 550 |
-
}
|
| 551 |
-
|
| 552 |
-
@app.get("/info")
|
| 553 |
-
def model_info():
|
| 554 |
-
return {
|
| 555 |
-
"model_name": "MTP-Intelligent",
|
| 556 |
-
"version": "2.0",
|
| 557 |
-
"architecture": config,
|
| 558 |
-
"parameters": sum(p.numel() for p in model.parameters()),
|
| 559 |
-
"device": DEVICE,
|
| 560 |
-
"nlp_enabled": True
|
| 561 |
-
}
|
| 562 |
-
|
| 563 |
-
@app.post("/analyze")
|
| 564 |
-
async def analyze_intent(req: PromptRequest):
|
| 565 |
-
"""Endpoint para analizar intención sin generar respuesta"""
|
| 566 |
-
intent = nlp.detect_intent(req.text)
|
| 567 |
-
return {
|
| 568 |
-
"text": req.text,
|
| 569 |
-
"intent": intent,
|
| 570 |
-
"confidence": 0.85 # Por ahora fijo, se puede mejorar
|
| 571 |
-
}
|
| 572 |
-
|
| 573 |
-
# ======================
|
| 574 |
-
# INTERFAZ WEB MEJORADA
|
| 575 |
-
# ======================
|
| 576 |
-
@app.get("/", response_class=HTMLResponse)
|
| 577 |
-
def chat_ui():
|
| 578 |
-
return """
|
| 579 |
<!DOCTYPE html>
|
| 580 |
-
<html
|
| 581 |
<head>
|
| 582 |
-
<
|
| 583 |
-
<meta
|
| 584 |
-
<
|
| 585 |
-
<
|
| 586 |
-
|
| 587 |
-
|
| 588 |
-
|
| 589 |
-
|
| 590 |
-
|
| 591 |
-
|
| 592 |
-
|
| 593 |
-
|
| 594 |
-
|
| 595 |
-
|
| 596 |
-
--success-color: #00c853;
|
| 597 |
-
}
|
| 598 |
-
* { box-sizing: border-box; outline: none; -webkit-tap-highlight-color: transparent; }
|
| 599 |
-
body {
|
| 600 |
-
margin: 0;
|
| 601 |
-
background-color: var(--bg-color);
|
| 602 |
-
font-family: 'Inter', sans-serif;
|
| 603 |
-
color: var(--text-primary);
|
| 604 |
-
height: 100dvh;
|
| 605 |
-
display: flex;
|
| 606 |
-
flex-direction: column;
|
| 607 |
-
overflow: hidden;
|
| 608 |
-
}
|
| 609 |
-
header {
|
| 610 |
-
padding: 12px 20px;
|
| 611 |
-
display: flex;
|
| 612 |
-
align-items: center;
|
| 613 |
-
justify-content: space-between;
|
| 614 |
-
background: rgba(19, 19, 20, 0.85);
|
| 615 |
-
backdrop-filter: blur(12px);
|
| 616 |
-
position: fixed;
|
| 617 |
-
top: 0;
|
| 618 |
-
width: 100%;
|
| 619 |
-
z-index: 50;
|
| 620 |
-
border-bottom: 1px solid rgba(255,255,255,0.05);
|
| 621 |
-
}
|
| 622 |
-
.brand-wrapper {
|
| 623 |
-
display: flex;
|
| 624 |
-
align-items: center;
|
| 625 |
-
gap: 12px;
|
| 626 |
-
cursor: pointer;
|
| 627 |
-
}
|
| 628 |
-
.brand-logo {
|
| 629 |
-
width: 32px;
|
| 630 |
-
height: 32px;
|
| 631 |
-
border-radius: 50%;
|
| 632 |
-
background: linear-gradient(135deg, #4a9eff, #00c853);
|
| 633 |
-
}
|
| 634 |
-
.brand-text {
|
| 635 |
-
font-weight: 500;
|
| 636 |
-
font-size: 1.05rem;
|
| 637 |
-
display: flex;
|
| 638 |
-
align-items: center;
|
| 639 |
-
gap: 8px;
|
| 640 |
-
}
|
| 641 |
-
.version-badge {
|
| 642 |
-
font-size: 0.75rem;
|
| 643 |
-
background: rgba(74, 158, 255, 0.15);
|
| 644 |
-
color: #8ab4f8;
|
| 645 |
-
padding: 2px 8px;
|
| 646 |
-
border-radius: 12px;
|
| 647 |
-
font-weight: 600;
|
| 648 |
-
}
|
| 649 |
-
.chat-scroll {
|
| 650 |
-
flex: 1;
|
| 651 |
-
overflow-y: auto;
|
| 652 |
-
padding: 80px 20px 40px 20px;
|
| 653 |
-
display: flex;
|
| 654 |
-
flex-direction: column;
|
| 655 |
-
gap: 30px;
|
| 656 |
-
max-width: 850px;
|
| 657 |
-
margin: 0 auto;
|
| 658 |
-
width: 100%;
|
| 659 |
-
scroll-behavior: smooth;
|
| 660 |
-
}
|
| 661 |
-
.msg-row {
|
| 662 |
-
display: flex;
|
| 663 |
-
gap: 16px;
|
| 664 |
-
width: 100%;
|
| 665 |
-
opacity: 0;
|
| 666 |
-
transform: translateY(10px);
|
| 667 |
-
animation: slideUpFade 0.4s cubic-bezier(0.2, 0.8, 0.2, 1) forwards;
|
| 668 |
-
}
|
| 669 |
-
.msg-row.user { justify-content: flex-end; }
|
| 670 |
-
.msg-row.bot { justify-content: flex-start; align-items: flex-start; }
|
| 671 |
-
.msg-content {
|
| 672 |
-
line-height: 1.6;
|
| 673 |
-
font-size: 1rem;
|
| 674 |
-
word-wrap: break-word;
|
| 675 |
-
max-width: 85%;
|
| 676 |
-
}
|
| 677 |
-
.user .msg-content {
|
| 678 |
-
background-color: var(--user-bubble);
|
| 679 |
-
padding: 10px 18px;
|
| 680 |
-
border-radius: 18px;
|
| 681 |
-
border-top-right-radius: 4px;
|
| 682 |
-
color: #fff;
|
| 683 |
-
}
|
| 684 |
-
.bot .msg-content-wrapper {
|
| 685 |
-
display: flex;
|
| 686 |
-
flex-direction: column;
|
| 687 |
-
gap: 8px;
|
| 688 |
-
width: 100%;
|
| 689 |
-
}
|
| 690 |
-
.bot .msg-text {
|
| 691 |
-
padding-top: 6px;
|
| 692 |
-
color: var(--text-primary);
|
| 693 |
-
}
|
| 694 |
-
.bot-avatar {
|
| 695 |
-
width: 34px;
|
| 696 |
-
height: 34px;
|
| 697 |
-
min-width: 34px;
|
| 698 |
-
border-radius: 50%;
|
| 699 |
-
background: linear-gradient(135deg, #4a9eff, #00c853);
|
| 700 |
-
box-shadow: 0 2px 6px rgba(0,0,0,0.2);
|
| 701 |
-
}
|
| 702 |
-
.bot-actions {
|
| 703 |
-
display: flex;
|
| 704 |
-
gap: 10px;
|
| 705 |
-
opacity: 0;
|
| 706 |
-
transition: opacity 0.3s;
|
| 707 |
-
margin-top: 5px;
|
| 708 |
-
}
|
| 709 |
-
.action-btn {
|
| 710 |
-
background: transparent;
|
| 711 |
-
border: none;
|
| 712 |
-
color: var(--text-secondary);
|
| 713 |
-
cursor: pointer;
|
| 714 |
-
padding: 4px;
|
| 715 |
-
border-radius: 4px;
|
| 716 |
-
display: flex;
|
| 717 |
-
align-items: center;
|
| 718 |
-
transition: color 0.2s, background 0.2s;
|
| 719 |
-
}
|
| 720 |
-
.action-btn:hover {
|
| 721 |
-
color: var(--text-primary);
|
| 722 |
-
background: rgba(255,255,255,0.08);
|
| 723 |
-
}
|
| 724 |
-
.action-btn svg { width: 16px; height: 16px; fill: currentColor; }
|
| 725 |
-
.typing-cursor::after {
|
| 726 |
-
content: '';
|
| 727 |
-
display: inline-block;
|
| 728 |
-
width: 10px;
|
| 729 |
-
height: 10px;
|
| 730 |
-
background: var(--accent-color);
|
| 731 |
-
border-radius: 50%;
|
| 732 |
-
margin-left: 5px;
|
| 733 |
-
vertical-align: middle;
|
| 734 |
-
animation: blink 1s infinite;
|
| 735 |
-
}
|
| 736 |
-
.footer-container {
|
| 737 |
-
padding: 0 20px 20px 20px;
|
| 738 |
-
background: linear-gradient(to top, var(--bg-color) 85%, transparent);
|
| 739 |
-
position: relative;
|
| 740 |
-
z-index: 60;
|
| 741 |
-
}
|
| 742 |
-
.input-box {
|
| 743 |
-
max-width: 850px;
|
| 744 |
-
margin: 0 auto;
|
| 745 |
-
background: var(--surface-color);
|
| 746 |
-
border-radius: 28px;
|
| 747 |
-
padding: 8px 10px 8px 20px;
|
| 748 |
-
display: flex;
|
| 749 |
-
align-items: center;
|
| 750 |
-
border: 1px solid rgba(255,255,255,0.1);
|
| 751 |
-
transition: border-color 0.2s, box-shadow 0.2s;
|
| 752 |
-
}
|
| 753 |
-
.input-box:focus-within {
|
| 754 |
-
border-color: rgba(74, 158, 255, 0.5);
|
| 755 |
-
box-shadow: 0 0 0 2px rgba(74, 158, 255, 0.1);
|
| 756 |
-
}
|
| 757 |
-
#userInput {
|
| 758 |
-
flex: 1;
|
| 759 |
-
background: transparent;
|
| 760 |
-
border: none;
|
| 761 |
-
color: white;
|
| 762 |
-
font-size: 1rem;
|
| 763 |
-
font-family: inherit;
|
| 764 |
-
padding: 10px 0;
|
| 765 |
-
}
|
| 766 |
-
#mainBtn {
|
| 767 |
-
background: var(--accent-color);
|
| 768 |
-
color: white;
|
| 769 |
-
border: none;
|
| 770 |
-
width: 36px;
|
| 771 |
-
height: 36px;
|
| 772 |
-
border-radius: 50%;
|
| 773 |
-
display: flex;
|
| 774 |
-
align-items: center;
|
| 775 |
-
justify-content: center;
|
| 776 |
-
cursor: pointer;
|
| 777 |
-
margin-left: 8px;
|
| 778 |
-
transition: transform 0.2s;
|
| 779 |
-
}
|
| 780 |
-
#mainBtn:hover { transform: scale(1.05); background: #3a7ed4; }
|
| 781 |
-
.disclaimer {
|
| 782 |
-
text-align: center;
|
| 783 |
-
font-size: 0.75rem;
|
| 784 |
-
color: #666;
|
| 785 |
-
margin-top: 12px;
|
| 786 |
-
}
|
| 787 |
-
@keyframes slideUpFade {
|
| 788 |
-
from { opacity: 0; transform: translateY(15px); }
|
| 789 |
-
to { opacity: 1; transform: translateY(0); }
|
| 790 |
-
}
|
| 791 |
-
@keyframes blink { 0%, 100% { opacity: 1; } 50% { opacity: 0; } }
|
| 792 |
-
@keyframes pulseAvatar {
|
| 793 |
-
0% { box-shadow: 0 0 0 0 rgba(74, 158, 255, 0.4); }
|
| 794 |
-
70% { box-shadow: 0 0 0 8px rgba(74, 158, 255, 0); }
|
| 795 |
-
100% { box-shadow: 0 0 0 0 rgba(74, 158, 255, 0); }
|
| 796 |
-
}
|
| 797 |
-
.pulsing { animation: pulseAvatar 1.5s infinite; }
|
| 798 |
-
.intent-badge {
|
| 799 |
-
font-size: 0.7rem;
|
| 800 |
-
background: rgba(0, 200, 83, 0.15);
|
| 801 |
-
color: #00c853;
|
| 802 |
-
padding: 2px 8px;
|
| 803 |
-
border-radius: 12px;
|
| 804 |
-
display: inline-block;
|
| 805 |
-
margin-top: 5px;
|
| 806 |
-
}
|
| 807 |
-
::-webkit-scrollbar { width: 8px; }
|
| 808 |
-
::-webkit-scrollbar-track { background: transparent; }
|
| 809 |
-
::-webkit-scrollbar-thumb { background: #333; border-radius: 4px; }
|
| 810 |
-
</style>
|
| 811 |
</head>
|
| 812 |
<body>
|
| 813 |
-
<
|
| 814 |
-
<div
|
| 815 |
-
<div class="
|
| 816 |
-
<div class="brand-text">
|
| 817 |
-
MTP <span class="version-badge">Inteligente</span>
|
| 818 |
-
</div>
|
| 819 |
-
</div>
|
| 820 |
-
</header>
|
| 821 |
-
<div id="chatScroll" class="chat-scroll">
|
| 822 |
-
<div class="msg-row bot" style="animation-delay: 0.1s;">
|
| 823 |
-
<div class="bot-avatar"></div>
|
| 824 |
-
<div class="msg-content-wrapper">
|
| 825 |
-
<div class="msg-text">
|
| 826 |
-
¡Hola! Soy MTP, tu asistente inteligente. ¿En qué puedo ayudarte hoy?
|
| 827 |
-
</div>
|
| 828 |
-
</div>
|
| 829 |
-
</div>
|
| 830 |
-
</div>
|
| 831 |
-
<div class="footer-container">
|
| 832 |
-
<div class="input-box">
|
| 833 |
-
<input type="text" id="userInput" placeholder="Escribe tu mensaje..." autocomplete="off">
|
| 834 |
-
<button id="mainBtn" onclick="handleBtnClick()">➤</button>
|
| 835 |
</div>
|
| 836 |
-
<div
|
| 837 |
-
|
|
|
|
| 838 |
</div>
|
| 839 |
-
<
|
| 840 |
-
|
| 841 |
-
const
|
| 842 |
-
const userInput = document.getElementById('userInput');
|
| 843 |
-
const mainBtn = document.getElementById('mainBtn');
|
| 844 |
-
let isGenerating = false;
|
| 845 |
-
let abortController = null;
|
| 846 |
-
let typingTimeout = null;
|
| 847 |
-
let lastUserPrompt = "";
|
| 848 |
-
|
| 849 |
-
function scrollToBottom() {
|
| 850 |
-
chatScroll.scrollTop = chatScroll.scrollHeight;
|
| 851 |
-
}
|
| 852 |
-
|
| 853 |
-
function setBtnState(state) {
|
| 854 |
-
if (state === 'sending') {
|
| 855 |
-
mainBtn.innerHTML = "⏹";
|
| 856 |
-
isGenerating = true;
|
| 857 |
-
} else {
|
| 858 |
-
mainBtn.innerHTML = "➤";
|
| 859 |
-
isGenerating = false;
|
| 860 |
-
abortController = null;
|
| 861 |
-
}
|
| 862 |
-
}
|
| 863 |
-
|
| 864 |
-
function handleBtnClick() {
|
| 865 |
-
if (isGenerating) {
|
| 866 |
-
stopGeneration();
|
| 867 |
-
} else {
|
| 868 |
-
sendMessage();
|
| 869 |
-
}
|
| 870 |
-
}
|
| 871 |
-
|
| 872 |
-
function stopGeneration() {
|
| 873 |
-
if (abortController) abortController.abort();
|
| 874 |
-
if (typingTimeout) clearTimeout(typingTimeout);
|
| 875 |
-
const activeCursor = document.querySelector('.typing-cursor');
|
| 876 |
-
if (activeCursor) activeCursor.classList.remove('typing-cursor');
|
| 877 |
-
const activeAvatar = document.querySelector('.pulsing');
|
| 878 |
-
if (activeAvatar) activeAvatar.classList.remove('pulsing');
|
| 879 |
-
setBtnState('idle');
|
| 880 |
-
userInput.focus();
|
| 881 |
-
}
|
| 882 |
-
|
| 883 |
-
async function sendMessage(textOverride = null) {
|
| 884 |
-
const text = textOverride || userInput.value.trim();
|
| 885 |
-
if (!text || isGenerating) return;
|
| 886 |
-
|
| 887 |
-
lastUserPrompt = text;
|
| 888 |
-
if (!textOverride) {
|
| 889 |
-
userInput.value = '';
|
| 890 |
-
addMessage(text, 'user');
|
| 891 |
-
}
|
| 892 |
-
|
| 893 |
-
setBtnState('sending');
|
| 894 |
-
abortController = new AbortController();
|
| 895 |
-
|
| 896 |
-
const botRow = document.createElement('div');
|
| 897 |
-
botRow.className = 'msg-row bot';
|
| 898 |
-
const avatar = document.createElement('div');
|
| 899 |
-
avatar.className = 'bot-avatar pulsing';
|
| 900 |
-
const wrapper = document.createElement('div');
|
| 901 |
-
wrapper.className = 'msg-content-wrapper';
|
| 902 |
-
const msgText = document.createElement('div');
|
| 903 |
-
msgText.className = 'msg-text';
|
| 904 |
-
wrapper.appendChild(msgText);
|
| 905 |
-
botRow.appendChild(avatar);
|
| 906 |
-
botRow.appendChild(wrapper);
|
| 907 |
-
chatScroll.appendChild(botRow);
|
| 908 |
-
scrollToBottom();
|
| 909 |
-
|
| 910 |
-
try {
|
| 911 |
-
const response = await fetch('/generate', {
|
| 912 |
-
method: 'POST',
|
| 913 |
-
headers: { 'Content-Type': 'application/json' },
|
| 914 |
-
body: JSON.stringify({
|
| 915 |
-
text: text,
|
| 916 |
-
max_tokens: 200,
|
| 917 |
-
temperature: 0.7,
|
| 918 |
-
top_k: 50,
|
| 919 |
-
top_p: 0.9
|
| 920 |
-
}),
|
| 921 |
-
signal: abortController.signal
|
| 922 |
-
});
|
| 923 |
-
|
| 924 |
-
const data = await response.json();
|
| 925 |
-
if (!isGenerating) return;
|
| 926 |
|
| 927 |
-
|
| 928 |
-
|
| 929 |
-
|
| 930 |
-
|
| 931 |
-
|
| 932 |
-
|
| 933 |
-
intentSpan.className = 'intent-badge';
|
| 934 |
-
intentSpan.textContent = `🎯 Intención: ${data.intent}`;
|
| 935 |
-
wrapper.appendChild(intentSpan);
|
| 936 |
}
|
| 937 |
|
| 938 |
-
|
| 939 |
-
|
| 940 |
-
|
| 941 |
-
|
| 942 |
-
|
| 943 |
-
|
| 944 |
-
|
| 945 |
-
|
| 946 |
-
|
| 947 |
-
|
| 948 |
-
|
| 949 |
-
|
| 950 |
-
|
| 951 |
-
|
| 952 |
-
|
| 953 |
-
}
|
| 954 |
-
|
| 955 |
-
|
| 956 |
-
|
| 957 |
-
|
| 958 |
-
|
| 959 |
-
|
| 960 |
-
|
| 961 |
-
row.appendChild(content);
|
| 962 |
-
chatScroll.appendChild(row);
|
| 963 |
-
scrollToBottom();
|
| 964 |
-
}
|
| 965 |
-
|
| 966 |
-
function typeWriter(element, text, speed = 10) {
|
| 967 |
-
return new Promise(resolve => {
|
| 968 |
-
let i = 0;
|
| 969 |
-
element.classList.add('typing-cursor');
|
| 970 |
-
function type() {
|
| 971 |
-
if (!isGenerating) {
|
| 972 |
-
element.classList.remove('typing-cursor');
|
| 973 |
-
resolve();
|
| 974 |
-
return;
|
| 975 |
-
}
|
| 976 |
-
if (i < text.length) {
|
| 977 |
-
element.textContent += text.charAt(i);
|
| 978 |
-
i++;
|
| 979 |
-
scrollToBottom();
|
| 980 |
-
typingTimeout = setTimeout(type, speed + Math.random() * 5);
|
| 981 |
-
} else {
|
| 982 |
-
element.classList.remove('typing-cursor');
|
| 983 |
-
resolve();
|
| 984 |
}
|
| 985 |
}
|
| 986 |
-
|
| 987 |
-
|
| 988 |
-
|
| 989 |
-
|
| 990 |
-
function addActions(wrapperElement, textToCopy) {
|
| 991 |
-
const actionsDiv = document.createElement('div');
|
| 992 |
-
actionsDiv.className = 'bot-actions';
|
| 993 |
-
|
| 994 |
-
const copyBtn = document.createElement('button');
|
| 995 |
-
copyBtn.className = 'action-btn';
|
| 996 |
-
copyBtn.innerHTML = `<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><rect x="9" y="9" width="13" height="13" rx="2" ry="2"></rect><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"></path></svg>`;
|
| 997 |
-
copyBtn.onclick = () => { navigator.clipboard.writeText(textToCopy); };
|
| 998 |
-
|
| 999 |
-
const regenBtn = document.createElement('button');
|
| 1000 |
-
regenBtn.className = 'action-btn';
|
| 1001 |
-
regenBtn.innerHTML = `<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="M23 4v6h-6"></path><path d="M1 20v-6h6"></path><path d="M3.51 9a9 9 0 0 1 14.85-3.36L23 10M1 14l4.64 4.36A9 9 0 0 0 20.49 15"></path></svg>`;
|
| 1002 |
-
regenBtn.onclick = () => { sendMessage(lastUserPrompt); };
|
| 1003 |
-
|
| 1004 |
-
actionsDiv.appendChild(copyBtn);
|
| 1005 |
-
actionsDiv.appendChild(regenBtn);
|
| 1006 |
-
wrapperElement.appendChild(actionsDiv);
|
| 1007 |
-
requestAnimationFrame(() => actionsDiv.style.opacity = "1");
|
| 1008 |
-
scrollToBottom();
|
| 1009 |
-
}
|
| 1010 |
-
|
| 1011 |
-
userInput.addEventListener('keydown', (e) => {
|
| 1012 |
-
if (e.key === 'Enter') handleBtnClick();
|
| 1013 |
-
});
|
| 1014 |
-
|
| 1015 |
-
window.onload = () => userInput.focus();
|
| 1016 |
-
</script>
|
| 1017 |
</body>
|
| 1018 |
</html>
|
| 1019 |
-
"""
|
| 1020 |
|
| 1021 |
if __name__ == "__main__":
|
| 1022 |
-
|
| 1023 |
-
print(f"\n🚀 MTP Inteligente iniciado en puerto {port}")
|
| 1024 |
-
print(f"🌐 http://0.0.0.0:{port}")
|
| 1025 |
-
|
| 1026 |
-
uvicorn.run(app, host="0.0.0.0", port=port, log_level="info")
|
|
|
|
| 1 |
import os
|
|
|
|
| 2 |
import torch
|
| 3 |
import json
|
|
|
|
|
|
|
| 4 |
import re
|
| 5 |
+
from fastapi import FastAPI
|
| 6 |
from fastapi.responses import HTMLResponse
|
| 7 |
from fastapi.middleware.cors import CORSMiddleware
|
| 8 |
from pydantic import BaseModel, Field
|
|
|
|
| 13 |
import torch.nn.functional as F
|
| 14 |
import sentencepiece as spm
|
| 15 |
|
| 16 |
+
# ====================== CONFIGURACIÓN ======================
|
| 17 |
+
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 18 |
+
print(f"📱 Dispositivo: {DEVICE}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
|
| 20 |
MODEL_REPO = "TeszenAI/MTP-3"
|
| 21 |
|
| 22 |
+
# ====================== ARQUITECTURA DEL MODELO ======================
|
|
|
|
|
|
|
| 23 |
class LayerNorm(nn.Module):
|
| 24 |
def __init__(self, d_model: int, eps: float = 1e-5):
|
| 25 |
super().__init__()
|
|
|
|
| 98 |
return x + self.pe[:, :x.size(1), :]
|
| 99 |
|
| 100 |
class MTPModel(nn.Module):
|
| 101 |
+
def __init__(self, vocab_size: int, d_model: int = 512, n_heads: int = 8,
|
| 102 |
+
n_layers: int = 8, d_ff: int = 2048, dropout: float = 0.1, max_len: int = 512):
|
| 103 |
super().__init__()
|
| 104 |
self.vocab_size = vocab_size
|
| 105 |
self.d_model = d_model
|
|
|
|
| 109 |
self.blocks = nn.ModuleList([TransformerBlock(d_model, n_heads, d_ff, dropout) for _ in range(n_layers)])
|
| 110 |
self.norm = LayerNorm(d_model)
|
| 111 |
self.lm_head = nn.Linear(d_model, vocab_size)
|
|
|
|
| 112 |
def forward(self, x, mask=None):
|
| 113 |
if mask is None:
|
| 114 |
mask = torch.tril(torch.ones(x.size(1), x.size(1))).unsqueeze(0).unsqueeze(0).to(x.device)
|
|
|
|
| 119 |
x = self.norm(x)
|
| 120 |
return self.lm_head(x)
|
| 121 |
|
| 122 |
+
# ====================== DESCARGA DEL MODELO ======================
|
| 123 |
+
print(f"📦 Cargando modelo...")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 124 |
|
| 125 |
if os.path.exists("mtp_repo") and os.path.exists("mtp_repo/mtp_model.pt"):
|
|
|
|
| 126 |
repo_path = "mtp_repo"
|
| 127 |
else:
|
| 128 |
try:
|
| 129 |
+
repo_path = snapshot_download(repo_id=MODEL_REPO, repo_type="model", local_dir="mtp_repo", resume_download=True)
|
| 130 |
except:
|
| 131 |
repo_path = "mtp_repo"
|
| 132 |
|
| 133 |
+
# Configuración
|
| 134 |
config_path = os.path.join(repo_path, "config.json")
|
| 135 |
if os.path.exists(config_path):
|
| 136 |
with open(config_path, "r") as f:
|
| 137 |
config = json.load(f)
|
| 138 |
else:
|
| 139 |
+
config = {"vocab_size": 10000, "d_model": 512, "n_heads": 8, "n_layers": 8, "d_ff": 2048, "dropout": 0.1, "max_len": 512}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 140 |
|
| 141 |
+
# Tokenizador
|
| 142 |
tokenizer_path = os.path.join(repo_path, "mtp_tokenizer.model")
|
| 143 |
if os.path.exists(tokenizer_path):
|
| 144 |
sp = spm.SentencePieceProcessor()
|
| 145 |
sp.load(tokenizer_path)
|
| 146 |
VOCAB_SIZE = sp.get_piece_size()
|
| 147 |
config["vocab_size"] = VOCAB_SIZE
|
|
|
|
| 148 |
else:
|
| 149 |
sp = None
|
| 150 |
+
VOCAB_SIZE = 10000
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 151 |
|
| 152 |
model = MTPModel(**config)
|
| 153 |
model.to(DEVICE)
|
| 154 |
|
|
|
|
| 155 |
model_path = os.path.join(repo_path, "mtp_model.pt")
|
| 156 |
if os.path.exists(model_path):
|
| 157 |
try:
|
| 158 |
state_dict = torch.load(model_path, map_location=DEVICE)
|
| 159 |
model.load_state_dict(state_dict)
|
| 160 |
+
print("✅ Modelo cargado")
|
| 161 |
+
except:
|
| 162 |
+
print("⚠️ Error cargando pesos")
|
|
|
|
| 163 |
model.eval()
|
| 164 |
|
| 165 |
+
# ====================== API ======================
|
| 166 |
+
app = FastAPI()
|
| 167 |
+
app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 168 |
|
| 169 |
class PromptRequest(BaseModel):
|
| 170 |
text: str = Field(..., max_length=2000)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 171 |
|
| 172 |
+
class TokenizerWrapper:
|
| 173 |
+
def __init__(self, sp_model):
|
| 174 |
+
self.sp = sp_model
|
| 175 |
+
def encode(self, text):
|
| 176 |
+
if self.sp is None:
|
| 177 |
+
return [ord(c) % 1000 for c in text[:200]]
|
| 178 |
+
return self.sp.encode(text)
|
| 179 |
+
def decode(self, tokens):
|
| 180 |
+
if self.sp is None:
|
| 181 |
+
return ''.join([chr(t % 128) if 32 <= t % 128 < 127 else ' ' for t in tokens])
|
| 182 |
+
return self.sp.decode(tokens)
|
| 183 |
+
def eos_id(self):
|
| 184 |
+
return self.sp.eos_id() if self.sp else 3
|
| 185 |
|
| 186 |
+
tokenizer = TokenizerWrapper(sp)
|
| 187 |
+
|
| 188 |
+
# Diccionario de respuestas por categoría (fallback cuando el modelo alucina)
|
| 189 |
+
RESPUESTAS_FALLBACK = {
|
| 190 |
+
"marketing": "El marketing digital incluye estrategias como SEO, marketing en redes sociales, email marketing, publicidad pagada y marketing de contenidos. ¿Te gustaría que profundice en alguna de estas áreas?",
|
| 191 |
+
"blackpink": "BLACKPINK es un grupo femenino de K-pop formado por Jisoo, Jennie, Rosé y Lisa. Tienen éxitos como 'Ddu-Du Ddu-Du', 'Kill This Love' y 'How You Like That'.",
|
| 192 |
+
"bts": "BTS es un grupo masculino de K-pop formado por RM, Jin, Suga, J-Hope, Jimin, V y Jungkook. Son conocidos por éxitos como 'Dynamite', 'Butter' y 'Boy With Luv'.",
|
| 193 |
+
"default": "Lo siento, no entendí bien tu pregunta. ¿Podrías reformularla? Estoy aquí para ayudarte con marketing, K-pop (BLACKPINK, BTS), tecnología, y más."
|
| 194 |
+
}
|
| 195 |
+
|
| 196 |
+
def detectar_tema(texto):
|
| 197 |
+
texto = texto.lower()
|
| 198 |
+
if "marketing" in texto or "seo" in texto or "publicidad" in texto or "redes sociales" in texto:
|
| 199 |
+
return "marketing"
|
| 200 |
+
if "blackpink" in texto or "jisoo" in texto or "jennie" in texto or "rosé" in texto or "lisa" in texto:
|
| 201 |
+
return "blackpink"
|
| 202 |
+
if "bts" in texto or "rm" in texto or "jin" in texto or "suga" in texto or "j-hope" in texto or "jimin" in texto or "v" in texto or "jungkook" in texto:
|
| 203 |
+
return "bts"
|
| 204 |
+
return None
|
| 205 |
+
|
| 206 |
+
def generar_respuesta(prompt, max_length=200, temperature=0.7):
|
| 207 |
+
formatted = f"### Instrucción:\n{prompt}\n\n### Respuesta:\n"
|
| 208 |
+
input_ids = tokenizer.encode(formatted)
|
| 209 |
generated = input_ids.copy()
|
| 210 |
eos_id = tokenizer.eos_id()
|
| 211 |
|
| 212 |
+
for _ in range(max_length):
|
| 213 |
+
input_tensor = torch.tensor([generated[-model.max_len:]], dtype=torch.long).to(DEVICE)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 214 |
with torch.no_grad():
|
| 215 |
logits = model(input_tensor)
|
| 216 |
next_logits = logits[0, -1, :] / temperature
|
| 217 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 218 |
probs = F.softmax(next_logits, dim=-1)
|
| 219 |
next_token = torch.multinomial(probs, 1).item()
|
| 220 |
|
|
|
|
| 221 |
if next_token == eos_id:
|
| 222 |
break
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 223 |
generated.append(next_token)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 224 |
|
|
|
|
| 225 |
response = tokenizer.decode(generated)
|
|
|
|
|
|
|
| 226 |
if "### Respuesta:" in response:
|
| 227 |
response = response.split("### Respuesta:")[-1].strip()
|
|
|
|
|
|
|
|
|
|
|
|
|
| 228 |
|
| 229 |
+
# Limpiar caracteres basura
|
| 230 |
+
response = re.sub(r'[^\w\s\u00C0-\u00FF.,!?¿¡\-:;"]+', ' ', response)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 231 |
response = re.sub(r'\s+', ' ', response).strip()
|
| 232 |
|
| 233 |
+
# Verificar si la respuesta es coherente
|
| 234 |
+
if len(response) < 5 or "kSq" in response or "%" in response or "{" in response:
|
| 235 |
+
tema = detectar_tema(prompt)
|
| 236 |
+
if tema:
|
| 237 |
+
response = RESPUESTAS_FALLBACK.get(tema, RESPUESTAS_FALLBACK["default"])
|
| 238 |
+
else:
|
| 239 |
+
response = RESPUESTAS_FALLBACK["default"]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 240 |
|
| 241 |
return response
|
| 242 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 243 |
@app.post("/generate")
|
| 244 |
async def generate(req: PromptRequest):
|
|
|
|
|
|
|
|
|
|
| 245 |
user_input = req.text.strip()
|
| 246 |
if not user_input:
|
| 247 |
+
return {"reply": ""}
|
|
|
|
|
|
|
|
|
|
|
|
|
| 248 |
|
| 249 |
try:
|
| 250 |
+
response = generar_respuesta(user_input)
|
| 251 |
+
return {"reply": response}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 252 |
except Exception as e:
|
| 253 |
+
print(f"Error: {e}")
|
| 254 |
+
tema = detectar_tema(user_input)
|
| 255 |
+
return {"reply": RESPUESTAS_FALLBACK.get(tema, RESPUESTAS_FALLBACK["default"])}
|
|
|
|
|
|
|
|
|
|
|
|
|
| 256 |
|
| 257 |
+
@app.get("/")
|
| 258 |
+
async def chat_ui():
|
| 259 |
+
return HTMLResponse("""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 260 |
<!DOCTYPE html>
|
| 261 |
+
<html>
|
| 262 |
<head>
|
| 263 |
+
<title>MTP Asistente</title>
|
| 264 |
+
<meta charset="UTF-8">
|
| 265 |
+
<meta name="viewport" content="width=device-width, initial-scale=1.0">
|
| 266 |
+
<style>
|
| 267 |
+
body { font-family: Arial, sans-serif; max-width: 800px; margin: 0 auto; padding: 20px; background: #131314; color: #e3e3e3; }
|
| 268 |
+
#chat { height: 500px; border: 1px solid #333; overflow-y: auto; padding: 10px; margin-bottom: 10px; border-radius: 10px; background: #1E1F20; }
|
| 269 |
+
.user { text-align: right; margin: 10px; }
|
| 270 |
+
.user span { background: #4a9eff; padding: 8px 15px; border-radius: 18px; display: inline-block; }
|
| 271 |
+
.bot { text-align: left; margin: 10px; }
|
| 272 |
+
.bot span { background: #282a2c; padding: 8px 15px; border-radius: 18px; display: inline-block; }
|
| 273 |
+
input { width: 80%; padding: 10px; border-radius: 25px; border: none; background: #1E1F20; color: white; }
|
| 274 |
+
button { padding: 10px 20px; border-radius: 25px; border: none; background: #4a9eff; color: white; cursor: pointer; }
|
| 275 |
+
.typing { opacity: 0.7; font-style: italic; }
|
| 276 |
+
</style>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 277 |
</head>
|
| 278 |
<body>
|
| 279 |
+
<h1>🤖 MTP Asistente</h1>
|
| 280 |
+
<div id="chat">
|
| 281 |
+
<div class="bot"><span>¡Hola! Soy MTP. Puedo ayudarte con marketing, K-pop (BLACKPINK, BTS), tecnología y más. ¿Qué necesitas?</span></div>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 282 |
</div>
|
| 283 |
+
<div>
|
| 284 |
+
<input type="text" id="input" placeholder="Escribe tu mensaje..." autocomplete="off">
|
| 285 |
+
<button onclick="sendMessage()">Enviar</button>
|
| 286 |
</div>
|
| 287 |
+
<script>
|
| 288 |
+
const chat = document.getElementById('chat');
|
| 289 |
+
const input = document.getElementById('input');
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 290 |
|
| 291 |
+
function addMessage(text, sender) {
|
| 292 |
+
const div = document.createElement('div');
|
| 293 |
+
div.className = sender;
|
| 294 |
+
div.innerHTML = `<span>${text}</span>`;
|
| 295 |
+
chat.appendChild(div);
|
| 296 |
+
chat.scrollTop = chat.scrollHeight;
|
|
|
|
|
|
|
|
|
|
| 297 |
}
|
| 298 |
|
| 299 |
+
async function sendMessage() {
|
| 300 |
+
const text = input.value.trim();
|
| 301 |
+
if (!text) return;
|
| 302 |
+
input.value = '';
|
| 303 |
+
addMessage(text, 'user');
|
| 304 |
+
|
| 305 |
+
const loadingDiv = document.createElement('div');
|
| 306 |
+
loadingDiv.className = 'bot';
|
| 307 |
+
loadingDiv.innerHTML = '<span class="typing">✍️ Pensando...</span>';
|
| 308 |
+
chat.appendChild(loadingDiv);
|
| 309 |
+
chat.scrollTop = chat.scrollHeight;
|
| 310 |
+
|
| 311 |
+
try {
|
| 312 |
+
const response = await fetch('/generate', {
|
| 313 |
+
method: 'POST',
|
| 314 |
+
headers: { 'Content-Type': 'application/json' },
|
| 315 |
+
body: JSON.stringify({ text: text })
|
| 316 |
+
});
|
| 317 |
+
const data = await response.json();
|
| 318 |
+
loadingDiv.remove();
|
| 319 |
+
addMessage(data.reply, 'bot');
|
| 320 |
+
} catch (error) {
|
| 321 |
+
loadingDiv.innerHTML = '<span class="typing">❌ Error de conexión</span>';
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 322 |
}
|
| 323 |
}
|
| 324 |
+
|
| 325 |
+
input.addEventListener('keypress', (e) => { if (e.key === 'Enter') sendMessage(); });
|
| 326 |
+
</script>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 327 |
</body>
|
| 328 |
</html>
|
| 329 |
+
""")
|
| 330 |
|
| 331 |
if __name__ == "__main__":
|
| 332 |
+
uvicorn.run(app, host="0.0.0.0", port=7860)
|
|
|
|
|
|
|
|
|
|
|
|