Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,593 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# ================================================================
|
| 2 |
+
# MTP-2.5 - app.py para Hugging Face Space (Gradio, CPU)
|
| 3 |
+
# Generacion 2: RoPE + SwiGLU (reemplaza position embeddings aprendidos y el
|
| 4 |
+
# GELU-MLP de MTP-1.x). Requiere un checkpoint entrenado con la Celda 1/2 de
|
| 5 |
+
# MTP-2.5; NO carga checkpoints de MTP-2.0 (RMSNorm cambia el state_dict) ni de MTP-1.x.
|
| 6 |
+
#
|
| 7 |
+
# OPTIMIZACIÓN DE VELOCIDAD (sin tocar el resto de la logica de muestreo):
|
| 8 |
+
# - KV-cache en la atención: en generación autoregresiva, cada paso
|
| 9 |
+
# antes recomputaba TODO el contexto desde cero (O(n^2) en total).
|
| 10 |
+
# Ahora se reutiliza lo ya calculado y solo se procesa el token
|
| 11 |
+
# nuevo (O(n) en total). Es el mismo cálculo matemático, solo que
|
| 12 |
+
# no se repite trabajo ya hecho.
|
| 13 |
+
# - F.scaled_dot_product_attention: kernel fusionado de PyTorch,
|
| 14 |
+
# mismo resultado que el softmax manual pero más rápido en CPU.
|
| 15 |
+
# Si la versión de PyTorch no lo trae, cae automáticamente al
|
| 16 |
+
# cálculo manual (fallback), así que no se rompe en ningún entorno.
|
| 17 |
+
# - repetition_penalty vectorizado + bloqueo de n-gramas repetidos
|
| 18 |
+
# (evita que la respuesta final copie literalmente un fragmento ya
|
| 19 |
+
# generado, sin que esto sea un "modelo de n-gramas": el modelo que
|
| 20 |
+
# predice sigue siendo 100% transformer).
|
| 21 |
+
# ================================================================
|
| 22 |
+
import os
|
| 23 |
+
import math
|
| 24 |
+
import time
|
| 25 |
+
import logging
|
| 26 |
+
import threading
|
| 27 |
+
import traceback
|
| 28 |
+
import torch
|
| 29 |
+
import torch.nn as nn
|
| 30 |
+
import torch.nn.functional as F
|
| 31 |
+
import gradio as gr
|
| 32 |
+
import sentencepiece as spm
|
| 33 |
+
from starlette.middleware import Middleware
|
| 34 |
+
from fastapi.middleware.cors import CORSMiddleware
|
| 35 |
+
from pydantic import BaseModel
|
| 36 |
+
from typing import Optional
|
| 37 |
+
from huggingface_hub import hf_hub_download
|
| 38 |
+
|
| 39 |
+
# ---------------- Logging ----------------
|
| 40 |
+
# print() se pierde facil entre el ruido de arranque de Gradio/Starlette en
|
| 41 |
+
# los logs de un Space; con logging queda todo con timestamp y nivel, y es
|
| 42 |
+
# mas facil de filtrar si algo sale mal en produccion.
|
| 43 |
+
logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
|
| 44 |
+
logger = logging.getLogger("mtp")
|
| 45 |
+
|
| 46 |
+
# ---------------- Dispositivo ----------------
|
| 47 |
+
# Autodetecta GPU si el Space corre en un tier con GPU (DEVICE=cuda por
|
| 48 |
+
# variable de entorno tambien fuerza el valor si hace falta). Si no hay
|
| 49 |
+
# GPU disponible, cae a CPU como siempre.
|
| 50 |
+
DEVICE = os.environ.get("DEVICE") or ("cuda" if torch.cuda.is_available() else "cpu")
|
| 51 |
+
|
| 52 |
+
if DEVICE == "cpu":
|
| 53 |
+
# Limita hilos a los núcleos disponibles (evita overhead en Spaces pequeños).
|
| 54 |
+
# No tiene sentido en GPU, donde el computo no lo hace la CPU.
|
| 55 |
+
torch.set_num_threads(max(1, os.cpu_count() or 1))
|
| 56 |
+
# set_num_interop_threads solo puede llamarse una vez y antes de cualquier
|
| 57 |
+
# operación paralela; lo protegemos por si el entorno ya lo fijó.
|
| 58 |
+
try:
|
| 59 |
+
torch.set_num_interop_threads(1)
|
| 60 |
+
except RuntimeError:
|
| 61 |
+
pass
|
| 62 |
+
|
| 63 |
+
torch.set_grad_enabled(False) # solo inferencia, nunca necesitamos gradientes
|
| 64 |
+
|
| 65 |
+
# Disponibilidad de scaled_dot_product_attention (PyTorch >= 2.0).
|
| 66 |
+
# Si no está disponible, usamos el softmax manual original como fallback.
|
| 67 |
+
_HAS_SDPA = hasattr(F, "scaled_dot_product_attention")
|
| 68 |
+
|
| 69 |
+
REPO_ID = os.environ.get("MTP_REPO_ID", "TeszenAI/MTP-2.7") # <-- ajusta al nombre real de tu repo/Space en el Hub
|
| 70 |
+
FILENAME = os.environ.get("MTP_FILENAME", "MTP2_7_MODEL.pt")
|
| 71 |
+
|
| 72 |
+
# Origenes permitidos para CORS. Por defecto "*" (como antes), pero se puede
|
| 73 |
+
# restringir en produccion con la variable de entorno MTP_ALLOWED_ORIGINS
|
| 74 |
+
# (separados por coma), por ejemplo: "https://teszen.com,https://www.teszen.com"
|
| 75 |
+
ALLOWED_ORIGINS = [
|
| 76 |
+
o.strip() for o in os.environ.get("MTP_ALLOWED_ORIGINS", "*").split(",") if o.strip()
|
| 77 |
+
] or ["*"]
|
| 78 |
+
|
| 79 |
+
# Techo de caracteres del input antes de tokenizar. No es por seguridad (el
|
| 80 |
+
# modelo igual solo "ve" los ultimos BLOCK_SIZE tokens), es para no perder
|
| 81 |
+
# tiempo tokenizando un texto absurdamente largo por error o abuso.
|
| 82 |
+
MAX_INPUT_CHARS = 4000
|
| 83 |
+
|
| 84 |
+
# Si es "1", los errores devueltos por /generate incluyen el detalle interno
|
| 85 |
+
# de la excepcion (util mientras desarrollas). En produccion, dejar en "0"
|
| 86 |
+
# para no filtrarle al cliente detalles internos del servidor.
|
| 87 |
+
DEBUG_ERRORS = os.environ.get("MTP_DEBUG_ERRORS", "0") == "1"
|
| 88 |
+
|
| 89 |
+
# Serializa las llamadas a generate(): sin esto, dos requests concurrentes
|
| 90 |
+
# (por ejemplo la UI de Gradio y el endpoint /generate al mismo tiempo, o
|
| 91 |
+
# varias visitas simultaneas al sitio) compiten por los mismos hilos de CPU
|
| 92 |
+
# y todas terminan mas lentas en vez de una rapida y la otra esperando. Con
|
| 93 |
+
# el lock, cada generacion corre de punta a punta antes de que empiece la
|
| 94 |
+
# siguiente -- mismo comportamiento de fondo que antes bajo carga baja, pero
|
| 95 |
+
# estable bajo carga alta en vez de degradarse.
|
| 96 |
+
_generation_lock = threading.Lock()
|
| 97 |
+
|
| 98 |
+
# ---------------- Arquitectura MTP-2.x: RoPE + SwiGLU, con KV-cache ----------------
|
| 99 |
+
def rotate_half(x):
|
| 100 |
+
x1, x2 = x.chunk(2, dim=-1)
|
| 101 |
+
return torch.cat((-x2, x1), dim=-1)
|
| 102 |
+
|
| 103 |
+
def apply_rope(q, k, cos, sin):
|
| 104 |
+
cos = cos.unsqueeze(0).unsqueeze(0)
|
| 105 |
+
sin = sin.unsqueeze(0).unsqueeze(0)
|
| 106 |
+
q_rot = (q * cos) + (rotate_half(q) * sin)
|
| 107 |
+
k_rot = (k * cos) + (rotate_half(k) * sin)
|
| 108 |
+
return q_rot, k_rot
|
| 109 |
+
|
| 110 |
+
class RotaryEmbedding(nn.Module):
|
| 111 |
+
def __init__(self, head_dim, max_seq_len, base=10000):
|
| 112 |
+
super().__init__()
|
| 113 |
+
inv_freq = 1.0 / (base ** (torch.arange(0, head_dim, 2).float() / head_dim))
|
| 114 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
| 115 |
+
self._build_cache(max_seq_len)
|
| 116 |
+
|
| 117 |
+
def _build_cache(self, seq_len):
|
| 118 |
+
t = torch.arange(seq_len, dtype=self.inv_freq.dtype, device=self.inv_freq.device)
|
| 119 |
+
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
| 120 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
| 121 |
+
self.register_buffer("cos_cached", emb.cos(), persistent=False)
|
| 122 |
+
self.register_buffer("sin_cached", emb.sin(), persistent=False)
|
| 123 |
+
self.max_seq_len_cached = seq_len
|
| 124 |
+
|
| 125 |
+
def forward(self, seq_len, device, dtype, offset=0):
|
| 126 |
+
# Con KV-cache, `offset` es cuantos tokens ya estan en la cache: el
|
| 127 |
+
# token nuevo necesita el angulo correspondiente a SU posicion
|
| 128 |
+
# absoluta, no a la posicion relativa dentro de este forward.
|
| 129 |
+
if offset + seq_len > self.max_seq_len_cached:
|
| 130 |
+
self._build_cache(offset + seq_len)
|
| 131 |
+
cos = self.cos_cached[offset:offset + seq_len].to(device=device, dtype=dtype)
|
| 132 |
+
sin = self.sin_cached[offset:offset + seq_len].to(device=device, dtype=dtype)
|
| 133 |
+
return cos, sin
|
| 134 |
+
|
| 135 |
+
|
| 136 |
+
class CausalSelfAttention(nn.Module):
|
| 137 |
+
def __init__(self, n_embd, n_head, block_size, dropout):
|
| 138 |
+
super().__init__()
|
| 139 |
+
self.n_head = n_head
|
| 140 |
+
self.head_dim = n_embd // n_head
|
| 141 |
+
self.qkv = nn.Linear(n_embd, 3 * n_embd)
|
| 142 |
+
self.proj = nn.Linear(n_embd, n_embd)
|
| 143 |
+
self.attn_dropout = nn.Dropout(dropout)
|
| 144 |
+
self.resid_dropout = nn.Dropout(dropout)
|
| 145 |
+
mask = torch.tril(torch.ones(block_size, block_size)).view(1, 1, block_size, block_size)
|
| 146 |
+
self.register_buffer("mask", mask)
|
| 147 |
+
|
| 148 |
+
def forward(self, x, cos, sin, past_kv=None, use_cache=False):
|
| 149 |
+
B, T, C = x.shape
|
| 150 |
+
qkv = self.qkv(x)
|
| 151 |
+
q, k, v = qkv.split(C, dim=2)
|
| 152 |
+
q = q.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
|
| 153 |
+
k = k.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
|
| 154 |
+
v = v.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
|
| 155 |
+
|
| 156 |
+
# RoPE se aplica ANTES de guardar en cache, con el angulo absoluto de
|
| 157 |
+
# cada token (pasado por `cos`/`sin`, ya calculado con el offset
|
| 158 |
+
# correcto en MTP.forward). Asi el k cacheado ya trae rotada su
|
| 159 |
+
# posicion real y no hay que re-rotar nada en pasos futuros.
|
| 160 |
+
q, k = apply_rope(q, k, cos, sin)
|
| 161 |
+
|
| 162 |
+
if past_kv is not None:
|
| 163 |
+
past_k, past_v = past_kv
|
| 164 |
+
k = torch.cat([past_k, k], dim=2)
|
| 165 |
+
v = torch.cat([past_v, v], dim=2)
|
| 166 |
+
|
| 167 |
+
present_kv = (k, v) if use_cache else None
|
| 168 |
+
|
| 169 |
+
is_causal = (past_kv is None) and (T > 1)
|
| 170 |
+
|
| 171 |
+
if _HAS_SDPA:
|
| 172 |
+
out = F.scaled_dot_product_attention(
|
| 173 |
+
q, k, v, attn_mask=None, dropout_p=0.0, is_causal=is_causal,
|
| 174 |
+
)
|
| 175 |
+
else:
|
| 176 |
+
att = (q @ k.transpose(-2, -1)) / math.sqrt(self.head_dim)
|
| 177 |
+
if is_causal:
|
| 178 |
+
Tk = k.size(-2)
|
| 179 |
+
causal_mask = torch.tril(torch.ones(T, Tk, device=x.device, dtype=torch.bool))
|
| 180 |
+
att = att.masked_fill(~causal_mask, float("-inf"))
|
| 181 |
+
att = F.softmax(att, dim=-1)
|
| 182 |
+
att = self.attn_dropout(att)
|
| 183 |
+
out = att @ v
|
| 184 |
+
|
| 185 |
+
out = out.transpose(1, 2).contiguous().view(B, T, C)
|
| 186 |
+
out = self.resid_dropout(self.proj(out))
|
| 187 |
+
return out, present_kv
|
| 188 |
+
|
| 189 |
+
|
| 190 |
+
class SwiGLU(nn.Module):
|
| 191 |
+
def __init__(self, n_embd, dropout):
|
| 192 |
+
super().__init__()
|
| 193 |
+
hidden = int(2 * (4 * n_embd) / 3)
|
| 194 |
+
hidden = ((hidden + 7) // 8) * 8
|
| 195 |
+
self.w_gate = nn.Linear(n_embd, hidden, bias=False)
|
| 196 |
+
self.w_up = nn.Linear(n_embd, hidden, bias=False)
|
| 197 |
+
self.w_down = nn.Linear(hidden, n_embd, bias=False)
|
| 198 |
+
self.dropout = nn.Dropout(dropout)
|
| 199 |
+
|
| 200 |
+
def forward(self, x):
|
| 201 |
+
return self.dropout(self.w_down(F.silu(self.w_gate(x)) * self.w_up(x)))
|
| 202 |
+
|
| 203 |
+
|
| 204 |
+
class RMSNorm(nn.Module):
|
| 205 |
+
"""Debe coincidir exactamente con la version de entrenamiento. No
|
| 206 |
+
necesita ningun cambio para funcionar con KV-cache: normaliza cada
|
| 207 |
+
posicion de forma independiente, igual que LayerNorm."""
|
| 208 |
+
def __init__(self, dim, eps=1e-6):
|
| 209 |
+
super().__init__()
|
| 210 |
+
self.eps = eps
|
| 211 |
+
self.weight = nn.Parameter(torch.ones(dim))
|
| 212 |
+
|
| 213 |
+
def forward(self, x):
|
| 214 |
+
norm = x * torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + self.eps)
|
| 215 |
+
return norm * self.weight
|
| 216 |
+
|
| 217 |
+
|
| 218 |
+
class Block(nn.Module):
|
| 219 |
+
def __init__(self, n_embd, n_head, block_size, dropout):
|
| 220 |
+
super().__init__()
|
| 221 |
+
self.ln1 = RMSNorm(n_embd)
|
| 222 |
+
self.attn = CausalSelfAttention(n_embd, n_head, block_size, dropout)
|
| 223 |
+
self.ln2 = RMSNorm(n_embd)
|
| 224 |
+
self.ff = SwiGLU(n_embd, dropout)
|
| 225 |
+
|
| 226 |
+
def forward(self, x, cos, sin, past_kv=None, use_cache=False):
|
| 227 |
+
attn_out, present_kv = self.attn(self.ln1(x), cos, sin, past_kv=past_kv, use_cache=use_cache)
|
| 228 |
+
x = x + attn_out
|
| 229 |
+
x = x + self.ff(self.ln2(x))
|
| 230 |
+
return x, present_kv
|
| 231 |
+
|
| 232 |
+
|
| 233 |
+
class MTP(nn.Module):
|
| 234 |
+
def __init__(self, vocab_size, block_size, n_layer, n_head, n_embd, dropout):
|
| 235 |
+
super().__init__()
|
| 236 |
+
self.block_size = block_size
|
| 237 |
+
self.head_dim = n_embd // n_head
|
| 238 |
+
self.tok_emb = nn.Embedding(vocab_size, n_embd)
|
| 239 |
+
self.rope = RotaryEmbedding(self.head_dim, max_seq_len=block_size)
|
| 240 |
+
self.drop = nn.Dropout(dropout)
|
| 241 |
+
self.blocks = nn.ModuleList([Block(n_embd, n_head, block_size, dropout) for _ in range(n_layer)])
|
| 242 |
+
self.ln_f = RMSNorm(n_embd)
|
| 243 |
+
self.lm_head = nn.Linear(n_embd, vocab_size, bias=False)
|
| 244 |
+
self.lm_head.weight = self.tok_emb.weight
|
| 245 |
+
|
| 246 |
+
def forward(self, idx, past_key_values=None, use_cache=False, pos_offset=0):
|
| 247 |
+
B, T = idx.shape
|
| 248 |
+
x = self.tok_emb(idx)
|
| 249 |
+
x = self.drop(x)
|
| 250 |
+
cos, sin = self.rope(T, idx.device, x.dtype, offset=pos_offset)
|
| 251 |
+
|
| 252 |
+
new_past = [] if use_cache else None
|
| 253 |
+
for i, block in enumerate(self.blocks):
|
| 254 |
+
past_kv = past_key_values[i] if past_key_values is not None else None
|
| 255 |
+
x, present_kv = block(x, cos, sin, past_kv=past_kv, use_cache=use_cache)
|
| 256 |
+
if use_cache:
|
| 257 |
+
new_past.append(present_kv)
|
| 258 |
+
|
| 259 |
+
x = self.ln_f(x)
|
| 260 |
+
logits = self.lm_head(x)
|
| 261 |
+
return logits, new_past
|
| 262 |
+
|
| 263 |
+
|
| 264 |
+
# ---------------- Carga del checkpoint (una sola vez, al iniciar el Space) ----------------
|
| 265 |
+
def _download_checkpoint_with_retries(repo_id, filename, max_retries=5):
|
| 266 |
+
"""Reintenta la descarga con backoff. hf_hub_download puede fallar por un
|
| 267 |
+
corte de red momentaneo o un problema del backend Xet de HF; sin esto,
|
| 268 |
+
un solo fallo transitorio tira abajo el arranque completo del Space."""
|
| 269 |
+
last_err = None
|
| 270 |
+
for attempt in range(1, max_retries + 1):
|
| 271 |
+
try:
|
| 272 |
+
logger.info(f"Descargando checkpoint desde el Hub (intento {attempt}/{max_retries})...")
|
| 273 |
+
return hf_hub_download(repo_id=repo_id, filename=filename)
|
| 274 |
+
except Exception as e:
|
| 275 |
+
last_err = e
|
| 276 |
+
logger.warning(f"Fallo la descarga del checkpoint: {e}")
|
| 277 |
+
if attempt < max_retries:
|
| 278 |
+
time.sleep(5 * attempt)
|
| 279 |
+
raise RuntimeError(f"No se pudo descargar el checkpoint tras {max_retries} intentos") from last_err
|
| 280 |
+
|
| 281 |
+
ckpt_path = _download_checkpoint_with_retries(REPO_ID, FILENAME)
|
| 282 |
+
checkpoint = torch.load(ckpt_path, map_location=DEVICE)
|
| 283 |
+
|
| 284 |
+
cfg = checkpoint["config"]
|
| 285 |
+
special = checkpoint["special_tokens"]
|
| 286 |
+
gen_defaults = checkpoint["generation_defaults"]
|
| 287 |
+
|
| 288 |
+
PAD_ID, BOS_ID, EOS_ID, UNK_ID = special["pad_id"], special["bos_id"], special["eos_id"], special["unk_id"]
|
| 289 |
+
|
| 290 |
+
# El tokenizer es BPE (SentencePiece) entrenado desde cero junto con el modelo.
|
| 291 |
+
# No es un modelo preentrenado externo: viene embebido como bytes dentro del
|
| 292 |
+
# mismo checkpoint que los pesos. Se carga directo desde memoria con
|
| 293 |
+
# load_from_serialized_proto, sin necesidad de escribirlo a disco primero.
|
| 294 |
+
sp = spm.SentencePieceProcessor()
|
| 295 |
+
sp.load_from_serialized_proto(checkpoint["spm_model_bytes"])
|
| 296 |
+
|
| 297 |
+
model = MTP(
|
| 298 |
+
vocab_size=cfg["vocab_size"], block_size=cfg["block_size"],
|
| 299 |
+
n_layer=cfg["n_layer"], n_head=cfg["n_head"],
|
| 300 |
+
n_embd=cfg["n_embd"], dropout=cfg["dropout"],
|
| 301 |
+
).to(DEVICE)
|
| 302 |
+
model.load_state_dict(checkpoint["model_state_dict"])
|
| 303 |
+
model.eval()
|
| 304 |
+
|
| 305 |
+
BLOCK_SIZE = cfg["block_size"]
|
| 306 |
+
|
| 307 |
+
logger.info(
|
| 308 |
+
f"MTP cargado ({checkpoint['meta']['model_name']}, "
|
| 309 |
+
f"entrenado con {checkpoint['meta']['trained_examples']} ejemplos) "
|
| 310 |
+
f"| device={DEVICE} | SDPA={'sí' if _HAS_SDPA else 'no (fallback manual)'}"
|
| 311 |
+
)
|
| 312 |
+
|
| 313 |
+
|
| 314 |
+
import re as _re_indent
|
| 315 |
+
|
| 316 |
+
def protect_indentation(text):
|
| 317 |
+
"""Debe coincidir exactamente con la funcion usada en el entrenamiento."""
|
| 318 |
+
lines = text.split("\n")
|
| 319 |
+
new_lines = []
|
| 320 |
+
for line in lines:
|
| 321 |
+
stripped = line.lstrip(" ")
|
| 322 |
+
n_spaces = len(line) - len(stripped)
|
| 323 |
+
n_levels = n_spaces // 4
|
| 324 |
+
remainder = n_spaces % 4
|
| 325 |
+
if n_levels > 0:
|
| 326 |
+
prefix = " " + " ".join(["<tab>"] * n_levels) + " " + " " * remainder
|
| 327 |
+
else:
|
| 328 |
+
prefix = " " * remainder
|
| 329 |
+
new_lines.append(prefix + stripped)
|
| 330 |
+
text = "\n".join(new_lines)
|
| 331 |
+
def _repl(m):
|
| 332 |
+
n = len(m.group())
|
| 333 |
+
return " " + " ".join(["<nl>"] * n) + " "
|
| 334 |
+
text = _re_indent.sub(r"\n+", _repl, text)
|
| 335 |
+
return text
|
| 336 |
+
|
| 337 |
+
|
| 338 |
+
def restore_indentation(text):
|
| 339 |
+
text = _re_indent.sub(r"(<nl>\s*)+", lambda m: "\n" * m.group().count("<nl>"), text)
|
| 340 |
+
text = _re_indent.sub(r"(<tab>\s*)+", lambda m: " " * m.group().count("<tab>"), text)
|
| 341 |
+
return text
|
| 342 |
+
|
| 343 |
+
|
| 344 |
+
def encode_text(s):
|
| 345 |
+
return sp.encode(protect_indentation(s), out_type=int)
|
| 346 |
+
|
| 347 |
+
|
| 348 |
+
def decode_ids(ids):
|
| 349 |
+
text = sp.decode([i for i in ids if i not in (PAD_ID, BOS_ID, EOS_ID)])
|
| 350 |
+
return restore_indentation(text)
|
| 351 |
+
|
| 352 |
+
|
| 353 |
+
def _block_repeated_ngrams(generated_ids, logits, ngram_size):
|
| 354 |
+
"""Prohibe repetir literalmente un n-grama ya generado en esta misma
|
| 355 |
+
respuesta (tecnica de decoding tipo GPT-2/3, no un modelo de n-gramas:
|
| 356 |
+
el modelo que predice sigue siendo 100% transformer con KV-cache)."""
|
| 357 |
+
if ngram_size <= 0 or len(generated_ids) < ngram_size:
|
| 358 |
+
return logits
|
| 359 |
+
prefix = tuple(generated_ids[-(ngram_size - 1):])
|
| 360 |
+
banned = set()
|
| 361 |
+
for i in range(len(generated_ids) - ngram_size + 1):
|
| 362 |
+
if tuple(generated_ids[i:i + ngram_size - 1]) == prefix:
|
| 363 |
+
banned.add(generated_ids[i + ngram_size - 1])
|
| 364 |
+
if banned:
|
| 365 |
+
logits[0, list(banned)] = float("-inf")
|
| 366 |
+
return logits
|
| 367 |
+
|
| 368 |
+
|
| 369 |
+
# ---------------- Generación (con KV-cache) ----------------
|
| 370 |
+
@torch.inference_mode()
|
| 371 |
+
def generate(idx, max_new_tokens, temperature, top_k, top_p, repetition_penalty, no_repeat_ngram_size=3):
|
| 372 |
+
past_key_values = None
|
| 373 |
+
cache_len = 0 # cuántos tokens del extremo derecho de `idx` ya están en la caché
|
| 374 |
+
|
| 375 |
+
for _ in range(max_new_tokens):
|
| 376 |
+
total_len = idx.shape[1]
|
| 377 |
+
|
| 378 |
+
if total_len <= BLOCK_SIZE:
|
| 379 |
+
if past_key_values is None:
|
| 380 |
+
# Primer paso: una sola pasada ("prefill") sobre todo el prompt.
|
| 381 |
+
logits, past_key_values = model(idx, use_cache=True)
|
| 382 |
+
cache_len = total_len
|
| 383 |
+
else:
|
| 384 |
+
# Pasos siguientes: solo se procesa el último token generado,
|
| 385 |
+
# reutilizando la caché de todo lo anterior.
|
| 386 |
+
last_token = idx[:, -1:]
|
| 387 |
+
logits, past_key_values = model(
|
| 388 |
+
last_token,
|
| 389 |
+
past_key_values=past_key_values,
|
| 390 |
+
use_cache=True,
|
| 391 |
+
pos_offset=cache_len,
|
| 392 |
+
)
|
| 393 |
+
cache_len += 1
|
| 394 |
+
logits = logits[:, -1, :]
|
| 395 |
+
else:
|
| 396 |
+
# Se superó block_size: mismo comportamiento que el modelo original
|
| 397 |
+
# (ventana deslizante recalculada por completo). Solo ocurre en
|
| 398 |
+
# respuestas muy largas; la caché se reinicia para esa ventana.
|
| 399 |
+
idx_cond = idx[:, -BLOCK_SIZE:]
|
| 400 |
+
logits, past_key_values = model(idx_cond, use_cache=True)
|
| 401 |
+
cache_len = BLOCK_SIZE
|
| 402 |
+
logits = logits[:, -1, :]
|
| 403 |
+
|
| 404 |
+
logits = logits / max(temperature, 1e-5)
|
| 405 |
+
|
| 406 |
+
if repetition_penalty and repetition_penalty != 1.0:
|
| 407 |
+
# Vectorizado: antes era `for token_id in set(idx[0].tolist())`,
|
| 408 |
+
# un bucle Python nuevo por cada token generado.
|
| 409 |
+
unique_ids = torch.unique(idx[0])
|
| 410 |
+
logits[0, unique_ids] /= repetition_penalty
|
| 411 |
+
|
| 412 |
+
logits = _block_repeated_ngrams(idx[0].tolist(), logits, no_repeat_ngram_size)
|
| 413 |
+
|
| 414 |
+
if top_k is not None and top_k > 0:
|
| 415 |
+
v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
|
| 416 |
+
logits[logits < v[:, [-1]]] = float("-inf")
|
| 417 |
+
|
| 418 |
+
probs = F.softmax(logits, dim=-1)
|
| 419 |
+
|
| 420 |
+
if top_p is not None and 0 < top_p < 1:
|
| 421 |
+
sorted_probs, sorted_idx = torch.sort(probs, descending=True)
|
| 422 |
+
cum_probs = torch.cumsum(sorted_probs, dim=-1)
|
| 423 |
+
cutoff = cum_probs > top_p
|
| 424 |
+
cutoff[:, 1:] = cutoff[:, :-1].clone()
|
| 425 |
+
cutoff[:, 0] = False
|
| 426 |
+
sorted_probs[cutoff] = 0.0
|
| 427 |
+
sorted_probs = sorted_probs / sorted_probs.sum(dim=-1, keepdim=True)
|
| 428 |
+
next_id = sorted_idx.gather(-1, torch.multinomial(sorted_probs, 1))
|
| 429 |
+
else:
|
| 430 |
+
next_id = torch.multinomial(probs, num_samples=1)
|
| 431 |
+
|
| 432 |
+
idx = torch.cat([idx, next_id], dim=1)
|
| 433 |
+
if next_id.item() == EOS_ID:
|
| 434 |
+
break
|
| 435 |
+
|
| 436 |
+
return idx
|
| 437 |
+
|
| 438 |
+
|
| 439 |
+
def run_inference(text, max_new_tokens=None, temperature=None, top_k=None, top_p=None, repetition_penalty=None, no_repeat_ngram_size=None):
|
| 440 |
+
"""Núcleo de generación, reutilizado por la UI de Gradio y por la API /generate.
|
| 441 |
+
No reduce calidad por estar en CPU: usa exactamente el mismo muestreo
|
| 442 |
+
(top_k + top_p + repetition_penalty + bloqueo de n-gramas repetidos) que
|
| 443 |
+
en la Celda 2 de entrenamiento, solo que ahora con KV-cache es notablemente
|
| 444 |
+
más rápido en respuestas largas."""
|
| 445 |
+
max_new_tokens = int(max_new_tokens) if max_new_tokens else gen_defaults["max_new_tokens"]
|
| 446 |
+
temperature = float(temperature) if temperature is not None else gen_defaults["temperature"]
|
| 447 |
+
top_k = int(top_k) if top_k is not None else gen_defaults["top_k"]
|
| 448 |
+
top_p = float(top_p) if top_p is not None else gen_defaults["top_p"]
|
| 449 |
+
repetition_penalty = float(repetition_penalty) if repetition_penalty is not None else gen_defaults["repetition_penalty"]
|
| 450 |
+
no_repeat_ngram_size = int(no_repeat_ngram_size) if no_repeat_ngram_size is not None else gen_defaults.get("no_repeat_ngram_size", 3)
|
| 451 |
+
|
| 452 |
+
# Techo máximo de generación: 4000 no era realista en CPU (cada token
|
| 453 |
+
# adicional cuesta tiempo real). 700 sigue siendo una respuesta larga y
|
| 454 |
+
# mantiene el tiempo de respuesta bajo control en el peor caso.
|
| 455 |
+
MAX_TOKENS_HARD_LIMIT = 700
|
| 456 |
+
max_new_tokens = max(1, min(max_new_tokens, MAX_TOKENS_HARD_LIMIT))
|
| 457 |
+
|
| 458 |
+
if len(text) > MAX_INPUT_CHARS:
|
| 459 |
+
logger.warning(f"Input de {len(text)} caracteres recortado a {MAX_INPUT_CHARS}")
|
| 460 |
+
text = text[:MAX_INPUT_CHARS]
|
| 461 |
+
|
| 462 |
+
prefix = f"Usuario: {text}\nMTP: "
|
| 463 |
+
ids = [BOS_ID] + encode_text(prefix)
|
| 464 |
+
idx = torch.tensor([ids], dtype=torch.long, device=DEVICE)
|
| 465 |
+
|
| 466 |
+
# Con el lock, si llegan varias generaciones al mismo tiempo (UI + API,
|
| 467 |
+
# o varios usuarios a la vez) se procesan una despues de otra en vez de
|
| 468 |
+
# pisarse los hilos de CPU entre si.
|
| 469 |
+
with _generation_lock:
|
| 470 |
+
out = generate(idx, max_new_tokens, temperature, top_k, top_p, repetition_penalty, no_repeat_ngram_size)
|
| 471 |
+
|
| 472 |
+
new_ids = out[0].tolist()[len(ids):]
|
| 473 |
+
return decode_ids(new_ids).strip()
|
| 474 |
+
|
| 475 |
+
|
| 476 |
+
def chat_fn(message, history, max_new_tokens, temperature, top_k, top_p, repetition_penalty):
|
| 477 |
+
return run_inference(message, max_new_tokens, temperature, top_k, top_p, repetition_penalty)
|
| 478 |
+
|
| 479 |
+
|
| 480 |
+
# ---------------- Interfaz Gradio (para probar el modelo desde el navegador) ----------------
|
| 481 |
+
with gr.Blocks(title="MTP-2.5 Chat") as demo:
|
| 482 |
+
gr.Markdown(f"# MTP-2.5\nModelo GPT (RoPE + SwiGLU + RMSNorm) entrenado desde cero, tokenizer BPE. Ejecutándose en {DEVICE.upper()}.")
|
| 483 |
+
|
| 484 |
+
with gr.Accordion("Parámetros de generación", open=False):
|
| 485 |
+
max_new_tokens_ui = gr.Slider(16, 4000, value=gen_defaults["max_new_tokens"], step=10, label="max_new_tokens")
|
| 486 |
+
temperature_ui = gr.Slider(0.1, 2.0, value=gen_defaults["temperature"], step=0.05, label="temperature")
|
| 487 |
+
top_k_ui = gr.Slider(0, 100, value=gen_defaults["top_k"], step=1, label="top_k")
|
| 488 |
+
top_p_ui = gr.Slider(0.1, 1.0, value=gen_defaults["top_p"], step=0.05, label="top_p")
|
| 489 |
+
repetition_penalty_ui = gr.Slider(1.0, 2.0, value=gen_defaults["repetition_penalty"], step=0.05,
|
| 490 |
+
label="repetition_penalty")
|
| 491 |
+
|
| 492 |
+
chatbot = gr.ChatInterface(
|
| 493 |
+
fn=chat_fn,
|
| 494 |
+
additional_inputs=[max_new_tokens_ui, temperature_ui, top_k_ui, top_p_ui, repetition_penalty_ui],
|
| 495 |
+
title=None,
|
| 496 |
+
examples=[
|
| 497 |
+
["Hola, ¿cómo estás?"],
|
| 498 |
+
["¿Cuánto es 8 + 5?"],
|
| 499 |
+
["Explícame qué es un algoritmo."],
|
| 500 |
+
],
|
| 501 |
+
cache_examples=False,
|
| 502 |
+
)
|
| 503 |
+
|
| 504 |
+
demo.queue(max_size=16)
|
| 505 |
+
|
| 506 |
+
# ---------------- API REST /generate (la que consume el PHP) ----------------
|
| 507 |
+
# El PHP hace: fetch(url, { method:'POST', body: JSON.stringify({text, max_tokens, temperature}) })
|
| 508 |
+
# y espera de vuelta: { "reply": "..." }
|
| 509 |
+
#
|
| 510 |
+
# IMPORTANTE:
|
| 511 |
+
# - ssr_mode=False: Gradio 6 usa un servidor Node.js aparte para SSR, que
|
| 512 |
+
# intentaba levantarse en el puerto 7861 y chocaba. Lo desactivamos porque
|
| 513 |
+
# no lo necesitamos para servir la API.
|
| 514 |
+
# - El middleware CORS se pasa vía app_kwargs ANTES de llamar a launch(),
|
| 515 |
+
# porque una vez que la app arranca, Starlette ya no permite añadir
|
| 516 |
+
# middleware (por eso fallaba con app.add_middleware() después).
|
| 517 |
+
|
| 518 |
+
class GenerateRequest(BaseModel):
|
| 519 |
+
text: str
|
| 520 |
+
max_tokens: Optional[int] = None
|
| 521 |
+
temperature: Optional[float] = None
|
| 522 |
+
top_k: Optional[int] = None
|
| 523 |
+
top_p: Optional[float] = None
|
| 524 |
+
repetition_penalty: Optional[float] = None
|
| 525 |
+
no_repeat_ngram_size: Optional[int] = None
|
| 526 |
+
|
| 527 |
+
|
| 528 |
+
PORT = int(os.environ.get("PORT", 7860))
|
| 529 |
+
demo.launch(
|
| 530 |
+
server_name="0.0.0.0",
|
| 531 |
+
server_port=PORT,
|
| 532 |
+
prevent_thread_lock=True,
|
| 533 |
+
ssr_mode=False,
|
| 534 |
+
app_kwargs={
|
| 535 |
+
"middleware": [
|
| 536 |
+
Middleware(CORSMiddleware, allow_origins=ALLOWED_ORIGINS, allow_methods=["*"], allow_headers=["*"]),
|
| 537 |
+
]
|
| 538 |
+
},
|
| 539 |
+
)
|
| 540 |
+
|
| 541 |
+
app = demo.app
|
| 542 |
+
|
| 543 |
+
|
| 544 |
+
@app.post("/generate")
|
| 545 |
+
def generate_endpoint(req: GenerateRequest):
|
| 546 |
+
if not req.text or not req.text.strip():
|
| 547 |
+
return {"reply": "Escribe algo para que pueda responder."}
|
| 548 |
+
try:
|
| 549 |
+
reply = run_inference(
|
| 550 |
+
req.text,
|
| 551 |
+
max_new_tokens=req.max_tokens,
|
| 552 |
+
temperature=req.temperature,
|
| 553 |
+
top_k=req.top_k,
|
| 554 |
+
top_p=req.top_p,
|
| 555 |
+
repetition_penalty=req.repetition_penalty,
|
| 556 |
+
no_repeat_ngram_size=req.no_repeat_ngram_size,
|
| 557 |
+
)
|
| 558 |
+
if not reply:
|
| 559 |
+
reply = "No pude generar una respuesta."
|
| 560 |
+
return {"reply": reply}
|
| 561 |
+
except Exception as e:
|
| 562 |
+
# El detalle completo va al log del Space (con traceback), no a la
|
| 563 |
+
# respuesta publica: devolver la excepcion cruda a quien llama podria
|
| 564 |
+
# filtrar rutas internas, nombres de variables, etc. Con
|
| 565 |
+
# MTP_DEBUG_ERRORS=1 se puede activar el detalle mientras desarrollas.
|
| 566 |
+
logger.error(f"Error generando respuesta: {e}\n{traceback.format_exc()}")
|
| 567 |
+
reply = f"Error del modelo: {e}" if DEBUG_ERRORS else "Ocurrió un error al generar la respuesta. Intenta de nuevo en un momento."
|
| 568 |
+
return {"reply": reply}
|
| 569 |
+
|
| 570 |
+
|
| 571 |
+
@app.get("/generate")
|
| 572 |
+
def generate_health():
|
| 573 |
+
# Solo para poder comprobar en el navegador que la ruta existe (GET no genera texto)
|
| 574 |
+
return {"status": "ok", "info": "Usa POST con JSON {text, max_tokens, temperature}"}
|
| 575 |
+
|
| 576 |
+
|
| 577 |
+
@app.get("/health")
|
| 578 |
+
def health():
|
| 579 |
+
# Health check real: confirma que el modelo esta cargado y listo, no
|
| 580 |
+
# solo que el proceso esta vivo. Util para monitoreo externo (uptime
|
| 581 |
+
# checks) o para que el PHP sepa si conviene reintentar mas tarde.
|
| 582 |
+
return {
|
| 583 |
+
"status": "ok",
|
| 584 |
+
"model": checkpoint["meta"]["model_name"],
|
| 585 |
+
"device": DEVICE,
|
| 586 |
+
"trained_examples": checkpoint["meta"]["trained_examples"],
|
| 587 |
+
}
|
| 588 |
+
|
| 589 |
+
|
| 590 |
+
# demo.launch(prevent_thread_lock=True) ya dejó el servidor corriendo en un
|
| 591 |
+
# hilo en segundo plano (un solo proceso, un solo puerto). Mantenemos vivo
|
| 592 |
+
# el hilo principal para que el contenedor del Space no termine.
|
| 593 |
+
demo.block_thread()
|