Create app.py

app.py

@@ -0,0 +1,74 @@
+# -*- coding: utf-8 -*-
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import gradio as gr
+import os
+
+# --- MISMOS HIPERPARÁMETROS ---
+embed_size = 128
+num_heads = 4
+num_layers = 3
+block_size = 64
+vocab_size = 256
+device = "cpu"
+
+# --- TU ARQUITECTURA ---
+class MiniGPT(nn.Module):
+    def __init__(self, v_size):
+        super().__init__()
+        self.token_embedding = nn.Embedding(v_size, embed_size)
+        self.pos_embedding = nn.Embedding(block_size, embed_size)
+        self.blocks = nn.ModuleList([
+            nn.TransformerEncoderLayer(d_model=embed_size, nhead=num_heads, 
+                                       dim_feedforward=embed_size*4, batch_first=True, 
+                                       dropout=0.1, norm_first=True) 
+            for _ in range(num_layers)
+        ])
+        self.ln = nn.LayerNorm(embed_size)
+        self.fc_out = nn.Linear(embed_size, v_size)
+
+    def forward(self, idx, targets=None):
+        B, T = idx.shape
+        tok_emb = self.token_embedding(idx)
+        pos = torch.arange(T, device=device)
+        pos_emb = self.pos_embedding(pos)[None, :, :]
+        x = tok_emb + pos_emb
+        mask = torch.triu(torch.ones(T, T, device=device), diagonal=1).bool()
+        for block in self.blocks: x = block(x, src_mask=mask)
+        x = self.ln(x)
+        logits = self.fc_out(x)
+        return logits, None
+
+# --- CARGAR EL MODELO ---
+model = MiniGPT(vocab_size).to(device)
+if os.path.exists("mini_gpt.pth"):
+    model.load_state_dict(torch.load("mini_gpt.pth", map_location=device))
+model.eval()
+
+# --- FUNCIÓN DE RESPUESTA ---
+def responder(mensaje, historial):
+    contexto = f"\nUsuario: {mensaje}\nIA: "
+    tokens = [ord(c) if ord(c) < 256 else 32 for c in contexto]
+    ai_txt = ""
+    
+    with torch.no_grad():
+        for _ in range(150):
+            idx = torch.tensor([tokens[-block_size:]], dtype=torch.long).to(device)
+            logits, _ = model(idx)
+            probs = F.softmax(logits[:, -1, :] / 0.8, dim=-1)
+            next_token = torch.multinomial(probs, num_samples=1).item()
+            char = chr(next_token)
+            
+            if char == "\n" or ai_txt.endswith("Usuario:"): break
+            tokens.append(next_token)
+            ai_txt += char
+            
+    return ai_txt.replace("Usuario:", "").strip()
+
+# --- INTERFAZ ---
+demo = gr.ChatInterface(fn=responder, title="Mi IA Personal", description="Modelo MiniGPT entrenado.")
+
+if __name__ == "__main__":
+    demo.l
+    aunch()