feat: hybrid engine with real CUDA GPU acceleration and streaming on Hugging Face ZeroGPU

app.py

@@ -1,46 +1,122 @@
+import os
+import threading
 import gradio as gr
-import spaces
+from typing import List, Dict, Any
+
+# Importar configuración y lógica existente de main.py
 from main import (
+    app as fastapi_app,
     graph_app,
     HumanMessage,
     AIMessage,
     SystemMessage,
-    DEFAULT_SYSTEM_PROMPT,
-    app as fastapi_app
+    DEFAULT_SYSTEM_PROMPT
 )
 
+# Detección del motor GPU (Hugging Face ZeroGPU)
+USE_CUDA_ENGINE = False
+
+try:
+    import spaces
+    import torch
+    from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
+    
+    # En Hugging Face ZeroGPU siempre está disponible torch y spaces
+    MODEL_ID = "Qwen/Qwen2.5-0.5B-Instruct"
+    print(f"[INFO] Inicializando motor Transformers en GPU para {MODEL_ID}...")
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+    model = AutoModelForCausalLM.from_pretrained(
+        MODEL_ID,
+        torch_dtype="auto",
+        device_map="auto"
+    )
+    USE_CUDA_ENGINE = True
+    print("[INFO] Motor Transformers GPU cargado exitosamente en tarjeta gráfica.")
+except Exception as e:
+    print(f"[INFO] Modo local/CPU detectado ({e}). Utilizando motor local LangGraph / llama-cpp-python.")
+    USE_CUDA_ENGINE = False
+    
+    # Stub para spaces.GPU en desarrollo local
+    try:
+        import spaces
+    except ImportError:
+        class spaces:
+            @staticmethod
+            def GPU(func=None, duration=None):
+                if func is not None:
+                    return func
+                def decorator(f):
+                    return f
+                return decorator
+
+
+# Función de generación con soporte de Streaming y aceleración por GPU
 @spaces.GPU
 def chat_fn(message, history):
-    messages = [SystemMessage(content=DEFAULT_SYSTEM_PROMPT)]
-    if history:
-        for turn in history:
-            if isinstance(turn, dict):
-                role = turn.get("role", "user")
-                content = turn.get("content", "")
-                if role == "assistant":
-                    messages.append(AIMessage(content=content))
-                else:
-                    messages.append(HumanMessage(content=content))
-            elif isinstance(turn, (list, tuple)) and len(turn) == 2:
-                user_msg, bot_msg = turn
-                if user_msg:
-                    messages.append(HumanMessage(content=str(user_msg)))
-                if bot_msg:
-                    messages.append(AIMessage(content=str(bot_msg)))
-                    
-    messages.append(HumanMessage(content=message))
-    output = graph_app.invoke({"messages": messages})
-    return output["messages"][-1].content
-
-# Interfaz interactiva de Chat de Gradio
+    if USE_CUDA_ENGINE:
+        # 1. Modo HF ZeroGPU (Aceleración real A100 con Streaming en tiempo real)
+        messages = [{"role": "system", "content": DEFAULT_SYSTEM_PROMPT}]
+        if history:
+            for turn in history:
+                if isinstance(turn, dict):
+                    messages.append(turn)
+                elif isinstance(turn, (list, tuple)) and len(turn) == 2:
+                    if turn[0]:
+                        messages.append({"role": "user", "content": str(turn[0])})
+                    if turn[1]:
+                        messages.append({"role": "assistant", "content": str(turn[1])})
+        messages.append({"role": "user", "content": message})
+
+        text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
+
+        streamer = TextIteratorStreamer(tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)
+        generate_kwargs = dict(
+            model_inputs,
+            streamer=streamer,
+            max_new_tokens=120,
+            temperature=0.3,
+            repetition_penalty=1.15
+        )
+        t = threading.Thread(target=model.generate, kwargs=generate_kwargs)
+        t.start()
+
+        partial_text = ""
+        for new_token in streamer:
+            partial_text += new_token
+            yield partial_text
+    else:
+        # 2. Modo Local (LangGraph + llama-cpp-python)
+        messages = [SystemMessage(content=DEFAULT_SYSTEM_PROMPT)]
+        if history:
+            for turn in history:
+                if isinstance(turn, dict):
+                    role = turn.get("role", "user")
+                    content = turn.get("content", "")
+                    if role == "assistant":
+                        messages.append(AIMessage(content=content))
+                    else:
+                        messages.append(HumanMessage(content=content))
+                elif isinstance(turn, (list, tuple)) and len(turn) == 2:
+                    if turn[0]:
+                        messages.append(HumanMessage(content=str(turn[0])))
+                    if turn[1]:
+                        messages.append(AIMessage(content=str(turn[1])))
+
+        messages.append(HumanMessage(content=message))
+        output = graph_app.invoke({"messages": messages})
+        yield output["messages"][-1].content
+
+
+# Interfaz Gradio con Streaming en tiempo real
 demo = gr.ChatInterface(
     fn=chat_fn,
     title="⚡ Qwen 2.5 (0.5B) • Intelecto Conciso",
-    description="Chatbot superinteligente, hiperpreciso y conciso en español. Orquestado con LangGraph y llama-cpp-python en Hugging Face Spaces (ZeroGPU).",
+    description="Chatbot superinteligente, hiperpreciso y conciso en español. Acelerado con GPU en Hugging Face Spaces (Streaming en vivo).",
     examples=["¿Qué es la entropía?", "¿Por qué el cielo es azul?", "¿Qué es la vida?"],
 )
 
-# Integrar las rutas de FastAPI (/chat, /healthz, /docs) en la misma app
+# Integrar las rutas de FastAPI (/chat, /healthz, /docs) en la misma aplicación
 demo.app.include_router(fastapi_app.router)
 
 if __name__ == "__main__":

requirements.txt

@@ -6,3 +6,5 @@ langgraph>=0.2.0
 langchain-core>=0.3.0
 huggingface_hub
 gradio
+transformers>=4.44.0
+accelerate>=0.33.0