Update app.py

app.py

@@ -55,6 +55,49 @@ def _build_prompt(prompt: str, system: Optional[str]) -> str:
     sys_part = (f"<<SYS>>\n{system.strip()}\n<</SYS>>\n\n" if system and system.strip() else "")
     return sys_part + f"### Instrucción\n{prompt}\n\n### Respuesta\n"
 
+@spaces.GPU  # usa GPU si está disponible (ZeroGPU)
+def _generate_with_tools(
+    messages: List[Dict[str, str]],
+    tools: List[Dict[str, Any]],
+    max_new_tokens: int = 512,
+    temperature: float = 0.7,
+    top_p: float = 0.95,
+) -> Dict[str, Any]:
+    tok, model = _lazy_load()
+    tools_md = _render_tools_md(tools)
+    prompt = _compose_chat_prompt(messages, tools_md)
+
+    inputs = tok(prompt, return_tensors="pt").to(DEVICE)
+    with torch.inference_mode():
+        out = model.generate(
+            **inputs,
+            max_new_tokens=int(max_new_tokens),
+            temperature=float(temperature),
+            top_p=float(top_p),
+            do_sample=True if temperature > 0 else False,
+            pad_token_id=tok.eos_token_id,
+            eos_token_id=tok.eos_token_id,
+        )
+    text = tok.decode(out[0], skip_special_tokens=True).strip()
+
+    # Si el modelo devuelve un bloque JSON con 'tool_calls', lo intentamos extraer.
+    tool_calls: List[Dict[str, Any]] = []
+    try:
+        # busca el último {...} que contenga "tool_calls"
+        matches = list(re.finditer(r"\{.*?\"tool_calls\".*?\}", text, flags=re.S))
+        if matches:
+            block = text[matches[-1].start():matches[-1].end()]
+            obj = json.loads(block)
+            tc = obj.get("tool_calls", [])
+            if isinstance(tc, list):
+                tool_calls = tc
+    except Exception:
+        pass
+
+    tool_results = maybe_execute_tool_calls(tool_calls) if tool_calls else []
+
+    return {"text": text, "tool_calls": tool_calls, "tool_results": tool_results}
+
 @spaces.GPU  # usa GPU si está disponible (ZeroGPU)
 def _generate(
     prompt: str,
@@ -113,4 +156,13 @@ with gr.Blocks(title="Salamandra 7B Instruct · ZeroGPU") as demo:
     out_engine = gr.Textbox(label="Respuesta (ENGINE)")
     gr.Button("Probar /predict").click(predict_for_engine, [in_prompt_engine], out_engine, api_name="predict", concurrency_limit=1)
 
+    with gr.Row():
+        prompt = gr.Textbox(label="prompt", lines=10)
+    with gr.Row():
+        btn2 = gr.Button("Generar", variant="primary")
+    with gr.Row():
+        out2 = gr.JSON(label="Salida")
+
+    btn2.click(salamandra_chat_endpoint, [prompt], out2, api_name="generate_out_from_prompt", concurrency_limit=1)
+
 demo.queue(max_size=16).launch()