Update app.py

app.py

@@ -1,3 +1,4 @@
+import time
 import gradio as gr
 import torch
 from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
@@ -25,11 +26,20 @@ def load_model():
 
 @torch.inference_mode()
 def translate_es_to_pt(text, beams, max_new_tokens):
+    """Single run translation + debug info."""
     if not text or not text.strip():
-        return ""
+        return "", "No input."
 
     tok, mdl = load_model()
 
+    beams = int(beams)
+    max_new_tokens = int(max_new_tokens)
+
+    # LOG: verify slider values are reaching the backend
+    print(f"[RUN] beams={beams} max_new_tokens={max_new_tokens} device={device}")
+
+    t0 = time.time()
+
     inputs = tok(
         text,
         return_tensors="pt",
@@ -39,22 +49,85 @@ def translate_es_to_pt(text, beams, max_new_tokens):
 
     outputs = mdl.generate(
         **inputs,
-        num_beams=int(beams),
-        max_new_tokens=int(max_new_tokens),
+        num_beams=beams,
+        max_new_tokens=max_new_tokens,
         length_penalty=1.0,
         early_stopping=True,
-        no_repeat_ngram_size=3,   # helps reduce repetition
+        no_repeat_ngram_size=3,
+    )
+
+    out_text = tok.decode(outputs[0], skip_special_tokens=True)
+
+    dt = time.time() - t0
+    debug = (
+        f"beams={beams} | max_new_tokens={max_new_tokens} | device={device} | "
+        f"gen_time={dt:.2f}s | output_chars={len(out_text)}"
     )
 
-    return tok.decode(outputs[0], skip_special_tokens=True)
+    return out_text, debug
+
+
+@torch.inference_mode()
+def quick_test_suite(text):
+    """
+    Runs the same input through multiple settings so you can compare
+    beams/tokens behavior in one shot.
+    """
+    if not text or not text.strip():
+        return "No input."
+
+    tok, mdl = load_model()
+
+    settings = [
+        (1, 64),
+        (3, 128),
+        (5, 128),
+        (8, 128),
+        (5, 32),   # truncation stress test
+        (5, 256),  # long generation stress test
+    ]
+
+    lines = []
+    lines.append(f"MODEL: {MODEL_ID}")
+    lines.append(f"DEVICE: {device}")
+    lines.append("-" * 60)
+
+    for beams, max_new in settings:
+        print(f"[TEST] beams={beams} max_new_tokens={max_new}")
+        t0 = time.time()
+
+        inputs = tok(
+            text,
+            return_tensors="pt",
+            truncation=True,
+            max_length=256,
+        ).to(device)
+
+        outputs = mdl.generate(
+            **inputs,
+            num_beams=beams,
+            max_new_tokens=max_new,
+            length_penalty=1.0,
+            early_stopping=True,
+            no_repeat_ngram_size=3,
+        )
+
+        out_text = tok.decode(outputs[0], skip_special_tokens=True)
+        dt = time.time() - t0
+
+        lines.append(f"[beams={beams}, max_new={max_new}] time={dt:.2f}s chars={len(out_text)}")
+        lines.append(out_text)
+        lines.append("-" * 60)
+
+    return "\n".join(lines)
 
 
 with gr.Blocks(theme=gr.themes.Soft(), title="ES → PT Translator") as demo:
     gr.Markdown(
-        """
+        f"""
 # ES → PT Translator (BART)
 
-**Model:** `liansheng06/bart-tatoeba-es-pt`  
+**Model:** `{MODEL_ID}`  
 **Dataset:** Helsinki-NLP/Tatoeba (es-pt)  
 **Metric:** chrF (beam search evaluation)
 
@@ -78,15 +151,26 @@ Tip: If output looks repetitive, try lowering **beams** or **max_new_tokens**.
                 btn_translate = gr.Button("Translate", variant="primary")
                 btn_clear = gr.Button("Clear")
 
+            debug_box = gr.Textbox(
+                label="Run info (debug)",
+                value="",
+                lines=2,
+                interactive=False,
+            )
+
         with gr.Column(scale=1):
             out = gr.Textbox(label="Portuguese (pt) output", lines=6)
 
+    with gr.Row():
+        btn_test = gr.Button("Run quick test suite (compare beams/tokens)")
+    test_report = gr.Textbox(label="Test suite report", lines=18)
+
     examples = gr.Examples(
         examples=[
             ["Las personas dicen que estoy loco."],
             ["¿Puedes ayudarme a traducir esta frase, por favor?"],
-            ["Mañana vamos al mercado a comprar frutas y pan."],
-            ["Si tuviera más tiempo, aprendería a programar mejor."],
+            ["Mañana vamos al mercado a comprar frutas y pan, y después visitaremos a mis abuelos."],
+            ["Si el proyecto termina hoy, podremos presentar los resultados mañana, pero necesitamos revisar los datos con cuidado para evitar errores."],
         ],
         inputs=inp,
         label="Examples",
@@ -95,11 +179,17 @@ Tip: If output looks repetitive, try lowering **beams** or **max_new_tokens**.
     btn_translate.click(
         fn=translate_es_to_pt,
         inputs=[inp, beams, max_new],
-        outputs=out,
+        outputs=[out, debug_box],
+    )
+
+    btn_test.click(
+        fn=quick_test_suite,
+        inputs=inp,
+        outputs=test_report,
     )
 
-    btn_clear.click(lambda: ("", ""), outputs=[inp, out])
+    btn_clear.click(lambda: ("", "", ""), outputs=[inp, out, debug_box])
 
 if __name__ == "__main__":
     demo.queue()
-    demo.launch(server_name="0.0.0.0", server_port=7860)
+    demo.launch(server_name="0.0.0.0", server_port=7860)