Update app.py

app.py

@@ -35,25 +35,17 @@ tokenizer = None
 device = None
 
 def load_model():
-    """Load MarianMT model once and cache it using Auto classes"""
+    """Load MarianMT model once and cache it"""
     global model, tokenizer, device
     
     if model is None:
         print("Loading MarianMT model...")
         device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
         
-        try:
-            # Try Auto classes first (more flexible)
-            tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=False)
-            model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_ID).to(device).eval()
-            print(f"Model loaded successfully on {device} using Auto classes")
-        except Exception as e:
-            print(f"Auto classes failed: {e}")
-            print("Trying legacy Marian classes...")
-            from transformers import MarianMTModel, MarianTokenizer
-            tokenizer = MarianTokenizer.from_pretrained(MODEL_ID, use_fast=False)
-            model = MarianMTModel.from_pretrained(MODEL_ID).to(device).eval()
-            print(f"Model loaded on {device} using legacy classes")
+        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=False)
+        model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_ID).to(device).eval()
+        
+        print(f"Model loaded successfully on {device}")
     
     return model, tokenizer, device
 
@@ -70,20 +62,20 @@ def translate_marian(text):
         inputs = {k: v.to(device) for k, v in inputs.items()}
         
         with torch.no_grad():
+            # Simple beam search without group beam search
             outputs = model.generate(
                 **inputs,
-                num_beams=6,
-                num_beam_groups=3,
-                diversity_penalty=1.2,
+                num_beams=4,
                 num_return_sequences=3,
                 max_length=128,
                 early_stopping=True,
+                do_sample=False,
             )
         
         translations = []
         for output in outputs:
             trans = tokenizer.decode(output, skip_special_tokens=True)
-            if trans and trans not in translations:  # Avoid duplicates and empty
+            if trans and trans not in translations:
                 translations.append(trans)
         
         return translations if translations else ["[Error: No translation generated]"]