fixes

app.py

@@ -48,11 +48,17 @@ def generate(req: Request):
             outputs = model.generate(
                 **inputs,
                 max_new_tokens=req.max_tokens,
+
                 do_sample=True,
-                temperature=req.temperature,
-                top_p=0.85,              # stability fix
-                top_k=40,                # stability fix
-                repetition_penalty=1.1   # prevents loops
+                temperature=0.8,
+                top_p=0.92,
+                top_k=40,
+
+                repetition_penalty=1.7,        # strong anti-loop
+                no_repeat_ngram_size=5,        # stronger phrase blocking
+
+                eos_token_id=tokenizer.eos_token_id,
+                pad_token_id=tokenizer.eos_token_id
             )
 
         text = tokenizer.decode(