fix: configure clean chat template and hybrid RIF sliding window on T4 GPU

app.py

@@ -62,15 +62,12 @@ def kalpana_generate(prompt: str, max_tokens: float = 256, temperature: float =
 
     _load_model(device, dtype)
 
-    messages = [
-        {"role": "system", "content": "You are a helpful, intelligent AI assistant. Always answer user queries clearly, accurately, and concisely in English."},
-        {"role": "user", "content": prompt.strip()}
-    ]
+    messages = [{"role": "user", "content": prompt.strip()}]
     fmt = _LOCAL_TOKENIZER.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
     inputs = _LOCAL_TOKENIZER(fmt, return_tensors="pt").to(device)
 
     num_layers = getattr(_LOCAL_QWEN_MODEL.config, "num_hidden_layers", 24)
-    cache = KalpanaDynamicCache(num_layers=num_layers, bands=2048)
+    cache = KalpanaDynamicCache(num_layers=num_layers, bands=2048, sliding_window=128)
 
     t0 = time.perf_counter()
     with torch.inference_mode():
@@ -78,11 +75,10 @@ def kalpana_generate(prompt: str, max_tokens: float = 256, temperature: float =
             **inputs,
             past_key_values=cache,
             max_new_tokens=int(max_tokens),
-            do_sample=True,
-            temperature=float(temperature) if float(temperature) > 0 else 0.6,
-            top_p=0.85,
-            top_k=40,
-            repetition_penalty=1.15,
+            do_sample=(float(temperature) > 0),
+            temperature=float(temperature) if float(temperature) > 0 else 1.0,
+            top_p=0.9,
+            repetition_penalty=1.1,
             pad_token_id=_LOCAL_TOKENIZER.eos_token_id,
         )
     t1 = time.perf_counter()