Update app.py

app.py

@@ -13,37 +13,31 @@ if not HUGGINGFACE_TOKEN:
 
 print("✅ HUGGINGFACE_TOKEN is set.")
 
-# Model Paths (Replace with your actual Hugging Face Model Names)
-BASE_MODEL = "meta-llama/Llama-3.2-1B-Instruct"
-QLORA_ADAPTER = "meta-llama/Llama-3.2-1B-Instruct-QLORA_INT4_EO8"
+# Model Paths
+QUANTIZED_MODEL = "meta-llama/Llama-3.2-1B-Instruct-QLORA_INT4_EO8"  # Directly using quantized model
 LLAMA_GUARD_NAME = "meta-llama/Llama-Guard-3-1B-INT4"
 
-# Function to load Llama model
-def load_llama_model(base_model=BASE_MODEL, adapter=None, is_guard=False):
-    print(f"🔄 Loading Model: {base_model}")
+# Function to load Llama model (without LoRA)
+def load_llama_model(model_name, is_guard=False):
+    print(f"🔄 Loading Model: {model_name}")
     
-    tokenizer = AutoTokenizer.from_pretrained(base_model, token=HUGGINGFACE_TOKEN)
+    tokenizer = AutoTokenizer.from_pretrained(model_name, token=HUGGINGFACE_TOKEN)
     model = AutoModelForCausalLM.from_pretrained(
-        base_model,
+        model_name,
         token=HUGGINGFACE_TOKEN,
-        torch_dtype=torch.float32,  # Using float32 for CPU compatibility
+        torch_dtype=torch.float16,  # Use float16 for optimized performance
         low_cpu_mem_usage=True
     )
 
-    if adapter and not is_guard:
-        print(f"🔄 Loading Adapter: {adapter}")
-        model = PeftModel.from_pretrained(model, adapter, token=HUGGINGFACE_TOKEN)
-        model = model.merge_and_unload()
-        print("✅ Adapter Loaded Successfully")
-
     model.eval()
+    print("✅ Model Loaded Successfully")
     return tokenizer, model
 
-# Load Llama 3.2 model
-tokenizer, model = load_llama_model(adapter=QLORA_ADAPTER)
+# Load the quantized Llama model
+tokenizer, model = load_llama_model(QUANTIZED_MODEL)
 
 # Load Llama Guard for content moderation
-guard_tokenizer, guard_model = load_llama_model(base_model=LLAMA_GUARD_NAME, is_guard=True)
+guard_tokenizer, guard_model = load_llama_model(LLAMA_GUARD_NAME, is_guard=True)
 
 # Define Prompt Templates
 PROMPTS = {