Update app.py

app.py

@@ -47,21 +47,17 @@ retriever = retriever_from_chroma(docs, hf, "mmr", 6)
 
 callback_manager = CallbackManager([StreamingStdOutCallbackHandler()])
 
-@st.cache_resource
-def load_llm(model_path):
-    return LlamaCpp(
-        model_path=model_path,
-        n_gpu_layers=0,
-        temperature=0.0,
-        top_p=0.5,
-        n_ctx=7000,
-        max_tokens=350,
-        repeat_penalty=1.7,
-        callback_manager=callback_manager,
-        verbose=False,
-    )
-
-llm = load_llm(model_path)
+llm = llamacpp.LlamaCpp(
+            model_path="/kaggle/working/phi-2-layla-v1-chatml-Q8_0.gguf",
+            n_gpu_layers=1,
+            temperature=0.1,
+            top_p = 0.9,
+            n_ctx=22000,
+            max_tokens=200,
+            repeat_penalty=1.7,
+            callback_manager = callback_manager,
+            verbose=False,
+            )
 
 contextualize_q_system_prompt = """Given a context, chat history and the latest user question
 which maybe reference context in the chat history, formulate a standalone question