update

app.py

@@ -40,7 +40,7 @@ index.add(embeddings)
 
 model_name = "TheBloke/zephyr-7B-beta-GPTQ"
 tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
-model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", trust_remote_code=False)
+model = AutoModelForCausalLM.from_pretrained(model_name, device_map="cpu", trust_remote_code=False, use_exllama=False))
 
 
 # Step 4: Define the Retrieval Function
@@ -67,7 +67,7 @@ def gradio_interface(query):
 
 iface = gr.Interface(
     fn=gradio_interface,
-    inputs=gr.inputs.Textbox(lines=2, placeholder="Enter your query here..."),
+    inputs=gr.Textbox(lines=2, placeholder="Enter your query here..."),
     outputs="text",
     title="RAG-based Course Search",
     description="Enter a query to search for relevant courses using Retrieval Augmented Generation."