Add application file

app.py

@@ -34,7 +34,7 @@ model_id = "microsoft/phi-2"
 
 tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
 model = AutoModelForCausalLM.from_pretrained(
-    model_id, torch_dtype=torch.float32, device_map="auto", trust_remote_code=True
+    model_id, torch_dtype=torch.float32, trust_remote_code=True
 )
 
 # sentence transformers to be used in vector store
@@ -92,7 +92,6 @@ def generate(question, answer, text_file, max_new_tokens):
         max_new_tokens=max_new_tokens,
         pad_token_id=tokenizer.eos_token_id,
         eos_token_id=tokenizer.eos_token_id,
-        device_map="auto",
         streamer=streamer,
     )