Update app.py

app.py

@@ -1,7 +1,8 @@
 import gradio as gr
+import os
+os.environ["UNSLOTH_DEVICE"] = "cuda"
 from unsloth import FastLanguageModel
 import torch
-import os
 HF_TOKEN = os.environ["HF_TOKEN"]
 # -------------------- Load Model --------------------
 model, tokenizer = FastLanguageModel.from_pretrained(
@@ -9,7 +10,7 @@ model, tokenizer = FastLanguageModel.from_pretrained(
     max_seq_length=1024,
     dtype=None,
     load_in_4bit=True,
-    device_map="cuda",
+    device_map="auto",
     token=HF_TOKEN
 )