Update app.py

app.py

@@ -10,13 +10,15 @@ import torch
 # filename = "simplescaling_s1-32B-Q4_K_S.gguf"
 model_name = "simplescaling/s1.1-32B"
 torch_dtype = torch.bfloat16 # could be torch.float16 or torch.bfloat16 torch.float32 too
+cache_dir = "/data"
 
 model = AutoModelForCausalLM.from_pretrained(
     model_name,
     # subfolder=subfolder,
     # gguf_file=filename,
     torch_dtype=torch_dtype,
-    device_map="auto"
+    device_map="auto",
+    cache_dir = cache_dir,
 )
 tokenizer = AutoTokenizer.from_pretrained(model_name
     , gguf_file=filename