Update app.py

app.py

@@ -87,7 +87,7 @@ model = AutoModel.from_pretrained(
     torch_dtype=torch.bfloat16,
     attn_implementation='flash_attention_2',
     low_cpu_mem_usage=True,
-    trust_remote_code=True).eval().cuda()
+    trust_remote_code=True, token=auth_token).eval().cuda()
 processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True, token=auth_token)
 streamer = TextIteratorStreamer(processor.tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=15)