update readme

README.md

@@ -13,6 +13,15 @@ and increases the expressive power of the model by dynamically composing attenti
 
 We recommend <strong>compiled version</strong> of DCFormer with *torch.compile* for inference acceleration. Please refer to QuickStart section for compile implementation.
 
+## Env
+
+You need to upgrade transformers to avoid [(loading problems)](https://github.com/huggingface/transformers/pull/29175).  
+ 
+```
+pip install transformers>=4.40.2
+```
+
+
 ## Quickstart
 
 ```

generation_demo.py

@@ -7,7 +7,7 @@ os.environ['TOKENIZERS_PARALLELISM'] = 'false'
 tokenizer = AutoTokenizer.from_pretrained("Caiyun-AI/DCFormer-2.8B")
 model = AutoModelForCausalLM.from_pretrained("Caiyun-AI/DCFormer-2.8B", trust_remote_code=True)
 
-device = torch.device('cuda:1')
+device = torch.device('cuda')
 MAX_BATCH_SIZE = 1
 MAX_SEQ_LENGTH = 2048
 NUM_TOKENS_TO_GENERATE = 100