Update app.py

app.py

@@ -30,7 +30,7 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 model_id = "CohereForAI/c4ai-command-r-plus-4bit"
 
 
-tokenizer = AutoTokenizer.from_pretrained(model_id, token= token, quantization=False)
+tokenizer = AutoTokenizer.from_pretrained(model_id, token= token,)
 
 model = AutoModelForCausalLM.from_pretrained(model_id, token= token, 
                                                  # torch_dtype=torch.bfloat16,    
@@ -38,7 +38,7 @@ model = AutoModelForCausalLM.from_pretrained(model_id, token= token,
                                              # low_cpu_mem_usage=True,
                                              # llm_int8_enable_fp32_cpu_offload=True,
                                              device_map="cuda",
-                                             quantization=False
+                                             
                                             )