chore: back to bf16

handler.py

@@ -55,7 +55,7 @@ class EndpointHandler:
         with init_empty_weights():
             self.model = AutoModelForCausalLM.from_pretrained(
                 "THUDM/cogvlm-chat-hf",
-                torch_dtype=torch.float16,
+                torch_dtype=torch.bfloat16,
                 low_cpu_mem_usage=True,
                 trust_remote_code=True,
             )