jiashenggu

4090 is not compatible with cuda 10.1 and when I install the environment with cuda11.8 there are some conflicts.

### 🚀 The feature, motivation and pitch similar features Fast LoRA loading and switching in onediff diffusers ### Alternatives _No response_ ### Additional context _No response_

I'd like to know why the implementation looks like this ``` if write: self.id_bank[cur_step] = [ hidden_states[: self.id_length], hidden_states[self.id_length :], ] else: encoder_hidden_states = torch.cat( ( self.id_bank[cur_step][0].to(self.device), hidden_states[:1], self.id_bank[cur_step][1].to(self.device), hidden_states[1:],...