Help needed on some Code Explanation

[yinzuojie]

Q1：
in peft_model.py / class PeftModel:

    def get_prompt(self, batch_size: int):
  ......
        if peft_config.peft_type == PeftType.PREFIX_TUNING:
            prompt_tokens = prompt_tokens[:, : peft_config.num_virtual_tokens]
            if peft_config.inference_mode:
                # █ here
                past_key_values = prompt_encoder.embedding.weight.repeat(batch_size, 1, 1)
            else:
                past_key_values = prompt_encoder(prompt_tokens)

My question is why here we distinguish inference_mode and do special treatment? can we combine 2 branches and just use "past_key_values = prompt_encoder(prompt_tokens)"?

Q2:
other.py

# needed for prefix-tuning of bloom model
def bloom_model_postprocess_past_key_value(past_key_values):
    past_key_values = torch.cat(past_key_values)
    total_layers, batch_size, num_attention_heads, num_virtual_tokens, head_dim = past_key_values.shape
    keys = past_key_values[: total_layers // 2] #[24, 8, 16, 30, 64]
    #██ here
    keys = keys.transpose(2, 3).reshape(
        total_layers // 2, batch_size * num_attention_heads, head_dim, num_virtual_tokens
    )

my question is if we can omit the transpose(2, 3) operation and just use keys.reshape(...)?
what's the difference?

I'm new comer to pytorch and PEFT code, any help is highly appreciated.
thanks all ~