Batch Inference Pipeline - Input/Output Sequence Handling and Prompt Failures guide

Hi, I have a question regarding **batch inference** in `lmdeploy` (using the Offline Inference Pipeline).

Suppose I set up a batch of, say, 32 prompts for inference:

```python
from lmdeploy import pipeline

pipe = pipeline('internlm/internlm2_5-7b-chat')
prompts = ['Hi, please introduce yourself', 'Shanghai is ...', ...] # batch of 32 prompts
response = pipe(prompts)
print(response)
```

My concern is about **tracking the output corresponding to each input prompt**:

1. Will the outputs be returned in the **same order** as the input prompts?
2. What happens if **one of the prompts fails** during inference—does it affect the order of outputs for the rest of the batch?
3. Is there a way to reliably **map each input prompt to its corresponding output**, even if some prompts fail?

Essentially, I want to ensure that for batch inference, I can always trace which output belongs to which input, without any sequence mismatch.

Could you clarify how `lmdeploy` handles this scenario?