Questions about installation steps and flash_attention_2 compatibility in training

[FanchengDing00]

Hi, thank you very much for releasing this great project and the detailed documentation!

I have two questions while following the README and preparing the training environment, and I would really appreciate your clarification：

1. About installing dependencies (requirements.txt)
   In the Getting Started section of the README, the installation guide ends at:

git clone https://github.com/OpenRobotLab/StreamVLN.git
cd StreamVLN

However, I noticed that there is a requirements.txt file in the repository.
For successfully running training, should we also install the dependencies listed there (e.g. via pip install -r requirements.txt), or are all required packages already covered by the previous steps (Habitat, habitat-lab, etc.)?
I just want to make sure I’m not missing any required installation steps.

2. About flash_attention_2 and Python / PyTorch compatibility
   I noticed that in TrainingArguments, the default value of attn_implementation is set to "flash_attention_2", and the training scripts also seem to rely on this default.
   However, as far as I understand, flash_attention_2 requires PyTorch ≥ 2.2.
   At the same time, the README specifies a testing environment of Python 3.9 and PyTorch 2.1.2.
   So I’m a bit confused about the intended setup here:
   For normal training, are we expected not to use flash_attention_2 (e.g. switch to another attention implementation)?
   Or is there a recommended workaround / compatible installation method that allows using flash_attention_2 under the stated environment?
   Any guidance on the recommended configuration would be very helpful.

Thanks again for your work and for taking the time to clarify these questions!
Best regards