Troubleshooting Guide

May 2, 2026 · View on GitHub

Q1. Issues with flash-attn Version Compatibility

  • Problem: Errors occur during the installation of the flash-attn package.

  • Cause: The flash-attn package has strict version requirements for PyTorch and CUDA. Incompatible versions can lead to installation failures.

  • Solution: Ensure that you are using compatible versions of PyTorch and CUDA. Our setup uses flash-attn 2.7.1.post4 with PyTorch 2.6.0 and CUDA 11.8. You can install flash-attn via the offline wheel file (Google Drive), and then run:

    # If the installation fails, run the following command to install flash-attn
pip install cache/flash-attn/flash_attn-2.7.1.post4+cu11torch2.6cxx11abiFALSE-cp310-cp310-linux_x86_64.whl

Q2. Compatibility of various ulcerative colitis labels

  • Problem:
    There may appear to be overlap between the general CLS#33 ulcerative colitis label and fine-grained labels (CLS#34~CLS#40) for ulcerative colitis grading.
  • Explanation: The CLS#33 ulcerative colitis label is used only for the classification task, whereas the fine-grained ulcerative colitis grading labels (CLS#34~CLS#40) are used only for the ulcerative colitis grading task. These labels are separated at the task-design level, which avoids direct label conflicts.