pip install requirtement and Building wheels for collected packages: flash-attn, simple-watchdog-timer, dictor cost so much time and doesnt finish

[violetsaber]

Using cached https://pypi.tuna.tsinghua.edu.cn/packages/1a/ab/3e941e3fcf1b7d3ab3d0233194d99d6a0ed6b24f8f956fc81e47edc8c079/backports.zoneinfo-0.2.1-cp38-cp38-manylinux1_x86_64.whl (74 kB)
WARNING: The candidate selected for download or install is a yanked version: 'yapf' candidate (version 0.33.0 at https://pypi.tuna.tsinghua.edu.cn/packages/32/ec/531851d561ecb656bd58dc102338d0aa07e086788f351c63a9f6b8a00fe6/yapf-0.33.0-py2.py3-none-any.whl (from https://pypi.tuna.tsinghua.edu.cn/simple/yapf/))
Reason for being yanked:
Building wheels for collected packages: flash-attn, simple-watchdog-timer, dictor
Building wheel for flash-attn (setup.py) ... |

pip installl required packages and don't know how to finish this build , is this necessary?

[dlrudco]

I am not the author, but to be a good Samaritan, here are a few workarounds I could think of:

Make sure you have ninja installed. This package enables multiprocessing when building Python packages. After installing ninja, you can control the level of parallelism with:

export MAX_JOBS=$(nproc)

On my 40-thread machine, the build process took about one hour.

You can also run the code without installing flash-attn. Remove all flash-attn–related imports and change
MultiheadFlashAttention to MultiheadAttention in the configuration file.

I tested this on my Titan RTX machine, since flash-attn does not support GPUs with architectures older than Ampere.