put flash attention in requirements instead

app.py

@@ -4,8 +4,8 @@ from image_generator import process_and_generate
 from gradio_interface import create_gradio_interface
 
 if __name__ == "__main__":
- import subprocess
- subprocess.run('pip install "flash-attn>=2.1.0" --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
+ # import subprocess
+ # subprocess.run('pip install "flash-attn>=2.1.0" --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
 
 
 

requirements.txt

@@ -4,4 +4,5 @@ Pillow
 gradio
 janus @ git+https://github.com/deepseek-ai/Janus
 transformers
-spaces
+spaces
+flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.6.3/flash_attn-2.6.3+cu123torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl