0.20 subprocess implementing flash_attn

app.py

@@ -4,12 +4,13 @@ import torch
 import gradio as gr
 import logging
 from huggingface_hub import login
-from flash_attn.flash_attention import FlashAttention
 
 import os
 import traceback
 
 from threading import Thread
+import subprocess
+subprocess.run('pip install -U flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
 
 # Status: Breaks during generation
 

requirements.txt

@@ -5,5 +5,4 @@ accelerate==0.33.0
 sentencepiece==0.2.0
 spaces==0.29.2
 gradio==4.39.0
-bitsandbytes==0.43.2
-flash-attn
+bitsandbytes==0.43.2