update requirements

app.py

@@ -20,11 +20,11 @@ import argparse
 from transformers import TextIteratorStreamer
 from threading import Thread
 
-import subprocess
-import sys
+# import subprocess
+# import sys
 
-print("Installing flash attention...")
-subprocess.check_call([sys.executable, "-m", "pip", "install", "flash-attn==2.4.2", "--no-build-isolation"])
+# print("Installing flash attention...")
+# subprocess.check_call([sys.executable, "-m", "pip", "install", "flash-attn==2.4.2", "--no-build-isolation"])
 
 no_change_btn = gr.Button()
 enable_btn = gr.Button(interactive=True)

requirements.txt

@@ -40,6 +40,7 @@ timm==0.9.11
 toml==0.10.2
 tqdm==4.66.1
 loguru==0.7.3
+https://github.com/Dao-AILab/flash-attention/releases/download/v2.4.2/flash_attn-2.4.2+cu122torch2.1cxx11abiTRUE-cp310-cp310-linux_x86_64.whl
 tenacity
 sqlitedict
 evaluate==0.4.3