Move flash attention install to runtime

app.py

@@ -6,6 +6,9 @@ from utils import get_assistant_message
 from functioncall import ModelInference
 from prompter import PromptManager
 
+# HACK
+import subprocess
+subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
 
 @st.cache_resource(show_spinner="Loading model..")
 def init_llm():

requirements.txt

@@ -1,4 +1,3 @@
-flash-attn==2.5.5
 ninja==1.11.1.1
 numpy==1.26.4
 orjson==3.10.3