Fix(MInference): fix the env

app.py

@@ -1,21 +1,12 @@
 import subprocess
+import os
 # Install flash attention, skipping CUDA build if necessary
-subprocess.run(
-    "export CPATH=$CPATH:/usr/local/cuda/include",
-    shell=True,
-)
-subprocess.run(
-    "export CUDA_HOME=/usr/local/cuda",
-    shell=True,
-)
-subprocess.run(
-    "export PATH=$PATH:$CUDA_HOME/bin",
-    shell=True,
-)
-subprocess.run(
-    "export LIBRARY_PATH=$LIBRARY_PATH:/usr/local/cuda/lib64",
-    shell=True,
-)
+
+os.environ["CPATH"] = "$CPATH:/usr/local/cuda/include"
+os.environ["CUDA_HOME"] = "/usr/local/cuda"
+os.environ["PATH"] = "$PATH:$CUDA_HOME/bin"
+os.environ["LIBRARY_PATH"] = "$LIBRARY_PATH:/usr/local/cuda/lib64"
+
 subprocess.run(
     "/home/user/.pyenv/shims/pip install flash-attn --no-build-isolation",
     env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},