Update src/florence/model.py

src/florence/model.py

@@ -1,6 +1,7 @@
 # Importing necessary libraries
 import os
 import sys
+import importlib
 import subprocess
 from typing import Optional
 from PIL import Image
@@ -16,17 +17,25 @@ from src.exception import CustomExceptionHandling
 
 # Install the required dependencies
 # NOTE: passing env={...} to subprocess.run REPLACES the entire environment
-# instead of extending it, which strips PATH/HOME/etc. and can break the
-# pip install (and therefore the whole Space) at import time. Copy the
-# current environment first and only add/override the flag we need.
+# instead of extending it, which strips PATH/HOME/etc. Copy the current
+# environment first and only add/override the flag we need.
 env = os.environ.copy()
 env["FLASH_ATTENTION_SKIP_CUDA_BUILD"] = "TRUE"
 subprocess.run(
     "pip install flash-attn --no-build-isolation",
     env=env,
     shell=True,
+    check=True,
 )
 
+# The pip install above happens in a subprocess AFTER this Python process
+# has already started, so the interpreter's import machinery has a stale
+# view of what's on disk (sys.path_importer_cache). Without invalidating
+# it, transformers' internal check_imports()/find_spec() lookup for
+# flash_attn reports "not found" even though it just installed
+# successfully -- this is what was causing the ImportError at model load.
+importlib.invalidate_caches()
+
 # Load model and processor from Hugging Face
 model_id = "microsoft/Florence-2-large-ft"
 try: