Apply patch BEFORE transformers import to fix is_torch_fx_available error

app.py

@@ -1,15 +1,17 @@
 import gradio as gr
 import torch
-from transformers import AutoTokenizer, AutoModelForCausalLM
 import os
-# Patch for missing is_torch_fx_available function
-try:
-    from transformers.utils import is_torch_fx_available
-except ImportError:
-    print("Patching missing is_torch_fx_available function...")
-    import transformers.utils
-    transformers.utils.is_torch_fx_available = lambda: False
 
+# CRITICAL: Patch must be applied BEFORE importing transformers
+import sys
+import importlib
+
+# Monkey-patch transformers.utils.import_utils to add missing function
+import transformers.utils.import_utils
+transformers.utils.import_utils.is_torch_fx_available = lambda: False
+print("Patched is_torch_fx_available function before transformers import")
+
+from transformers import AutoTokenizer, AutoModelForCausalLM
 
 print("Starting model loading...")
 print(f"CUDA available: {torch.cuda.is_available()}")
@@ -26,7 +28,7 @@ tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
 # Load model with automatic device mapping for multi-GPU support
 model = AutoModelForCausalLM.from_pretrained(
     model_name,
-    torch_dtype=torch.float16,
+    dtype=torch.float16,
     device_map="auto",
     trust_remote_code=True
 )