Update app.py

app.py

@@ -1,17 +1,40 @@
-# --- MUST be first: disable Hugging Face Spaces ZeroGPU monkey-patch ---
-import os
+
+import os, sys, importlib
+
+# Disable Spaces ZeroGPU by env and by calling its disable/unpatch if preloaded
 os.environ["SPACES_ZERO_DISABLED"] = "1"
 
-# (optional but helpful) steer PyTorch to math attention kernels (no Flash/MemEfficient)
+def _hard_disable_spaces_zero():
+    # Hit common modules and try disable/unpatch/deactivate if present
+    candidates = [
+        "spaces.zero", "spaces.zero.torch.patching", "spaces.zero.torch",
+        "spaces.zero.patch", "spaces.zero.patching"
+    ]
+    for modname in candidates:
+        try:
+            m = sys.modules.get(modname) or importlib.import_module(modname)
+        except Exception:
+            continue
+        for attr in ("disable", "unpatch", "deactivate"):
+            fn = getattr(m, attr, None)
+            if callable(fn):
+                try:
+                    fn()
+                except Exception:
+                    pass
+
+_hard_disable_spaces_zero()
+
+# Force Transformers to use eager attention globally (affects all future loads)
+os.environ["TRANSFORMERS_ATTENTION_IMPLEMENTATION"] = "eager"
+
+# Prefer simple math SDP kernels (avoid vmap-heavy paths)
 try:
     import torch
     torch.backends.cuda.sdp_kernel(enable_math=True, enable_flash=False, enable_mem_efficient=False)
 except Exception:
     pass
-
-# If you truly need Spaces, import it AFTER disabling the patch.
-import spaces
-
+# -------------------------------------------------------------------------------
 import sys
 from huggingface_hub import hf_hub_download
 import pickle