Monkey-patch torch.autocast to natively support T4 Turing GPUs by intercepting bfloat16 demands

app.py

@@ -31,6 +31,17 @@ if not torch.cuda.is_available():
                 return __orig_fn(*args, **kwargs)
             setattr(torch, name, patched_fn)
 
+# Intercept Meta's hardcoded BFloat16 autocast (T4 Turing GPUs don't support BFloat16 hardware math)
+original_autocast = torch.autocast
+class PatchedAutocast(original_autocast):
+    def __init__(self, device_type, dtype=None, *args, **kwargs):
+        if dtype == torch.bfloat16 and torch.cuda.is_available() and not torch.cuda.is_bf16_supported():
+            dtype = torch.float16  # Fallback to fp16, supported perfectly by T4 Turing NVidia cards
+        if device_type == 'cuda' and not torch.cuda.is_available():
+            device_type = 'cpu'
+        super().__init__(device_type, dtype, *args, **kwargs)
+torch.autocast = PatchedAutocast
+
 # ── SAM 3 Imports ────────────────────────────────────────────────
 try:
     from sam3.model_builder import build_sam3_image_model