HOTFIX: Remove global bfloat16 alias to unbreak torch.load checkpoint parsing

app.py

@@ -11,17 +11,13 @@ import io
 # can *emulate* bfloat16 in software, but the actual kernels crash on mixed
 # dtype operations (linear, conv2d). We MUST patch unconditionally.
 if torch.cuda.is_available():
-    # 1. Globally alias bfloat16 → float16 so all future lookups resolve to fp16
-    _original_bf16 = torch.bfloat16
-    torch.bfloat16 = torch.float16
-    
-    # 2. Intercept ALL autocast entry points to force float16
+    # Intercept ALL autocast entry points to force float16
     import torch.amp.autocast_mode
     _OriginalAutocast = torch.amp.autocast_mode.autocast
     class _Fp16Autocast(_OriginalAutocast):
         def __init__(self, device_type, dtype=None, *args, **kwargs):
-            # Intercept any bfloat16 request (original C enum or aliased)
-            if dtype is not None and dtype in (_original_bf16, torch.float16):
+            # Intercept Meta's bfloat16 request and force float16 for Turing support
+            if dtype == torch.bfloat16:
                 dtype = torch.float16
             super().__init__(device_type, dtype=dtype, *args, **kwargs)