Switch to model.half() (float16) for native T4 acceleration with correct Meta checkpoint loading

app.py

@@ -86,17 +86,15 @@ processor = None
 if model_installed:
     print(f"Loading SAM 3 onto {device}...")
     model = build_sam3_image_model(checkpoint_path=ckpt_path)
-    # Let model stay in its NATIVE dtype (bfloat16 from checkpoint).
-    # Our F.linear/F.conv2d interceptors handle dtype mismatches dynamically.
-    # DO NOT cast to float32 — it was causing zero mask outputs!
+    # Cast to float16 — T4 has native float16 Tensor Core acceleration.
+    # bfloat16 hangs (software emulated on Turing), float32 produced zero masks.
+    model.half()
     
-    # Diagnostic: print parameter dtypes to verify checkpoint loaded correctly
-    param_dtypes = set()
-    for name, p in model.named_parameters():
-        param_dtypes.add(str(p.dtype))
-    print(f"Model parameter dtypes: {param_dtypes}", flush=True)
+    # Diagnostic: verify checkpoint loaded correctly
     total_params = sum(p.numel() for p in model.parameters())
     print(f"Total parameters: {total_params:,}", flush=True)
+    sample_dtype = next(model.parameters()).dtype
+    print(f"Model dtype: {sample_dtype}", flush=True)
         
     processor = Sam3Processor(model)
     if not torch.cuda.is_available():