DIAGNOSTIC: Remove float32 cast, let model run in native bfloat16 with interceptors. Print param dtypes

app.py

@@ -86,9 +86,17 @@ processor = None
 if model_installed:
     print(f"Loading SAM 3 onto {device}...")
     model = build_sam3_image_model(checkpoint_path=ckpt_path)
-    # NOTE: build_sam3_image_model() internally loads weights, remaps keys,
-    # and calls model.cuda().eval(). We just need to cast to float32 for T4.
-    model.to(torch.float32)
+    # 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!
+    
+    # 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)
+    total_params = sum(p.numel() for p in model.parameters())
+    print(f"Total parameters: {total_params:,}", flush=True)
         
     processor = Sam3Processor(model)
     if not torch.cuda.is_available():