v4.4.2 update. (#3104)

examples/python/CuTeDSL/blackwell/dense_gemm_persistent.py

@@ -1655,9 +1655,11 @@ def run(
             .to(dtype=torch_dtype(c_dtype))
             .to(dtype=torch.float32)
         )
-        # Read back the result from CuTe tensor (c_storage was updated in-place)
         torch.testing.assert_close(
-            c_storage.to(dtype=torch.float32), ref, atol=tolerance, rtol=1e-03
+            c_storage.view(torch_dtype(c_dtype)).to(dtype=torch.float32),
+            ref,
+            atol=tolerance,
+            rtol=1e-03,
         )
 
     if not benchmark:

examples/python/CuTeDSL/blackwell/dense_gemm_persistent_dynamic.py

@@ -1725,9 +1725,11 @@ def run(
             .to(dtype=torch_dtype(c_dtype))
             .to(dtype=torch.float32)
         )
-        # Read back the result from CuTe tensor (c_storage was updated in-place)
         torch.testing.assert_close(
-            c_storage.to(dtype=torch.float32), ref, atol=tolerance, rtol=1e-03
+            c_storage.view(torch_dtype(c_dtype)).to(dtype=torch.float32),
+            ref,
+            atol=tolerance,
+            rtol=1e-03,
         )
 
     if not benchmark:

examples/python/CuTeDSL/blackwell/fmha.py

@@ -2546,8 +2546,6 @@ def run(
         slices = tuple(slice(s, e) for s, e in zip(padding, shape_))
         torch_tensor = torch_tensor_full[slices].detach()
         f32_torch_tensor = f32_torch_tensor_full[slices].detach()
-        torch_tensor._keep_alive = torch_tensor_full
-        f32_torch_tensor._keep_alive = f32_torch_tensor_full
 
         # Create dtype cute tensor with offset (gpu)
         cute_tensor = from_dlpack(torch_tensor, assumed_align=16)