CuTe DSL documentation recommends caching CuTe tensors to avoid the overhead of repeated from_dlpack calls.

Experiments with caching show that the original torch tensors are not being garbage collected.
For example, this program:

from cutlass.cute.runtime import from_dlpack
import torch

cache = {}
for i in range(0, 5):
    torch_tensor = torch.empty(1024, dtype=torch.int8, device="cuda")
    cute_tensor = from_dlpack(torch_tensor)
    cache[i] = cute_tensor
    del torch_tensor # explicit deletion to ensure GC opportunity
    print(f"allocated {torch.cuda.memory_allocated()} bytes")

produces:

allocated 1024 bytes
allocated 2048 bytes
allocated 3072 bytes
allocated 4096 bytes
allocated 5120 bytes

When the line cache[i] = cute_tensor is commented out, the program prints:

allocated 1024 bytes
allocated 1024 bytes
allocated 1024 bytes
allocated 1024 bytes
allocated 1024 bytes

This suggests that the CuTe tensor returned by from_dlpack retains an internal reference to the original torch tensor.
Is this the intended behavior?