Repository navigation
TensorAccessor on GPU tensor leaks memory #1517
Description
Describe the bug
Creating a TensorAccessor on a GPU tensor leaks a CPU tensor after the accessor has been disposed.
To Reproduce
let test () =
use scope = torch.NewDisposeScope()
do
use tensor = torch.tensor([|1.0f; 2.0f; 3.0f|], device=torch.CUDA)
use accessor = tensor.data<float32>()
printfn "%A" <| accessor.ToArray()
printfn "Disposables: %A" scope.DisposablesCount
for disposable in scope.DisposablesView do
printfn "%A" disposableExpected behavior
Output should be:
[|1.0f; 2.0f; 3.0f|]
Disposables: 0
Actual output:
[|1.0f; 2.0f; 3.0f|]
Disposables: 1
[3], type = Float32, device = cpu
Note that setting device=torch.CPU when creating the original tensor does result in the expected output of 0 disposables.
Please complete the following information:
- OS: Windows
- Package Type: torchsharp-cuda-windows
- Version: 0.105.2
Activity
https://github.com/dotnet/TorchSharp/issues/1517
ds5678 commentedon Mar 23, 2026
TorchSharp/src/TorchSharp/Tensor/Tensor.cs
Line 434 in b2a2f20
litn2018 commentedon May 31, 2026
running GPU memory problem too. Not release memory on time. Should we expose method to releae memory explicitly? I ran into the problem with SGD + IRLS chain to train logistic model, after SGD, GPU memory is not released on time.