Torchx - Memory allocation / release problem (Only on Windows?)

[m0rt3nlund]

Hi!

As discussed in https://elixirforum.com/t/torchx-oom-on-windows/73005/6 thread there is a problem on Windows where the allocations for tensors never gets released.

My setup was on Elixir 1.18.3, OTP 27
I have tried both TorchX with libtorch 2.7.1 and 2.8.0 with same results.

Code to reproduce:

defmodule Mix.Tasks.Simple do
  @impl Mix.Task
  def run(_args) do
    10000
    |> runner()
  end

  defp runner(0), do: :ok

  defp runner(ittr) do
    Task.async(fn ->
      tensor =
        Nx.broadcast(0, {1, 3, 640, 640})
        |> Nx.backend_transfer(Nx.default_backend())
    end)
    |> Task.await()

    Process.sleep(50)

    runner(ittr - 1)
  end
end

[m0rt3nlund]

I have tried to debug the execution using MTuner to pinpoint the leak but getting all the required debug symbols was easier said then done!

And also, reading the code i cannot quite understand why this would be a Windows-only problem…

[polvalente]

From a very cursory analysis with Claude, it looks like windows might handle memory management differently to linux and mac. I'll try to set up my windows machine for native compilation of Torchx and this will hopefully let us debug in more depth

[m0rt3nlund]

Yes, my research tells the same. Seems to be a torchlib problem, even today 2025.

[m0rt3nlund]

This looks very promising so far. I have tested this in production now and it would appear that the leak is gone.
I will continue to monitor and report back if I find anything that does not work.