[PT] detach tensor in get_const_data

[AlexanderDokuchaev]

Changes

Return detached tensor from get_const_data function

Reason for changes

Memory leak

[AlexanderDokuchaev]

Reproducer

import torch
results = []
for _  in range(100):
    a = torch.rand((1,3,1000, 1000),requires_grad=True)
    b = torch.rand((1,3,1000, 1000),requires_grad=True)
    r = torch.sum(a + b)
    results.append(r)

import torch
results = []
for _  in range(100):
    a = torch.rand((1,3,1000, 1000),requires_grad=True)
    b = torch.rand((1,3,1000, 1000),requires_grad=True)
    r = torch.sum(a.detach() + b.detach())
    results.append(r)

[alexsu52]

Do you have an idea how to check that you covered all places in NNCF code?

[AlexanderDokuchaev]

No, but problem actual only for case when original tensor is replaced to new in the model which was calculated by using original.