Memory leak DNSMOS functional interface

[augmented-albert]

🐛 Bug

CUDA memory overhang when using the DNSMOS functional interface with onnxruntime-gpu for computing batched metrics on various standard voice enhancement test sets.

To Reproduce

import torch
from torchmetrics.functional.audio import deep_noise_suppression_mean_opinion_score as DNSMOS

enhanced = torch.nn.utils.rnn.pad_sequence([torch.randn(1, torch.randint(48000, 160000, (1,))) for _ in range(8)], batch_first=True).to("cuda" if torch.cuda.is_available() else "cpu")

dnsmos_scores_1 = DNSMOS(preds=enhanced, fs=16_000, personalized=False)
dnsmos_scores_2 = DNSMOS(preds=enhanced, fs=16_000, personalized=False)
dnsmos_scores_3 = DNSMOS(preds=enhanced, fs=16_000, personalized=False)

Expected behavior

CUDA memory is expected to be cleared after use. Upon further inspection, an onnxruntime session is started but never terminated, hence the overhang.

Environment

• torch==2.5.1
• torchaudio==2.5.1
• onnxruntime-gpu==1.20.1
• torchmetrics==1.6.1
• Python & PyTorch Version (e.g., 1.0): 3.10.x and 2.5.1
• UBUNTU

[github-actions]

Hi! thanks for your contribution!, great first issue!

[Borda]

hi, @augmented-albert could you pls add what TM version you are using?

[augmented-albert]

hi, @augmented-albert could you pls add what TM version you are using?

torchmetrics==1.6.1