RAM memory leaks for EncDecCTCModelBPE at inference

[vikcost]

Numbers suggests that EncDecCTCModelBPE has RAM memory leaks.
I'm running inference with batch_size=1 on a directory with 10k small audio files with total duration of around 9.5 hours.

Does any body know where this excessive RAM usage comes from?

Here is a sample script:

import time
from pathlib import Path

import psutil
import torch
from nemo.collections.asr.models import EncDecCTCModelBPE

torch.set_num_threads(1)
proc_usage = psutil.Process()
cur_usage_rss = proc_usage.memory_info().rss / 1024 / 1024
print(f"Initial rss usage:{cur_usage_rss}")


dir_name = "/mnt/data/subset_10k"
audio_paths = Path(dir_name)


def find_leaks():
    model = EncDecCTCModelBPE.from_pretrained(model_name="stt_en_conformer_ctc_medium")
    res_file_path = f"memory_usage3_start_{time.time()}.txt"
    fd = open(res_file_path, "w")

    files = audio_paths.glob("*.wav")
    for idx, a_path in enumerate(files):
        res = model.transcribe(str(a_path), verbose=False)

        if idx % 100 == 0:
            cur_usage_rss = proc_usage.memory_info().rss / 1024 / 1024
            msg = f"{idx: <10} rss usage:{cur_usage_rss}"
            print(msg)
            print(msg, file=fd)

    fd.close()


if __name__ == "__main__":
    find_leaks()

And a sample output:

0          rss usage:1433.66796875
100        rss usage:1561.16796875
200        rss usage:1584.16796875
300        rss usage:1596.66796875
400        rss usage:1607.16796875
500        rss usage:1629.9375
600        rss usage:1636.4375
700        rss usage:1660.2890625
800        rss usage:1660.2890625
900        rss usage:1661.2890625
1000       rss usage:1672.7890625
1100       rss usage:1680.7890625
1200       rss usage:1688.7890625
1300       rss usage:1695.7890625
1400       rss usage:1697.2890625
1500       rss usage:1706.2890625
1600       rss usage:1718.09765625
1700       rss usage:1718.09765625
1800       rss usage:1721.09765625
1900       rss usage:1725.09765625
2000       rss usage:1738.59765625

Expected behavior

Constant memory consumption over time.

Environment overview (please complete the following information)

• Environment location: Bare-metal
• Method of NeMo install: python -m pip install git+https://github.com/NVIDIA/NeMo.git@{2.0.0rc0}#egg=nemo_toolkit[all]

Environment details

• OS version: Ubuntu 22.04, kernel 6.5.0-35-generic
• PyTorch version: 2.2.2
• Python version: 3.10.12