[Bug] Memory Leak while using VITS model

[abhijitpal1247]

Describe the bug

While querying the vits model for a number of texts, there seems to be a memory leakage. The memory consumption almost linearly increases as the loop progresses.

To Reproduce

 import os
 import torch
 from TTS.api import TTS
 
 
 # Get device
 device = "cpu"
 
 tts = TTS("tts_models/en/vctk/vits").to(device)
 
 
 texts_to_test = [
 "some 112 pieces of text, each consisting of multiple lines ...."
 ]
 
 
 speaker = tts.speakers[0]
 num_requests = 0
 for idx, txt in enumerate(texts_to_test):
     wav = tts.tts(text=txt, speaker=speaker, speed=0.8)

To reproduce the issue just replace texts_to_test with a list of good number of texts (>100) and each text should consist of 2-3 lines.
I performed this experiment on a c2-standard-4 machine (gcp).

Expected behavior

The consumption of RAM should be stable and should not increase while running the given code.

Logs

Environment

{
     "CUDA": {
         "GPU": [],
         "available": false,
         "version": "12.4"
     },
     "Packages": {
         "PyTorch_debug": false,
         "PyTorch_version": "2.6.0+cu124",
         "TTS": "0.25.3",
         "numpy": "1.26.4"
     },
     "System": {
         "OS": "Linux",
         "architecture": [
             "64bit",
             "ELF"
         ],
         "processor": "x86_64",
         "python": "3.10.16",
         "version": "#23~22.04.1-Ubuntu SMP Thu Jan 16 02:17:57 UTC 2025"
     }
 }

Additional context

No response

[juangon]

I think using:

import gc gc.collect() torch.cuda.empty_cache()

fixes this for me

[ROBERT-MCDOWELL]

"good number of texts (>100)"
do you mean 100 chars or words?
if it's words so it goes beyond the limit of an acceptable sentence and should be reduced by 2 (from my experience)