Possible memory leak executing inference multiple times

[insanebytes]

Hello when i run inference by this code, i see that memory increases in usage without freeing.

`var cwd = Directory.GetCurrentDirectory();
string modelDirPath = Path.Join(cwd, "Assets", "Voice");
string modelPath = Path.Join(modelDirPath, "vits-vctk.int8.onnx");

OfflineTtsVitsModelConfig modelConfigVits = new OfflineTtsVitsModelConfig();
modelConfigVits.Model = modelPath;
modelConfigVits.Lexicon = Path.Join(modelDirPath, "lexicon.txt");
modelConfigVits.Tokens = Path.Join(modelDirPath, "tokens.txt");

OfflineTtsModelConfig modelConfig = new OfflineTtsModelConfig();
modelConfig.Vits = modelConfigVits;
modelConfig.Provider = "cuda";

OfflineTtsConfig config = new OfflineTtsConfig();
config.Model = modelConfig;

var offlineTts = new SherpaOnnx.OfflineTts(config);

var audioDevice = new WasapiOut();
var waveFormat = WaveFormat.CreateIeeeFloatWaveFormat((offlineTts.SampleRate), 1);
var waveFileStream = new MemoryStream(waveFormat.ConvertLatencyToByteSize(30 * 1000)); //pre allocate 30 seconds
var rawSourceWaveStream = new RawSourceWaveStream(waveFileStream, waveFormat);

audioDevice.Init(rawSourceWaveStream);

while (true)
{
var input = Console.ReadLine();

waveFileStream.Seek(0, SeekOrigin.Begin);
waveFileStream.SetLength(0);

Task.Run(() =>
{
    var result = offlineTts.Generate(input, 1f, 1);
    if (result.Samples.Length > 0)
    {
        var wave = new byte[result.NumSamples * sizeof(float)];
        Buffer.BlockCopy(result.Samples, 0, wave, 0, wave.Length);

        waveFileStream.Write(wave, 0, wave.Length);
        waveFileStream.Position = 0;

        if (audioDevice.PlaybackState == PlaybackState.Stopped)
        {
            audioDevice.Stop();
        }
        audioDevice.Play();

        result.Dispose();
    }
});

}`

when first executed waiting:

274 MB RAM

executed: hello world how are you

341 MB RAM

second execution same phrase:

342 MB RAM

third execution same phrase:

343 MB RAM

Im am disposing the result struct.

Is there something more to dispose? or that 1MB extra per inference is a memory leak?

execution screenshot:

Thank you

[insanebytes]

Any update about this??

[csukuangfj]

please run for 10 minutes and post the result

[insanebytes]

runned 10 min and memory stabilized sorry, but in other machine: D:\a\sherpa-onnx\sherpa-onnx\sherpa-onnx\csrc\session.cc:GetSessionOptionsImpl:176 Please compile with -DSHERPA_ONNX_ENABLE_GPU=ON. Available providers: AzureExecutionProvider, CPUExecutionProvider, . Fallback to cpu!

and i have installed cuda:

[csukuangfj]

Please follow.our doc and search in sherpa-onnx's issues about running sherpa-onnx with gpu.

[csukuangfj]

and i have installed cuda:

@insanebytes Please see #1954