Regarding the issue of GPU memory usage in models

[w-Jelf]

Hello, my steps are as follows.
First, use VAD (Voice Activity Detection) to process a WAV audio file with a duration of about 70 seconds and split it into audio segments no longer than 8 seconds. Then, use your model to perform Chinese audio recognition.
During this process, the GPU memory usage for audio is around 6000MB. However, after running, the GPU memory is not released, and even gradually increases (approximately 1MB increase for processing 1000 audio files on top of the original usage). Is there any solution?

您好，我的步骤是这样的。
先对一个时长约70秒的wav音频使用vad处理切分成不超过8秒的音频片段，之后用您的模型进行中文音频识别。
在这个过程中，音频的GPU显存占用为6000MB左右，但运行过后没有释放对应的显存，甚至在逐渐增高（大约是处理1000条音频在原本基础上增加1MB）请问有解决思路吗？