In the same process, decoding and reasoning will cause GPU video memory leakage

[SubMarineas]

The problem is that when I add resnet or yolo to VPF as a thread, I find that GPU increases rapidly until out of memory. The code is:

class VPF_decode_inference(object):
     
    def run(self):
        self.rtsp_list = ["rtmp://192.168.1.229/live/mytest",]
        # cuda.init()
        decode_thread_pool = []
        gpu_id = 0
        yolo_thread_pool = []
        for i in range(0, 25):
            thread = Thread(target = VPF_decode_inference.decode,args=[self,gpu_id, "rtmp://192.168.1.229/live/mytest",i])
            thread.start()
            decode_thread_pool.append(thread)
        for thread in decode_thread_pool:
            thread.join()
        for i in range(1):
            vpf_yolo_obj = yolo_inference(gpu_id)
            yolo_thread  = Thread(target = yolo_inference.inference,args=[vpf_yolo_obj])
            yolo_thread.start()
            yolo_thread.join()


class yolo_inference(object):
    pass


if __name__ == "__main__":
    gpu_id = 0
    
    for i in range(1):
        vpf_decode = VPF_decode_inference(0)
        p = Process(target=VPF_decode_inference.run,args=(vpf_decode,))
        p.start()
    time.sleep()

[SubMarineas]

However, if I put yolo under the main function without starting the process, it will not leak, roughly as follows:

class yolo_inference(object):
    pass


if __name__ == "__main__":
    gpu_id = 0
    
    vpf_yolo_obj = yolo_inference(gpu_id)
    yolo_thread  = Thread(target = yolo_inference.inference,args=[vpf_yolo_obj])
    yolo_thread.start()
    vpf_decode = VPF_decode_inference(0)
    vpf_decode.run()

    yolo_thread.join()

Why? I don't understand the memory allocation of GPU.The API I decoded to the inference uses pnvc.DptrToTensor.Image is not connected to the CPU

[SubMarineas]

oh i found out i made a nasty mistake