Memory over 100% with decoupled dali video model

[wq9]

Description Triton uses over 100% of physical memory and freezes the server when using a decoupled dali model with a long video input.

Triton Information Docker nvcr.io/nvidia/tritonserver:24.03-py3

To Reproduce resize_384/1/dali.py

import nvidia.dali as dali
from nvidia.dali.plugin.triton import autoserialize #must include

@dali.plugin.triton.autoserialize
@dali.pipeline_def(batch_size=1, num_threads=1, output_dtype=dali.types.FLOAT, output_ndim=3)
def pipeline():
    vid = dali.fn.experimental.inputs.video(name="INPUT", sequence_length=1, device='mixed')

    height = 384
    width = 640

    vid = dali.fn.resize(vid, resize_x=width, resize_y=height, mode="default", interp_type=dali.types.DALIInterpType.INTERP_CUBIC) #resize
    vid = dali.fn.crop(vid, crop_w=width, crop_h=height, out_of_bounds_policy="pad") #pad
    vid = dali.fn.squeeze(vid, axes=0) #remove sequence dim
    #vid = dali.fn.color_space_conversion(vid, image_type=dali.types.BGR, output_type=dali.types.RGB) #BGR to RGB
    vid = dali.fn.cast(vid, dtype=dali.types.FLOAT) #UINT8 to FP32
    vid = vid / dali.types.Constant(255)
    vid = dali.fn.transpose(vid, perm=[2, 0, 1], name="OUTPUT") #HWC to CHW
    return vid

resize_384/config.pbtxt

backend: "dali"
max_batch_size: 0

model_transaction_policy {
  decoupled: True
}

instance_group [
  {
    count: 1
    kind: KIND_GPU
    gpus: [ 1 ]
  }
]

client.py

#!/usr/bin/env python

import argparse
import numpy as np
import tritonclient.grpc
from functools import partial
import queue
import cv2
import time
import gc
import torch

def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument('-u', '--url', type=str, required=False, default='localhost:8001',
                        help='Inference server URL. Default is localhost:8001.')
    parser.add_argument('-v', '--video', type=str, required=False, default='/home/ubuntu/test_videos/office.mp4',
                        help='Path to video file')
    return parser.parse_args()

requests = queue.Queue()

def callback(queue, result, error):
    if error:
        print(error)
        queue.put(error)
    else:
        queue.put(result)

def main():
    FLAGS = parse_args()

    cam = cv2.VideoCapture(FLAGS.video)
    frame_count = int(cam.get(cv2.CAP_PROP_FRAME_COUNT))
    fps = cam.get(cv2.CAP_PROP_FPS)
    del cam
    gc.collect()

    with tritonclient.grpc.InferenceServerClient(url=FLAGS.url) as triton_client:
        # Start stream
        triton_client.start_stream(callback=partial(callback, requests))

        # Config output
        outputs = []
        outputs.append(tritonclient.grpc.InferRequestedOutput("OUTPUT"))

        # Config input 0: video
        video_raw = np.expand_dims(np.fromfile(FLAGS.video, dtype=np.uint8), axis=0)
        inputs = [
            tritonclient.grpc.InferInput("INPUT", video_raw.shape, "UINT8"),
        ]
        inputs[0].set_data_from_numpy(video_raw)

        print(f"{time.perf_counter() - tic:0.4f} |", FLAGS.video, frame_count, video_raw.shape)
        triton_client.async_stream_infer(model_name="resize_384", inputs=inputs, outputs=outputs)

        # Collect results
        frame_num=0

        while True:
            data_item = requests.get()
            bbox = torch.from_numpy(data_item.as_numpy("OUTPUT").copy())

            if frame_num%1000==0:
                print(f"{time.perf_counter() - tic:0.4f} | {frame_num}/{frame_count}")

            frame_num+=1

            if frame_num>=frame_count:
                break

        triton_client.close()

if __name__ == '__main__':
    tic = time.perf_counter()
    main()
    print(f"{time.perf_counter() - tic:0.4f} | Done")

Run triton server with docker run -it --gpus=all --net=host --rm -v /home/ubuntu:/home/ubuntu nvcr.io/nvidia/tritonserver:24.03-py3 tritonserver --model-repository=/home/ubuntu/model_repository, then run client.py with video of ~18min.

Triton server memory usage grows to 75GB of RAM.

If using a longer video, RAM memory usage goes over 100% and the server freezes.

Expected behavior RAM usage should not go over 100% and should not depend on input video length.

[statiraju]

[6864]. created for tracking

[k18coldwl]

I have the same problem. It is probably caused by the dali backend dali.fn.experimental.inputs.video API, because I have tried using the minimal code and the problem also happened.

version : 23.07