Memory Leak in python backend decouple mode #6270
Labels
Comments
rainyfly
changed the title
|
Hello @rainyfly, thank you for reporting this issue. We are aware of this bug and are currently working on a fix that we hope to have out soon. |
|
Hi, I use the latest container for r23.08, the problem still exists. You can follow the instructions to reproduce. ** Host Environment ** ** Image **
docker run --name test_r23.08 -v `pwd`:/work --network=host --shm-size=10g -it nvcr.io/nvidia/tritonserver:23.08-py3 /bin/bash
cd /work
mkdir debug_memory_leak
cd debug_memory_leak
touch client.py
mkdir models
cd models
mkdir square_int32
cd square_int32
touch config.pbtxt
mkdir 1
cd 1
touch model.pyCode for config.pbtxt Code for model.py # Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
import json
import threading
import time
import numpy as np
# triton_python_backend_utils is available in every Triton Python model. You
# need to use this module to create inference requests and responses. It also
# contains some utility functions for extracting information from model_config
# and converting Triton input/output types to numpy types.
import triton_python_backend_utils as pb_utils
class TritonPythonModel:
"""Your Python model must use the same class name. Every Python model
that is created must have "TritonPythonModel" as the class name.
This model demonstrates how to write a decoupled model where each
request can generate 0 to many responses.
This model has one input and one output. The model can support batching,
with constraint that each request must be batch-1 request, but the shapes
described here refer to the non-batch portion of the shape.
- Input 'IN' must have shape [1] and datatype INT32.
- Output 'OUT' must have shape [1] and datatype INT32.
For a request, the backend will sent 'n' responses where 'n' is the
element in IN. For each response, OUT will equal the element of IN.
"""
def initialize(self, args):
"""`initialize` is called only once when the model is being loaded.
Implementing `initialize` function is optional. This function allows
the model to initialize any state associated with this model.
Parameters
----------
args : dict
Both keys and values are strings. The dictionary keys and values are:
* model_config: A JSON string containing the model configuration
* model_instance_kind: A string containing model instance kind
* model_instance_device_id: A string containing model instance device ID
* model_repository: Model repository path
* model_version: Model version
* model_name: Model name
"""
# You must parse model_config. JSON string is not parsed here
self.model_config = model_config = json.loads(args["model_config"])
using_decoupled = pb_utils.using_decoupled_model_transaction_policy(
model_config
)
if not using_decoupled:
raise pb_utils.TritonModelException(
"""the model `{}` can generate any number of responses per request,
enable decoupled transaction policy in model configuration to
serve this model""".format(
args["model_name"]
)
)
# Get IN configuration
in_config = pb_utils.get_input_config_by_name(model_config, "IN")
# Validate the shape and data type of IN
in_shape = in_config["dims"]
if (len(in_shape) != 1) or (in_shape[0] != 1):
raise pb_utils.TritonModelException(
"""the model `{}` requires the shape of 'IN' to be
[1], got {}""".format(
args["model_name"], in_shape
)
)
if in_config["data_type"] != "TYPE_INT32":
raise pb_utils.TritonModelException(
"""the model `{}` requires the data_type of 'IN' to be
'TYPE_INT32', got {}""".format(
args["model_name"], in_config["data_type"]
)
)
# Get OUT configuration
out_config = pb_utils.get_output_config_by_name(model_config, "OUT")
# Validate the shape and data type of OUT
out_shape = out_config["dims"]
if (len(out_shape) != 1) or (out_shape[0] != 1):
raise pb_utils.TritonModelException(
"""the model `{}` requires the shape of 'OUT' to be
[1], got {}""".format(
args["model_name"], out_shape
)
)
if out_config["data_type"] != "TYPE_INT32":
raise pb_utils.TritonModelException(
"""the model `{}` requires the data_type of 'OUT' to be
'TYPE_INT32', got {}""".format(
args["model_name"], out_config["data_type"]
)
)
self.inflight_thread_count = 0
self.inflight_thread_count_lck = threading.Lock()
def execute(self, requests):
"""`execute` MUST be implemented in every Python model. `execute`
function receives a list of pb_utils.InferenceRequest as the only
argument. This function is called when an inference request is made
for this model. The request.get_response_sender() must be used to
get an InferenceResponseSender object associated with the request.
Use the InferenceResponseSender.send(response=<infer response object>,
flags=<flags>) to send responses.
In the final response sent using the response sender object, you must
set the flags argument to TRITONSERVER_RESPONSE_COMPLETE_FINAL to
indicate no responses will be sent for the corresponding request. If
there is an error, you can set the error argument when creating a
pb_utils.InferenceResponse. Setting the flags argument is optional and
defaults to zero. When the flags argument is set to
TRITONSERVER_RESPONSE_COMPLETE_FINAL providing the response argument is
optional.
Parameters
----------
requests : list
A list of pb_utils.InferenceRequest
Returns
-------
None
"""
# Visit individual request to start processing them. Note that execute
# function is not required to wait for all the requests of the current
# batch to be processed before returning.
for request in requests:
self.process_request(request)
# Unlike in non-decoupled model transaction policy, execute function
# here returns no response. A return from this function only notifies
# Triton that the model instance is ready to receive another batch of
# requests. As we are not waiting for the response thread to complete
# here, it is possible that at any give time the model may be processing
# multiple batches of requests. Depending upon the request workload,
# this may lead to a lot of requests being processed by a single model
# instance at a time. In real-world models, the developer should be
# mindful of when to return from execute and be willing to accept next
# request batch.
return None
def process_request(self, request):
# Start a separate thread to send the responses for the request. The
# sending back the responses is delegated to this thread.
thread = threading.Thread(
target=self.response_thread,
args=(
request.get_response_sender(),
pb_utils.get_input_tensor_by_name(request, "IN").as_numpy(),
),
)
# A model using decoupled transaction policy is not required to send all
# responses for the current request before returning from the execute.
# To demonstrate the flexibility of the decoupled API, we are running
# response thread entirely independent of the execute thread.
thread.daemon = True
with self.inflight_thread_count_lck:
self.inflight_thread_count += 1
thread.start()
def response_thread(self, response_sender, in_input):
# The response_sender is used to send response(s) associated with the
# corresponding request.
for idx in range(in_input[0]):
out_output = pb_utils.Tensor("OUT", np.array([in_input[0]], np.int32))
response = pb_utils.InferenceResponse(output_tensors=[out_output])
response_sender.send(response)
# We must close the response sender to indicate to Triton that we are
# done sending responses for the corresponding request. We can't use the
# response sender after closing it. The response sender is closed by
# setting the TRITONSERVER_RESPONSE_COMPLETE_FINAL.
response_sender.send(flags=pb_utils.TRITONSERVER_RESPONSE_COMPLETE_FINAL)
with self.inflight_thread_count_lck:
self.inflight_thread_count -= 1
def finalize(self):
"""`finalize` is called only once when the model is being unloaded.
Implementing `finalize` function is OPTIONAL. This function allows
the model to perform any necessary clean ups before exit.
Here we will wait for all response threads to complete sending
responses.
"""
print("Finalize invoked")
inflight_threads = True
cycles = 0
logging_time_sec = 5
sleep_time_sec = 0.1
cycle_to_log = logging_time_sec / sleep_time_sec
while inflight_threads:
with self.inflight_thread_count_lck:
inflight_threads = self.inflight_thread_count != 0
if cycles % cycle_to_log == 0:
print(
f"Waiting for {self.inflight_thread_count} response threads to complete..."
)
if inflight_threads:
time.sleep(sleep_time_sec)
cycles += 1
print("Finalize complete...")Code for client.py # Copyright 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# * Neither the name of NVIDIA CORPORATION nor the names of its
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
# PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
import queue
import sys
from functools import partial
import numpy as np
import tritonclient.grpc as grpcclient
from tritonclient.utils import *
class UserData:
def __init__(self):
self._completed_requests = queue.Queue()
def callback(user_data, result, error):
if error:
user_data._completed_requests.put(error)
else:
user_data._completed_requests.put(result)
# This client sends a 4 requests to the model with the
# input as: [4], [2], [0] and [1] respectively. In
# compliance with the behavior of square_int32 model,
# it will expect the 4 responses for the 1st request
# each with output [4], 2 responses for 2nd request
# each with output [2], no response for the 3rd request
# and finally 1 response for the 4th request with output
# [1]
model_name = "square_int32"
in_values = [10000, 10001, 10002, 10003]
inputs = [grpcclient.InferInput("IN", [1], np_to_triton_dtype(np.int32))]
outputs = [grpcclient.InferRequestedOutput("OUT")]
user_data = UserData()
with grpcclient.InferenceServerClient(
url="localhost:8001", verbose=True
) as triton_client:
# Establish stream
triton_client.start_stream(callback=partial(callback, user_data))
# Send specified many requests in parallel
for i in range(len(in_values)):
in_data = np.array([in_values[i]], dtype=np.int32)
inputs[0].set_data_from_numpy(in_data)
triton_client.async_stream_infer(
model_name=model_name,
inputs=inputs,
request_id=str(i),
outputs=outputs,
)
# Retrieve results...
recv_count = 0
expected_count = sum(in_values)
result_dict = {}
while recv_count < expected_count:
data_item = user_data._completed_requests.get()
if type(data_item) == InferenceServerException:
raise data_item
else:
this_id = data_item.get_response().id
if this_id not in result_dict.keys():
result_dict[this_id] = []
result_dict[this_id].append((recv_count, data_item))
recv_count += 1
# Validate results...
for i in range(len(in_values)):
this_id = str(i)
if in_values[i] != 0 and this_id not in result_dict.keys():
print("response for request id {} not received".format(this_id))
sys.exit(1)
elif in_values[i] == 0 and this_id in result_dict.keys():
print("received unexpected response for request id {}".format(this_id))
sys.exit(1)
if in_values[i] != 0:
if len(result_dict[this_id]) != in_values[i]:
print(
"expected {} many responses for request id {}, got {}".format(
in_values[i], this_id, result_dict[this_id]
)
)
sys.exit(1)
if in_values[i] != 0:
result_list = result_dict[this_id]
expected_data = np.array([in_values[i]], dtype=np.int32)
for j in range(len(result_list)):
this_data = result_list[j][1].as_numpy("OUT")
if not np.array_equal(expected_data, this_data):
print(
"incorrect data: expected {}, got {}".format(
expected_data, this_data
)
)
sys.exit(1)
print("PASS: square_int32")
sys.exit(0)
cd /work/debug_memory_leak
tritonserver --model-repository models
python3 -m pip install tritonclient[grpc]
python3 client.py
watch -n 1 ps -auxRecord the RSS for process python backend stub every time after run client.py. Repeat serveral times and record the variant. This example is nothing matters with GPU. I think maybe it is easily reproduced in any GPU environment. |
|
Hello @rainyfly, As Jian has communicated to you, this is not the same issue I previously thought. Thank you for your thorough reproduction steps. I have created a ticket (DLIS-5435) to investigate the matter further. CC @Tabrizian @krishung5 for vis. |
fpetrini15
added
bug
|
@fpetrini15 What is the current progress? Is there an expected completion time? |
2 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Description
I noticed that when I use python backend decouple mode to serve, the memory continues to increase.
I use the official examples in [https://github.com/triton-inference-server/python_backend/tree/main/examples/decoupled](Decoupled Model Examples). I use square_model to serve and use square_client.py to request the server.
I change the in value to 10000. And repeat run the client for several times. Here is the VSZ and RSS data for python backend stub I record from
ps -aux. We can see the RSS data is increasing every time after I run the square_client.py.Triton Information
What version of Triton are you using?
r23.06
The text was updated successfully, but these errors were encountered: