server: [Libtorch] Triton server produces inconsistent results when hosting multiple models in one GPU

Description With the triton server hosting multiple model on one GPU, we see inconsistent result when GPU is heavily used (same inputs, different and incorrect outputs).

We created a minimal example in the following repository to reproduce this issue: https://github.com/zmy1116/tritoninferenceissues In this example, we create two TensorRT models from TorchVision’s resnet18 and inceptionv3 using NVIDIA torch2trt. We launch 4 inference jobs for each of the two models with the same inputs. There is significant discrepancy in the results using the same model on the same input.

So far in our test, this error seems only happen when multiple models are running within the same GPU heavily. This issues seems to not model dependent. We have seen the issues on our production models with various backbone:

• resnet
• inceptionnet
• efficientnet
• lstm

Triton Information What version of Triton are you using? We use triton 2.5.0 Are you using the Triton container or did you build it yourself? We directly use the NGC container nvcr.io/nvidia/tritonserver:20.11-py3

Aditional system information

The problem occure on AWS VM instance type g4dn.2xlarge It contains one T4 GPU. The models in model repository are all TensoRT model, created in the correspondent NGC container nvcr.io/nvidia/tensorrt:20.11-py3

To Reproduce

We created a minimal example in the following repository to reproduce this issue: https://github.com/zmy1116/tritoninferenceissues

To summarize the steps:

• donwload and decompress the data folder trt_issues_data from the google drive link https://drive.google.com/file/d/1hBOpXMxSbeYbPltR23oJuzfdCxg6t1eo/view?usp=sharing

  The data folder consists of:

  • model repository for triton server model_repository_2011, two tensorrt models resnet18 and inceptionv3 are included
  • example inputs testing_inputs.p

• download the repository https://github.com/zmy1116/tritoninferenceissues

• start the triton server

# assume the model repository is at /home/ubuntu/trt_issues_data/model_repository_2011
docker run -d --gpus=all --rm --shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864 -p8000:8000 -p8001:8001 -p8002:8002 -v /home/ubuntu/trt_issues_data/model_repository_2011:/models -e CUDA_VISIBLE_DEVICES=0 nvcr.io/nvidia/tritonserver:20.11-py3 tritonserver --model-repository=/models --strict-model-config=false

• use python grpc client to invoke 2 models multiple times and store multiple runs results. For one T4 machine, I run:
  • each model 4 jobs
  • each job runs 64 times of inputs with size 64x3x224x224

nohup python /home/ubuntu/tritoninferenceissues/triton_inference.py resnet18 /home/ubuntu/trt_issues_data/testing_inputs.p 64 outputs1.p &
nohup python /home/ubuntu/tritoninferenceissues/triton_inference.py inceptionv3 /home/ubuntu/trt_issues_data/testing_inputs.p 64 outputs5.p &
nohup python /home/ubuntu/tritoninferenceissues/triton_inference.py resnet18 /home/ubuntu/trt_issues_data/testing_inputs.p 64 outputs2.p &
nohup python /home/ubuntu/tritoninferenceissues/triton_inference.py inceptionv3 /home/ubuntu/trt_issues_data/testing_inputs.p 64 outputs6.p &
nohup python /home/ubuntu/tritoninferenceissues/triton_inference.py resnet18 /home/ubuntu/trt_issues_data/testing_inputs.p 64 outputs3.p &
nohup python /home/ubuntu/tritoninferenceissues/triton_inference.py inceptionv3 /home/ubuntu/trt_issues_data/testing_inputs.p 64 outputs7.p &
nohup python /home/ubuntu/tritoninferenceissues/triton_inference.py resnet18 /home/ubuntu/trt_issues_data/testing_inputs.p 64 outputs4.p &
nohup python /home/ubuntu/tritoninferenceissues/triton_inference.py inceptionv3 /home/ubuntu/trt_issues_data/testing_inputs.p 64 outputs8.p &

I think the important thing here is to run GPU intensively enough (so that GPU reaches at least 80%+ utils).

In the above specific setup, outputs 1-4 are from model resnet18, output 5-8 are from model inceptionv3, since i passed the same inputs they should return the same outputs. However, when comparing the results you should see signifcant discrepancies:

f1 = pickle.load(open('/home/ubuntu/outputs1.p','rb'))
f2 = pickle.load(open('/home/ubuntu/outputs2.p','rb'))
f3 = pickle.load(open('/home/ubuntu/outputs3.p','rb'))
f4 = pickle.load(open('/home/ubuntu/outputs4.p','rb'))

f5 = pickle.load(open('/home/ubuntu/outputs5.p','rb'))
f6 = pickle.load(open('/home/ubuntu/outputs6.p','rb'))
f7 = pickle.load(open('/home/ubuntu/outputs7.p','rb'))
f8 = pickle.load(open('/home/ubuntu/outputs8.p','rb'))
for entry in [f2,f3,f4]:
    print(np.max(np.abs(f1-entry)))
for entry in [f7,f6,f8]:
    print(np.max(np.abs(f5-entry)))

You can certainly generate the inputs data TensorRT models yourself, the inputs data is just a random generated array of size 64x3x224x224, the two models are from torchvision directly

import torch
import torchvision

model = torchvision.models.resnet18(pretrained=True).cuda().half().eval()
data = torch.randn((1, 3, 224, 224)).cuda().half()
with open('/workspace/ubuntu/model_repository_2011/resnet18/1/model', "wb") as f:
    f.write(model_trt.engine.serialize())

model = torchvision.models.inception_v3(pretrained=True).cuda().half().eval()
data = torch.randn((1, 3, 224, 224)).cuda().half()
with open('/workspace/ubuntu/model_repository_2011/inceptionv3/1/model', "wb") as f:
    f.write(model_trt.engine.serialize())

Expected behavior

I think we should expect for the same model (of course, that is supposed to give deterministics result) should return correct and consistent result for the same inputs.

What we see in our actual work is that the discrepancies are significantly large. The more models we host, the larger magnitude the errors are. Without using TritonServer, we can confirm that these models produce consistent results when running only one model on GPU.

Let me know if you need additional information.