kmkolasinski / triton-saved-model

Triton Python Backend for Multimodel/Signatures Inference Demo

• This repository contains code for running multiple models on Triton Inference Server with python backend
• The python backend code can handle multiple models and multiple signatures
• This project was prepared for educational purposes, to show how we can use Triton Inference Server with python backend to simulate similar API as TFServing

docker compose build

Running notebooks

• Notebooks were tested with python version 3.11.4, see requirements.txt
• Use export-classifier.ipynb to export various classifiers.
• Triton and TFServing will be reading these models from models.conf configuration file.
• Then use run-client.ipynb to run the client and benchmark the performance
• See docker-compose.yml for the available services

Running servers

• Firstly, use export-classifier.ipynb to export various classifiers
• To start triton server, run the following command

docker compose up triton_server 

• To start the Tensorflow Serving server, run the following command

docker compose up tf_serving_server

Benchmark results

• Two types of architectures were tested and exported to the classifier Modules used by servers:

  • ResNet50
    • standard SavedModel
    • SavedModel compiled with XLA and AMP
  • EfficientNetB0
    • standard SavedModel
    • SavedModel compiled with XLA and AMP

• The benchmarks were performed on NVIDIA RTX A4000 GPU with 8GB of memory

• Each benchmark was run for 500 iterations to predict batch of 100 images of size 224x224 (50k images in total)

• I benchmarked only the images signature, which accepts images tensor of shape [batch, 224, 224, 3]

• When running the models locally with TF python API, I got following results:

Model	Architecture	Time [s]
ResNet50	SavedModel	57
ResNet50	SavedModel XLA/AMP	25
EfficientNetB0	SavedModel	52
EfficientNetB0	SavedModel XLA/AMP	13

• Running same benchmark but with Triton Inference Server (4 client threads, 1 server instance), I got following results:

Model	Architecture	Time [s]
ResNet50	SavedModel	73
ResNet50	SavedModel XLA/AMP	23
EfficientNetB0	SavedModel	54
EfficientNetB0	SavedModel XLA/AMP	17

• Running same benchmark but with Triton Inference Server (4 client threads, 2 server instances), I got following results:

Model	Architecture	Time [s]
ResNet50	SavedModel	76
ResNet50	SavedModel XLA/AMP	20
EfficientNetB0	SavedModel	54
EfficientNetB0	SavedModel XLA/AMP	13

• For Tensorflow Serving I was not able to text XLA/AMP models, I got following error when trying to serve them:

UNIMPLEMENTED: Could not find compiler for platform CUDA: NOT_FOUND: could not find registered compiler for platform CUDA

• The results for TF Serving were as follows (excluding XLA/AMP models):

Model	Architecture	Time [s]
ResNet50	SavedModel	59
ResNet50	SavedModel XLA/AMP	CUDA: NOT_FOUND error
EfficientNetB0	SavedModel	51
EfficientNetB0	SavedModel XLA/AMP	CUDA: NOT_FOUND error

Also, I noticed that when using TFServing, GPU memory was higher than when using Triton Inference Server. I was getting OOM when allocating tensor with shape[100,56,56,256] when using num_workers=10