DeepRec Serving is a high-performance serving system for DeepRec based on TensorFlow Serving. DeepRec Serving could highly improve performance and cpu/gpu utilization in inference, such as SessionGroup, CUDA multi-stream, etc.

Few features in DeepRec Serving:

• Support SessionGroup which is shared-variable (only variables shared) architecture for mutliple session in serving process.
• Support CUDA Multiple Stream, could highly improve QPS and GPU Utilization in GPU Inference.

CPU Dev Docker

GPU(cuda11.6) Dev Docker

Develop Branch: master, Latest Release Branch: deeprec2302

Build Package Builder-CPU

bazel build -c opt tensorflow_serving/...

Build CPU Package Builder with OneDNN + Eigen Threadpool

bazel build  -c opt --config=mkl_threadpool --define build_with_mkl_dnn_v1_only=true tensorflow_serving/...

Build Package Builder-GPU

bazel build -c opt --config=cuda tensorflow_serving/...

Build Package

bazel-bin/tensorflow_serving/tools/pip_package/build_pip_package /tmp/tf_serving_client_whl

Server Bin

Server Bin would generated in following directory:

bazel-bin/tensorflow_serving/model_servers/tensorflow_model_server

• SessionGroup
• CUDA MultiStream
• Device Placement Optimization