triton-grpc-proxy-rs

Proxy server for triton gRPC server that inferences embedding model in Rust.

• it refines the request and response formats of the Triton server.
• without tritonclient dependency.
• fast & easy to use.

Build

1. Convert the embedding model to onnx

• BAAI/bge-m3 is used for an example.
• It'll convert Pytorch into onnx model with the cls pooling + l2 normalization layers, and save it to ./model_repository/embedding/1/model.onnx.
  • if you don't want to add the pooling + l2 normalization layers, then need to change the config.pbtxt properly.
• Currently, max_batch_size is limited to 256 due to OOM. You can change this value to fit your environment.

python3 convert.py

2. Run docker-compose

• I'll run both Triton inference server and the proxy server.
• You need to edit the absolute path of the volume (where pointed to the ./model_repository) in docker-compose.yml.

make run-docker-compose

Build & run a proxy server only

• You can also build and run a triton proxy server with the below command.

export RUSTFLAGS="-C target-cpu=native"
make server

make build-docker

Build & run triton inference server only

docker run --gpus all --rm --ipc=host --shm-size=8g --ulimit memlock=-1 --ulimit stack=67108864 -p8000:8000 -p8001:8001 -p8002:8002 -v$(pwd)triton-grpc-proxy-rs/model_repository:/models nvcr.io/nvidia/tritonserver:24.07-py3 bash -c "LD_PRELOAD=/usr/lib/$(uname -m)-linux-gnu/libtcmalloc.so.4:${LD_PRELOAD} && pip install transformers tokenizers && tritonserver --model-repository=/models"

Architecture

1. recieve request(s) from the user.
   • list of text (String) in this case.
2. request the Triton gRPC server to get embeddings.
3. post-process (cast and reshape) the embeddings and returns to the users.

API Specs

• Swagger : http://localhost:8080/explorer/

Configs

• parse configuration from the env variables.

• SERVER_PORT: proxy server port. default 8080.

• TRITON_SERVER_URL: triton inference gRPC server url. default http://triton-server.

• TRITON_SERVER_GRPC_PORT: triton inference gRPC server port. default 8001.

• MODEL_VERSION: model version. default 1.

• MODEL_NAME: model name. default model.

• INPUT_NAME: input name. default text.

• OUTPUT_NAME: output name. default embedding.

• EMBEDDING_SIZE: size of the embedding. default 1024.

health

• GET /health

curl -i http://127.0.0.1:8080/health

HTTP/1.1 200 OK
content-length: 2
date: Sun, 08 Oct 2023 06:33:53 GMT

ok

metrics

get prometheus metrics

• GET /metrics

curl -i http://127.0.0.1:8080/metrics

embedding

• POST /v1/embedding
• Request Body : [{'query': 'input'}, ... ]

curl -H "Content-type:application/json" -X POST http://127.0.0.1:8080/v1/embedding -d "[{\"query\": \"asdf\"}, {\"query\": \"asdf asdf\"}, {\"query\": \"asdf asdf asdf\"}, {\"query\": \"asdf asdf asdf asdf\"}]"

• Response Body : [{'embedding': '1024 f32 vector'}, ...]

[{"embedding": [-0.8067292,-0.004603,-0.24123234,0.59398544,-0.5583446,...]}, ...]

Benchmark

• Environment
  • CPU : i7-7700K (not overclocked)
  • GPU : GTX 1060 6 GB
  • Rust : v1.79.0 stable
  • Triton Server : 24-07-py3
    • backend : onnxruntime-gpu
    • allocator : tcmalloc
  • model : BAAI/bge-m3 w/ fp32
• payload : [{'query': 'asdf' * 126}] * batch_size (asdf * 126 == 255 tokens)
• stages
  • model : only triton gRPC server latency (preprocess + tokenize + model)
  • processing : end-to-end latency (service-side)
    • json de/serialization
    • payload serialization (byte string, float vector)
    • cast & reshape 2d vectors

batch size	model (p90)	processing (p90)
8	1428.20 ms	0.044 ms
16	2915.01 ms	0.051 ms
32	5626.15 ms	0.055 ms

To-Do

• optimize the processing performance and memory usage
• support /metrics endpoint to get prometheus metrics
• OpenAPI specifications and Swagger UI
• add Dockerfile and docker-compose to easily deploy the servers
• triton inference server
  • add model converter script.
  • configurations
• move hard-coded configs to env
• optimize the proxy server performance
• README
• move tokenizer part from triton server into proxy-server

Maintainer

@kozistr