openai_trtllm - OpenAI-compatible API for TensorRT-LLM

Provide TensorRT-LLM and NVIDIA Triton Inference Server with an OpenAI-compatible API. This allows you to integrate with langchain

Quick overview

Get started

Follow the tensorrtllm_backend tutorial to build your TensorRT engine, and launch a triton server. We provide an Baichuan example below to follow.

You need to clone the repository with dependencies to build the project.

git clone --recursive https://github.com/npuichigo/openai_trtllm.git

Build with Docker

Make sure you have Docker and Docker Compose installed.

docker compose up --build

Build locally

cargo run --release

The parameters can be set with environment variables or command line arguments:

./target/release/openai_trtllm --help
Usage: openai_trtllm [OPTIONS]

Options:
  -H, --host <HOST>                        Host to bind to [default: 0.0.0.0]
  -p, --port <PORT>                        Port to bind to [default: 3000]
  -t, --triton-endpoint <TRITON_ENDPOINT>  [default: http://localhost:16001]
  -o, --otlp-endpoint <OTLP_ENDPOINT>      Endpoint of OpenTelemetry collector
  -h, --help                               Print help

Example

We provide a model template in models/Baichuan to let you follow. Since we're unknown of your hardware, we don't provide a pre-built TensorRT engine. You need to follow the steps below to build your own engine.

Download the Baichuan model from HuggingFace.

# Make sure you have git-lfs installed (https://git-lfs.com)
git lfs install
git clone https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat models/download/Baichuan2-13B-Chat

We provide a pre-built docker which is slightly newer than v0.6.1. You are free to test on other versions.

docker run --rm -it --shm-size=2g --ulimit memlock=-1 --ulimit stack=67108864 --gpus=all -v /models:/models npuichigo/tritonserver-trtllm:711a28d bash

Follow the tutorial here to build your engine.

# int8 for example [with inflight batching]
python /app/tensorrt_llm/examples/baichuan/build.py \
  --model_version v2_13b \
  --model_dir /models/download/Baichuan2-13B-Chat \
  --output_dir /models/baichuan/tensorrt_llm/1 \
  --max_input_len 4096 \
  --max_output_len 1024 \
  --dtype float16 \
  --use_gpt_attention_plugin float16 \
  --use_gemm_plugin float16 \
  --enable_context_fmha \
  --use_weight_only \
  --use_inflight_batching

After the build, the engine will be saved to /models/baichuan/tensorrt_llm/1/ to be used by Triton.

Make sure the models/baichuan/preprocessing/config.pbtxt and models/baichuan/postprocessing/config.pbtxt refer to the correct tokenizer directory. For example:
```
parameters {
  key: "tokenizer_dir"
  value: {
    string_value: "/models/download/Baichuan2-13B-Chat"
  }
}
```
Go ahead to launch the server, better with docker-compose.

Tracing

We are tracing performance metrics using tracing, tracing-opentelemetry and opentelemetry-otlp crates.

Here is an example of tracing with Tempo on a k8s cluster:

Let's say you are running a Jaeger instance locally, you can run the following command to start it:

docker run --rm --name jaeger \
  -p 6831:6831/udp \
  -p 6832:6832/udp \
  -p 5778:5778 \
  -p 16686:16686 \
  -p 4317:4317 \
  -p 4318:4318 \
  -p 14250:14250 \
  -p 14268:14268 \
  -p 14269:14269 \
  -p 9411:9411 \
  jaegertracing/all-in-one:1.51

To enable tracing, set the OTLP_ENDPOINT environment variable or --otlp-endpoint command line argument to the endpoint of your OpenTelemetry collector.

OPENAI_TRTLLM_OTLP_ENDPOINT=http://localhost:4317 cargo run --release

References

cria

janhq / openai_trtllm