OpenAI-Compatible Frontend for Triton Inference Server (Beta)#

[!NOTE] The OpenAI-Compatible API is currently in BETA. Its features and functionality are subject to change as we collect feedback. We’re excited to hear any thoughts you have and what features you’d like to see!

Pre-requisites#

Docker + NVIDIA Container Runtime
A correctly configured HF_TOKEN for access to HuggingFace models.
- The current examples and testing primarily use the meta-llama/Meta-Llama-3.1-8B-Instruct model, but you can manually bring your own models and adjust accordingly.

VLLM#

Launch the container and install dependencies:

Mounts the ~/.huggingface/cache for re-use of downloaded models across runs, containers, etc.
Sets the HF_TOKEN environment variable to access gated models, make sure this is set in your local environment if needed.

docker run -it --net=host --gpus all --rm \
  -v ${HOME}/.cache/huggingface:/root/.cache/huggingface \
  -e HF_TOKEN \
  nvcr.io/nvidia/tritonserver:25.11-vllm-python-py3

Launch the OpenAI-compatible Triton Inference Server:

cd /opt/tritonserver/python/openai

# NOTE: Adjust the --tokenizer based on the model being used
python3 openai_frontend/main.py --model-repository tests/vllm_models --tokenizer meta-llama/Meta-Llama-3.1-8B-Instruct

Example output

...
+-----------------------+---------+--------+
| Model                 | Version | Status |
+-----------------------+---------+--------+
| llama-3.1-8b-instruct | 1       | READY  | <- Correct Model Loaded in Triton
+-----------------------+---------+--------+
...
Found model: name='llama-3.1-8b-instruct', backend='vllm'
[WARNING] Adding CORS for the following origins: ['http://localhost']
INFO:     Started server process [126]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:9000 (Press CTRL+C to quit) <- OpenAI Frontend Started Successfully

Send a /v1/chat/completions request:

Note the use of jq is optional, but provides a nicely formatted output for JSON responses.

MODEL="llama-3.1-8b-instruct"
curl -s http://localhost:9000/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "'${MODEL}'",
  "messages": [{"role": "user", "content": "Say this is a test!"}]
}' | jq

Example output

{
  "id": "cmpl-0242093d-51ae-11f0-b339-e7480668bfbe",
  "choices": [
    {
      "finish_reason": "stop",
      "index": 0,
      "message":
      {
        "content": "This is only a test.",
        "tool_calls": null,
        "role": "assistant",
        "function_call": null
      },
      "logprobs": null
    }
  ],
  "created": 1750846825,
  "model": "llama-3.1-8b-instruct",
  "system_fingerprint": null,
  "object": "chat.completion",
  "usage": {
    "completion_tokens": 7,
    "prompt_tokens": 42,
    "total_tokens": 49
  }
}

Send a /v1/completions request:

Note the use of jq is optional, but provides a nicely formatted output for JSON responses.

MODEL="llama-3.1-8b-instruct"
curl -s http://localhost:9000/v1/completions -H 'Content-Type: application/json' -d '{
  "model": "'${MODEL}'",
  "prompt": "Machine learning is"
}' | jq

Example output

{
  "id": "cmpl-58fba3a0-51ae-11f0-859d-e7480668bfbe",
  "choices": [
    {
      "finish_reason": "stop",
      "index": 0,
      "logprobs": null,
      "text": " an amazing field that can truly understand the hidden patterns that exist in the data,"
    }
  ],
  "created": 1750846970,
  "model": "llama-3.1-8b-instruct",
  "system_fingerprint": null,
  "object": "text_completion",
  "usage": {
    "completion_tokens": 16,
    "prompt_tokens": 4,
    "total_tokens": 20
  }
}

Benchmark with genai-perf:

To install genai-perf in this container, see the instructions here
Or try using genai-perf from the SDK container

MODEL="llama-3.1-8b-instruct"
TOKENIZER="meta-llama/Meta-Llama-3.1-8B-Instruct"
genai-perf profile \
  --model ${MODEL} \
  --tokenizer ${TOKENIZER} \
  --service-kind openai \
  --endpoint-type chat \
  --url localhost:9000 \
  --streaming

Example output

2024-10-14 22:43 [INFO] genai_perf.parser:82 - Profiling these models: llama-3.1-8b-instruct
2024-10-14 22:43 [INFO] genai_perf.wrapper:163 - Running Perf Analyzer : 'perf_analyzer -m llama-3.1-8b-instruct --async --input-data artifacts/llama-3.1-8b-instruct-openai-chat-concurrency1/inputs.json -i http --concurrency-range 1 --endpoint v1/chat/completions --service-kind openai -u localhost:9000 --measurement-interval 10000 --stability-percentage 999 --profile-export-file artifacts/llama-3.1-8b-instruct-openai-chat-concurrency1/profile_export.json'
                              NVIDIA GenAI-Perf | LLM Metrics
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━━┳━━━━━━━━┳━━━━━━━━┳━━━━━━━━┳━━━━━━━━┓
┃                         Statistic ┃    avg ┃    min ┃    max ┃    p99 ┃    p90 ┃    p75 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━━╇━━━━━━━━╇━━━━━━━━╇━━━━━━━━╇━━━━━━━━┩
│          Time to first token (ms) │  71.66 │  64.32 │  86.52 │  76.13 │  74.92 │  73.26 │
│          Inter token latency (ms) │  18.47 │  18.25 │  18.72 │  18.67 │  18.61 │  18.53 │
│              Request latency (ms) │ 348.00 │ 274.60 │ 362.27 │ 355.41 │ 352.29 │ 350.66 │
│            Output sequence length │  15.96 │  12.00 │  16.00 │  16.00 │  16.00 │  16.00 │
│             Input sequence length │ 549.66 │ 548.00 │ 551.00 │ 550.00 │ 550.00 │ 550.00 │
│ Output token throughput (per sec) │  45.84 │    N/A │    N/A │    N/A │    N/A │    N/A │
│      Request throughput (per sec) │   2.87 │    N/A │    N/A │    N/A │    N/A │    N/A │
└───────────────────────────────────┴────────┴────────┴────────┴────────┴────────┴────────┘
2024-10-14 22:44 [INFO] genai_perf.export_data.json_exporter:62 - Generating artifacts/llama-3.1-8b-instruct-openai-chat-concurrency1/profile_export_genai_perf.json
2024-10-14 22:44 [INFO] genai_perf.export_data.csv_exporter:71 - Generating artifacts/llama-3.1-8b-instruct-openai-chat-concurrency1/profile_export_genai_perf.csv

Use the OpenAI python client directly:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:9000/v1",
    api_key="EMPTY",
)

model = "llama-3.1-8b-instruct"
completion = client.chat.completions.create(
    model=model,
    messages=[
        {
            "role": "system",
            "content": "You are a helpful assistant.",
        },
        {"role": "user", "content": "What are LLMs?"},
    ],
    max_completion_tokens=256,
)

print(completion.choices[0].message.content)

Run tests (NOTE: The server should not be running, the tests will handle starting/stopping the server as necessary):

cd /opt/tritonserver/python/openai/
pip install -r requirements-test.txt

pytest -v tests/

LoRA Adapters#

If the command line argument --lora-separator=<separator_string> is provided when starting the OpenAI Frontend, a vLLM LoRA adaptor listed on the multi_lora.json may be selected by appending the LoRA name to the model name, separated by the LoRA separator, on the inference request in <model_name><separator_string><lora_name> format.

For example

# start server with model named gemma-2b
python3 openai_frontend/main.py --lora-separator=_lora_ ...

# inference without LoRA
curl -s http://localhost:9000/v1/completions -H 'Content-Type: application/json' -d '{
  "model": "gemma-2b",
  "temperature": 0,
  "prompt": "When was the wheel invented?"
}'
{
  ...
  "choices":[{..."text":"\n\nThe wheel was invented by the Sumerians in Mesopotamia around 350"}],
  ...
}

# inference with LoRA named doll
curl -s http://localhost:9000/v1/completions -H 'Content-Type: application/json' -d '{
  "model": "gemma-2b_lora_doll",
  "temperature": 0,
  "prompt": "When was the wheel invented?"
}'
{
  ...
  "choices":[{..."text":"\n\nThe wheel was invented in Mesopotamia around 3500 BC.\n\n"}],
  ...
}

# inference with LoRA named sheep
curl -s http://localhost:9000/v1/completions -H 'Content-Type: application/json' -d '{
  "model": "gemma-2b_lora_sheep",
  "temperature": 0,
  "prompt": "When was the wheel invented?"
}'
{
  ...
  "choices":[{..."text":"\n\nThe wheel was invented around 3000 BC in Mesopotamia.\n\n"}],
  ...
}

When listing or retrieving model(s), the model id will include the LoRA name in the same <model_name><separator_string><lora_name> format for each LoRA adapter listed on the multi_lora.json. Note: The LoRA name inclusion is limited to locally stored models, inference requests are not limited though.

See the vLLM documentation on how to serve a model with LoRA adapters.

Embedding Models#

Currently, OpenAI-Compatible Frontend supports loading embedding models and embeddings endpoints via vLLM backend. Check vLLM supported models for all supported embedding models from vLLM.

Launch the container and install dependencies:

Mounts the ~/.huggingface/cache for re-use of downloaded models across runs, containers, etc.
Sets the HF_TOKEN environment variable to access gated models, make sure this is set in your local environment if needed.

docker run -it --net=host --gpus all --rm \
  -v ${HOME}/.cache/huggingface:/root/.cache/huggingface \
  -e HF_TOKEN \
  nvcr.io/nvidia/tritonserver:25.11-vllm-python-py3

Launch the OpenAI-compatible Triton Inference Server:

cd /opt/tritonserver/python/openai

# NOTE: Embeddings endpoint does not require "--tokenizer"
python3 openai_frontend/main.py --model-repository tests/vllm_embedding_models

Example output

...
+------------------+---------+--------+
| Model            | Version | Status |
+------------------+---------+--------+
| all-MiniLM-L6-v2 | 1       | READY  | <- Correct Model Loaded in Triton
+------------------+---------+--------+
...
Found model: name='all-MiniLM-L6-v2', backend='vllm'
[WARNING] Adding CORS for the following origins: ['http://localhost']
INFO:     Started server process [133]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:9000 (Press CTRL+C to quit) <- OpenAI Frontend Started Successfully

Send a /v1/embeddings request:

Note the use of jq is optional, but provides a nicely formatted output for JSON responses.

MODEL="all-MiniLM-L6-v2"
curl -s http://localhost:9000/v1/embeddings \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "'${MODEL}'",
    "input": "The food was delicious and the waiter...",
    "dimensions": 10,
    "encoding_format": "float"
  }' | jq

Example output

{
  "object": "list",
  "data": [
    {
      "object": "embedding",
      "embedding": [
        -0.1914404183626175,
        0.4000193178653717,
        0.058502197265625,
        0.18909454345703125,
        -0.4690297544002533,
        0.004936536308377981,
        0.45893096923828125,
        -0.31141534447669983,
        0.18299102783203125,
        -0.4907582700252533
      ],
      "index": 0
    }
  ],
  "model": "all-MiniLM-L6-v2",
  "usage": {
    "prompt_tokens": 12,
    "total_tokens": 12
  }
}

TensorRT-LLM#

Prepare your model repository for a TensorRT-LLM model, build the engine, etc. You can try any of the following options:

Launch the container:

Mounts the ~/.huggingface/cache for re-use of downloaded models across runs, containers, etc.
Sets the HF_TOKEN environment variable to access gated models, make sure this is set in your local environment if needed.

docker run -it --net=host --gpus all --rm \
  -v ${HOME}/.cache/huggingface:/root/.cache/huggingface \
  -e HF_TOKEN \
  -e TRTLLM_ORCHESTRATOR=1 \
  nvcr.io/nvidia/tritonserver:24.11-trtllm-python-py3

Install dependencies inside the container:

# Install python bindings for tritonserver and tritonfrontend
pip install /opt/tritonserver/python/triton*.whl

# Install application requirements
git clone https://github.com/triton-inference-server/server.git
cd server/python/openai/
pip install -r requirements.txt

Launch the OpenAI server:

# NOTE: Adjust the --tokenizer based on the model being used
python3 openai_frontend/main.py --model-repository path/to/models --tokenizer meta-llama/Meta-Llama-3.1-8B-Instruct

Send a /v1/chat/completions request:

Note the use of jq is optional, but provides a nicely formatted output for JSON responses.

# MODEL should be the client-facing model name in your model repository for a pipeline like TRT-LLM.
# For example, this could also be "ensemble", or something like "gpt2" if generated from Triton CLI
MODEL="tensorrt_llm_bls"
curl -s http://localhost:9000/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "'${MODEL}'",
  "messages": [{"role": "user", "content": "Say this is a test!"}]
}' | jq

Example output

{
  "id": "cmpl-5ad4f860-bf13-11f0-b137-b75b7f0a8586",
  "choices": [
    {
      "finish_reason": "stop",
      "index": 0,
      "message": {
        "content": "It looks like you're ready to see if I'm functioning properly. What would",
        "tool_calls": null,
        "role": "assistant",
        "function_call": null
      },
      "logprobs": null
    }
  ],
  "created": 1762875029,
  "model": "tensorrt_llm_bls",
  "system_fingerprint": null,
  "object": "chat.completion",
  "usage": {
    "prompt_tokens": 42,
    "total_tokens": 58,
    "completion_tokens": 16
  }
}

The other examples should be the same as vLLM, except that you should set MODEL="tensorrt_llm_bls" or MODEL="ensemble", everywhere applicable as seen in the example request above.

KServe Frontends#

To support serving requests through both the OpenAI-Compatible and KServe Predict v2 frontends to the same running Triton Inference Server, the tritonfrontend python bindings are included for optional use in this application as well.

You can opt-in to including these additional frontends, assuming tritonfrontend is installed, with --enable-kserve-frontends like below:

python3 openai_frontend/main.py \
  --model-repository tests/vllm_models \
  --tokenizer meta-llama/Meta-Llama-3.1-8B-Instruct \
  --enable-kserve-frontends

See python3 openai_frontend/main.py --help for more information on the available arguments and default values.

For more information on the tritonfrontend python bindings, see the docs here.

Model Parallelism Support#

[x] vLLM (EngineArgs)
- ex: Configure tensor_parallel_size: 2 in the model.json
[x] TensorRT-LLM (Orchestrator Mode)
- Set the following environment variable: export TRTLLM_ORCHESTRATOR=1
[ ] TensorRT-LLM (Leader Mode)
- Not currently supported

Tool Calling#

The OpenAI frontend supports tools and tool_choice in the v1/chat/completions API. Please refer to the OpenAI API reference for more details about these parameters: tools, tool_choice

To enable the tool-calling feature, add the --tool-call-parser {parser_name} flag when starting the server. The two available parsers are llama3 and mistral. The llama3 parser supports tool-calling features for LLaMA 3.1, 3.2, and 3.3 models, while the mistral parser supports tool-calling features for the Mistral Instruct model.

Example for launching the OpenAI frontend with a tool call parser:

python3 openai_frontend/main.py \
  --model-repository tests/vllm_models \
  --tokenizer meta-llama/Meta-Llama-3.1-8B-Instruct \
  --tool-call-parser llama3

Example for making a tool calling request:

import json
from openai import OpenAI


def get_current_weather(city: str, state: str, unit: "str"):
    return (
        "The weather in Dallas, Texas is 85 degrees fahrenheit. It is "
        "partly cloudly, with highs in the 90's."
    )

available_tools = {"get_current_weather": get_current_weather}

openai_api_key = "EMPTY"
openai_api_base = "http://localhost:9000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

model = "llama-3.1-8b-instruct" # change this to the model in the repository

tools = [
    {
        "type": "function",
        "function": {
            "name": "get_current_weather",
            "description": "Get the current weather in a given location",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {
                        "type": "string",
                        "description": "The city to find the weather for, e.g. 'San Francisco'",
                    },
                    "state": {
                        "type": "string",
                        "description": "the two-letter abbreviation for the state that the city is"
                        " in, e.g. 'CA' which would mean 'California'",
                    },
                    "unit": {
                        "type": "string",
                        "description": "The unit to fetch the temperature in",
                        "enum": ["celsius", "fahrenheit"],
                    },
                },
                "required": ["city", "state", "unit"],
            },
        },
    }
]

messages = [
    {
        "role": "system",
        "content": "You're a helpful assistant! Answer the users question best you can.",
    },
    {"role": "user", "content": "What is the weather in Dallas, Texas in Fahrenheit?"},
]

tool_calls = client.chat.completions.create(
    messages=messages, model=model, tools=tools, max_completion_tokens=128
)
function_name = tool_calls.choices[0].message.tool_calls[0].function.name
function_arguments = tool_calls.choices[0].message.tool_calls[0].function.arguments

print(f"function name: " f"{function_name}")
print(f"function arguments: {function_arguments}")
print(f"tool calling result: {available_tools[function_name](https://github.com/triton-inference-server/server/blob/main/docs/client_guide/**json.loads(function_arguments))}")

Example output:

function name: get_current_weather
function arguments: {"city": "Dallas", "state": "TX", "unit": "fahrenheit"}
tool calling result: The weather in Dallas, Texas is 85 degrees fahrenheit. It is partly cloudly, with highs in the 90's.

Named Tool Calling#

The OpenAI frontend supports named function calling, utilizing structured outputs in the vLLM backend and guided decoding in TensorRT-LLM backend. Users can specify one of the tools in tool_choice to force the model to select a specific tool for function calling.

[!NOTE] For instructions on enabling guided decoding in the TensorRT-LLM backend, please refer to this guide

Example for making a named tool calling request:

import json
from openai import OpenAI


def get_current_weather(city: str, state: str, unit: "str"):
    return (
        "The weather in Dallas, Texas is 85 degrees fahrenheit. It is "
        "partly cloudly, with highs in the 90's."
    )

def get_n_day_weather_forecast(city: str, state: str, unit: str, num_days: int):
    return (
        f"The weather in Dallas, Texas is 85 degrees fahrenheit in next {num_days} days."
    )

available_tools = {"get_current_weather": get_current_weather,
                  "get_n_day_weather_forecast": get_n_day_weather_forecast}

openai_api_key = "EMPTY"
openai_api_base = "http://localhost:9000/v1"
client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)
model = "llama-3.1-8b-instruct" # change this to the model in the repository
tools = [
    {
        "type": "function",
        "function": {
            "name": "get_current_weather",
            "description": "Get the current weather in a given location",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {
                        "type": "string",
                        "description": "The city to find the weather for, e.g. 'San Francisco'",
                    },
                    "state": {
                        "type": "string",
                        "description": "the two-letter abbreviation for the state that the city is"
                        " in, e.g. 'CA' which would mean 'California'",
                    },
                    "unit": {
                        "type": "string",
                        "description": "The unit to fetch the temperature in",
                        "enum": ["celsius", "fahrenheit"],
                    },
                },
                "required": ["city", "state", "unit"],
            },
        },
    },
    {
        "type": "function",
        "function": {
            "name": "get_n_day_weather_forecast",
            "description": "Get an N-day weather forecast",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {
                        "type": "string",
                        "description": "The city to find the weather for, "
                        "e.g. 'San Francisco'",
                    },
                    "state": {
                        "type": "string",
                        "description": "must the two-letter abbreviation for the state "
                        "that the city is in, e.g. 'CA' which would "
                        "mean 'California'",
                    },
                    "unit": {
                        "type": "string",
                        "description": "The unit to fetch the temperature in",
                        "enum": ["celsius", "fahrenheit"],
                    },
                    "num_days": {
                        "type": "integer",
                        "description": "The number of days to forecast",
                    },
                },
                "required": ["city", "state", "unit", "num_days"],
            },
        },
     }
]

tool_choice = {"function": {"name": "get_n_day_weather_forecast"}, "type": "function"}

messages = [
    {
        "role": "system",
        "content": "You're a helpful assistant! Answer the users question best you can.",
    },
    {"role": "user", "content": "What is the weather in Dallas, Texas in Fahrenheit?"},
]

tool_calls = client.chat.completions.create(
    messages=messages, model=model, tools=tools, tool_choice=tool_choice, max_completion_tokens=128
)
function_name = tool_calls.choices[0].message.tool_calls[0].function.name
function_arguments = tool_calls.choices[0].message.tool_calls[0].function.arguments

print(f"function name: {function_name}")
print(f"function arguments: {function_arguments}")
print(f"tool calling result: {available_tools[function_name](https://github.com/triton-inference-server/server/blob/main/docs/client_guide/**json.loads(function_arguments))}")

Example output:

function name: get_n_day_weather_forecast
function arguments: {"city": "Dallas", "state": "TX", "unit": "fahrenheit", num_days: 1}
tool calling result: The weather in Dallas, Texas is 85 degrees fahrenheit in next 1 days.

Limit Endpoint Access#

The OpenAI-compatible server supports restricting access to specific API endpoints through authentication headers. This feature allows you to protect sensitive endpoints while keeping others publicly accessible.

Configuration#

Use the --openai-restricted-api command-line argument to configure endpoint restrictions:

--openai-restricted-api <API_1>,<API_2>,... <restricted-key> <restricted-value>

API: A comma-separated list of APIs to be included in this group. Note that currently a given API is not allowed to be included in multiple groups. The following protocols / APIs are recognized:
- inference: Chat completions and text completions endpoints
  - POST /v1/chat/completions
  - POST /v1/completions
- embedding: Embedding endpoint
  - POST /v1/embeddings
- model-repository: Model listing and information endpoints
  - GET /v1/models
  - GET /v1/models/{model_name}
- metrics: Server metrics endpoint
  - GET /metrics
- health: Health check endpoint
  - GET /health/ready
restricted-key: The HTTP request header to be checked when a request is received.
restricted-value: The header value required to access the specified protocols.

Examples#

Restrict Inference API Endpoints Only#

--openai-restricted-api "inference api-key my-secret-key"

Clients must include the header:

curl -H "api-key: my-secret-key" \
     -X POST http://localhost:9000/v1/chat/completions \
     -d '{"model": "my-model", "messages": [{"role": "user", "content": "Hello"}]}'

Restrict Multiple API Endpoints#

# Different authentication for different APIs
--openai-restricted-api "inference user-key user-secret" \
--openai-restricted-api "model-repository admin-key admin-secret"

# Multiple APIs in single argument with shared authentication
--openai-restricted-api "inference,model-repository shared-key shared-secret"