kani-ext-vllm 0.3.1

vLLM backend for kani

kani-ext-vllm

This Kani extension repository adds 3 engines for using vLLM to deploy LLMs on local hardware.

vLLM is an LLM deployment platform optimized for GPU memory efficiency and throughput. This extension adds Kani engines to use vLLM engines in offline mode, manage a vLLM server, or connect to an existing vLLM server depending on the use case.

To install this package, you can install it from PyPI:

$ pip install kani-ext-vllm

Alternatively, you can install it using the git source:

$ pip install git+https://github.com/zhudotexe/kani-ext-vllm.git@main

See https://docs.vllm.ai/en/latest/index.html for more information on vLLM.

Usage

This package provides 3 main methods of serving models with vLLM:

• Offline mode
• vLLM-Native API mode
• OpenAI-Compatible API mode

These are generally equivalent, but offer slightly different options for each mode:

Mode	Communication	Multiple Parallel Models?	Prompt Template/Parsing	Best For
Offline	Local	No	kani	Low-level control over the model
vLLM API	HTTP	Yes	kani	Running multiple different models in parallel
OpenAI API	HTTP	Yes	vLLM	Fast iteration and testing multiple models; multimodal models

Offline Mode

from kani import Kani, chat_in_terminal
from kani.ext.vllm import VLLMEngine

engine = VLLMEngine(model_id="meta-llama/Meta-Llama-3-8B-Instruct")
ai = Kani(engine)
chat_in_terminal(ai)

vLLM-Native API Mode

The API mode can be used to connect to an existing running vLLM server or to start a managed vLLM server.

Connecting to a Running Server

from kani import Kani, chat_in_terminal
from kani.ext.vllm import VLLMServerEngine

engine = VLLMServerEngine(
    model_id="meta-llama/Meta-Llama-3-8B-Instruct",
    vllm_host="127.0.0.1",
    vllm_port=8000,
    use_managed_server=False,
)
ai = Kani(engine)
chat_in_terminal(ai)

Managed Server

[!NOTE] The vLLM server will be started on a random free port. It will not be exposed to the wider internet (i.e, it binds to localhost).

from kani import Kani, chat_in_terminal
from kani.ext.vllm import VLLMServerEngine

engine = VLLMServerEngine(model_id="meta-llama/Meta-Llama-3-8B-Instruct")
ai = Kani(engine)
chat_in_terminal(ai)

OpenAI-Compatible API Mode

Connecting to a Running Server

from kani import Kani, chat_in_terminal
from kani.ext.vllm import VLLMOpenAIEngine

engine = VLLMOpenAIEngine(
    model_id="meta-llama/Meta-Llama-3-8B-Instruct",
    vllm_host="127.0.0.1",
    vllm_port=8000,
    use_managed_server=False,
)
ai = Kani(engine)
chat_in_terminal(ai)

Managed Server

[!NOTE] The vLLM server will be started on a random free port. It will not be exposed to the wider internet (i.e, it binds to localhost).

from kani import Kani, chat_in_terminal
from kani.ext.vllm import VLLMOpenAIEngine

engine = VLLMOpenAIEngine(model_id="meta-llama/Meta-Llama-3-8B-Instruct")
ai = Kani(engine)
chat_in_terminal(ai)

Using Multiple GPUs

For multi-GPU support (probably needed), add model_load_kwargs={"tensor_parallel_size": 4}. Replace "4" with the number of GPUs you have available.

[!NOTE] If you are loading in an API mode, use vllm_args={"tensor_parallel_size": 4} instead.

Examples

Offline Mode

from kani.ext.vllm import VLLMEngine
from vllm import SamplingParams

model = VLLMEngine(
    model_id="mistralai/Mistral-Small-Instruct-2409",
    model_load_kwargs={"tensor_parallel_size": 2, "tokenizer_mode": "auto"},
    sampling_params=SamplingParams(temperature=0, max_tokens=2048),
)

vLLM-Native API Mode

from kani.ext.vllm import VLLMServerEngine

model = VLLMServerEngine(
    model_id="mistralai/Mistral-Small-Instruct-2409",
    vllm_args={"tensor_parallel_size": 2, "tokenizer_mode": "auto"},
    # note that these should not be wrapped in SamplingParams!
    temperature=0,
    max_tokens=2048,
)

See https://docs.vllm.ai/en/stable/serving/openai_compatible_server.html#completions-api_1 for a list of valid decoding parameters that can be specified in the engine constructor.

See https://docs.vllm.ai/en/stable/cli/serve/ for a list of valid arguments to vllm_args.

OpenAI-Compatible API Mode

from kani.ext.vllm import VLLMOpenAIEngine

model = VLLMOpenAIEngine(
    model_id="Qwen/Qwen3-Omni-30B-A3B-Instruct",
    vllm_args={"tensor_parallel_size": 2, "allowed_local_media_path": "/"},
    # note that these should not be wrapped in SamplingParams!
    temperature=0,
    max_tokens=2048,
)

See https://docs.vllm.ai/en/stable/serving/openai_compatible_server.html#chat-api_1 for a list of valid decoding parameters that can be specified in the engine constructor.

See https://docs.vllm.ai/en/stable/cli/serve/ for a list of valid arguments to vllm_args.