modis-client 0.1.4

Python client for MoDIS Service Nodes

MoDIS Python Client

modis-client is the Python client package for MoDIS Service Nodes (MSNs). It exposes MoDIS-native request and response objects for direct Python callers.

The PyPI distribution name is modis-client. pymodis and modis are already used by unrelated packages, while modis-client currently appears available.

Install

From a checkout:

python -m pip install -e .

The import package is modis_client:

from modis_client import SingleNodeClient

with SingleNodeClient(base_url="http://localhost:8000") as client:
    response = client.chat_completion({
        "model": "qwen3.6-27b",
        "messages": [{"role": "user", "content": "Hello MoDIS."}],
    })

print(response.content.content)

The package supports sync and async single-node clients, routing clients, batch requests, explicit streaming methods, MoDIS tool-call fields, and Pydantic v2 models with MoDIS wire aliases.

The default client timeout is 900 seconds because a MoDIS request may include a cold model load before inference begins. Pass timeout= to any client constructor when a shorter or longer budget is required.

Async Client

from modis_client import AsyncSingleNodeClient

async with AsyncSingleNodeClient(base_url="http://localhost:8000") as client:
    response = await client.responses({
        "model": "qwen3.6-27b",
        "input": "Summarize MoDIS in one sentence.",
    })

Batch Requests

responses = client.chat_completion([
    {"model": "qwen3.6-27b", "messages": [{"role": "user", "content": "one"}]},
    {"model": "qwen3.6-27b", "messages": [{"role": "user", "content": "two"}]},
])

Streaming

for event in client.chat_completion_stream({
    "model": "qwen3.6-27b",
    "messages": [{"role": "user", "content": "Count to three."}],
}):
    if event.event == "delta":
        print(event.payload)

Image streams preserve progress and final image payload events:

for event in client.image_generation_stream({
    "model": "qwen-image",
    "prompt": "A clean product photo of a compact AI workstation.",
    "numOutputs": 1,
}):
    if event.event == "progress":
        print(event.payload["step"], event.payload.get("total"))
    if event.event == "completed":
        images = event.payload["images"]

Text To Speech

Non-streaming speech requests return a Pydantic response with audio_b64, audio_url, and mime_type:

import base64

response = client.text_to_speech({
    "model": "qwen-tts-custom-voice",
    "text": "Hello from MoDIS.",
    "speaker": "Vivian",
    "responseFormat": "wav",
})

audio_bytes = base64.b64decode(response.audio_b64)

PCM streaming is exposed separately because /audio/speech returns raw audio bytes, not SSE events, when stream: true is used:

with open("speech.pcm", "wb") as output:
    for chunk in client.text_to_speech_pcm_stream({
        "model": "qwen-tts-custom-voice",
        "text": "Hello from MoDIS.",
        "speaker": "Vivian",
    }):
        output.write(chunk)

Reference voices can be passed per request with referenceAudio and referenceText. Use hosted URLs when the vLLM-Omni server can fetch the audio, or inline data URLs/base64 when the reference is not hosted.

Speech To Text

Non-streaming ASR requests call /audio/transcriptions and return a transcript in response.text:

response = client.speech_to_text({
    "model": "qwen3-asr",
    "audioUrl": "https://example.com/audio.wav",
    "language": "en",
})

print(response.text)

The typed SpeechToTextRequest model also accepts audioB64, audioPath, audio, url, prompt, responseFormat, and timeoutSeconds. The current MoDIS ASR route is non-streaming; uploaded-audio output streaming and realtime WebSocket transcription are service follow-ups.

Structured Output

response = client.chat_completion({
    "model": "qwen3.6-27b",
    "messages": [{"role": "user", "content": "Return one rating."}],
    "format": {
        "title": "Rating",
        "type": "object",
        "properties": {"score": {"type": "integer"}},
        "required": ["score"],
    },
})

format: "json" requests JSON-object output. A JSON Schema object is forwarded to MoDIS text runtimes as a structured response schema.

Reasoning Output

response = client.chat_completion({
    "model": "qwen3.6-27b",
    "messages": [{"role": "user", "content": "Solve 25^(1/2)."}],
    "includeReasoning": True,
})

print(response.content.reasoning)
print(response.content.content)

includeReasoning asks MoDIS to include parsed runtime reasoning when the text backend exposes it separately from the final assistant message.

Tool Calls

response = client.chat_completion({
    "model": "qwen3.6-27b",
    "messages": [{"role": "user", "content": "What time is it?"}],
    "tools": [
        {
            "type": "function",
            "function": {
                "name": "get_time",
                "description": "Return the current local time.",
                "parameters": {"type": "object", "properties": {}},
            },
        }
    ],
    "toolChoice": "auto",
    "parallelToolCalls": True,
})

Routing

from modis_client import RoutingClient

client = RoutingClient(
    nodes=[
        "http://msn-a:8000",
        {"base_url": "http://msn-b:8000", "priority": 5, "node_id": "msn-b"},
    ],
    strategy="balanced",
)

Pydantic Models

Responses are returned as Pydantic models by default. Use to_wire() or model_dump(by_alias=True, exclude_none=True) when you need the MoDIS JSON shape:

wire_payload = response.to_wire()

ModisClientError is raised for transport failures, typed MoDIS errors, and stream protocol errors.

Development

python -m pip install -e ".[dev]"
pytest
ruff check .
ruff format --check .
mypy src
python -m build

Live Service Tests

Live tests are opt-in and require a running modis-service or compatible MoDIS endpoint. They are skipped unless MODIS_LIVE_SERVICE_URL is set.

MODIS_LIVE_SERVICE_URL=http://localhost:8000 \
MODIS_LIVE_TEXT_MODEL=qwen3.6-27b \
MODIS_LIVE_IMAGE_MODEL=qwen-image \
.venv/bin/python -m pytest -m live

Optional variables:

Variable	Default	Purpose
MODIS_LIVE_TIMEOUT	900	Per-request timeout in seconds.
MODIS_LIVE_CONCURRENCY	4	Concurrent async text requests.
MODIS_LIVE_IMAGE_WIDTH	1024	Image test width.
MODIS_LIVE_IMAGE_HEIGHT	1024	Image test height.
MODIS_LIVE_IMAGE_STEPS	20	Image inference steps.