inferential 1.4.0

Robotics-aware inference orchestration and scheduling

Inferential Python SDK

Python client and server SDK for Inferential inference orchestration. The Python package includes the client SDK (for sending observations and receiving results) and the server (scheduling and dispatch to local models or Ray Serve).

Install

# Client + edge server (pyzmq, protobuf, numpy, pydantic — no Ray)
pip install inferential

# With Ray Serve for distributed serving
pip install inferential[ray]

# Development
pip install inferential[dev]

Quick Start

See the full Quick Start guide for step-by-step setup.

Edge Server (LocalDispatcher — no Ray)

import asyncio
import numpy as np
from inferential import Server, LocalDispatcher

def my_policy(obs: dict) -> dict:
    dim = 7
    for v in obs.values():
        if isinstance(v, np.ndarray) and v.ndim == 1:
            dim = v.shape[0]
            break
    return {"actions": np.random.randn(dim).astype(np.float32)}

dispatcher = LocalDispatcher({"policy-v2": my_policy})
server = Server(bind="tcp://*:5555", dispatcher=dispatcher)

@server.on_metric
def log(name, value, labels):
    if name == "inference_latency_ms":
        print(f"Client {labels.get('client')}: {value:.1f}ms")

asyncio.run(server.run())

Ray Serve Server (distributed)

import asyncio
import numpy as np
from ray import serve
from inferential import Server

@serve.deployment
class MockPolicy:
    def infer(self, obs: dict) -> dict:
        dim = 7
        for v in obs.values():
            if isinstance(v, np.ndarray) and v.ndim == 1:
                dim = v.shape[0]
                break
        return {"actions": np.random.randn(dim).astype(np.float32)}

serve.run(MockPolicy.bind(), name="policy-v2")

server = Server(bind="tcp://*:5555", models=["policy-v2"])  # defaults to RayDispatcher
asyncio.run(server.run())

Client (sync)

import numpy as np
from inferential import Connection

conn = Connection(server="tcp://localhost:5555", client_id="agent-01", client_type="franka")
model = conn.model("policy-v2", latency_budget_ms=30.0)

state = np.random.randn(7).astype(np.float32)
model.observe(urgency=0.8, state=state)

result = model.get_result(timeout_ms=50)
if result is not None:
    actions = result["actions"]  # np.ndarray

conn.close()

Client (async)

import asyncio
import numpy as np
from inferential import AsyncConnection

async def main():
    async with AsyncConnection(server="tcp://localhost:5555", client_id="agent-01") as conn:
        model = conn.model("policy-v2", latency_budget_ms=30.0)

        state = np.random.randn(7).astype(np.float32)
        await model.observe(urgency=0.8, state=state)

        result = await model.get_result(timeout_ms=50)
        if result is not None:
            actions = result["actions"]  # np.ndarray

asyncio.run(main())

API Reference

Connection(server, client_id, client_type, reconnect_ivl_ms=100, reconnect_max_ms=5000)

Creates a ZMQ DEALER connection to the server. The server address can be with or without the tcp:// prefix.

AsyncConnection(server, client_id, client_type, ...)

Async variant using zmq.asyncio.Context. Supports async with for automatic cleanup.

conn.model(model_id, latency_budget_ms=50.0, priority=1) → Model / AsyncModel

Creates a handle to a specific model on the server.

model.observe(urgency=0.0, steps_remaining=None, **kwargs)

Sends an observation to the server. Keyword arguments are automatically dispatched:

• np.ndarray values → serialized as tensors (dtype/shape preserved)
• str values → passed as metadata key-value pairs
• urgency (float, 0.0–1.0) → scheduling priority hint
• steps_remaining (int) → remaining steps in trajectory

model.observe(
    urgency=0.5,
    steps_remaining=120,
    state_vector=np.zeros(7, dtype=np.float32),
    image=np.zeros((3, 224, 224), dtype=np.uint8),
    prompt="describe the scene",  # → metadata
)

model.get_result(timeout_ms=100) → dict | None

Waits for a response. Returns a dict mapping tensor keys to numpy arrays, or None on timeout. Also includes response_id, model_id, inference_latency_ms, and any metadata from the server.

conn.close()

Closes the ZMQ socket. Called automatically by AsyncConnection.__aexit__.

Server Configuration

See Architecture for full details on schedulers, queue management, metrics, and configuration schema.

Documentation

• Quick Start — Install, run server + client, get your first result
• Architecture — System design, wire protocol, schedulers, metrics
• Examples — Multi-language client demos, server extensions
• Contributing — Commit conventions, branching, code style