aserpc 0.1.3

ASE - Remote Procedure Call

ASE - Remote Procedure Call

A lightweight RPC framework for serving ASE (Atomic Simulation Environment) calculators over ZeroMQ.

This package is primarily designed to run MLIP calculators in dedicated python environments to handle dependency conflicts or for multi-GPU distribution.

Quick Start

1. Start the broker:

aserpc broker

2. Check available calculators:

aserpc list
# or
aserpc list --registry tmp/registry.py:CALCULATORS

A registry file (or module via pck.registry:CALCULATORS) example:

# tmp/registry.py
from ase.calculators.lj import LennardJones

# Registry: name -> calculator factory
CALCULATORS = {
    "LJ": LennardJones,
}

2. Start a worker:

aserpc worker LJ # optional --registry tmp/registry.py:CALCULATORS

3. Use the remote calculator:

from aserpc import RemoteCalculator
from ase.build import molecule

water = molecule("H2O")
water.calc = RemoteCalculator("LJ")
energy = water.get_potential_energy()

# With retries for fault tolerance (retries=1 means 1 retry on failure, 2 total attempts)
water.calc = RemoteCalculator("LJ", retries=2)

Calculator Discovery

Workers discover calculators through Python entry point providers. This allows packages to register calculators that aserpc can automatically find.

Provider Functions

Register a provider function that returns calculator metadata:

# pyproject.toml
[project.entry-points."aserpc.calculators"]
mypackage = "mypackage.aserpc:get_calculators"

# mypackage/aserpc.py
import importlib.util

def get_calculators() -> dict[str, dict]:
    """Return metadata for available calculators.

    Each entry is: {"factory": "module:class", "args": [...], "kwargs": {...}}
    Use importlib.util.find_spec() to check availability without importing.
    """
    calcs = {}

    # Simple calculators (always available with ASE)
    calcs["LJ"] = {"factory": "ase.calculators.lj:LennardJones"}
    calcs["EMT"] = {"factory": "ase.calculators.emt:EMT"}

    # Check if mace is installed (without importing torch!)
    if importlib.util.find_spec("mace") is not None:
        calcs["mace_mp"] = {
            "factory": "mace.calculators:mace_mp",
            "kwargs": {"model": "medium"},
        }

    # Check if chgnet is installed
    if importlib.util.find_spec("chgnet") is not None:
        calcs["chgnet"] = {"factory": "chgnet.model:CHGNetCalculator"}

    return calcs

After installing the package, the new calculators should show up on aserpc list and can be started as workers:

aserpc worker LJ
aserpc worker mace_mp

Using a Registry File

Alternatively, specify a registry file with calculator factories:

# registry.py
from ase.calculators.lj import LennardJones

CALCULATORS = {
    "LJ": LennardJones,
}

aserpc worker LJ --registry registry.py:CALCULATORS

Direct Module Import

You can also specify a calculator factory directly:

# Direct module:callable syntax
aserpc worker ase.calculators.lj:LennardJones
aserpc worker mace.calculators:mace_mp

# With kwargs for the factory
aserpc worker mace.calculators:mace_mp --kwargs '{"model": "medium"}'
aserpc worker mace.calculators:mace_mp --kwargs '{"model": "large", "device": "cuda"}'

# With a custom name for broker registration
aserpc worker mace.calculators:mace_mp --name mace_medium --kwargs '{"model": "medium"}'
aserpc worker mace.calculators:mace_mp --name mace_large --kwargs '{"model": "large"}'

# With positional args
aserpc worker my.module:Calculator --args '[1, 2]' --kwargs '{"key": "value"}'

This is particularly useful with uvx for running calculators without pre-installing them:

# Run MACE calculator directly with uvx
uvx --with mace-torch aserpc worker mace.calculators:mace_mp --kwargs '{"model": "medium"}'

# Run CHGNet
uvx --with chgnet aserpc worker chgnet.model:CHGNetCalculator

Listing Available Calculators

List all discoverable calculators:

# From entry points
aserpc list

# From a registry file
aserpc list --registry registry.py:CALCULATORS

# From a running broker
aserpc list --broker ipc:///tmp/aserpc/frontend.ipc

Scale-to-Zero with Manager

The Manager enables automatic worker spawning on demand, allowing scale-to-zero deployments where workers are only running when needed.

How It Works

1. Client requests a calculator with no available workers
2. Broker forwards spawn request to Manager
3. Manager spawns a worker process using configured command
4. Worker registers with broker and handles the request
5. Worker auto-shuts down after idle timeout

flowchart TD
    Client -->|REQUEST| Broker
    Broker -->|SPAWN_REQUEST| Manager
    Manager -->|spawns| Worker
    Worker -->|READY| Broker
    Broker -->|routes request| Worker
    Worker -->|RESPONSE| Broker
    Broker -->|RESPONSE| Client
    Worker -->|idle timeout| Worker
    Worker -->|DISCONNECT| Broker

Setup

1. Create spawn configurations:

# mypackage/aserpc.py
from aserpc.spawn import SpawnConfig

def get_spawn_configs() -> dict[str, SpawnConfig]:
    return {
        "mace_mp": SpawnConfig(
            name="mace_mp",
            command=[
                "uvx", "--with", "mace-torch", "aserpc", "worker",
                "mace.calculators:mace_mp", "--kwargs", '{"model": "medium"}',
            ],
            idle_timeout=120.0,
        ),
        "lj": SpawnConfig(
            name="lj",
            command=["aserpc", "worker", "ase.calculators.lj:LennardJones", "--name", "lj"],
            idle_timeout=60.0,
        ),
    }

2. Register via entry points:

# pyproject.toml
[project.entry-points."aserpc.spawn"]
mypackage = "mypackage.aserpc:get_spawn_configs"

3. Start the broker and manager:

# Terminal 1: Start broker
aserpc broker

# Terminal 2: Start manager (discovers from entry points)
aserpc manager

# Or with explicit registry
aserpc manager --registry mypackage.aserpc:get_spawn_configs

4. Use as normal - workers spawn automatically:

from aserpc import RemoteCalculator
from ase.build import molecule

# No workers running initially
water = molecule("H2O")
water.calc = RemoteCalculator("mace_mp")
energy = water.get_potential_energy()  # Manager spawns worker on demand

SpawnConfig Options

SpawnConfig(
    name="calc_name",           # Calculator name (must match worker registration)
    command=["cmd", "args"],    # Command to spawn worker
    env={"KEY": "value"},       # Additional environment variables
    cwd="/path/to/workdir",     # Working directory
    idle_timeout=300.0,         # Worker idle timeout in seconds
)

Command templates support variables:

• {name} - Calculator name
• {broker} - Broker backend address
• {idle_timeout} - Idle timeout value

Example with templates:

SpawnConfig(
    name="mace",
    command=["aserpc", "worker", "{name}", "--broker", "{broker}", "--idle-timeout", "{idle_timeout}"],
)

Without a Manager

If no manager is running, the broker returns an error when no workers are available for a requested calculator.

Architecture

flowchart TD

Client <--> Broker <--> Worker1[Worker 1]
Broker <--> Worker2[Worker 2]
Broker <--> WorkerN[Worker N]
Manager -.->|spawns on demand| Worker1
Manager -.->|spawns on demand| Worker2
Broker <-->|SPAWN_REQUEST| Manager

• Broker: Routes requests to available workers, handles load balancing
• Worker: Runs calculator computations, sends heartbeats
• Client: RemoteCalculator acts as a drop-in ASE calculator
• Manager: Spawns workers on demand for scale-to-zero deployments

Configuration

Configuration can be set via environment variables or pyproject.toml. Environment variables take precedence.

Environment Variables

Variable	Default	Description
ASERPC_IPC_DIR	.aserpc (cwd)	Directory for IPC sockets
ASERPC_IPC_FRONTEND	ipc://{IPC_DIR}/frontend.ipc	Client socket address
ASERPC_IPC_BACKEND	ipc://{IPC_DIR}/backend.ipc	Worker socket address
ASERPC_WORKER_TIMEOUT	30.0	Seconds before broker considers worker dead
ASERPC_HEARTBEAT_INTERVAL	5.0	Seconds between heartbeats
ASERPC_IDLE_TIMEOUT	300.0	Seconds before idle worker shuts down
ASERPC_REQUEST_QUEUE_TIMEOUT	60.0	Seconds before queued request expires
ASERPC_CLIENT_TIMEOUT_MS	60000	Client timeout in milliseconds

pyproject.toml

[tool.aserpc]
ipc_dir = "/tmp/aserpc"
worker_timeout = 30.0
heartbeat_interval = 5.0
idle_timeout = 300.0
request_queue_timeout = 60.0
client_timeout_ms = 60000