Implementation:Pyro ppl Pyro Settings

Property	Value
Module	pyro.settings
Source	pyro/settings.py
Lines	164
Functions	get, set, context, register
Dependencies	(none -- pure Python)

Overview

This module provides a global settings registry for Pyro. It allows any Pyro module to register a configurable setting, and provides a uniform interface for users to get, set, and temporarily override settings. The registry stores settings by alias (a user-friendly name) and maps them to their actual location in the codebase (module + attribute path).

Key features:

• Get/set individual or all settings.
• Context manager for temporary overrides (also works as a decorator).
• Registration with optional validator functions.
• Auto-documentation of defaults in the module docstring.

The most commonly used setting is cholesky_relative_jitter, which controls the amount of jitter added during Cholesky decomposition in pyro.ops.tensor_utils.safe_cholesky.

Code Reference

Function: get(alias=None)

Gets one or all settings.

• If alias is None, returns a dict of all registered settings and their current values.
• If alias is a string, looks up the registered module and deep attribute path, imports the module, and traverses the attribute path to get the current value.

Function: set(**kwargs)

Sets one or more settings by alias.

• Looks up the module and deep attribute path.
• Runs the optional validator (if registered).
• Imports the module and sets the attribute.

Function: context(**kwargs)

Context manager (and decorator) for temporarily overriding settings.

• Saves old values on enter.
• Applies new values.
• Restores old values on exit (even if an exception occurs).

Function: register(alias, modulename, deepname, validator=None)

Registers a new setting in the global registry.

• alias: User-friendly name (snake_case).
• modulename: Python module path (typically __name__).
• deepname: Dot-separated attribute path within the module (e.g., "MyClass.my_attr").
• validator: Optional callable that validates values and raises on error.

Can be used as a decorator on a validator function:

@pyro.settings.register("my_setting", __name__, "MY_SETTING")
def _validate(value):
    assert isinstance(value, float)

Updates the module docstring with the new default value after registration.

Internal Registry

The _REGISTRY dictionary maps each alias to a tuple (modulename, deepname, validator).

I/O Contract

Function	Input	Output
get(alias=None)	Optional str	Single value or Dict[str, Any]
set(**kwargs)	alias=value keyword arguments	None
context(**kwargs)	alias=value keyword arguments	Context manager / decorator
register(alias, modulename, deepname, validator)	Strings, optional callable	Callable (decorator or validator)

Usage Examples

import pyro

# View all settings
all_settings = pyro.settings.get()
print(all_settings)

# Get a specific setting
jitter = pyro.settings.get("cholesky_relative_jitter")
print(f"Current jitter: {jitter}")

# Set a setting
pyro.settings.set(cholesky_relative_jitter=0.5)

# Temporarily override (context manager)
with pyro.settings.context(cholesky_relative_jitter=1e-3):
    # Inside this block, jitter is 1e-3
    result = run_model()
# jitter is restored to previous value

# Temporarily override (decorator)
@pyro.settings.context(cholesky_relative_jitter=1e-6)
def my_precise_computation():
    return run_model()

# Register a custom setting in a Pyro module
# In your_module.py:
MY_THRESHOLD = 0.1
pyro.settings.register(
    "my_threshold",    # alias
    __name__,           # module name
    "MY_THRESHOLD",     # attribute name
)

# Register with a validator
@pyro.settings.register("my_threshold", __name__, "MY_THRESHOLD")
def _validate_threshold(value):
    assert isinstance(value, float) and value > 0

Related Pages

• Pyro_ppl_Pyro_TensorUtils -- Uses cholesky_relative_jitter in safe_cholesky
• Pyro_ppl_Pyro_Util -- General Pyro utility functions