DSPy: The framework for programming—not prompting—language models
Python
clean up assertions
manishshettymeea3b2c0768f
diff --git a/dspy/primitives/assertions.py b/dspy/primitives/assertions.pyindex e945603dc..c4b65ffc2 100644--- a/dspy/primitives/assertions.py+++ b/dspy/primitives/assertions.py@@ -1,13 +1,11 @@-# Note(shangyin): What abstractions we want to make about assertions?-# things to consider: what we want to assert, and what would be the syntax?-# One possible starting point would be constraints on the output of a module-from typing import Any+from typing import Any, Callableimport dspimport dspyclass DSPyAssertionError(AssertionError):- """DSPy custom error message."""+ """Custom exception raised when a DSPy `Assert` fails.+ """def __init__(self, msg: str, state: Any = None) -> None:super().__init__(msg)@@ -16,9 +14,9 @@ class DSPyAssertionError(AssertionError):class Assert:- """Compile time assertion."""+ """DSPy Assertion"""- def __init__(self, assert_fun, *args, **kwargs):+ def __init__(self, assert_fun: Callable, *args, **kwargs):self.assert_fun = assert_funself.args = args@@ -31,7 +29,6 @@ class Assert:self.kwargs = kwargsself.__call__()- # assert fun should always return booldef __call__(self) -> bool:result = self.assert_fun(*self.args, **self.kwargs)if isinstance(result, bool):@@ -43,118 +40,12 @@ class Assert:raise ValueError("Assertion function should always return [bool]")-# we could possibly to have runtime assertions as well-# class RuntimeAssert(Assert): ...---############################# ASSERTION AND BACKTRACKING POLICIES #############################---def assert_transform(max_backtracks=2):- """Decorator that simply re-runs the function if assertion fails,- up to `max_backtracks` times."""-- def wrapper(func):- def inner(*args, **kwargs):- for i in range(max_backtracks + 1):- try:- return func(*args, **kwargs)- except DSPyAssertionError as e:- print(f"{e.msg}")- if dsp.settings.trace:- pass- else:- print(- "UNREACHABLE: No trace available, this should not happen. Is this run time?"- )-- return inner-- return wrapper---def assert_update_transform(max_backtracks=2):- """Decorator that simply re-runs the function with updated temperature- if assertion fails, up to `max_backtracks` times."""-- def wrapper(func):- def inner(*args, **kwargs):- for i in range(max_backtracks + 1):- lm = dsp.settings.lm-- if i > 0 and backtrack_to is not None:- # print(f"updating temperature to {0.71 + 0.002 * i}")- lm = lm.copy(temperature=0.71 + 0.002 * i)-- new_settings = dict(lm=lm) if i > 0 else {}-- try:- with dsp.settings.context(**new_settings):- return func(*args, **kwargs)- except DSPyAssertionError as e:- print(f"{e.msg}")-- if dsp.settings.trace:- backtrack_to = id(dsp.settings.trace[-1][0])- else:- print(- "UNREACHABLE: No trace available, this should not happen. Is this run time?"- )-- return inner-- return wrapper---def assert_latest_transform(max_backtracks=2):- """Decorator that simply re-runs the function but updates the temperature- only for the latest predictor in the trace if assertion fails, up to `max_backtracks` times.-- # NOTE (@manish): the previous assert_update_transform applies temperature change to all predictors in the pipeline.- # Here, we udpate/pass the new temperature to the *backtrack_to* predictor *only* --> cause a cache miss --> re-run- # For other predictors, settings remain the same --> cache hit --> no re-run- """-- def wrapper(func):- def inner(*args, **kwargs):- backtrack_to = None-- for i in range(max_backtracks + 1):- if i > 0 and backtrack_to is not None:- print(f"rewinding to {id(backtrack_to)}")-- # udpate temperature via "config" attribute of the `backtrack_to` predictor- predictor_id = id(backtrack_to)- print(- f"prev temp @ {predictor_id}: {backtrack_to.get_config().get('temperature', 'default')}"- )- backtrack_to.update_config(temperature=0.7 + 0.001 * i)- print(- f"new temp @ {predictor_id}: {backtrack_to.get_config()['temperature']}"- )- try:- return func(*args, **kwargs)- except DSPyAssertionError as e:- print(f"{e.msg}")-- if dsp.settings.trace:- backtrack_to = dsp.settings.trace[-1][0]- else:- print(- "UNREACHABLE: No trace available, this should not happen. Is this run time?"- )-- return inner-- return wrapper--def assert_latest_feedback_transform(max_backtracks=2):- """ "Decorator that simply re-runs the function but updates the signature of- the latest predictor in the trace if assertion fails, up to `max_backtracks` times.-- The updated signature passes a new input to the predictor, which is the feedback- from the failed assertion. This feedback will be used by the predictor to self-repair.+ """Decorator that defines the backtracking policy for assertions.++ current policy: re-run the latest predictor up to `max_backtracks` times,+ with updated signature if assertion fails. updated signature adds a new+ input field to the signature, which is the feedback."""def wrapper(func):@@ -211,8 +102,6 @@ def assert_latest_feedback_transform(max_backtracks=2):for i in range(max_backtracks + 1):if i > 0 and backtrack_to is not None:- print(f"rewinding to {id(backtrack_to)}")-# create a new feedback fieldfeedback = dspy.InputField(prefix="Instruction:", desc="Some instructions you must satisfy"@@ -251,7 +140,6 @@ def assert_latest_feedback_transform(max_backtracks=2):predictor,original_forward,) in extended_predictors_to_original_forward.items():- print(f"reverting predictor {id(predictor)}")_revert_predictor_signature(predictor, "feedback")predictor.forward = original_forward