Skip to content

armgit / dspy

Public Imported from GitHub
0

DSPy: The framework for programming—not prompting—language models

Python

[bug fix] revert extended predictors after whole program run; update to feedback field

manishshettymdb4b51cfaeab
diff --git a/dspy/primitives/assertions.py b/dspy/primitives/assertions.py
index 1a9c39636..d18bd264d 100644
--- a/dspy/primitives/assertions.py
+++ b/dspy/primitives/assertions.py
@@ -11,7 +11,7 @@ class DSPyAssertionError(AssertionError):
def __init__(self, msg: str, state: Any = None) -> None:
super().__init__(msg)
- self.msg = f"Assertion Error: {msg}"
+ self.msg = msg
self.state = state
@@ -50,13 +50,13 @@ class Assert:
############################# ASSERTION AND BACKTRACKING POLICIES #############################
-def assert_transform(backtrack=2):
+def assert_transform(max_backtracks=2):
"""Decorator that simply re-runs the function if assertion fails,
- up to `backtrack` times."""
+ up to `max_backtracks` times."""
def wrapper(func):
def inner(*args, **kwargs):
- for i in range(backtrack):
+ for i in range(max_backtracks+1):
try:
return func(*args, **kwargs)
except DSPyAssertionError as e:
@@ -73,13 +73,13 @@ def assert_transform(backtrack=2):
return wrapper
-def assert_update_transform(backtrack=2):
+def assert_update_transform(max_backtracks=2):
"""Decorator that simply re-runs the function with updated temperature
- if assertion fails, up to `backtrack` times."""
+ if assertion fails, up to `max_backtracks` times."""
def wrapper(func):
def inner(*args, **kwargs):
- for i in range(backtrack):
+ for i in range(max_backtracks+1):
lm = dsp.settings.lm
if i > 0 and backtrack_to is not None:
@@ -106,9 +106,9 @@ def assert_update_transform(backtrack=2):
return wrapper
-def assert_latest_transform(backtrack=2):
+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 `backtrack` times.
+ 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
@@ -119,7 +119,7 @@ def assert_latest_transform(backtrack=2):
def inner(*args, **kwargs):
backtrack_to = None
- for i in range(backtrack):
+ for i in range(max_backtracks+1):
if i > 0 and backtrack_to is not None:
print(f"rewinding to {id(backtrack_to)}")
@@ -149,9 +149,9 @@ def assert_latest_transform(backtrack=2):
return wrapper
-def assert_latest_feedback_transform(backtrack=2):
+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 `backtrack` times.
+ 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.
@@ -161,7 +161,8 @@ def assert_latest_feedback_transform(backtrack=2):
def inner(*args, **kwargs):
backtrack_to = None
error_msg = None
- extended_backtrack_to = None
+ result = None
+ extended_predictors_to_original_forward = {}
def _extend_predictor_signature(predictor, **kwargs):
"""Update the signature of a predictor instance with specified fields."""
@@ -198,27 +199,31 @@ def assert_latest_feedback_transform(backtrack=2):
predictor.forward = new_forward
- for i in range(backtrack):
+ for i in range(max_backtracks+1):
if i > 0 and backtrack_to is not None:
print(f"rewinding to {id(backtrack_to)}")
- # udpate the signature of the `backtrack_to` predictor with a feedback field
+ # create a new feedback field
feedback = dspy.InputField(
- prefix="Note:", desc="common mistakes to avoid"
+ prefix="Instruction:", desc="Some instructions you must satisfy"
)
- extended_backtrack_to = backtrack_to
- original_forward = extended_backtrack_to.forward
+
+ # save the original forward function to revert back to
+ extended_predictors_to_original_forward[backtrack_to] = backtrack_to.forward
+
+ # extend signature with feedback field
_extend_predictor_signature(backtrack_to, feedback=feedback)
- # set the feedback field as a default kwarg to the predictor's forward function
+ # set feedback field as a default kwarg to the predictor's forward function
_wrap_forward_with_set_fields(
backtrack_to, default_args={"feedback": error_msg}
)
try:
- return func(*args, **kwargs)
+ result = func(*args, **kwargs)
+ break
except DSPyAssertionError as e:
- print(f"{e.msg}")
+ print(f"AssertionError: {e.msg}")
error_msg = e.msg
if dsp.settings.trace:
@@ -227,10 +232,15 @@ def assert_latest_feedback_transform(backtrack=2):
print(
"UNREACHABLE: No trace available, this should not happen. Is this run time?"
)
- finally:
- if extended_backtrack_to:
- _revert_predictor_signature(extended_backtrack_to, "feedback")
- extended_backtrack_to.forward = original_forward
+
+ # revert any extended predictors to their originals
+ if extended_predictors_to_original_forward:
+ for 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
+
+ return result
return inner