more effort values

Author: mohammedahmed18

codeflash/code_utils/config_consts.py

@@ -1,4 +1,4 @@
-from enum import Enum
+from enum import StrEnum, auto
 
 MAX_TEST_RUN_ITERATIONS = 5
 INDIVIDUAL_TESTCASE_TIMEOUT = 15
@@ -13,18 +13,11 @@
 REPEAT_OPTIMIZATION_PROBABILITY = 0.1
 DEFAULT_IMPORTANCE_THRESHOLD = 0.001
 
-# Refinement
-REFINE_ALL_THRESHOLD = 2  # when valid optimizations count is 2 or less, refine all optimizations
 REFINED_CANDIDATE_RANKING_WEIGHTS = (2, 1)  # (runtime, diff), runtime is more important than diff by a factor of 2
-TOP_N_REFINEMENTS = 0.45  # top 45% of valid optimizations (based on the weighted score) are refined
 
 # LSP-specific
 TOTAL_LOOPING_TIME_LSP = 10.0  # Kept same timing for LSP mode to avoid in increase in performance reporting
 
-# Code repair
-REPAIR_UNMATCHED_PERCENTAGE_LIMIT = 0.4  # if the percentage of unmatched tests is greater than this, we won't fix it (lowering this value makes the repair more stricted)
-MAX_REPAIRS_PER_TRACE = 4  # maximum number of repairs we will do for each function
-
 try:
     from codeflash.lsp.helpers import is_LSP_enabled
 
@@ -37,36 +30,49 @@
 MAX_CONTEXT_LEN_REVIEW = 1000
 
 
-class EffortLevel(str, Enum):
-    LOW = "low"
-    MEDIUM = "medium"
-    HIGH = "high"
+class EffortLevel(StrEnum):
+    LOW = auto()
+    MEDIUM = auto()
+    HIGH = auto()
 
 
-class Effort:
-    @staticmethod
-    def get_number_of_optimizer_candidates(effort: str) -> int:
-        if effort == EffortLevel.LOW.value:
-            return 3
-        if effort == EffortLevel.MEDIUM.value:
-            return 4
-        if effort == EffortLevel.HIGH.value:
-            return 5
-        msg = f"Invalid effort level: {effort}"
-        raise ValueError(msg)
+class EffortKeys(StrEnum):
+    N_OPTIMIZER_CANDIDATES = auto()
+    N_OPTIMIZER_LP_CANDIDATES = auto()
+    N_GENERATED_TESTS = auto()
+    MAX_CODE_REPAIRS_PER_TRACE = auto()
+    REPAIR_UNMATCHED_PERCENTAGE_LIMIT = auto()
+    REFINE_ALL_THRESHOLD = auto()
+    TOP_VALID_CANDIDATES_FOR_REFINEMENT = auto()
+
 
-    @staticmethod
-    def get_number_of_optimizer_lp_candidates(effort: str) -> int:
-        if effort == EffortLevel.LOW.value:
-            return 3
-        if effort == EffortLevel.MEDIUM.value:
-            return 5
-        if effort == EffortLevel.HIGH.value:
-            return 6
+EFFORT_VALUES: dict[str, dict[EffortLevel, any]] = {
+    EffortKeys.N_OPTIMIZER_CANDIDATES.value: {EffortLevel.LOW: 3, EffortLevel.MEDIUM: 4, EffortLevel.HIGH: 5},
+    EffortKeys.N_OPTIMIZER_LP_CANDIDATES.value: {EffortLevel.LOW: 3, EffortLevel.MEDIUM: 5, EffortLevel.HIGH: 6},
+    # we don't use effort with generated tests for now
+    EffortKeys.N_GENERATED_TESTS.value: {EffortLevel.LOW: 2, EffortLevel.MEDIUM: 2, EffortLevel.HIGH: 2},
+    # maximum number of repairs we will do for each function
+    EffortKeys.MAX_CODE_REPAIRS_PER_TRACE.value: {EffortLevel.LOW: 2, EffortLevel.MEDIUM: 4, EffortLevel.HIGH: 5},
+    # if the percentage of unmatched tests is greater than this, we won't fix it (lowering this value makes the repair more stricted)
+    # on the low effort we lower the limit to 20% to be more strict (less repairs)
+    EffortKeys.REPAIR_UNMATCHED_PERCENTAGE_LIMIT.value: {
+        EffortLevel.LOW: 0.2,
+        EffortLevel.MEDIUM: 0.4,
+        EffortLevel.HIGH: 0.5,
+    },
+    # when valid optimizations count is N or less, refine all optimizations
+    EffortKeys.REFINE_ALL_THRESHOLD.value: {EffortLevel.LOW: 2, EffortLevel.MEDIUM: 3, EffortLevel.HIGH: 4},
+    # Top valid candidates percentage for refinements
+    EffortKeys.TOP_VALID_CANDIDATES_FOR_REFINEMENT: {EffortLevel.LOW: 2, EffortLevel.MEDIUM: 3, EffortLevel.HIGH: 4},
+}
+
+
+def get_effort_value(key: EffortKeys, effort: EffortLevel) -> any:
+    key_str = key.value
+    if key_str in EFFORT_VALUES:
+        if effort in EFFORT_VALUES[key_str]:
+            return EFFORT_VALUES[key_str][effort]
         msg = f"Invalid effort level: {effort}"
         raise ValueError(msg)
-
-    @staticmethod
-    def get_number_of_generated_tests(effort: str) -> int:  # noqa: ARG004
-        # we don't use effort with generated tests for now
-        return 2
+    msg = f"Invalid key: {key_str}"
+    raise ValueError(msg)

codeflash/optimization/function_optimizer.py

@@ -45,14 +45,11 @@
 from codeflash.code_utils.config_consts import (
     COVERAGE_THRESHOLD,
     INDIVIDUAL_TESTCASE_TIMEOUT,
-    MAX_REPAIRS_PER_TRACE,
-    REFINE_ALL_THRESHOLD,
     REFINED_CANDIDATE_RANKING_WEIGHTS,
-    REPAIR_UNMATCHED_PERCENTAGE_LIMIT,
     REPEAT_OPTIMIZATION_PROBABILITY,
-    TOP_N_REFINEMENTS,
     TOTAL_LOOPING_TIME_EFFECTIVE,
-    Effort,
+    EffortKeys,
+    get_effort_value,
 )
 from codeflash.code_utils.deduplicate_code import normalize_code
 from codeflash.code_utils.edit_generated_tests import (
@@ -191,8 +188,16 @@ def refine_optimizations(self, request: list[AIServiceRefinerRequest]) -> concur
     def _process_refinement_results(self) -> OptimizedCandidate | None:
         """Process refinement results and add to queue. We generate a weighted ranking based on the runtime and diff lines and select the best (round of 45%) of valid optimizations to be refined."""
         future_refinements: list[concurrent.futures.Future] = []
+        top_n_candidates = int(
+            min(
+                get_effort_value(EffortKeys.TOP_VALID_CANDIDATES_FOR_REFINEMENT, self.args.effort),
+                len(self.all_refinements_data),
+            )
+        )
 
-        if len(self.all_refinements_data) <= REFINE_ALL_THRESHOLD:
+        if top_n_candidates == len(self.all_refinements_data) or len(self.all_refinements_data) <= get_effort_value(
+            EffortKeys.REFINE_ALL_THRESHOLD, self.args.effort
+        ):
             for data in self.all_refinements_data:
                 future_refinements.append(self.refine_optimizations([data]))  # noqa: PERF401
         else:
@@ -209,7 +214,6 @@ def _process_refinement_results(self) -> OptimizedCandidate | None:
             diffs_norm = normalize_by_max(diff_lens_list)
             # the lower the better
             score_dict = create_score_dictionary_from_metrics(weights, runtime_norm, diffs_norm)
-            top_n_candidates = int((TOP_N_REFINEMENTS * len(runtimes_list)) + 0.5)
             top_indecies = sorted(score_dict, key=score_dict.get)[:top_n_candidates]
 
             for idx in top_indecies:
@@ -310,7 +314,7 @@ def __init__(
         self.function_benchmark_timings = function_benchmark_timings if function_benchmark_timings else {}
         self.total_benchmark_timings = total_benchmark_timings if total_benchmark_timings else {}
         self.replay_tests_dir = replay_tests_dir if replay_tests_dir else None
-        n_tests = Effort.get_number_of_generated_tests(args.effort)
+        n_tests = get_effort_value(EffortKeys.N_GENERATED_TESTS, args.effort)
         self.executor = concurrent.futures.ThreadPoolExecutor(
             max_workers=n_tests + 3 if self.experiment_id is None else n_tests + 4
         )
@@ -360,7 +364,7 @@ def generate_and_instrument_tests(
         str,
     ]:
         """Generate and instrument tests for the function."""
-        n_tests = Effort.get_number_of_generated_tests(self.args.effort)
+        n_tests = get_effort_value(EffortKeys.N_GENERATED_TESTS, self.args.effort)
         generated_test_paths = [
             get_test_file_path(
                 self.test_cfg.tests_root, self.function_to_optimize.function_name, test_index, test_type="unit"
@@ -925,7 +929,7 @@ def determine_best_candidate(
             dependency_code=code_context.read_only_context_code,
             trace_id=self.get_trace_id(exp_type),
             line_profiler_results=original_code_baseline.line_profile_results["str_out"],
-            num_candidates=Effort.get_number_of_optimizer_lp_candidates(self.args.effort),
+            num_candidates=get_effort_value(EffortKeys.N_OPTIMIZER_LP_CANDIDATES, self.args.effort),
             experiment_metadata=ExperimentMetadata(
                 id=self.experiment_id, group="control" if exp_type == "EXP0" else "experiment"
             )
@@ -1290,7 +1294,7 @@ def generate_tests(
         generated_perf_test_paths: list[Path],
     ) -> Result[tuple[int, GeneratedTestsList, dict[str, set[FunctionCalledInTest]], str], str]:
         """Generate unit tests and concolic tests for the function."""
-        n_tests = Effort.get_number_of_generated_tests(self.args.effort)
+        n_tests = get_effort_value(EffortKeys.N_GENERATED_TESTS, self.args.effort)
         assert len(generated_test_paths) == n_tests
 
         # Submit test generation tasks
@@ -1352,7 +1356,7 @@ def generate_optimizations(
         run_experiment: bool = False,  # noqa: FBT001, FBT002
     ) -> Result[tuple[OptimizationSet, str], str]:
         """Generate optimization candidates for the function."""
-        n_candidates = Effort.get_number_of_optimizer_candidates(self.args.effort)
+        n_candidates = get_effort_value(EffortKeys.N_OPTIMIZER_CANDIDATES, self.args.effort)
 
         future_optimization_candidates = self.executor.submit(
             self.aiservice_client.optimize_python_code,
@@ -1919,8 +1923,9 @@ def repair_if_possible(
         test_results_count: int,
         exp_type: str,
     ) -> None:
-        if self.repair_counter >= MAX_REPAIRS_PER_TRACE:
-            logger.debug(f"Repair counter reached {MAX_REPAIRS_PER_TRACE}, skipping repair")
+        max_repairs = get_effort_value(EffortKeys.MAX_CODE_REPAIRS_PER_TRACE, self.args.effort)
+        if self.repair_counter >= max_repairs:
+            logger.debug(f"Repair counter reached {max_repairs}, skipping repair")
             return
         if candidate.source not in (OptimizedCandidateSource.OPTIMIZE, OptimizedCandidateSource.OPTIMIZE_LP):
             # only repair the first pass of the candidates for now
@@ -1930,7 +1935,7 @@ def repair_if_possible(
             logger.debug("No diffs found, skipping repair")
             return
         result_unmatched_perc = len(diffs) / test_results_count
-        if result_unmatched_perc > REPAIR_UNMATCHED_PERCENTAGE_LIMIT:
+        if result_unmatched_perc > get_effort_value(EffortKeys.REPAIR_UNMATCHED_PERCENTAGE_LIMIT, self.args.effort):
             logger.debug(f"Result unmatched percentage is {result_unmatched_perc * 100}%, skipping repair")
             return