[Enhancement] Stop looping when runtime is stable (CF-934)

codeflash/code_utils/config_consts.py

@@ -14,6 +14,13 @@
 DEFAULT_IMPORTANCE_THRESHOLD = 0.001
 N_CANDIDATES_LP = 6
 
+# pytest loop stability
+STABILITY_WARMUP_LOOPS = 4
+STABILITY_WINDOW_SIZE = 6
+STABILITY_CENTER_TOLERANCE = 0.01  # ±1% around median
+STABILITY_SPREAD_TOLERANCE = 0.02  # 2% window spread
+STABILITY_SLOPE_TOLERANCE = 0.01  # 1% improvement allowed
+
 # 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

codeflash/code_utils/env_utils.py

@@ -19,7 +19,6 @@
 def check_formatter_installed(formatter_cmds: list[str], exit_on_failure: bool = True) -> bool:  # noqa
     if not formatter_cmds or formatter_cmds[0] == "disabled":
         return True
-
     first_cmd = formatter_cmds[0]
     cmd_tokens = shlex.split(first_cmd) if isinstance(first_cmd, str) else [first_cmd]
 

codeflash/models/models.py

@@ -10,6 +10,7 @@
 from codeflash.lsp.helpers import is_LSP_enabled, report_to_markdown_table
 from codeflash.lsp.lsp_message import LspMarkdownMessage
 from codeflash.models.test_type import TestType
+from codeflash.result.best_summed_runtime import calculate_best_summed_runtime
 
 if TYPE_CHECKING:
     from collections.abc import Iterator
@@ -817,9 +818,7 @@ def total_passed_runtime(self) -> int:
         :return: The runtime in nanoseconds.
         """
         # TODO this doesn't look at the intersection of tests of baseline and original
-        return sum(
-            [min(usable_runtime_data) for _, usable_runtime_data in self.usable_runtime_data_by_test_case().items()]
-        )
+        return calculate_best_summed_runtime(self.usable_runtime_data_by_test_case())
 
     def file_to_no_of_tests(self, test_functions_to_remove: list[str]) -> Counter[Path]:
         map_gen_test_file_to_no_of_tests = Counter()

codeflash/optimization/function_optimizer.py

@@ -1887,7 +1887,6 @@ def establish_original_code_baseline(
                 benchmarking_results, self.function_to_optimize.function_name
             )
             logger.debug(f"Original async function throughput: {async_throughput} calls/second")
-            console.rule()
 
         if self.args.benchmark:
             replay_benchmarking_test_results = benchmarking_results.group_by_benchmarks(

codeflash/result/best_summed_runtime.py

@@ -0,0 +1,2 @@
+def calculate_best_summed_runtime(grouped_runtime_info: dict[any, list[int]]) -> int:
+    return sum([min(usable_runtime_data) for _, usable_runtime_data in grouped_runtime_info.items()])

codeflash/verification/pytest_plugin.py

@@ -2,23 +2,32 @@
 
 import contextlib
 import inspect
-
-# System Imports
 import logging
 import os
 import platform
 import re
 import sys
 import time as _time_module
 import warnings
+
+# System Imports
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable
+from typing import TYPE_CHECKING, Any, Callable, Optional
 from unittest import TestCase
 
 # PyTest Imports
 import pytest
 from pluggy import HookspecMarker
 
+from codeflash.code_utils.config_consts import (
+    STABILITY_CENTER_TOLERANCE,
+    STABILITY_SLOPE_TOLERANCE,
+    STABILITY_SPREAD_TOLERANCE,
+    STABILITY_WARMUP_LOOPS,
+    STABILITY_WINDOW_SIZE,
+)
+from codeflash.result.best_summed_runtime import calculate_best_summed_runtime
+
 if TYPE_CHECKING:
     from _pytest.config import Config, Parser
     from _pytest.main import Session
@@ -77,6 +86,7 @@ class UnexpectedError(Exception):
 # Store references to original functions before any patching
 _ORIGINAL_TIME_TIME = _time_module.time
 _ORIGINAL_PERF_COUNTER = _time_module.perf_counter
+_ORIGINAL_PERF_COUNTER_NS = _time_module.perf_counter_ns
 _ORIGINAL_TIME_SLEEP = _time_module.sleep
 
 
@@ -260,6 +270,71 @@ def pytest_configure(config: Config) -> None:
     _apply_deterministic_patches()
 
 
+def get_runtime_from_stdout(stdout: str) -> Optional[int]:
+    marker_start = "!######"
+    marker_end = "######!"
+
+    if not stdout:
+        return None
+
+    end = stdout.rfind(marker_end)
+    if end == -1:
+        return None
+
+    start = stdout.rfind(marker_start, 0, end)
+    if start == -1:
+        return None
+
+    payload = stdout[start + len(marker_start) : end]
+    last_colon = payload.rfind(":")
+    if last_colon == -1:
+        return None
+
+    return int(payload[last_colon + 1 :])
+
+
+_NODEID_BRACKET_PATTERN = re.compile(r"\s*\[\s*\d+\s*\]\s*$")
+
+
+def should_stop(
+    runtimes: list[int],
+    warmup: int = STABILITY_WARMUP_LOOPS,
+    window: int = STABILITY_WINDOW_SIZE,
+    center_rel_tol: float = STABILITY_CENTER_TOLERANCE,
+    spread_rel_tol: float = STABILITY_SPREAD_TOLERANCE,
+    slope_rel_tol: float = STABILITY_SLOPE_TOLERANCE,
+) -> bool:
+    if len(runtimes) < warmup + window:
+        return False
+
+    recent = runtimes[-window:]
+
+    # Use sorted array for faster median and min/max operations
+    recent_sorted = sorted(recent)
+    mid = window // 2
+    m = recent_sorted[mid] if window % 2 else (recent_sorted[mid - 1] + recent_sorted[mid]) / 2
+
+    # 1) All recent points close to the median
+    centered = True
+    for r in recent:
+        if abs(r - m) / m > center_rel_tol:
+            centered = False
+            break
+
+    # 2) Window spread is small
+    r_min, r_max = recent_sorted[0], recent_sorted[-1]
+    spread_ok = (r_max - r_min) / r_min <= spread_rel_tol
+
+    # 3) No strong downward trend (still improving)
+    # Compare first half vs second half
+    half = window // 2
+    first = sum(recent[:half]) / half
+    second = sum(recent[half:]) / (window - half)
+    slope_ok = (first - second) / first <= slope_rel_tol
+
+    return centered and spread_ok and slope_ok
+
+
 class PytestLoops:
     name: str = "pytest-loops"
 
@@ -268,6 +343,15 @@ def __init__(self, config: Config) -> None:
         level = logging.DEBUG if config.option.verbose > 1 else logging.INFO
         logging.basicConfig(level=level)
         self.logger = logging.getLogger(self.name)
+        self.usable_runtime_data_by_test_case: dict[str, list[int]] = {}
+
+    @pytest.hookimpl
+    def pytest_runtest_logreport(self, report: pytest.TestReport) -> None:
+        if report.when == "call" and report.passed:
+            duration_ns = get_runtime_from_stdout(report.capstdout)
+            if duration_ns:
+                clean_id = _NODEID_BRACKET_PATTERN.sub("", report.nodeid)
+                self.usable_runtime_data_by_test_case.setdefault(clean_id, []).append(duration_ns)
 
     @hookspec(firstresult=True)
     def pytest_runtestloop(self, session: Session) -> bool:
@@ -283,11 +367,9 @@ def pytest_runtestloop(self, session: Session) -> bool:
         total_time: float = self._get_total_time(session)
 
         count: int = 0
-
-        while total_time >= SHORTEST_AMOUNT_OF_TIME:  # need to run at least one for normal tests
+        runtimes = []
+        while total_time >= SHORTEST_AMOUNT_OF_TIME:
             count += 1
-            total_time = self._get_total_time(session)
-
             for index, item in enumerate(session.items):
                 item: pytest.Item = item  # noqa: PLW0127, PLW2901
                 item._report_sections.clear()  # clear reports for new test  # noqa: SLF001
@@ -304,8 +386,17 @@ def pytest_runtestloop(self, session: Session) -> bool:
                     raise session.Failed(session.shouldfail)
                 if session.shouldstop:
                     raise session.Interrupted(session.shouldstop)
+
+            best_runtime_until_now = calculate_best_summed_runtime(self.usable_runtime_data_by_test_case)
+            if best_runtime_until_now > 0:
+                runtimes.append(best_runtime_until_now)
+
+            if should_stop(runtimes):
+                break
+
             if self._timed_out(session, start_time, count):
-                break  # exit loop
+                break
+
             _ORIGINAL_TIME_SLEEP(self._get_delay_time(session))
         return True