Merge branch 'main' into inquirer

Author: KRRT7

codeflash/api/aiservice.py

@@ -4,6 +4,7 @@
 import os
 import platform
 import time
+from itertools import count
 from typing import TYPE_CHECKING, Any, cast
 
 import requests
@@ -12,7 +13,6 @@
 from codeflash.cli_cmds.console import console, logger
 from codeflash.code_utils.code_replacer import is_zero_diff
 from codeflash.code_utils.code_utils import unified_diff_strings
-from codeflash.code_utils.config_consts import N_CANDIDATES_EFFECTIVE, N_CANDIDATES_LP_EFFECTIVE
 from codeflash.code_utils.env_utils import get_codeflash_api_key
 from codeflash.code_utils.git_utils import get_last_commit_author_if_pr_exists, get_repo_owner_and_name
 from codeflash.code_utils.time_utils import humanize_runtime
@@ -40,6 +40,11 @@ class AiServiceClient:
     def __init__(self) -> None:
         self.base_url = self.get_aiservice_base_url()
         self.headers = {"Authorization": f"Bearer {get_codeflash_api_key()}", "Connection": "close"}
+        self.llm_call_counter = count(1)
+
+    def get_next_sequence(self) -> int:
+        """Get the next LLM call sequence number."""
+        return next(self.llm_call_counter)
 
     def get_aiservice_base_url(self) -> str:
         if os.environ.get("CODEFLASH_AIS_SERVER", default="prod").lower() == "local":
@@ -106,6 +111,7 @@ def _get_valid_candidates(
                     optimization_id=opt["optimization_id"],
                     source=source,
                     parent_id=opt.get("parent_id", None),
+                    model=opt.get("model"),
                 )
             )
         return candidates
@@ -115,7 +121,6 @@ def optimize_python_code(  # noqa: D417
         source_code: str,
         dependency_code: str,
         trace_id: str,
-        num_candidates: int = 10,
         experiment_metadata: ExperimentMetadata | None = None,
         *,
         is_async: bool = False,
@@ -127,46 +132,49 @@ def optimize_python_code(  # noqa: D417
         - source_code (str): The python code to optimize.
         - dependency_code (str): The dependency code used as read-only context for the optimization
         - trace_id (str): Trace id of optimization run
-        - num_candidates (int): Number of optimization variants to generate. Default is 10.
         - experiment_metadata (Optional[ExperimentalMetadata, None]): Any available experiment metadata for this optimization
+        - is_async (bool): Whether the function being optimized is async
 
         Returns
         -------
         - List[OptimizationCandidate]: A list of Optimization Candidates.
 
         """
+        logger.info("Generating optimized candidates…")
+        console.rule()
         start_time = time.perf_counter()
         git_repo_owner, git_repo_name = safe_get_repo_owner_and_name()
 
         payload = {
             "source_code": source_code,
             "dependency_code": dependency_code,
-            "num_variants": num_candidates,
             "trace_id": trace_id,
             "python_version": platform.python_version(),
             "experiment_metadata": experiment_metadata,
             "codeflash_version": codeflash_version,
             "current_username": get_last_commit_author_if_pr_exists(None),
             "repo_owner": git_repo_owner,
             "repo_name": git_repo_name,
-            "n_candidates": N_CANDIDATES_EFFECTIVE,
             "is_async": is_async,
+            "lsp_mode": is_LSP_enabled(),
+            "call_sequence": self.get_next_sequence(),
         }
+        logger.debug(f"Sending optimize request: trace_id={trace_id}, lsp_mode={payload['lsp_mode']}")
 
-        logger.info("!lsp|Generating optimized candidates…")
-        console.rule()
         try:
             response = self.make_ai_service_request("/optimize", payload=payload, timeout=60)
         except requests.exceptions.RequestException as e:
             logger.exception(f"Error generating optimized candidates: {e}")
             ph("cli-optimize-error-caught", {"error": str(e)})
+            console.rule()
             return []
 
         if response.status_code == 200:
             optimizations_json = response.json()["optimizations"]
-            console.rule()
             end_time = time.perf_counter()
             logger.debug(f"!lsp|Generating possible optimizations took {end_time - start_time:.2f} seconds.")
+            logger.info(f"!lsp|Received {len(optimizations_json)} optimization candidates.")
+            console.rule()
             return self._get_valid_candidates(optimizations_json, OptimizedCandidateSource.OPTIMIZE)
         try:
             error = response.json()["error"]
@@ -183,54 +191,53 @@ def optimize_python_code_line_profiler(  # noqa: D417
         dependency_code: str,
         trace_id: str,
         line_profiler_results: str,
-        num_candidates: int = 10,
         experiment_metadata: ExperimentMetadata | None = None,
     ) -> list[OptimizedCandidate]:
-        """Optimize the given python code for performance by making a request to the Django endpoint.
+        """Optimize the given python code for performance using line profiler results.
 
         Parameters
         ----------
         - source_code (str): The python code to optimize.
         - dependency_code (str): The dependency code used as read-only context for the optimization
         - trace_id (str): Trace id of optimization run
-        - num_candidates (int): Number of optimization variants to generate. Default is 10.
+        - line_profiler_results (str): Line profiler output to guide optimization
         - experiment_metadata (Optional[ExperimentalMetadata, None]): Any available experiment metadata for this optimization
 
         Returns
         -------
         - List[OptimizationCandidate]: A list of Optimization Candidates.
 
         """
+        if line_profiler_results == "":
+            logger.info("No LineProfiler results were provided, Skipping optimization.")
+            return []
+
+        logger.info("Generating optimized candidates with line profiler…")
+        console.rule()
+
         payload = {
             "source_code": source_code,
             "dependency_code": dependency_code,
-            "num_variants": num_candidates,
             "line_profiler_results": line_profiler_results,
             "trace_id": trace_id,
             "python_version": platform.python_version(),
             "experiment_metadata": experiment_metadata,
             "codeflash_version": codeflash_version,
             "lsp_mode": is_LSP_enabled(),
-            "n_candidates_lp": N_CANDIDATES_LP_EFFECTIVE,
+            "call_sequence": self.get_next_sequence(),
         }
 
-        console.rule()
-        if line_profiler_results == "":
-            logger.info("No LineProfiler results were provided, Skipping optimization.")
-            console.rule()
-            return []
         try:
             response = self.make_ai_service_request("/optimize-line-profiler", payload=payload, timeout=60)
         except requests.exceptions.RequestException as e:
             logger.exception(f"Error generating optimized candidates: {e}")
             ph("cli-optimize-error-caught", {"error": str(e)})
+            console.rule()
             return []
 
         if response.status_code == 200:
             optimizations_json = response.json()["optimizations"]
-            logger.info(
-                f"!lsp|Generated {len(optimizations_json)} candidate optimizations using line profiler information."
-            )
+            logger.info(f"!lsp|Received {len(optimizations_json)} line profiler optimization candidates.")
             console.rule()
             return self._get_valid_candidates(optimizations_json, OptimizedCandidateSource.OPTIMIZE_LP)
         try:
@@ -268,6 +275,7 @@ def optimize_python_code_refinement(self, request: list[AIServiceRefinerRequest]
                 "trace_id": opt.trace_id,
                 "function_references": opt.function_references,
                 "python_version": platform.python_version(),
+                "call_sequence": self.get_next_sequence(),
             }
             for opt in request
         ]
@@ -402,6 +410,7 @@ def get_new_explanation(  # noqa: D417
             "throughput_improvement": throughput_improvement,
             "function_references": function_references,
             "codeflash_version": codeflash_version,
+            "call_sequence": self.get_next_sequence(),
         }
         logger.info("loading|Generating explanation")
         console.rule()
@@ -564,6 +573,7 @@ def generate_regression_tests(  # noqa: D417
             "python_version": platform.python_version(),
             "codeflash_version": codeflash_version,
             "is_async": function_to_optimize.is_async,
+            "call_sequence": self.get_next_sequence(),
         }
         try:
             response = self.make_ai_service_request("/testgen", payload=payload, timeout=90)
@@ -650,6 +660,7 @@ def get_optimization_review(
             "codeflash_version": codeflash_version,
             "calling_fn_details": calling_fn_details,
             "python_version": platform.python_version(),
+            "call_sequence": self.get_next_sequence(),
         }
         console.rule()
         try:

codeflash/cli_cmds/cli.py

@@ -79,6 +79,9 @@ def parse_args() -> Namespace:
     parser.add_argument(
         "--no-pr", action="store_true", help="Do not create a PR for the optimization, only update the code locally."
     )
+    parser.add_argument(
+        "--no-gen-tests", action="store_true", help="Do not generate tests, use only existing tests for optimization."
+    )
     parser.add_argument("--staging-review", action="store_true", help="Upload optimizations to staging for review")
     parser.add_argument(
         "--verify-setup",

codeflash/code_utils/config_consts.py

@@ -14,6 +14,12 @@
 DEFAULT_IMPORTANCE_THRESHOLD = 0.001
 N_CANDIDATES_LP = 6
 
+# pytest loop stability
+# For now, we use strict thresholds (large windows and low tolerances), since this is still experimental.
+STABILITY_WINDOW_SIZE = 0.35  # 35% of total window
+STABILITY_CENTER_TOLERANCE = 0.0025  # ±0.25% around median
+STABILITY_SPREAD_TOLERANCE = 0.0025  # 0.25% window spread
+
 # 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/code_utils/formatter.py

@@ -46,18 +46,13 @@ def apply_formatter_cmds(
     print_status: bool,  # noqa
     exit_on_failure: bool = True,  # noqa
 ) -> tuple[Path, str, bool]:
-    should_make_copy = False
-    file_path = path
-
-    if test_dir_str:
-        should_make_copy = True
-        file_path = Path(test_dir_str) / "temp.py"
-
     if not path.exists():
         msg = f"File {path} does not exist. Cannot apply formatter commands."
         raise FileNotFoundError(msg)
 
-    if should_make_copy:
+    file_path = path
+    if test_dir_str:
+        file_path = Path(test_dir_str) / "temp.py"
         shutil.copy2(path, file_path)
 
     file_token = "$file"  # noqa: S105

codeflash/discovery/discover_unit_tests.py

@@ -751,6 +751,7 @@ def process_test_files(
 
     tests_cache = TestsCache(project_root_path)
     logger.info("!lsp|Discovering tests and processing unit tests")
+    console.rule()
     with test_files_progress_bar(total=len(file_to_test_map), description="Processing test files") as (
         progress,
         task_id,

codeflash/models/models.py

@@ -46,6 +46,7 @@ class AIServiceRefinerRequest:
     original_line_profiler_results: str
     optimized_line_profiler_results: str
     function_references: str | None = None
+    call_sequence: int | None = None
 
 
 class TestDiffScope(str, Enum):
@@ -464,6 +465,7 @@ class OptimizedCandidate:
     optimization_id: str
     source: OptimizedCandidateSource
     parent_id: str | None = None
+    model: str | None = None  # Which LLM model generated this candidate
 
 
 @dataclass(frozen=True)