let's add diversity to our optimizations

codeflash/api/aiservice.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import concurrent.futures
 import json
 import os
 import platform
@@ -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
@@ -35,6 +35,8 @@
     from codeflash.models.models import AIServiceCodeRepairRequest, AIServiceRefinerRequest
     from codeflash.result.explanation import Explanation
 
+multi_model_executor = concurrent.futures.ThreadPoolExecutor(max_workers=10, thread_name_prefix="multi_model")
+
 
 class AiServiceClient:
     def __init__(self) -> None:
@@ -92,7 +94,7 @@ def make_ai_service_request(
         return response
 
     def _get_valid_candidates(
-        self, optimizations_json: list[dict[str, Any]], source: OptimizedCandidateSource
+        self, optimizations_json: list[dict[str, Any]], source: OptimizedCandidateSource, model: str | None = None
     ) -> list[OptimizedCandidate]:
         candidates: list[OptimizedCandidate] = []
         for opt in optimizations_json:
@@ -106,6 +108,7 @@ def _get_valid_candidates(
                     optimization_id=opt["optimization_id"],
                     source=source,
                     parent_id=opt.get("parent_id", None),
+                    model=model,
                 )
             )
         return candidates
@@ -115,10 +118,11 @@ 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,
+        model: str | None = None,
+        call_sequence: int | None = None,
     ) -> list[OptimizedCandidate]:
         """Optimize the given python code for performance by making a request to the Django endpoint.
 
@@ -127,8 +131,9 @@ 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
+        - model (str | None): Model name to use ("gpt-4.1" or "claude-sonnet-4-5"). Default is None (server default).
+        - call_sequence (int | None): Sequence number for multi-model calls (1, 2, 3...). Default is None.
 
         Returns
         -------
@@ -141,20 +146,19 @@ def optimize_python_code(  # noqa: D417
         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,
+            "model": model,
+            "call_sequence": call_sequence,
         }
+        logger.debug(f"Sending optimize request: model={model}, trace_id={trace_id}, call_sequence={call_sequence}")
 
-        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:
@@ -164,17 +168,16 @@ def optimize_python_code(  # noqa: D417
 
         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.")
-            return self._get_valid_candidates(optimizations_json, OptimizedCandidateSource.OPTIMIZE)
+            logger.debug(f"Backend returned {len(optimizations_json)} optimization(s)")
+            return self._get_valid_candidates(optimizations_json, OptimizedCandidateSource.OPTIMIZE, model=model)
         try:
             error = response.json()["error"]
         except Exception:
             error = response.text
         logger.error(f"Error generating optimized candidates: {response.status_code} - {error}")
         ph("cli-optimize-error-response", {"response_status_code": response.status_code, "error": error})
-        console.rule()
         return []
 
     def optimize_python_code_line_profiler(  # noqa: D417
@@ -183,8 +186,9 @@ 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,
+        model: str | None = None,
+        call_sequence: int | None = None,
     ) -> list[OptimizedCandidate]:
         """Optimize the given python code for performance by making a request to the Django endpoint.
 
@@ -193,8 +197,9 @@ def optimize_python_code_line_profiler(  # 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
+        - model (str | None): Model name to use ("gpt-4.1" or "claude-sonnet-4-5"). Default is None (server default).
+        - call_sequence (int | None): Sequence number for multi-model calls (1, 2, 3...). Default is None.
 
         Returns
         -------
@@ -204,20 +209,18 @@ def optimize_python_code_line_profiler(  # noqa: D417
         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,
+            "model": model,
+            "call_sequence": call_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)
@@ -228,20 +231,113 @@ def optimize_python_code_line_profiler(  # noqa: D417
 
         if response.status_code == 200:
             optimizations_json = response.json()["optimizations"]
-            logger.info(
-                f"!lsp|Generated {len(optimizations_json)} candidate optimizations using line profiler information."
-            )
-            console.rule()
-            return self._get_valid_candidates(optimizations_json, OptimizedCandidateSource.OPTIMIZE_LP)
+            logger.debug(f"Backend returned {len(optimizations_json)} LP optimization(s)")
+            return self._get_valid_candidates(optimizations_json, OptimizedCandidateSource.OPTIMIZE_LP, model=model)
         try:
             error = response.json()["error"]
         except Exception:
             error = response.text
         logger.error(f"Error generating optimized candidates: {response.status_code} - {error}")
         ph("cli-optimize-error-response", {"response_status_code": response.status_code, "error": error})
-        console.rule()
         return []
 
+    def optimize_python_code_multi_model(
+        self,
+        source_code: str,
+        dependency_code: str,
+        base_trace_id: str,
+        model_distribution: list[tuple[str, int]],
+        experiment_metadata: ExperimentMetadata | None = None,
+        *,
+        is_async: bool = False,
+        sequence_offset: int = 0,
+    ) -> tuple[list[OptimizedCandidate], int]:
+        """Generate optimizations using multiple models in parallel."""
+        logger.info("Generating optimized candidates…")
+        console.rule()
+
+        futures: list[tuple[concurrent.futures.Future[list[OptimizedCandidate]], str]] = []
+
+        call_index = 0
+        for model_name, num_calls in model_distribution:
+            for _ in range(num_calls):
+                call_trace_id = f"{base_trace_id[:-3]}0{call_index:02x}"
+                call_sequence = sequence_offset + call_index + 1
+                call_index += 1
+                future = multi_model_executor.submit(
+                    self.optimize_python_code,
+                    source_code,
+                    dependency_code,
+                    call_trace_id,
+                    experiment_metadata,
+                    is_async=is_async,
+                    model=model_name,
+                    call_sequence=call_sequence,
+                )
+                futures.append((future, model_name))
+
+        concurrent.futures.wait([f for f, _ in futures])
+
+        all_candidates: list[OptimizedCandidate] = []
+        for future, model_name in futures:
+            try:
+                candidates = future.result()
+                all_candidates.extend(candidates)
+            except Exception as e:
+                logger.warning(f"Model {model_name} call failed: {e}")
+                continue
+
+        console.rule()
+        return all_candidates, call_index
+
+    def optimize_python_code_line_profiler_multi_model(
+        self,
+        source_code: str,
+        dependency_code: str,
+        base_trace_id: str,
+        line_profiler_results: str,
+        model_distribution: list[tuple[str, int]],
+        experiment_metadata: ExperimentMetadata | None = None,
+        sequence_offset: int = 0,
+    ) -> tuple[list[OptimizedCandidate], int]:
+        """Generate line profiler optimizations using multiple models in parallel."""
+        logger.info("Generating optimized candidates with line profiler…")
+        console.rule()
+
+        futures: list[tuple[concurrent.futures.Future[list[OptimizedCandidate]], str]] = []
+
+        call_index = 0
+        for model_name, num_calls in model_distribution:
+            for _ in range(num_calls):
+                call_trace_id = f"{base_trace_id[:-3]}1{call_index:02x}"
+                call_sequence = sequence_offset + call_index + 1
+                call_index += 1
+                future = multi_model_executor.submit(
+                    self.optimize_python_code_line_profiler,
+                    source_code,
+                    dependency_code,
+                    call_trace_id,
+                    line_profiler_results,
+                    experiment_metadata,
+                    model_name,
+                    call_sequence,
+                )
+                futures.append((future, model_name))
+
+        concurrent.futures.wait([f for f, _ in futures])
+
+        all_candidates: list[OptimizedCandidate] = []
+        for future, model_name in futures:
+            try:
+                candidates = future.result()
+                all_candidates.extend(candidates)
+            except Exception as e:
+                logger.warning(f"Line profiler model {model_name} call failed: {e}")
+                continue
+
+        console.rule()
+        return all_candidates, call_index
+
     def optimize_python_code_refinement(self, request: list[AIServiceRefinerRequest]) -> list[OptimizedCandidate]:
         """Optimize the given python code for performance by making a request to the Django endpoint.
 
@@ -268,6 +364,7 @@ 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": opt.call_sequence,
             }
             for opt in request
         ]
@@ -357,6 +454,7 @@ def get_new_explanation(  # noqa: D417
         throughput_improvement: str | None = None,
         function_references: str | None = None,
         codeflash_version: str = codeflash_version,
+        call_sequence: int | None = None,
     ) -> str:
         """Optimize the given python code for performance by making a request to the Django endpoint.
 
@@ -402,6 +500,7 @@ def get_new_explanation(  # noqa: D417
             "throughput_improvement": throughput_improvement,
             "function_references": function_references,
             "codeflash_version": codeflash_version,
+            "call_sequence": call_sequence,
         }
         logger.info("loading|Generating explanation")
         console.rule()
@@ -529,6 +628,7 @@ def generate_regression_tests(  # noqa: D417
         test_timeout: int,
         trace_id: str,
         test_index: int,
+        call_sequence: int | None = None,
     ) -> tuple[str, str, str] | None:
         """Generate regression tests for the given function by making a request to the Django endpoint.
 
@@ -564,6 +664,7 @@ def generate_regression_tests(  # noqa: D417
             "python_version": platform.python_version(),
             "codeflash_version": codeflash_version,
             "is_async": function_to_optimize.is_async,
+            "call_sequence": call_sequence,
         }
         try:
             response = self.make_ai_service_request("/testgen", payload=payload, timeout=90)
@@ -604,6 +705,7 @@ def get_optimization_review(
         replay_tests: str,
         concolic_tests: str,  # noqa: ARG002
         calling_fn_details: str,
+        call_sequence: int | None = None,
     ) -> str:
         """Compute the optimization review of current Pull Request.
 
@@ -650,6 +752,7 @@ def get_optimization_review(
             "codeflash_version": codeflash_version,
             "calling_fn_details": calling_fn_details,
             "python_version": platform.python_version(),
+            "call_sequence": call_sequence,
         }
         console.rule()
         try:

codeflash/code_utils/config_consts.py

@@ -32,6 +32,20 @@
 MAX_N_CANDIDATES = 5
 MAX_N_CANDIDATES_LP = 6
 
+# Multi-model diversity configuration
+# Each tuple is (model_name, num_calls) where each call returns 1 candidate
+# Standard mode: 3 GPT-4.1 + 2 Claude Sonnet = 5 candidates
+MODEL_DISTRIBUTION: list[tuple[str, int]] = [("gpt-4.1", 3), ("claude-sonnet-4-5", 2)]
+
+# LSP mode: fewer candidates for faster response
+MODEL_DISTRIBUTION_LSP: list[tuple[str, int]] = [("gpt-4.1", 2), ("claude-sonnet-4-5", 1)]
+
+# Line profiler mode: 6 candidates total
+MODEL_DISTRIBUTION_LP: list[tuple[str, int]] = [("gpt-4.1", 4), ("claude-sonnet-4-5", 2)]
+
+# Line profiler LSP mode
+MODEL_DISTRIBUTION_LP_LSP: list[tuple[str, int]] = [("gpt-4.1", 2), ("claude-sonnet-4-5", 1)]
+
 try:
     from codeflash.lsp.helpers import is_LSP_enabled
 
@@ -43,5 +57,7 @@
 N_CANDIDATES_LP_EFFECTIVE = min(N_CANDIDATES_LP_LSP if _IS_LSP_ENABLED else N_CANDIDATES_LP, MAX_N_CANDIDATES_LP)
 N_TESTS_TO_GENERATE_EFFECTIVE = N_TESTS_TO_GENERATE_LSP if _IS_LSP_ENABLED else N_TESTS_TO_GENERATE
 TOTAL_LOOPING_TIME_EFFECTIVE = TOTAL_LOOPING_TIME_LSP if _IS_LSP_ENABLED else TOTAL_LOOPING_TIME
+MODEL_DISTRIBUTION_EFFECTIVE = MODEL_DISTRIBUTION_LSP if _IS_LSP_ENABLED else MODEL_DISTRIBUTION
+MODEL_DISTRIBUTION_LP_EFFECTIVE = MODEL_DISTRIBUTION_LP_LSP if _IS_LSP_ENABLED else MODEL_DISTRIBUTION_LP
 
 MAX_CONTEXT_LEN_REVIEW = 1000

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)