OpenEvolve / data /openevolve /process_parallel.py
introvoyz041's picture
Migrated from GitHub
5e4510c verified
"""
Process-based parallel controller for true parallelism
"""
import asyncio
import logging
import multiprocessing as mp
import pickle
import signal
import time
from concurrent.futures import Future, ProcessPoolExecutor
from concurrent.futures import TimeoutError as FutureTimeoutError
from dataclasses import asdict, dataclass
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple
from openevolve.config import Config
from openevolve.database import Program, ProgramDatabase
from openevolve.utils.metrics_utils import safe_numeric_average
logger = logging.getLogger(__name__)
@dataclass
class SerializableResult:
"""Result that can be pickled and sent between processes"""
child_program_dict: Optional[Dict[str, Any]] = None
parent_id: Optional[str] = None
iteration_time: float = 0.0
prompt: Optional[Dict[str, str]] = None
llm_response: Optional[str] = None
artifacts: Optional[Dict[str, Any]] = None
iteration: int = 0
error: Optional[str] = None
def _worker_init(config_dict: dict, evaluation_file: str, parent_env: dict = None) -> None:
"""Initialize worker process with necessary components"""
import os
# Set environment from parent process
if parent_env:
os.environ.update(parent_env)
global _worker_config
global _worker_evaluation_file
global _worker_evaluator
global _worker_llm_ensemble
global _worker_prompt_sampler
# Store config for later use
# Reconstruct Config object from nested dictionaries
from openevolve.config import (
Config,
DatabaseConfig,
EvaluatorConfig,
LLMConfig,
LLMModelConfig,
PromptConfig,
)
# Reconstruct model objects
models = [LLMModelConfig(**m) for m in config_dict["llm"]["models"]]
evaluator_models = [LLMModelConfig(**m) for m in config_dict["llm"]["evaluator_models"]]
# Create LLM config with models
llm_dict = config_dict["llm"].copy()
llm_dict["models"] = models
llm_dict["evaluator_models"] = evaluator_models
llm_config = LLMConfig(**llm_dict)
# Create other configs
prompt_config = PromptConfig(**config_dict["prompt"])
database_config = DatabaseConfig(**config_dict["database"])
evaluator_config = EvaluatorConfig(**config_dict["evaluator"])
_worker_config = Config(
llm=llm_config,
prompt=prompt_config,
database=database_config,
evaluator=evaluator_config,
**{
k: v
for k, v in config_dict.items()
if k not in ["llm", "prompt", "database", "evaluator"]
},
)
_worker_evaluation_file = evaluation_file
# These will be lazily initialized on first use
_worker_evaluator = None
_worker_llm_ensemble = None
_worker_prompt_sampler = None
def _lazy_init_worker_components():
"""Lazily initialize expensive components on first use"""
global _worker_evaluator
global _worker_llm_ensemble
global _worker_prompt_sampler
if _worker_llm_ensemble is None:
from openevolve.llm.ensemble import LLMEnsemble
_worker_llm_ensemble = LLMEnsemble(_worker_config.llm.models)
if _worker_prompt_sampler is None:
from openevolve.prompt.sampler import PromptSampler
_worker_prompt_sampler = PromptSampler(_worker_config.prompt)
if _worker_evaluator is None:
from openevolve.evaluator import Evaluator
from openevolve.llm.ensemble import LLMEnsemble
from openevolve.prompt.sampler import PromptSampler
# Create evaluator-specific components
evaluator_llm = LLMEnsemble(_worker_config.llm.evaluator_models)
evaluator_prompt = PromptSampler(_worker_config.prompt)
evaluator_prompt.set_templates("evaluator_system_message")
_worker_evaluator = Evaluator(
_worker_config.evaluator,
_worker_evaluation_file,
evaluator_llm,
evaluator_prompt,
database=None, # No shared database in worker
suffix=getattr(_worker_config, "file_suffix", ".py"),
)
def _run_iteration_worker(
iteration: int, db_snapshot: Dict[str, Any], parent_id: str, inspiration_ids: List[str]
) -> SerializableResult:
"""Run a single iteration in a worker process"""
try:
# Lazy initialization
_lazy_init_worker_components()
# Reconstruct programs from snapshot
programs = {pid: Program(**prog_dict) for pid, prog_dict in db_snapshot["programs"].items()}
parent = programs[parent_id]
inspirations = [programs[pid] for pid in inspiration_ids if pid in programs]
# Get parent artifacts if available
parent_artifacts = db_snapshot["artifacts"].get(parent_id)
# Get island-specific programs for context
parent_island = parent.metadata.get("island", db_snapshot["current_island"])
island_programs = [
programs[pid] for pid in db_snapshot["islands"][parent_island] if pid in programs
]
# Sort by metrics for top programs
island_programs.sort(
key=lambda p: p.metrics.get("combined_score", safe_numeric_average(p.metrics)),
reverse=True,
)
# Use config values for limits instead of hardcoding
# Programs for LLM display (includes both top and diverse for inspiration)
programs_for_prompt = island_programs[
: _worker_config.prompt.num_top_programs + _worker_config.prompt.num_diverse_programs
]
# Best programs only (for previous attempts section, focused on top performers)
best_programs_only = island_programs[: _worker_config.prompt.num_top_programs]
# Build prompt
prompt = _worker_prompt_sampler.build_prompt(
current_program=parent.code,
parent_program=parent.code,
program_metrics=parent.metrics,
previous_programs=[p.to_dict() for p in best_programs_only],
top_programs=[p.to_dict() for p in programs_for_prompt],
inspirations=[p.to_dict() for p in inspirations],
language=_worker_config.language,
evolution_round=iteration,
diff_based_evolution=_worker_config.diff_based_evolution,
program_artifacts=parent_artifacts,
feature_dimensions=db_snapshot.get("feature_dimensions", []),
)
iteration_start = time.time()
# Generate code modification (sync wrapper for async)
try:
llm_response = asyncio.run(
_worker_llm_ensemble.generate_with_context(
system_message=prompt["system"],
messages=[{"role": "user", "content": prompt["user"]}],
)
)
except Exception as e:
logger.error(f"LLM generation failed: {e}")
return SerializableResult(error=f"LLM generation failed: {str(e)}", iteration=iteration)
# Check for None response
if llm_response is None:
return SerializableResult(error="LLM returned None response", iteration=iteration)
# Parse response based on evolution mode
if _worker_config.diff_based_evolution:
from openevolve.utils.code_utils import apply_diff, extract_diffs, format_diff_summary
diff_blocks = extract_diffs(llm_response, _worker_config.diff_pattern)
if not diff_blocks:
return SerializableResult(
error=f"No valid diffs found in response", iteration=iteration
)
child_code = apply_diff(parent.code, llm_response, _worker_config.diff_pattern)
changes_summary = format_diff_summary(diff_blocks)
else:
from openevolve.utils.code_utils import parse_full_rewrite
new_code = parse_full_rewrite(llm_response, _worker_config.language)
if not new_code:
return SerializableResult(
error=f"No valid code found in response", iteration=iteration
)
child_code = new_code
changes_summary = "Full rewrite"
# Check code length
if len(child_code) > _worker_config.max_code_length:
return SerializableResult(
error=f"Generated code exceeds maximum length ({len(child_code)} > {_worker_config.max_code_length})",
iteration=iteration,
)
# Evaluate the child program
import uuid
child_id = str(uuid.uuid4())
child_metrics = asyncio.run(_worker_evaluator.evaluate_program(child_code, child_id))
# Get artifacts
artifacts = _worker_evaluator.get_pending_artifacts(child_id)
# Create child program
child_program = Program(
id=child_id,
code=child_code,
language=_worker_config.language,
parent_id=parent.id,
generation=parent.generation + 1,
metrics=child_metrics,
iteration_found=iteration,
metadata={
"changes": changes_summary,
"parent_metrics": parent.metrics,
"island": parent_island,
},
)
iteration_time = time.time() - iteration_start
return SerializableResult(
child_program_dict=child_program.to_dict(),
parent_id=parent.id,
iteration_time=iteration_time,
prompt=prompt,
llm_response=llm_response,
artifacts=artifacts,
iteration=iteration,
)
except Exception as e:
logger.exception(f"Error in worker iteration {iteration}")
return SerializableResult(error=str(e), iteration=iteration)
class ProcessParallelController:
"""Controller for process-based parallel evolution"""
def __init__(
self,
config: Config,
evaluation_file: str,
database: ProgramDatabase,
evolution_tracer=None,
file_suffix: str = ".py",
):
self.config = config
self.evaluation_file = evaluation_file
self.database = database
self.evolution_tracer = evolution_tracer
self.file_suffix = file_suffix
self.executor: Optional[ProcessPoolExecutor] = None
self.shutdown_event = mp.Event()
self.early_stopping_triggered = False
# Number of worker processes
self.num_workers = config.evaluator.parallel_evaluations
self.num_islands = config.database.num_islands
logger.info(f"Initialized process parallel controller with {self.num_workers} workers")
def _serialize_config(self, config: Config) -> dict:
"""Serialize config object to a dictionary that can be pickled"""
# Manual serialization to handle nested objects properly
# The asdict() call itself triggers the deepcopy which tries to serialize novelty_llm. Remove it first.
config.database.novelty_llm = None
return {
"llm": {
"models": [asdict(m) for m in config.llm.models],
"evaluator_models": [asdict(m) for m in config.llm.evaluator_models],
"api_base": config.llm.api_base,
"api_key": config.llm.api_key,
"temperature": config.llm.temperature,
"top_p": config.llm.top_p,
"max_tokens": config.llm.max_tokens,
"timeout": config.llm.timeout,
"retries": config.llm.retries,
"retry_delay": config.llm.retry_delay,
},
"prompt": asdict(config.prompt),
"database": asdict(config.database),
"evaluator": asdict(config.evaluator),
"max_iterations": config.max_iterations,
"checkpoint_interval": config.checkpoint_interval,
"log_level": config.log_level,
"log_dir": config.log_dir,
"random_seed": config.random_seed,
"diff_based_evolution": config.diff_based_evolution,
"max_code_length": config.max_code_length,
"language": config.language,
"file_suffix": self.file_suffix,
}
def start(self) -> None:
"""Start the process pool"""
# Convert config to dict for pickling
# We need to be careful with nested dataclasses
config_dict = self._serialize_config(self.config)
# Pass current environment to worker processes
import os
import sys
current_env = dict(os.environ)
executor_kwargs = {
"max_workers": self.num_workers,
"initializer": _worker_init,
"initargs": (config_dict, self.evaluation_file, current_env),
}
if sys.version_info >= (3, 11):
logger.info(f"Set max {self.config.max_tasks_per_child} tasks per child")
executor_kwargs["max_tasks_per_child"] = self.config.max_tasks_per_child
elif self.config.max_tasks_per_child is not None:
logger.warn(
"max_tasks_per_child is only supported in Python 3.11+. "
"Ignoring max_tasks_per_child and using spawn start method."
)
executor_kwargs["mp_context"] = mp.get_context("spawn")
# Create process pool with initializer
self.executor = ProcessPoolExecutor(**executor_kwargs)
logger.info(f"Started process pool with {self.num_workers} processes")
def stop(self) -> None:
"""Stop the process pool"""
self.shutdown_event.set()
if self.executor:
self.executor.shutdown(wait=True)
self.executor = None
logger.info("Stopped process pool")
def request_shutdown(self) -> None:
"""Request graceful shutdown"""
logger.info("Graceful shutdown requested...")
self.shutdown_event.set()
def _create_database_snapshot(self) -> Dict[str, Any]:
"""Create a serializable snapshot of the database state"""
# Only include necessary data for workers
snapshot = {
"programs": {pid: prog.to_dict() for pid, prog in self.database.programs.items()},
"islands": [list(island) for island in self.database.islands],
"current_island": self.database.current_island,
"feature_dimensions": self.database.config.feature_dimensions,
"artifacts": {}, # Will be populated selectively
}
# Include artifacts for programs that might be selected
# IMPORTANT: This limits artifacts (execution outputs/errors) to first 100 programs only.
# This does NOT affect program code - all programs are fully serialized above.
# With max_artifact_bytes=20KB and population_size=1000, artifacts could be 20MB total,
# which would significantly slow worker process initialization. The limit of 100 keeps
# artifact data under 2MB while still providing execution context for recent programs.
# Workers can still evolve properly as they have access to ALL program code.
for pid in list(self.database.programs.keys())[:100]:
artifacts = self.database.get_artifacts(pid)
if artifacts:
snapshot["artifacts"][pid] = artifacts
return snapshot
async def run_evolution(
self,
start_iteration: int,
max_iterations: int,
target_score: Optional[float] = None,
checkpoint_callback=None,
):
"""Run evolution with process-based parallelism"""
if not self.executor:
raise RuntimeError("Process pool not started")
total_iterations = start_iteration + max_iterations
logger.info(
f"Starting process-based evolution from iteration {start_iteration} "
f"for {max_iterations} iterations (total: {total_iterations})"
)
# Track pending futures by island to maintain distribution
pending_futures: Dict[int, Future] = {}
island_pending: Dict[int, List[int]] = {i: [] for i in range(self.num_islands)}
batch_size = min(self.num_workers * 2, max_iterations)
# Submit initial batch - distribute across islands
batch_per_island = max(1, batch_size // self.num_islands) if batch_size > 0 else 0
current_iteration = start_iteration
# Round-robin distribution across islands
for island_id in range(self.num_islands):
for _ in range(batch_per_island):
if current_iteration < total_iterations:
future = self._submit_iteration(current_iteration, island_id)
if future:
pending_futures[current_iteration] = future
island_pending[island_id].append(current_iteration)
current_iteration += 1
next_iteration = current_iteration
completed_iterations = 0
# Early stopping tracking
early_stopping_enabled = self.config.early_stopping_patience is not None
if early_stopping_enabled:
best_score = float("-inf")
iterations_without_improvement = 0
logger.info(
f"Early stopping enabled: patience={self.config.early_stopping_patience}, "
f"threshold={self.config.convergence_threshold}, "
f"metric={self.config.early_stopping_metric}"
)
else:
logger.info("Early stopping disabled")
# Process results as they complete
while (
pending_futures
and completed_iterations < max_iterations
and not self.shutdown_event.is_set()
):
# Find completed futures
completed_iteration = None
for iteration, future in list(pending_futures.items()):
if future.done():
completed_iteration = iteration
break
if completed_iteration is None:
await asyncio.sleep(0.01)
continue
# Process completed result
future = pending_futures.pop(completed_iteration)
try:
# Use evaluator timeout + buffer to gracefully handle stuck processes
timeout_seconds = self.config.evaluator.timeout + 30
result = future.result(timeout=timeout_seconds)
if result.error:
logger.warning(f"Iteration {completed_iteration} error: {result.error}")
elif result.child_program_dict:
# Reconstruct program from dict
child_program = Program(**result.child_program_dict)
# Add to database (will auto-inherit parent's island)
# No need to specify target_island - database will handle parent island inheritance
self.database.add(child_program, iteration=completed_iteration)
# Store artifacts
if result.artifacts:
self.database.store_artifacts(child_program.id, result.artifacts)
# Log evolution trace
if self.evolution_tracer:
# Retrieve parent program for trace logging
parent_program = (
self.database.get(result.parent_id) if result.parent_id else None
)
if parent_program:
# Determine island ID
island_id = child_program.metadata.get(
"island", self.database.current_island
)
self.evolution_tracer.log_trace(
iteration=completed_iteration,
parent_program=parent_program,
child_program=child_program,
prompt=result.prompt,
llm_response=result.llm_response,
artifacts=result.artifacts,
island_id=island_id,
metadata={
"iteration_time": result.iteration_time,
"changes": child_program.metadata.get("changes", ""),
},
)
# Log prompts
if result.prompt:
self.database.log_prompt(
template_key=(
"full_rewrite_user"
if not self.config.diff_based_evolution
else "diff_user"
),
program_id=child_program.id,
prompt=result.prompt,
responses=[result.llm_response] if result.llm_response else [],
)
# Island management
# get current program island id
island_id = child_program.metadata.get(
"island", self.database.current_island
)
#use this to increment island generation
self.database.increment_island_generation(island_idx=island_id)
# Check migration
if self.database.should_migrate():
logger.info(f"Performing migration at iteration {completed_iteration}")
self.database.migrate_programs()
self.database.log_island_status()
# Log progress
logger.info(
f"Iteration {completed_iteration}: "
f"Program {child_program.id} "
f"(parent: {result.parent_id}) "
f"completed in {result.iteration_time:.2f}s"
)
if child_program.metrics:
metrics_str = ", ".join(
[
f"{k}={v:.4f}" if isinstance(v, (int, float)) else f"{k}={v}"
for k, v in child_program.metrics.items()
]
)
logger.info(f"Metrics: {metrics_str}")
# Check if this is the first program without combined_score
if not hasattr(self, "_warned_about_combined_score"):
self._warned_about_combined_score = False
if (
"combined_score" not in child_program.metrics
and not self._warned_about_combined_score
):
avg_score = safe_numeric_average(child_program.metrics)
logger.warning(
f"⚠️ No 'combined_score' metric found in evaluation results. "
f"Using average of all numeric metrics ({avg_score:.4f}) for evolution guidance. "
f"For better evolution results, please modify your evaluator to return a 'combined_score' "
f"metric that properly weights different aspects of program performance."
)
self._warned_about_combined_score = True
# Check for new best
if self.database.best_program_id == child_program.id:
logger.info(
f"🌟 New best solution found at iteration {completed_iteration}: "
f"{child_program.id}"
)
# Checkpoint callback
# Don't checkpoint at iteration 0 (that's just the initial program)
if (
completed_iteration > 0
and completed_iteration % self.config.checkpoint_interval == 0
):
logger.info(
f"Checkpoint interval reached at iteration {completed_iteration}"
)
self.database.log_island_status()
if checkpoint_callback:
checkpoint_callback(completed_iteration)
# Check target score
if target_score is not None and child_program.metrics:
if (
"combined_score" in child_program.metrics
and child_program.metrics["combined_score"] >= target_score
):
logger.info(
f"Target score {target_score} reached at iteration {completed_iteration}"
)
break
# Check early stopping
if early_stopping_enabled and child_program.metrics:
# Get the metric to track for early stopping
current_score = None
if self.config.early_stopping_metric in child_program.metrics:
current_score = child_program.metrics[self.config.early_stopping_metric]
elif self.config.early_stopping_metric == "combined_score":
# Default metric not found, use safe average (standard pattern)
current_score = safe_numeric_average(child_program.metrics)
else:
# User specified a custom metric that doesn't exist
logger.warning(
f"Early stopping metric '{self.config.early_stopping_metric}' not found, using safe numeric average"
)
current_score = safe_numeric_average(child_program.metrics)
if current_score is not None and isinstance(current_score, (int, float)):
# Check for improvement
improvement = current_score - best_score
if improvement >= self.config.convergence_threshold:
best_score = current_score
iterations_without_improvement = 0
logger.debug(
f"New best score: {best_score:.4f} (improvement: {improvement:+.4f})"
)
else:
iterations_without_improvement += 1
logger.debug(
f"No improvement: {iterations_without_improvement}/{self.config.early_stopping_patience}"
)
# Check if we should stop
if (
iterations_without_improvement
>= self.config.early_stopping_patience
):
self.early_stopping_triggered = True
logger.info(
f"🛑 Early stopping triggered at iteration {completed_iteration}: "
f"No improvement for {iterations_without_improvement} iterations "
f"(best score: {best_score:.4f})"
)
break
except FutureTimeoutError:
logger.error(
f"⏰ Iteration {completed_iteration} timed out after {timeout_seconds}s "
f"(evaluator timeout: {self.config.evaluator.timeout}s + 30s buffer). "
f"Canceling future and continuing with next iteration."
)
# Cancel the future to clean up the process
future.cancel()
except Exception as e:
logger.error(f"Error processing result from iteration {completed_iteration}: {e}")
completed_iterations += 1
# Remove completed iteration from island tracking
for island_id, iteration_list in island_pending.items():
if completed_iteration in iteration_list:
iteration_list.remove(completed_iteration)
break
# Submit next iterations maintaining island balance
for island_id in range(self.num_islands):
if (
len(island_pending[island_id]) < batch_per_island
and next_iteration < total_iterations
and not self.shutdown_event.is_set()
):
future = self._submit_iteration(next_iteration, island_id)
if future:
pending_futures[next_iteration] = future
island_pending[island_id].append(next_iteration)
next_iteration += 1
break # Only submit one iteration per completion to maintain balance
# Handle shutdown
if self.shutdown_event.is_set():
logger.info("Shutdown requested, canceling remaining evaluations...")
for future in pending_futures.values():
future.cancel()
# Log completion reason
if self.early_stopping_triggered:
logger.info("✅ Evolution completed - Early stopping triggered due to convergence")
elif self.shutdown_event.is_set():
logger.info("✅ Evolution completed - Shutdown requested")
else:
logger.info("✅ Evolution completed - Maximum iterations reached")
return self.database.get_best_program()
def _submit_iteration(
self, iteration: int, island_id: Optional[int] = None
) -> Optional[Future]:
"""Submit an iteration to the process pool, optionally pinned to a specific island"""
try:
# Use specified island or current island
target_island = island_id if island_id is not None else self.database.current_island
# Use thread-safe sampling that doesn't modify shared state
# This fixes the race condition from GitHub issue #246
parent, inspirations = self.database.sample_from_island(
island_id=target_island, num_inspirations=self.config.prompt.num_top_programs
)
# Create database snapshot
db_snapshot = self._create_database_snapshot()
db_snapshot["sampling_island"] = target_island # Mark which island this is for
# Submit to process pool
future = self.executor.submit(
_run_iteration_worker,
iteration,
db_snapshot,
parent.id,
[insp.id for insp in inspirations],
)
return future
except Exception as e:
logger.error(f"Error submitting iteration {iteration}: {e}")
return None