evolve / openevolve-fixed /scripts /gen_phase_configs.py
zhouk777's picture
Upload folder using huggingface_hub
ea8c728 verified
Raw
History Blame Contribute Delete
3.98 kB
#!/usr/bin/env python3
"""Generate phase-1 and phase-2 YAML configs for a problem.
Reads the problem's existing config.yaml to extract the system_message and
evaluator timeout, then writes two new configs pointing at a local vLLM server.
Usage:
python scripts/gen_phase_configs.py examples/.../config.yaml output_dir/ \
--model Qwen/Qwen2.5-7B-Instruct --api-base http://localhost:8000/v1
"""
import argparse
import yaml
import os
PHASE_TEMPLATE = {
"log_level": "INFO",
"checkpoint_interval": 5,
"diff_based_evolution": False,
"allow_full_rewrites": True,
}
PHASE_SETTINGS = {
1: {
"temperature": 0.7,
"num_top_programs": 3,
"population_size": 50,
"archive_size": 25,
"num_islands": 4,
"exploitation_ratio": 0.7,
"migration_interval": 30,
"cascade_thresholds": [0.3, 0.6],
},
2: {
"temperature": 0.8,
"num_top_programs": 4,
"population_size": 60,
"archive_size": 30,
"num_islands": 5,
"exploitation_ratio": 0.55,
"migration_interval": 25,
"cascade_thresholds": [0.4, 0.7],
"max_code_length": 100000,
},
}
PHASE2_ADDENDUM = (
"\n\nPHASE 2 INSTRUCTIONS: The evolution has reached a plateau. "
"Try fundamentally different approaches rather than incremental tweaks. "
"Consider alternative mathematical formulations, different optimization "
"strategies, or novel algorithmic ideas."
)
def build_config(base_cfg, phase, model, api_base, iterations):
s = PHASE_SETTINGS[phase]
system_msg = base_cfg.get("prompt", {}).get(
"system_message",
"You are an expert mathematician and algorithm designer.",
)
if phase == 2:
system_msg += PHASE2_ADDENDUM
eval_timeout = base_cfg.get("evaluator", {}).get("timeout", 360)
cfg = {
**PHASE_TEMPLATE,
"max_iterations": iterations,
"llm": {
"models": [{"name": model, "weight": 1.0}],
"api_base": api_base,
"api_key": "EMPTY",
"temperature": s["temperature"],
"top_p": 0.95,
"max_tokens": 8192,
"timeout": 300,
},
"prompt": {
"system_message": system_msg,
"num_top_programs": s["num_top_programs"],
"use_template_stochasticity": True,
},
"database": {
"population_size": s["population_size"],
"archive_size": s["archive_size"],
"num_islands": s["num_islands"],
"elite_selection_ratio": 0.3,
"exploitation_ratio": s["exploitation_ratio"],
"migration_interval": s["migration_interval"],
},
"evaluator": {
"timeout": eval_timeout,
"cascade_evaluation": True,
"cascade_thresholds": s["cascade_thresholds"],
"parallel_evaluations": 2,
"use_llm_feedback": False,
},
}
if "max_code_length" in s:
cfg["max_code_length"] = s["max_code_length"]
return cfg
def main():
p = argparse.ArgumentParser(description=__doc__)
p.add_argument("base_config", help="Path to problem's config.yaml")
p.add_argument("output_dir", help="Directory to write phase configs into")
p.add_argument("--model", required=True)
p.add_argument("--api-base", required=True)
p.add_argument("--phase1-iter", type=int, default=100)
p.add_argument("--phase2-iter", type=int, default=200)
args = p.parse_args()
with open(args.base_config) as f:
base = yaml.safe_load(f)
for phase, iters in [(1, args.phase1_iter), (2, args.phase2_iter)]:
cfg = build_config(base, phase, args.model, args.api_base, iters)
out = os.path.join(args.output_dir, f"config_phase_{phase}.yaml")
with open(out, "w") as f:
yaml.dump(cfg, f, default_flow_style=False, sort_keys=False)
if __name__ == "__main__":
main()