File size: 9,120 Bytes
30bf68a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 | from __future__ import annotations
import argparse
import asyncio
import os
from typing import Any
from openai import OpenAI
from env.environment import CICDDebuggerEnvironment, REQUIRED_TOOLS
from inference.metrics import EpisodeMetrics
from inference.model_wrapper import ModelWrapper, score_action_candidate
from inference.prompts import heuristic_action
from inference.visualize import save_metrics_json, save_reward_curve, save_success_rate_history
API_BASE_URL = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1")
MODEL_NAME = os.getenv("MODEL_NAME", "Qwen/Qwen2.5-72B-Instruct")
HF_TOKEN = os.getenv("HF_TOKEN")
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
API_KEY = OPENAI_API_KEY or HF_TOKEN
LOCAL_IMAGE_NAME = os.getenv("LOCAL_IMAGE_NAME")
DEFAULT_TASK_ID = os.getenv("MY_ENV_V4_TASK", "easy-command-typo")
DEFAULT_BENCHMARK = os.getenv("MY_ENV_V4_BENCHMARK", "cicd_debugger_env")
MAX_STEPS_DEFAULT = int(os.getenv("MAX_STEPS", "8"))
TEMPERATURE = float(os.getenv("TEMPERATURE", "0.2"))
MAX_TOKENS = int(os.getenv("MAX_TOKENS", "120"))
OFFLINE_INFERENCE = os.getenv("OFFLINE_INFERENCE", "0") == "1"
SUCCESS_SCORE_THRESHOLD = float(os.getenv("SUCCESS_SCORE_THRESHOLD", "0.1"))
def log_start(task: str, env_name: str, model: str) -> None:
print(f"[START] task={_single_line(task)} env={_single_line(env_name)} model={_single_line(model)}", flush=True)
def log_step(step: int, action: str, reward: float, done: bool, error: str | None) -> None:
done_val = str(done).lower()
error_val = _single_line(error) if error else "null"
action_val = _single_line(action)
print(f"[STEP] step={step} action={action_val} reward={reward:.2f} done={done_val} error={error_val}", flush=True)
def log_end(success: bool, steps: int, score: float, rewards: list[float]) -> None:
rewards_str = ",".join(f"{value:.2f}" for value in rewards)
print(f"[END] success={str(success).lower()} steps={steps} score={score:.3f} rewards={rewards_str}", flush=True)
def _single_line(value: Any) -> str:
return " ".join(str(value).replace("\n", " ").replace("\r", " ").split())
def _is_hacking_action(action_text: str) -> bool:
value = (action_text or "").lower()
patterns = (
"if: false",
"when: never",
"echo \"tests passed\"",
"echo 'tests passed'",
"exit 0",
"force success",
"status: success",
)
return any(token in value for token in patterns)
def _extract_error(info: dict[str, Any] | None) -> str | None:
if not info:
return None
error = info.get("error")
return str(error) if error else None
def _extract_observation_fields(observation: dict[str, Any]) -> tuple[str, str, list[str]]:
config_text = str(observation.get("config") or "")
error_message = str(observation.get("error_message") or "")
tools = [str(item) for item in (observation.get("available_tools") or REQUIRED_TOOLS)]
return config_text, error_message, tools
def _tool_from_action(action_text: str) -> str:
return str(action_text or "").split(":", 1)[0].strip().lower()
def _is_action_allowed(action_text: str, available_tools: list[str]) -> bool:
return _tool_from_action(action_text) in {tool.lower() for tool in available_tools}
def _normalize_action(action_text: str, available_tools: list[str], fallback: str) -> str:
action = str(action_text or "").strip()
if not action:
return fallback
aliases = {
"run_stage": "run_pipeline_stage",
"validate": "validate_fix",
"submit": "submit_solution",
"submit_fix": "submit_solution",
}
tool = _tool_from_action(action)
normalized_tool = aliases.get(tool, tool)
if normalized_tool != tool:
suffix = action.split(":", 1)[1].strip() if ":" in action else ""
action = f"{normalized_tool}: {suffix}" if suffix else normalized_tool
if _is_action_allowed(action, available_tools):
return action
return fallback
def _select_action(
model_wrapper: ModelWrapper,
step: int,
config_text: str,
error_message: str,
history: list[str],
available_actions: list[str],
policy_mode: str,
trajectories: int,
) -> str:
mode = (policy_mode or "imp").lower()
fallback = heuristic_action(config_text, error_message, available_actions, history)
if mode == "sft":
return _normalize_action(fallback, available_actions, fallback)
if mode == "direct":
action = model_wrapper.generate_action(
step=step,
config_text=config_text,
error_message=error_message,
history=history,
available_actions=available_actions,
)
return _normalize_action(action, available_actions, fallback)
candidates = model_wrapper.generate_candidates(
step=step,
config_text=config_text,
error_message=error_message,
history=history,
count=max(1, int(trajectories)),
available_actions=available_actions,
)
if not candidates:
return _normalize_action(fallback, available_actions, fallback)
observation_text = f"{config_text}\n{error_message}"
best = max(candidates, key=lambda item: score_action_candidate(observation_text, item, _is_hacking_action))
return _normalize_action(best, available_actions, fallback)
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="Run CI/CD debugger inference loop")
parser.add_argument("--max-steps", type=int, default=MAX_STEPS_DEFAULT)
parser.add_argument("--task", default=DEFAULT_TASK_ID)
parser.add_argument("--benchmark", default=DEFAULT_BENCHMARK)
parser.add_argument("--difficulty", choices=["easy", "medium", "hard"], default=None)
parser.add_argument("--offline", action="store_true", default=OFFLINE_INFERENCE)
parser.add_argument("--force-local-env", action="store_true", default=False)
parser.add_argument("--policy-mode", choices=["sft", "imp", "direct"], default="imp")
parser.add_argument("--trajectories", type=int, default=3)
return parser.parse_args()
async def run_episode(args: argparse.Namespace) -> int:
history: list[str] = []
steps_taken = 0
success = False
episode_completed_cleanly = False
metrics = EpisodeMetrics()
env = CICDDebuggerEnvironment(max_steps=max(1, int(args.max_steps)))
offline_mode = bool(args.offline or not API_KEY)
client: OpenAI | None = None
if not offline_mode:
try:
client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)
except Exception:
client = None
offline_mode = True
model_wrapper = ModelWrapper(
client=client,
model_name=MODEL_NAME,
temperature=TEMPERATURE,
max_tokens=MAX_TOKENS,
offline=offline_mode,
)
log_start(task=str(args.task), env_name=str(args.benchmark), model=MODEL_NAME)
try:
observation = await env.reset(task_id=str(args.task), difficulty=args.difficulty)
for step in range(1, max(1, int(args.max_steps)) + 1):
config_text, error_message, available_tools = _extract_observation_fields(observation)
action_text = _select_action(
model_wrapper=model_wrapper,
step=step,
config_text=config_text,
error_message=error_message,
history=history,
available_actions=available_tools,
policy_mode=str(args.policy_mode),
trajectories=max(1, int(args.trajectories)),
)
observation, reward, done, info = await env.step(action_text)
step_error = _extract_error(info)
metrics.add_step(action=action_text, reward=float(reward), error=step_error, done=bool(done))
steps_taken = step
log_step(step=step, action=action_text, reward=float(reward), done=bool(done), error=step_error)
history.append(f"step={step} action={_single_line(action_text)} reward={float(reward):.2f}")
if done:
episode_completed_cleanly = step_error is None and not _is_hacking_action(action_text)
break
except Exception as exc:
success = False
if not metrics.rewards:
metrics.add_step(action="system_error", reward=0.0, error=str(exc), done=True)
finally:
score = max(0.0, min(1.0, float(metrics.average_reward)))
success = episode_completed_cleanly and score >= SUCCESS_SCORE_THRESHOLD
try:
save_reward_curve(metrics.rewards)
save_metrics_json(metrics.summary())
save_success_rate_history([success])
except Exception:
pass
try:
await env.close()
except Exception:
pass
log_end(success=success, steps=steps_taken, score=score, rewards=metrics.rewards)
return 0
def main() -> int:
args = parse_args()
return asyncio.run(run_episode(args))
if __name__ == "__main__":
raise SystemExit(main())
|