File size: 13,387 Bytes
7f611c5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 | """Containerized evaluator: runs evaluate.sh inside a persistent Docker container."""
import asyncio
import json
import logging
import os
import subprocess
import time
import uuid
from typing import Dict, List, Optional, Tuple
from skydiscover.config import EvaluatorConfig
from skydiscover.evaluation.evaluation_result import EvaluationResult
from skydiscover.utils.async_utils import TaskPool
from skydiscover.utils.metrics import format_metrics
logger = logging.getLogger(__name__)
class ContainerizedEvaluator:
"""Evaluates programs by running them inside a persistent Docker container.
The benchmark directory must contain:
- Dockerfile
- evaluate.sh (called as: evaluate.sh <solution_path> <mode>)
Any data files or other resources needed by evaluate.sh, such as a
requirements.txt or data files, are the benchmark's own concern — the
framework imposes no structure on them.
evaluate.sh receives two arguments:
1. ``<solution_path>`` — absolute path to the candidate program inside
the container (e.g. ``/tmp/candidate_abc123.py``).
2. ``<mode>`` — either ``"train"`` or ``"test"``.
- **train**: called during the optimization loop in the process
of iterating towards a single solution. This may be called multiple
times per program, thus should be relatively fast.
- **test**: called at publish time (e.g. end-of-run best program).
Should be the authoritative, full evaluation, which will be used
for reporting and leaderboard ranking.
Evaluators that don't need the distinction can ignore the mode.
evaluate.sh writes a single JSON object to stdout::
{
"status": "success" | "error" | "timeout",
"combined_score": <float>,
"metrics": {<str>: <float>},
"artifacts": {<str>: <str>} // optional
}
Exit codes:
0 — evaluation completed (score may still reflect failure)
1 — evaluator itself crashed (infrastructure problem)
The image is built once at init time (Docker's layer cache makes
subsequent builds near-instant when nothing changed).
A single container is started at init time and reused across evaluations.
Each evaluation injects its candidate file via stdin (no host filesystem
dependency) and runs evaluate.sh with docker exec. Concurrent evaluations
are safe because each uses a unique path inside the container's /tmp.
Design note: ``_run_single_in_container`` is intentionally a plain method
(not async) so it can be overridden by adapters targeting other container
interfaces (e.g. Harbor's /solution + /logs/verifier/reward.json).
"""
def __init__(
self,
benchmark_dir: str,
config: EvaluatorConfig,
max_concurrent: int = 4,
env_vars: Optional[Dict[str, str]] = None,
):
self.benchmark_dir = os.path.abspath(benchmark_dir)
self.config = config
self.program_suffix = config.file_suffix
self.task_pool = TaskPool(max_concurrency=max_concurrent)
self.llm_judge = None
self.env_vars = dict(env_vars or {})
if self.env_vars:
logger.info(
f"Passing {len(self.env_vars)} environment variables to container: {list(self.env_vars.keys())}"
)
self.image_tag = self._build_image()
self.container_id = self._start_container()
logger.info(f"ContainerizedEvaluator ready: container={self.container_id[:12]}")
def close(self):
"""Stop and remove the persistent container."""
cid = getattr(self, "container_id", None)
if cid:
try:
logger.info(f"Stopping container {cid[:12]}...")
subprocess.run(
["docker", "stop", cid],
capture_output=True,
timeout=30,
check=True,
)
except subprocess.TimeoutExpired:
logger.warning(f"Timed out stopping container {cid[:12]}, killing...")
try:
subprocess.run(["docker", "kill", cid], capture_output=True, timeout=10)
except Exception:
logger.warning(f"Failed to kill container {cid[:12]}", exc_info=True)
except Exception:
logger.warning(f"Failed to stop container {cid[:12]}", exc_info=True)
finally:
self.container_id = None
def __enter__(self):
return self
def __exit__(self, exc_type, exc_val, exc_tb):
self.close()
def __del__(self):
"""Safety net: stop the container if close() was never called."""
try:
self.close()
except Exception:
pass
# ------------------------------------------------------------------
# Public API — mirrors Evaluator's interface
# ------------------------------------------------------------------
async def evaluate_program(
self,
program_solution: str,
program_id: str = "",
mode: str = "train",
) -> EvaluationResult:
"""Evaluate one candidate program and return scores.
Args:
program_solution: Source code (or path, for image mode) of the candidate.
program_id: Optional identifier for logging.
mode: ``"train"`` for hot-loop evaluation, ``"test"`` for
authoritative/publish evaluation.
"""
start_time = time.time()
label = f" {program_id}" if program_id else ""
last_exception = None
for attempt in range(self.config.max_retries + 1):
try:
result = await asyncio.wait_for(
asyncio.get_running_loop().run_in_executor(
None, self._run_container, program_solution, mode
),
timeout=self.config.timeout,
)
elapsed = time.time() - start_time
logger.info(
f"Evaluated program{label} [{mode}] in {elapsed:.2f}s:"
f" {format_metrics(result.metrics)}"
)
return result
except asyncio.TimeoutError:
logger.error(f"Container timed out after {self.config.timeout}s{label}")
return EvaluationResult(metrics={"error": 0.0, "timeout": True})
except Exception as e:
last_exception = e
logger.warning(
f"Attempt {attempt + 1}/{self.config.max_retries + 1} failed{label}: {e}"
)
if attempt < self.config.max_retries:
await asyncio.sleep(1.0)
logger.error(f"All attempts failed{label}: {last_exception}")
return EvaluationResult(metrics={"error": 0.0})
async def evaluate_batch(
self,
programs: List[Tuple[str, str]],
) -> List[EvaluationResult]:
"""Evaluate multiple programs concurrently.
Args:
programs: List of (solution, program_id) tuples.
Returns:
Results in the same order as *programs*.
"""
return await self.task_pool.gather(
coros=[self.evaluate_program] * len(programs),
args_list=list(programs),
)
# ------------------------------------------------------------------
# Container interaction — override for alternative interfaces
# ------------------------------------------------------------------
def _run_container(self, program_solution: str, mode: str) -> EvaluationResult:
"""Inject the candidate program and run evaluate.sh inside the container.
Uses a unique /tmp path per call so concurrent evaluations don't collide.
Override this method to target a different container interface
(e.g. Harbor: cp to /solution/, read reward from /logs/verifier/reward.json).
"""
candidate_path = self._inject_file(program_solution, self.program_suffix)
try:
return self._run_single_in_container(candidate_path, mode)
finally:
self._remove_file(candidate_path)
def _run_single_in_container(self, candidate_path: str, mode: str) -> EvaluationResult:
"""Execute evaluate.sh inside the container and parse its JSON output."""
try:
# Build docker exec command with environment variables
cmd = ["docker", "exec"]
for key, value in self.env_vars.items():
cmd.extend(["-e", f"{key}={value}"])
cmd.extend(
[
self.container_id,
"/benchmark/evaluate.sh",
candidate_path,
mode,
]
)
proc = subprocess.run(
cmd,
capture_output=True,
text=True,
timeout=self.config.timeout,
)
except subprocess.TimeoutExpired:
logger.error(f"docker exec timed out after {self.config.timeout}s")
return EvaluationResult(
metrics={"error": 0.0, "timeout": True},
artifacts={"error": f"docker exec timed out after {self.config.timeout}s"},
)
if proc.returncode != 0:
logger.error(f"Evaluator exited with code {proc.returncode}:\n{proc.stderr}")
return EvaluationResult(
metrics={"error": 0.0},
artifacts={"stderr": proc.stderr, "exit_code": str(proc.returncode)},
)
result = self._parse_output(proc.stdout)
# Always surface stderr (e.g. warnings, partial tracebacks) even on
# successful exit — the evaluator may have caught the error internally
# and returned valid JSON, but stderr still has useful context.
if proc.stderr.strip():
result.artifacts.setdefault("stderr", proc.stderr)
return result
# ------------------------------------------------------------------
# Helpers
# ------------------------------------------------------------------
def _inject_file(self, content: str, suffix: str) -> str:
"""Write content to a unique temp file inside the container via stdin."""
path = f"/tmp/{uuid.uuid4().hex}{suffix}"
inject = subprocess.run(
["docker", "exec", "-i", self.container_id, "tee", path],
input=content.encode(),
capture_output=True,
)
if inject.returncode != 0:
raise RuntimeError(f"Failed to inject file into container: {inject.stderr.decode()}")
return path
def _remove_file(self, path: str) -> None:
"""Remove a file inside the container."""
subprocess.run(
["docker", "exec", self.container_id, "rm", "-f", path],
capture_output=True,
)
def _parse_output(self, stdout: str) -> EvaluationResult:
try:
data = json.loads(stdout.strip())
except json.JSONDecodeError as e:
logger.error(f"Failed to parse evaluator JSON: {e}\nOutput: {stdout!r}")
return EvaluationResult(
metrics={"error": 0.0},
artifacts={"raw_output": stdout},
)
status = data.get("status", "error")
combined_score = float(data.get("combined_score", 0.0))
metrics = {
k: float(v) for k, v in data.get("metrics", {}).items() if isinstance(v, (int, float))
}
if "combined_score" not in metrics:
metrics["combined_score"] = combined_score
artifacts = {k: str(v) for k, v in data.get("artifacts", {}).items()}
if status != "success":
artifacts.setdefault("status", status)
return EvaluationResult(metrics=metrics, artifacts=artifacts)
def _start_container(self) -> str:
"""Start a persistent container and return its ID."""
# Build docker run command with environment variables
cmd = ["docker", "run", "-d", "--rm"]
for key, value in self.env_vars.items():
cmd.extend(["-e", f"{key}={value}"])
cmd.extend(["--entrypoint", "sleep", self.image_tag, "infinity"])
result = subprocess.run(
cmd,
capture_output=True,
text=True,
check=True,
)
return result.stdout.strip()
def _build_image(self) -> str:
norm = os.path.normpath(self.benchmark_dir)
name = os.path.basename(norm)
# Include parent dir to avoid tag collisions when multiple benchmarks
# share the same leaf directory name (e.g. "evaluator").
parent = os.path.basename(os.path.dirname(norm))
if parent and name == "evaluator":
name = f"{parent}-{name}"
tag = f"skydiscover-{name}:latest"
logger.info(f"Building Docker image: {tag} (from {self.benchmark_dir})")
result = subprocess.run(
["docker", "build", "-t", tag, self.benchmark_dir],
capture_output=True,
text=True,
)
if result.returncode != 0:
raise RuntimeError(f"Docker build failed for {self.benchmark_dir}:\n{result.stderr}")
return tag
|