andylizf's picture
Upload folder using huggingface_hub
5fed0fc verified
"""
SkyPilot runner for research problems.
Runs evaluations on cloud VMs via SkyPilot.
Supports two result storage modes:
- scp (legacy): Fetch results via scp after job completes
- bucket: Write results directly to S3/GCS bucket during job execution
"""
import hashlib
import json
import shutil
import subprocess
import tempfile
import textwrap
import time
from datetime import datetime
from pathlib import Path
from typing import Optional, Tuple
from .base import Runner, EvaluationResult, EvaluationStatus
from ..config import load_runtime_config
def _sanitize_name(name: str) -> str:
"""Sanitize a name for use as cluster name."""
cleaned = []
valid = "abcdefghijklmnopqrstuvwxyz0123456789-"
last_dash = False
for ch in name.lower():
if ch in valid:
cleaned.append(ch)
last_dash = ch == "-"
else:
if not last_dash:
cleaned.append("-")
last_dash = True
return "".join(cleaned).strip("-") or "job"
class SkyPilotRunner(Runner):
"""
Runner for research problems using SkyPilot.
Executes evaluations on cloud VMs with support for:
- Auto-scaling resources based on problem requirements
- GPU provisioning
- Custom Docker images (from config.yaml)
"""
DEFAULT_CLOUD = "gcp"
DEFAULT_CPUS = "8+"
DEFAULT_MEMORY = "16+"
DEFAULT_DISK_SIZE = 50
DEFAULT_GPU = "L4:1"
DEFAULT_TIMEOUT = 1800 # 30 minutes
def __init__(
self,
base_dir: Optional[Path] = None,
cloud: str = DEFAULT_CLOUD,
region: Optional[str] = None,
keep_cluster: bool = False,
bucket_url: Optional[str] = None,
):
"""
Initialize SkyPilotRunner.
Args:
base_dir: Base directory of Frontier-CS repo
cloud: Cloud provider (gcp, aws, azure)
region: Cloud region (optional)
keep_cluster: Keep cluster running after evaluation
bucket_url: Optional bucket URL for result storage (s3://... or gs://...)
If provided, results are written to bucket instead of fetched via scp
"""
self.base_dir = base_dir or self._find_base_dir()
self.research_dir = self.base_dir / "research"
self.cloud = cloud
self.region = region
self.keep_cluster = keep_cluster
self.bucket_url = bucket_url
def _find_base_dir(self) -> Path:
"""Find the Frontier-CS base directory."""
candidates = [
Path(__file__).parents[4],
Path.cwd(),
Path.cwd().parent,
]
for candidate in candidates:
if (candidate / "research").is_dir() and (candidate / "pyproject.toml").exists():
return candidate
raise RuntimeError("Could not find Frontier-CS base directory")
def get_problem_path(self, problem_id: str) -> Path:
"""Get the path to a research problem directory."""
return self.research_dir / "problems" / problem_id
def evaluate(
self,
problem_id: str,
solution_code: str,
*,
timeout: Optional[int] = None,
solution_id: Optional[str] = None,
) -> EvaluationResult:
"""
Evaluate a solution using SkyPilot.
Args:
problem_id: Problem ID (e.g., "flash_attn")
solution_code: Python solution code
timeout: Optional timeout in seconds
solution_id: Optional solution ID for bucket storage (forms pair_id with problem_id)
Returns:
EvaluationResult with score and status
"""
problem_path = self.get_problem_path(problem_id)
if not problem_path.exists():
return EvaluationResult(
problem_id=problem_id,
status=EvaluationStatus.ERROR,
message=f"Problem not found: {problem_path}",
)
# Create temp directory with solution
with tempfile.TemporaryDirectory(prefix="frontier_sky_") as temp_dir:
temp_path = Path(temp_dir)
solution_path = temp_path / "solution.py"
solution_path.write_text(solution_code, encoding="utf-8")
return self._run_evaluation(problem_id, problem_path, solution_path, timeout, solution_id)
def evaluate_file(
self,
problem_id: str,
solution_path: Path,
*,
timeout: Optional[int] = None,
solution_id: Optional[str] = None,
) -> EvaluationResult:
"""Evaluate a solution file using SkyPilot."""
if not solution_path.exists():
return EvaluationResult(
problem_id=problem_id,
status=EvaluationStatus.ERROR,
message=f"Solution file not found: {solution_path}",
)
problem_path = self.get_problem_path(problem_id)
if not problem_path.exists():
return EvaluationResult(
problem_id=problem_id,
status=EvaluationStatus.ERROR,
message=f"Problem not found: {problem_path}",
)
return self._run_evaluation(problem_id, problem_path, solution_path, timeout, solution_id)
def _run_evaluation(
self,
problem_id: str,
problem_path: Path,
solution_path: Path,
timeout: Optional[int],
solution_id: Optional[str] = None,
) -> EvaluationResult:
"""Run evaluation on SkyPilot."""
import sky
start_time = time.time()
# Load config from config.yaml
runtime_config = load_runtime_config(problem_path)
docker_config = runtime_config.docker
res = runtime_config.resources
# Determine resources
accelerators = res.accelerators
if not accelerators and docker_config.gpu:
accelerators = self.DEFAULT_GPU
if not accelerators and runtime_config.requires_gpu:
accelerators = self.DEFAULT_GPU
# Determine timeout
effective_timeout = timeout or runtime_config.timeout_seconds or self.DEFAULT_TIMEOUT
# Create cluster name
digest = hashlib.md5(problem_id.encode()).hexdigest()[:8]
cluster_name = _sanitize_name(f"eval-{problem_id}-{digest}")[:63]
# Build pair_id for bucket storage
pair_id = f"{solution_id}:{problem_id}" if solution_id else None
# Create workspace and task
with tempfile.TemporaryDirectory(prefix="frontier_sky_workspace_") as workspace_dir:
workspace = Path(workspace_dir)
file_mounts = self._setup_mounts(workspace, problem_id, problem_path, solution_path)
# Add bucket mount if using bucket storage
if self.bucket_url:
results_url = f"{self.bucket_url.rstrip('/')}/results"
file_mounts["~/results_bucket"] = {
"source": results_url,
"mode": "MOUNT",
}
# Build SkyPilot resources
resources = sky.Resources(
cloud=res.cloud or self.cloud,
region=res.region or self.region,
cpus=res.cpus or self.DEFAULT_CPUS,
memory=res.memory or self.DEFAULT_MEMORY,
accelerators=accelerators,
disk_size=res.disk_size or self.DEFAULT_DISK_SIZE,
instance_type=res.instance_type,
image_id=res.image_id,
)
# Build task
run_script = self._get_run_script(
problem_id,
docker_config.image,
docker_config.gpu,
docker_config.dind,
pair_id=pair_id if self.bucket_url else None,
)
task = sky.Task(
name=cluster_name,
setup=self._get_setup_script(),
run=run_script,
file_mounts=file_mounts,
)
task.set_resources(resources)
# Launch and wait
try:
request_id = sky.launch(task, cluster_name=cluster_name)
result = sky.stream_and_get(request_id)
job_id = result[0] if isinstance(result, tuple) and len(result) > 0 else None
handle = result[1] if isinstance(result, tuple) and len(result) > 1 else None
# Wait for completion
exit_code = 0
if job_id is not None:
exit_code = sky.tail_logs(cluster_name, job_id, follow=True)
duration = time.time() - start_time
if exit_code != 0:
return EvaluationResult(
problem_id=problem_id,
status=EvaluationStatus.ERROR,
message=f"Remote job failed with exit code {exit_code}",
duration_seconds=duration,
)
# Fetch results (bucket mode writes directly, scp mode fetches after)
if self.bucket_url:
# Results already written to bucket by run script
# Return placeholder - caller should read from bucket
return EvaluationResult(
problem_id=problem_id,
status=EvaluationStatus.SUCCESS,
message="Results written to bucket",
duration_seconds=duration,
)
else:
# Legacy scp mode
score, logs = self._fetch_results(cluster_name, handle)
return EvaluationResult(
problem_id=problem_id,
score=score,
status=EvaluationStatus.SUCCESS,
logs=logs,
duration_seconds=duration,
)
except Exception as e:
return EvaluationResult(
problem_id=problem_id,
status=EvaluationStatus.ERROR,
message=str(e),
duration_seconds=time.time() - start_time,
)
finally:
if not self.keep_cluster:
try:
down_request = sky.down(cluster_name)
sky.stream_and_get(down_request)
except Exception:
pass
def _setup_mounts(
self,
workspace: Path,
problem_id: str,
problem_path: Path,
solution_path: Path,
) -> dict:
"""Set up file mounts for SkyPilot."""
mounts = {}
remote_base = "~/sky_workdir"
# Mount problem
mounts[f"{remote_base}/research/{problem_id}"] = str(problem_path.resolve())
# Mount common directories
parts = problem_id.split("/")
for i in range(1, len(parts)):
parent = "/".join(parts[:i])
common_dir = self.research_dir / "problems" / parent / "common"
if common_dir.is_dir():
mounts[f"{remote_base}/research/{parent}/common"] = str(common_dir.resolve())
# Mount solution
solution_dir = workspace / "solution"
solution_dir.mkdir(parents=True)
shutil.copy2(solution_path, solution_dir / "solution.py")
mounts[f"{remote_base}/solution"] = str(solution_dir.resolve())
return mounts
def _get_setup_script(self) -> str:
"""Get setup script for SkyPilot task."""
return textwrap.dedent("""\
set -euo pipefail
# Install Docker
if ! command -v docker &>/dev/null; then
curl -fsSL https://get.docker.com | sudo sh
sudo usermod -aG docker $USER
sudo systemctl start docker
fi
# Make scripts executable
find ~/sky_workdir -name '*.sh' -exec chmod +x {} \\; 2>/dev/null || true
""")
def _get_run_script(
self,
problem_id: str,
docker_image: str,
gpu: bool,
dind: bool,
pair_id: Optional[str] = None,
) -> str:
"""Get run script for SkyPilot task."""
gpu_flags = "--gpus all" if gpu else ""
dind_flags = '-v /var/run/docker.sock:/var/run/docker.sock' if dind else ""
# Build bucket write command if pair_id is provided
if pair_id:
# Escape pair_id for shell and generate safe filename
safe_pair_id = pair_id.replace(":", "__")
bucket_write = textwrap.dedent(f'''
# Write result to bucket as JSON
SCORE=$(cat /results/score.txt 2>/dev/null || echo "")
TIMESTAMP=$(date -Is)
cat > ~/results_bucket/{safe_pair_id}.json << RESULT_EOF
{{
"pair_id": "{pair_id}",
"score": ${{SCORE:-null}},
"status": "success",
"message": null,
"duration_seconds": $SECONDS,
"timestamp": "$TIMESTAMP",
"logs": null
}}
RESULT_EOF
echo "Result written to bucket: {safe_pair_id}.json"
''')
else:
bucket_write = ""
return textwrap.dedent(f"""\
set -euo pipefail
SECONDS=0
cd ~/sky_workdir
# Create results directory
mkdir -p results
# Run evaluation in Docker
docker run --rm {gpu_flags} {dind_flags} \\
-v "$(pwd):/workspace:ro" \\
-v "$(pwd)/results:/results" \\
-w /work \\
"{docker_image}" \\
bash -c '
set -euo pipefail
cp -r /workspace/* /work/
cd /work/research/{problem_id}
# Setup environment
if [ -f set_up_env.sh ]; then
chmod +x set_up_env.sh
./set_up_env.sh
fi
# Copy solution to execution environment
mkdir -p /work/execution_env/solution_env
cp /work/solution/solution.py /work/execution_env/solution_env/
# Run evaluation
chmod +x evaluate.sh
./evaluate.sh | tee /results/output.txt
# Extract score (last numeric line)
grep -E "^-?[0-9]+\\.?[0-9]*$" /results/output.txt | tail -1 > /results/score.txt || true
'
{bucket_write}
""")
def _fetch_results(self, cluster_name: str, handle: object) -> Tuple[Optional[float], Optional[str]]:
"""Fetch results from remote cluster via scp."""
score = None
logs = None
with tempfile.TemporaryDirectory(prefix="frontier_results_") as temp_dir:
temp_path = Path(temp_dir)
# Try to scp results
try:
result = subprocess.run(
["scp", "-r", "-o", "StrictHostKeyChecking=no",
f"{cluster_name}:~/sky_workdir/results", str(temp_path)],
capture_output=True,
text=True,
timeout=60,
)
if result.returncode == 0:
results_dir = temp_path / "results"
# Read score
score_file = results_dir / "score.txt"
if score_file.exists():
score_text = score_file.read_text().strip()
if score_text:
try:
score = float(score_text)
except ValueError:
pass
# Read logs
output_file = results_dir / "output.txt"
if output_file.exists():
logs = output_file.read_text()
except (subprocess.TimeoutExpired, Exception):
pass
return score, logs