Recover active training loop checkpoints
This commit is contained in:
@@ -0,0 +1,98 @@
|
||||
#!/usr/bin/env python3
|
||||
"""Keep a checkpoint-aware training-loop parent alive across container recreation."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import subprocess
|
||||
import time
|
||||
from datetime import UTC, datetime
|
||||
from pathlib import Path
|
||||
|
||||
TERMINAL_STATUSES = {"training_complete"}
|
||||
|
||||
|
||||
def read_loop_status(path: Path) -> str | None:
|
||||
if not path.is_file():
|
||||
return None
|
||||
try:
|
||||
value = json.loads(path.read_text(encoding="utf-8"))
|
||||
except (OSError, json.JSONDecodeError):
|
||||
return "invalid"
|
||||
return value.get("status") if isinstance(value, dict) else "invalid"
|
||||
|
||||
|
||||
def process_active(container: str, marker: str) -> bool:
|
||||
result = subprocess.run(
|
||||
["docker", "top", container, "-eo", "pid,args"],
|
||||
capture_output=True, text=True, check=False,
|
||||
)
|
||||
return result.returncode == 0 and any(marker in line for line in result.stdout.splitlines())
|
||||
|
||||
|
||||
def write_state(path: Path, state: dict) -> None:
|
||||
state["updated_at"] = datetime.now(UTC).isoformat()
|
||||
temporary = path.with_suffix(path.suffix + ".tmp")
|
||||
temporary.write_text(json.dumps(state, indent=2), encoding="utf-8")
|
||||
temporary.replace(path)
|
||||
|
||||
|
||||
def main() -> int:
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--container", required=True)
|
||||
parser.add_argument("--process-marker", required=True)
|
||||
parser.add_argument("--loop-state", type=Path, required=True)
|
||||
parser.add_argument("--command-json", type=Path, required=True)
|
||||
parser.add_argument("--host-script", type=Path, required=True)
|
||||
parser.add_argument("--container-script", required=True)
|
||||
parser.add_argument("--state", type=Path, required=True)
|
||||
parser.add_argument("--poll-seconds", type=int, default=30)
|
||||
parser.add_argument("--max-launches", type=int, default=20)
|
||||
parser.add_argument("--once", action="store_true")
|
||||
args = parser.parse_args()
|
||||
state = {"schema_version": 1, "status": "monitoring", "launch_count": 0}
|
||||
if args.state.is_file():
|
||||
state.update(json.loads(args.state.read_text(encoding="utf-8")))
|
||||
while True:
|
||||
loop_status = read_loop_status(args.loop_state)
|
||||
state["loop_status"] = loop_status
|
||||
if loop_status in TERMINAL_STATUSES:
|
||||
state["status"] = "loop_finished"
|
||||
write_state(args.state, state)
|
||||
return 0
|
||||
if loop_status == "invalid":
|
||||
state["status"] = "invalid_loop_state"
|
||||
write_state(args.state, state)
|
||||
return 2
|
||||
if not process_active(args.container, args.process_marker):
|
||||
if int(state["launch_count"]) >= args.max_launches:
|
||||
state["status"] = "launch_budget_exhausted"
|
||||
write_state(args.state, state)
|
||||
return 3
|
||||
command = json.loads(args.command_json.read_text(encoding="utf-8"))
|
||||
if not isinstance(command, list) or not command or not all(isinstance(x, str) and x for x in command):
|
||||
state["status"] = "invalid_command"
|
||||
write_state(args.state, state)
|
||||
return 4
|
||||
copied = subprocess.run(
|
||||
["docker", "cp", str(args.host_script), f"{args.container}:{args.container_script}"],
|
||||
check=False,
|
||||
)
|
||||
launched = subprocess.run(command, check=False) if copied.returncode == 0 else copied
|
||||
if launched.returncode != 0:
|
||||
state["status"] = "launch_failed"
|
||||
write_state(args.state, state)
|
||||
return 5
|
||||
state["launch_count"] = int(state["launch_count"]) + 1
|
||||
state["status"] = "loop_relaunched"
|
||||
else:
|
||||
state["status"] = "monitoring"
|
||||
write_state(args.state, state)
|
||||
if args.once:
|
||||
return 0
|
||||
time.sleep(args.poll_seconds)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
raise SystemExit(main())
|
||||
Reference in New Issue
Block a user