File size: 2,783 Bytes
9de6be7 | 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 | """Replay example actions over /ws and assert every episode ends done=True with reward 0.5.
Usage: python replay.py ws://127.0.0.1:8000/ws [log.json]
"""
import asyncio
import json
import sys
from datetime import datetime, timezone
import websockets
TASK_IDS = ["null"]
ACTIONS = [
{"answer": "done"}, # example action (submission example_actions[0])
{}, # empty object
{"answer": "42", "tool": "calc", "x": 1}, # stray keys a policy may emit
{"answer": 0}, # non-string answer
{"answer": None}, # null answer
]
# Replies that fail the schema never reach step(): the arena ends the episode and
# scores it through finish_action (guide, "How the model sees your environment").
REWARD = 0.5
async def episode(url, task_id, action):
async with websockets.connect(url) as ws:
await ws.send(json.dumps({"type": "reset", "data": {"task_id": task_id}}))
reset = json.loads(await ws.recv())
await ws.send(json.dumps({"type": "step", "data": action}))
step = json.loads(await ws.recv())
await ws.send(json.dumps({"type": "close"}))
return {"task_id": task_id, "action": action, "reset": reset, "step": step}
async def main(url, log_path):
results = []
for task_id in TASK_IDS:
for action in ACTIONS:
r = await episode(url, task_id, action)
data = r["step"].get("data", {})
r["ok"] = (r["reset"].get("type") == "observation" and r["step"].get("type") == "observation"
and data.get("done") is True and data.get("reward") == REWARD)
results.append(r)
print(("OK " if r["ok"] else "FAIL"), task_id, json.dumps(action), "->", json.dumps(data))
# One GRPO group: 5 rollouts of the same task in parallel sessions.
group = await asyncio.gather(*(episode(url, TASK_IDS[0], ACTIONS[0]) for _ in range(5)))
rewards = [g["step"].get("data", {}).get("reward") for g in group]
group_ok = all(g["step"].get("data", {}).get("done") is True for g in group) and rewards == [REWARD] * 5
print(("OK " if group_ok else "FAIL"), "concurrent group of 5 rewards", rewards)
results.append({"task_id": TASK_IDS[0], "concurrent_group_rewards": rewards, "ok": group_ok})
log = {"url": url, "at_utc": datetime.now(timezone.utc).isoformat(timespec="seconds"),
"episodes": results, "all_ok": all(r["ok"] for r in results)}
if log_path:
with open(log_path, "w") as f:
json.dump(log, f, indent=1)
f.write("\n")
return 0 if log["all_ok"] else 1
if __name__ == "__main__":
sys.exit(asyncio.run(main(sys.argv[1], sys.argv[2] if len(sys.argv) > 2 else None)))
|