ounce100m-code / kernels /p4_probe2_soak_runner.py
Cion-lab's picture
p4_probe2_soak_runner: the wrapper as submitted for probe chain v8 / soak / rehearsal v4
7fa7ceb verified
Raw History Blame Contribute Delete
1.13 kB
# Kaggle kernel body for P4 probe 2, the soak (2x T4, ~1.2 h). Same pinned body as the smoke, one
# environment difference: MODE=soak, which runs 180 steps at the real step shape without gradient
# checkpointing, through three push/verify/prune cycles, a forced cold resume and the horizon's
# validation pass -- and gates on the memory of the leg that never evaluates. That gate is D-018.
import hashlib, os, runpy, subprocess, sys
os.chdir("/kaggle/working")
sys.path.insert(0, "/kaggle/working")
REV = "add04ae7a24bd9925115773a93e870fc1dd01f59"
SHA = "f55bd29c94deff0a9309e41be2818c49db6720bf871cd60cd43de654810bfaf2"
MODE = "soak"
url = f"https://huggingface.co/Cion-lab/ounce100m-code/resolve/{REV}/kernels/p4_stop_probe.py"
assert subprocess.run(["curl", "-sfL", url, "-o", "p4_stop_probe.py"]).returncode == 0, "fetch probe"
got = hashlib.sha256(open("p4_stop_probe.py", "rb").read()).hexdigest()
assert got == SHA, ("PROBE SHA MISMATCH", got[:16], SHA[:16])
os.environ["P4_PROBE_MODE"] = MODE
print("PROBE_BOOTSTRAP_OK", REV[:12], got[:12], "mode", MODE, flush=True)
runpy.run_path("p4_stop_probe.py", run_name="__main__")