single_pickplace / scripts /check_checkpoints.py
simpk's picture
Update runbook and tooling
84af0a7 verified
Raw History Blame Contribute Delete
6.46 kB
"""Inventory and integrity-check the Orbax checkpoints written by a pi0.5 training run.
python3 scripts/check_checkpoints.py /data/checkpoints/pi05_g1_pickplace/pickplace_full_v1
Pure stdlib, no GPU, no JAX -- safe to run while training still holds the A100. It answers the
questions you cannot answer from `ls` alone:
* which steps exist, and which are resumable (train_state present) vs params-only (pruned)
* whether any save is INCOMPLETE -- an aborted or in-flight write leaves a
`.orbax-checkpoint-tmp-*` directory, or a step dir with no commit marker
* whether `assets/norm_stats.json` is present, because a checkpoint without it serves
garbage actions while looking perfectly healthy
* zero-byte files, which is what a checkpoint written into a full disk looks like
Exit code is 0 only if every complete checkpoint passes. `--json` emits machine-readable output.
"""
import argparse
import json
import pathlib
import sys
# Orbax drops one of these in a step directory once the write has been committed. Which one
# depends on the version, so presence of ANY is treated as "committed".
COMMIT_MARKERS = ("commit_success.txt", "_CHECKPOINT_METADATA", "_METADATA")
GIB = 1024**3
def human(n: int) -> str:
"""Adaptive units -- a 12.5 GiB params dir and a 4 KiB assets dir both need to read."""
for unit, div in (("T", 1024**4), ("G", GIB), ("M", 1024**2), ("K", 1024)):
if n >= div:
return f"{n / div:.1f}{unit}"
return f"{n}B"
def dir_stats(path: pathlib.Path) -> tuple[int, int, int]:
"""(total bytes, file count, zero-byte file count) over a directory tree."""
total = files = empty = 0
for f in path.rglob("*"):
if not f.is_file():
continue
try:
size = f.stat().st_size
except OSError:
continue
total += size
files += 1
empty += size == 0
return total, files, empty
def inspect(step_dir: pathlib.Path) -> dict:
r: dict = {"step": int(step_dir.name), "path": str(step_dir), "problems": []}
r["committed"] = any((step_dir / m).exists() for m in COMMIT_MARKERS) or any(
(step_dir / sub / m).exists() for sub in ("params",) for m in COMMIT_MARKERS
)
for sub in ("params", "train_state", "assets"):
d = step_dir / sub
if d.is_dir():
size, files, empty = dir_stats(d)
r[sub] = {"bytes": size, "files": files, "empty_files": empty}
if empty:
r["problems"].append(f"{sub}/ has {empty} zero-byte file(s) -- disk full?")
if files == 0:
r["problems"].append(f"{sub}/ exists but is empty")
else:
r[sub] = None
if r["params"] is None:
r["problems"].append("no params/ -- nothing deployable here")
# assets/ carries norm_stats.json. Without it the policy denormalises with the wrong
# statistics and the robot does nothing while every log looks fine.
hits = list((step_dir / "assets").rglob("norm_stats.json")) if (step_dir / "assets").is_dir() else []
r["norm_stats"] = str(hits[0].relative_to(step_dir)) if hits else None
if not hits:
r["problems"].append("no assets/**/norm_stats.json -- served policy would emit garbage")
r["resumable"] = r["train_state"] is not None
return r
def main() -> int:
ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("root", type=pathlib.Path, help="the exp directory, e.g. .../pickplace_full_v1")
ap.add_argument("--frames", type=int, default=161_440, help="dataset frames, for the epoch column")
ap.add_argument("--batch-size", type=int, default=32)
ap.add_argument("--json", action="store_true")
a = ap.parse_args()
if not a.root.is_dir():
print(f"FAIL: {a.root} is not a directory", file=sys.stderr)
return 2
partial = sorted({p.name for p in a.root.glob("*orbax-checkpoint-tmp-*")})
steps = sorted((p for p in a.root.iterdir() if p.is_dir() and p.name.isdigit()),
key=lambda p: int(p.name))
if not steps and not partial:
print(f"No checkpoints in {a.root} yet.")
return 0
rows = [inspect(s) for s in steps]
spe = a.frames / a.batch_size # steps per epoch
if a.json:
print(json.dumps({"root": str(a.root), "in_flight": partial, "checkpoints": rows}, indent=2))
else:
print(f"{a.root}\n")
hdr = f"{'step':>7} {'epoch':>5} {'params':>8} {'tstate':>8} {'norm_stats':>10} state"
print(hdr)
print("-" * len(hdr))
for r in rows:
p = human(r["params"]["bytes"]) if r["params"] else "--"
t = human(r["train_state"]["bytes"]) if r["train_state"] else "--"
n = "yes" if r["norm_stats"] else "NO"
state = "resumable" if r["resumable"] else "params only"
if r["problems"]:
state = "PROBLEM"
print(f"{r['step']:>7} {r['step']/spe:>5.2f} {p:>8} {t:>8} {n:>10} {state}")
total = sum((r["params"]["bytes"] if r["params"] else 0)
+ (r["train_state"]["bytes"] if r["train_state"] else 0) for r in rows)
print(f"\n{len(rows)} checkpoint(s), {human(total)} total")
if partial:
print(f"\n{len(partial)} save(s) IN FLIGHT or aborted (Orbax temp dirs):")
for p in partial:
print(f" {p}")
print(" A save in progress is normal. One that persists after training exits is not --")
print(" delete it, that step never committed.")
uncommitted = [r["step"] for r in rows if not r["committed"]]
if uncommitted:
print(f"\nNOTE: no commit marker found for step(s) {uncommitted}. Orbax layout varies by")
print("version, so this is weak evidence on its own -- trust the load test instead.")
bad = [r for r in rows if r["problems"]]
if bad:
print("\nPROBLEMS")
for r in bad:
for p in r["problems"]:
print(f" step {r['step']}: {p}")
else:
print("\nAll checkpoints structurally OK. Structure is not the same as loadable --")
print("run the load test in the runbook (Step 11.1) on at least one of them.")
return 1 if any(r["problems"] for r in rows) else 0
if __name__ == "__main__":
sys.exit(main())