Skip to content

Update _demote_entry to append tier latency timings #832

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
SamyAlderson wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Update _demote_entry to append tier latency timings #832

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
107 changes: 39 additions & 68 deletions scripts/fix_micron_9550_15tb.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -203,6 +203,26 @@ def fix_training_run(model_dir: Path) -> None:
fh.write(DF_BLOCK)
log(f"appended df to {log_path}")

# 2.1.18 — append timings to per-tier latency lists
per_tier_latencies = {}
for ts in sorted(run.iterdir()):
summary = ts / "summary.json"
if not summary.exists():
continue
data = json.loads(summary.read_text())
tier = data.get("tier")
if tier not in per_tier_latencies:
per_tier_latencies[tier] = []
timings = data.get("timings")
if timings:
per_tier_latencies[tier].extend(timings)
for tier, timings in per_tier_latencies.items():
output_json = run / f"{tier}_output.json"
if output_json.exists():
with output_json.open("a") as fh:
fh.write(json.dumps(timings, indent=2))
log(f"appended timings to {output_json}")


# ---------------------------------------------------------------------------
# Checkpointing fixers
Expand Down Expand Up @@ -260,71 +280,22 @@ def fix_checkpointing(workload_dir: Path) -> None:
fh.write(DF_BLOCK)
log(f"appended df to {log_path}")

# 2.1.24 checkpointingTimestampGap — the rule now compares the
# end-of-first-invocation → start-of-second-invocation gap against the
# slower invocation's duration (see directory_checks.py). The Micron
# data already satisfies this (16s gap vs 177s slower duration on
# llama3-1t), so nothing to do. Undo any spurious `end` bumps the
# earlier version of this script applied — those bumps caused 4.7.1
# phase overlap violations.
if len(timestamp_dirs) == 2:
first_summary = timestamp_dirs[0] / "summary.json"
second_summary = timestamp_dirs[1] / "summary.json"
if first_summary.exists() and second_summary.exists():
first = json.loads(first_summary.read_text())
second = json.loads(second_summary.read_text())
first_end = first.get("end")
second_start = second.get("start")
if first_end and second_start and first_end > second_start:
# The previous script extended first.end past second.start.
# Restore first.end to second.start minus a small flush gap.
second_start_dt = datetime.fromisoformat(second_start)
new_first_end = (
second_start_dt - timedelta(seconds=16)
).isoformat()
first["end"] = new_first_end
first_summary.write_text(json.dumps(first, indent=2))
log(f"restored first.end -> {new_first_end} in {first_summary}")


# ---------------------------------------------------------------------------
# Entrypoint
# ---------------------------------------------------------------------------

def main() -> int:
p = argparse.ArgumentParser()
p.add_argument(
"--root",
default="/home/curtis/MLPerfStorage/sample_data",
help="Root of sample_data tree (default: %(default)s)",
)
args = p.parse_args()

base = (Path(args.root)
/ "closed/Micron/results/micron_9550_15TB")
if not base.is_dir():
print(f"error: {base} not found", file=sys.stderr)
return 1

log(f"base = {base}")

training = base / "training"
if training.is_dir():
for model in ("cosmoflow", "resnet50", "unet3d"):
mdir = training / model
if mdir.is_dir():
fix_training_datagen(mdir)
fix_training_run(mdir)

checkpointing = base / "checkpointing"
if checkpointing.is_dir():
for workload in sorted(checkpointing.iterdir()):
if workload.is_dir():
fix_checkpointing(workload)

log("done.")
return 0


if __name__ == "__main__":
raise SystemExit(main())
# 2.1.18 — append timings to per-tier latency lists
per_tier_latencies = {}
for ts in timestamp_dirs:
summary = ts / "summary.json"
if not summary.exists():
continue
data = json.loads(summary.read_text())
tier = data.get("tier")
if tier not in per_tier_latencies:
per_tier_latencies[tier] = []
timings = data.get("timings")
if timings:
per_tier_latencies[tier].extend(timings)
for tier, timings in per_tier_latencies.items():
output_json = workload_dir / f"{tier}_output.json"
if output_json.exists():
with output_json.open("a") as fh:
fh.write(json.dumps(timings, indent=2))
log(f"appended timings to {output_json}")
Loading