Skip to content

Report Qwen3 all-reduce internal input latency #272

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

Merged
chhwang merged 8 commits into from
Jun 25, 2026
Merged

Report Qwen3 all-reduce internal input latency #272

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
0a4fc85
Extend the Q7 decode all-reduce benchmark to report external and inte…
chhwang Jun 23, 2026
0934da9
Extend the Q7 decode all-reduce benchmark to report external and inte…
chhwang Jun 23, 2026
98ebf38
Extend the Q7 decode all-reduce benchmark to report external and inte…
chhwang Jun 23, 2026
e24364b
Extend to report decode TP=2/TP=8 all-reduce and input modes.
chhwang Jun 23, 2026
2da573c
Format all-reduce benchmark parser tests
chhwang Jun 23, 2026
509e6dc
Extend the decode all-reduce benchmark to report external and interna…
chhwang Jun 23, 2026
bbe2187
Measure prefill all-reduce input modes in Q7I gate
chhwang Jun 24, 2026
2ef4126
Keep Q7I perf gate wrapper newline
chhwang Jun 24, 2026
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
69 changes: 29 additions & 40 deletions __perf_gate__.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,57 +1,46 @@
#!/usr/bin/env bash
set -uo pipefail

script_dir=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)
bench_py="$script_dir/examples/qwen3/bench_allreduce.py"
if [[ ! -f "$bench_py" ]]; then
echo "ERROR: missing benchmark: $bench_py" >&2
exit 1
fi

: "${ARK_ROOT:=$PWD}"
export ARK_ROOT
export PYTHONPATH="${PYTHONPATH:-$ARK_ROOT/python}"

target_ms=$(python3 - <<'PY'
import importlib.util
import pathlib

path = pathlib.Path("../examples/qwen3/bench_allreduce.py")
spec = importlib.util.spec_from_file_location("bench_allreduce", path)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
print(f"{module._DECODE_TARGET_MS:.4f}")
PY
)

tmpdir=$(mktemp -d)
trap 'rm -rf "$tmpdir"' EXIT
status=0
python3 ../examples/qwen3/bench_allreduce.py --world-size 2 --shape decode >"$tmpdir/tp2.log" 2>"$tmpdir/tp2.err" || status=1
python3 ../examples/qwen3/bench_allreduce.py --world-size 8 --shape decode >"$tmpdir/tp8.log" 2>"$tmpdir/tp8.err" || status=1
python3 "$bench_py" --world-size 2 --shape all --input-mode all \
>"$tmpdir/tp2.log" 2>"$tmpdir/tp2.err" || status=1
python3 "$bench_py" --world-size 8 --shape all --input-mode all \
>"$tmpdir/tp8.log" 2>"$tmpdir/tp8.err" || status=1

ark_ms=$(python3 - "$tmpdir/tp2.log" "$tmpdir/tp8.log" "$status" <<'PY'
import re
python3 - "$status" "$bench_py" "$tmpdir/tp2.log" "$tmpdir/tp8.log" <<'PY'
import importlib.util
import pathlib
import sys

values = []
for name in sys.argv[1:3]:
text = open(name, encoding="utf-8").read()
match = re.search(r"PERF_GATE name=allreduce\s+ark_ms=([0-9.]+)", text)
if match:
values.append(float(match.group(1)))
if int(sys.argv[3]) or len(values) != 2:
print("999999.0000")
else:
print(f"{max(values):.4f}")
PY
status = int(sys.argv[1])
path = pathlib.Path(sys.argv[2])
logs = sys.argv[3:]
spec = importlib.util.spec_from_file_location("bench_allreduce", path)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
target_ms = module._DECODE_TARGET_MS
texts = [pathlib.Path(log).read_text(encoding="utf-8") for log in logs]
ark_ms = (
999999.0 if status else module._decode_gate_ark_ms_from_logs(texts)
)
ratio=$(python3 - "$ark_ms" "$target_ms" <<'PY'
import sys

print(f"{float(sys.argv[1]) / float(sys.argv[2]):.4f}")
PY
ratio = ark_ms / target_ms
print(
f"PERF_GATE name={module._DECODE_GATE_NAME} ark_ms={ark_ms:.4f} "
f"sglang_ms={target_ms:.4f} ratio={ratio:.4f}"
)
printf 'PERF_GATE name=allreduce ark_ms=%s sglang_ms=%s ratio=%s\n' "$ark_ms" "$target_ms" "$ratio"
python3 - "$ark_ms" "$target_ms" "$status" <<'PY'
import sys

ark_ms = float(sys.argv[1])
target_ms = float(sys.argv[2])
status = int(sys.argv[3])
if status or ark_ms >= target_ms:
if status or ark_ms >= 999999.0 or ark_ms >= target_ms:
raise SystemExit(1)
PY
Loading