Ascend/msagent / skills/profiler/msagent-profiler-breakdown
msagent-profiler-breakdown
VerifiedDB格式Profiling数据场景化性能拆解:自动完成场景确认(框架/任务类型判定)、阶段拆解与单次执行边界拆解,并给出报告。Use when 需要对昇腾 Profiling DB 做阶段拆解、step 归因或生成性能报告, or when the user says "帮我按任务阶段拆解这个Profiling数据", "多层次性能耗时拆解", "分析这个 vllm/sglang/verl 数据,主要耗时都花费在哪里了"。
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 8.0 KB |
| references/breakdown-schema.md | markdown | 4.1 KB |
| references/check-rules.md | markdown | 1.5 KB |
| references/db-usage.md | markdown | 5.8 KB |
| references/frameworks/README.md | markdown | 5.5 KB |
| references/frameworks/verl.md | markdown | 11.9 KB |
| references/frameworks/vllm-instrumentation.md | markdown | 13.6 KB |
| references/frameworks/vllm-profiler-data.md | markdown | 4.6 KB |
| references/frameworks/vllm.md | markdown | 5.6 KB |
| references/report-template.md | markdown | 5.0 KB |
| references/scenario-taxonomy.md | markdown | 5.7 KB |
| resources/scenarios/inference/sglang.json | config | 549 B |
| resources/scenarios/inference/vllm.json | config | 912 B |
| resources/scenarios/rl/slime.json | config | 468 B |
| resources/scenarios/rl/verl.json | config | 1.4 KB |
| resources/scenarios/training/fsdp.json | config | 303 B |
| resources/scenarios/training/megatron.json | config | 556 B |
| resources/scenarios/training/mindspeed.json | config | 318 B |
| scripts/compare_runs.py | script | 12.4 KB |
| scripts/db_query.py | script | 12.7 KB |
| scripts/decompose.py | script | 47.7 KB |
scripts/compare_runs.py
View on GitHub#!/usr/bin/python3
# -*- coding: utf-8 -*-
# -------------------------------------------------------------------------
# This file is part of the MindStudio project.
# Copyright (c) 2026 Huawei Technologies Co.,Ltd.
#
# MindStudio is licensed under Mulan PSL v2.
# You can use this software according to the terms and conditions of the Mulan PSL v2.
# You may obtain a copy of Mulan PSL v2 at:
#
# http://license.coscl.org.cn/MulanPSL2
#
# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
# See the Mulan PSL v2 for more details.
# -------------------------------------------------------------------------
"""compare_runs.py —— 多 run 对比(A/B 基线 vs 当前)
消费 msagent-profiler-breakdown 产出的统一拆解树,产出多 sheet 的 xlsx 对比表(阶段、稳态 step、
层间 wall_time、层内子模块占比)与一句话结论,服务于「改前 vs 改后」的量化差异判定。
用法:
python compare_runs.py --base-dir <runA> --comp-dir <runB> \
[--labels "baseline,current"] [--output compare.xlsx]
- 每个目录内读取固定产物名 breakdown.json(统一拆解树;stage.json 为同内容别名;
旧格式 stage.json + steps.json 兼容回退)、layers.json 可选。
- 依赖 openpyxl 写 xlsx;缺省(ImportError)时降级为逐 sheet 的 CSV + 控制台汇总。
设计依据:设计方案 §3.7 / §5(原 breakdown-compare 下沉),report-template.md「多 run 对比 xlsx」。
"""
import argparse
import csv
import json
import os
try:
from openpyxl import Workbook
from openpyxl.styles import Font, PatternFill
HAS_OPENPYXL = True
except ImportError: # pragma: no cover - 离线无第三方依赖时降级
HAS_OPENPYXL = False
def load_json(path):
with open(path, encoding="utf-8") as f:
return json.load(f)
def load_json_opt(path):
"""可选读取:文件不存在(如 layer-decompose 暂缓无 layers.json)返回空 dict。"""
if not os.path.isfile(path):
return {}
return load_json(path)
def load_run(rundir):
"""读取一个 run 的产物,返回规范化 dict。
- 新格式:breakdown.json(统一拆解树,顶层端到端 -> children 阶段 + 单次执行分支;
stage.json 为同内容别名)。
- 旧格式(仅 stage.json + steps.json)兼容回退。
layers 可选(layer-decompose 暂缓期间为空)。
"""
tree_path = os.path.join(rundir, "breakdown.json")
stage_path = os.path.join(rundir, "stage.json")
if os.path.isfile(tree_path):
tree = load_json(tree_path)
else:
tree = load_json(stage_path)
children = tree.get("children", [])
stages = [c for c in children if c.get("level") == "阶段"]
step_node = next((c for c in children if c.get("level") == "单次执行"), None)
steps = step_node or load_json_opt(os.path.join(rundir, "steps.json"))
return {
"dir": rundir,
"stage": tree,
"stages": stages,
"steps": steps,
"layers": load_json_opt(os.path.join(rundir, "layers.json")),
}
def ms(ns):
return ns / 1e6 if isinstance(ns, (int, float)) else None
def diff_pct(base, comp):
"""(comp - base) / base * 100;base 为 0/None 或 comp 为 None 返回 None。"""
if base in (None, 0) or comp is None:
return None
return (comp - base) / base * 100.0
def fmt_ms(v, nd=2):
x = ms(v)
return ("%.*f" % (nd, x)) if x is not None else "-"
def _csv_cell(v):
"""CSV 单元格安全化:转字符串;以 = + - @ 开头(Excel 公式注入向量)时加单引号前缀。"""
s = str(v)
return "'" + s if s.startswith(("=", "+", "-", "@")) else s
def run_metrics(r):
"""抽取可对比指标。"""
stage = r["stage"]
stages = r["stages"]
steps = r["steps"]
layers = r["layers"]
stage_map = {c["name"]: c.get("self_time", 0) for c in stages}
steady = [s for s in steps.get("steps", []) if not s.get("is_warmup")
and not s.get("is_prefill")]
ss = steps.get("stats", {})
best = min(steady, key=lambda s: s.get("wall_time", 0)) if steady else None
worst = max(steady, key=lambda s: s.get("wall_time", 0)) if steady else None
layer_walls = {layer["name"]: layer.get("wall_time", 0)
for layer in layers.get("layers", [])}
layer_names = [layer["name"] for layer in layers.get("layers", [])]
sub_pct = {s["name"]: s.get("pct", 0) for s in layers.get("submodule_totals", [])}
return {
"session_wall": stage.get("wall_time"),
"stage_map": stage_map,
"steady": steady,
"stats": ss,
"best_step": best,
"worst_step": worst,
"layer_walls": layer_walls,
"layer_names": layer_names,
"sub_pct": sub_pct,
"num_layers": layers.get("num_layers"),
}
def union_keys(a, b):
keys = []
for k in list(a.keys()) + list(b.keys()):
if k not in keys:
keys.append(k)
return keys
def style_header(ws, row, ncols):
for c in range(1, ncols + 1):
cell = ws.cell(row=row, column=c)
cell.font = Font(bold=True)
cell.fill = PatternFill("solid", fgColor="F2F2F2")
def _num_row(ws, label, a, b, fmt_a, fmt_b, pct_fmt="%.2f%%"):
"""追加一行数值对比:标量 a/b,展示时格式化,Δ% 用原始值计算。"""
dp = diff_pct(a, b)
ws.append([label, fmt_a(a), fmt_b(b),
(pct_fmt % dp) if dp is not None else "-"])
def write_xlsx(ma, mb, base_label, comp_label, out_path):
wb = Workbook()
# Sheet 1 总览
ws = wb.active
ws.title = "总览"
ws.append(["指标", base_label, comp_label, "Δ%"])
_num_row(ws, "端到端 wall_time(ms)", ma["session_wall"], mb["session_wall"],
lambda v: fmt_ms(v), lambda v: fmt_ms(v))
sa, sb = ma["stats"], mb["stats"]
_num_row(ws, "稳态 step P50(ms)", sa.get("p50"), sb.get("p50"),
lambda v: fmt_ms(v), lambda v: fmt_ms(v))
_num_row(ws, "稳态 step P99(ms)", sa.get("p99"), sb.get("p99"),
lambda v: fmt_ms(v), lambda v: fmt_ms(v))
_num_row(ws, "稳态 step mean(ms)", sa.get("mean"), sb.get("mean"),
lambda v: fmt_ms(v), lambda v: fmt_ms(v))
_num_row(ws, "稳态 step CV", sa.get("cv"), sb.get("cv"),
lambda v: "%.4f" % (v or 0), lambda v: "%.4f" % (v or 0))
ws.append(["层数", ma["num_layers"], mb["num_layers"], "-"])
style_header(ws, 1, 4)
# Sheet 2 阶段对比
ws2 = wb.create_sheet("阶段对比")
ws2.append(["阶段", base_label + "(ms)", comp_label + "(ms)", "Δ%"])
for k in union_keys(ma["stage_map"], mb["stage_map"]):
ws2.append([k, fmt_ms(ma["stage_map"].get(k)), fmt_ms(mb["stage_map"].get(k)),
("%.2f%%" % diff_pct(ma["stage_map"].get(k), mb["stage_map"].get(k)))
if diff_pct(ma["stage_map"].get(k), mb["stage_map"].get(k)) is not None else "-"])
style_header(ws2, 1, 4)
# Sheet 3 稳态 step(best/worst + 序列对齐)
ws3 = wb.create_sheet("稳态step")
ws3.append(["指标", base_label + "(ms)", comp_label + "(ms)", "Δ%"])
bb, bw = ma["best_step"], ma["worst_step"]
cb, cw = mb["best_step"], mb["worst_step"]
for label, a, b in (("最好 step", bb, cb), ("最差 step", bw, cw)):
av = a.get("wall_time") if a else None
bv = b.get("wall_time") if b else None
_num_row(ws3, label, av, bv, lambda v: fmt_ms(v), lambda v: fmt_ms(v))
ws3.append(["稳态 count", sa.get("count"), sb.get("count"), "-"])
style_header(ws3, 1, 4)
ws3.append([])
ws3.append(["step", base_label + "(ms)", comp_label + "(ms)", "Δ%"])
n = min(len(ma["steady"]), len(mb["steady"]))
for i in range(n):
a = ma["steady"][i].get("wall_time")
b = mb["steady"][i].get("wall_time")
dp = diff_pct(a, b)
ws3.append(["#%d" % i, fmt_ms(a), fmt_ms(b),
("%.2f%%" % dp) if dp is not None else "-"])
# Sheet 4 层对比(层名取两侧并集,与阶段/子模块占比口径一致)
ws4 = wb.create_sheet("层对比")
ws4.append(["层", base_label + "(ms)", comp_label + "(ms)", "Δ%"])
for k in union_keys(ma["layer_walls"], mb["layer_walls"]):
a = ma["layer_walls"].get(k)
b = mb["layer_walls"].get(k)
dp = diff_pct(a, b)
ws4.append([k, fmt_ms(a), fmt_ms(b), ("%.2f%%" % dp) if dp is not None else "-"])
style_header(ws4, 1, 4)
# Sheet 5 子模块占比
ws5 = wb.create_sheet("子模块占比")
ws5.append(["子模块", base_label + "(%)", comp_label + "(%)", "Δ(pp)"])
for k in union_keys(ma["sub_pct"], mb["sub_pct"]):
a = ma["sub_pct"].get(k)
b = mb["sub_pct"].get(k)
ws5.append([k, ("%.1f" % a) if a is not None else "-",
("%.1f" % b) if b is not None else "-",
("%.2f" % (b - a)) if (a is not None and b is not None) else "-"])
style_header(ws5, 1, 4)
wb.save(out_path)
return out_path
def write_csv_fallback(ma, mb, base_label, comp_label, out_prefix):
"""openpyxl 缺失时降级:逐 sheet 写 CSV。"""
sheets = {
"overview": [["metric", base_label, comp_label],
["session_wall_ms", fmt_ms(ma["session_wall"]), fmt_ms(mb["session_wall"])],
["step_p50_ms", fmt_ms(ma["stats"].get("p50")), fmt_ms(mb["stats"].get("p50"))],
["step_p99_ms", fmt_ms(ma["stats"].get("p99")), fmt_ms(mb["stats"].get("p99"))],
["step_cv", "%.4f" % (ma["stats"].get("cv") or 0),
"%.4f" % (mb["stats"].get("cv") or 0)]],
"stage": [["stage", base_label + "(ms)", comp_label + "(ms)"]] +
[[k, fmt_ms(ma["stage_map"].get(k)), fmt_ms(mb["stage_map"].get(k))]
for k in union_keys(ma["stage_map"], mb["stage_map"])],
"layer": [["layer", base_label + "(ms)", comp_label + "(ms)"]] +
[[k, fmt_ms(ma["layer_walls"].get(k)), fmt_ms(mb["layer_walls"].get(k))]
for k in union_keys(ma["layer_walls"], mb["layer_walls"])],
"submodule": [["submodule", base_label + "(%)", comp_label + "(%)"]] +
[[k, ma["sub_pct"].get(k), mb["sub_pct"].get(k)]
for k in union_keys(ma["sub_pct"], mb["sub_pct"])],
}
paths = []
for name, rows in sheets.items():
path = out_prefix + "_" + name + ".csv"
with open(path, "w", encoding="utf-8", newline="") as f:
w = csv.writer(f)
for row in rows:
w.writerow([_csv_cell(c) for c in row])
paths.append(path)
return paths
def main(argv=None):
p = argparse.ArgumentParser(description="多 run 对比(A/B)")
p.add_argument("--base-dir", required=True, help="基线 run 目录(含 stage/steps/layers.json)")
p.add_argument("--comp-dir", required=True, help="对比 run 目录(含 stage/steps/layers.json)")
p.add_argument("--labels", default="baseline,current", help="两个 run 的标签,逗号分隔")
p.add_argument("--output", default="compare.xlsx", help="输出 xlsx 路径(openpyxl 缺失时降级 CSV 前缀)")
args = p.parse_args(argv)
labels = [s.strip() for s in args.labels.split(",")]
base_label = labels[0] if len(labels) > 0 else "baseline"
comp_label = labels[1] if len(labels) > 1 else "current"
base = load_run(args.base_dir)
comp = load_run(args.comp_dir)
ma = run_metrics(base)
mb = run_metrics(comp)
a_p50 = ms(ma["stats"].get("p50"))
b_p50 = ms(mb["stats"].get("p50"))
if a_p50 is not None and b_p50 is not None:
d = diff_pct(ma["stats"]["p50"], mb["stats"]["p50"])
trend = "变慢" if d > 0 else ("变快" if d < 0 else "持平")
line = "稳态 step P50:%s=%.2fms → %s=%.2fms(%+.2f%%,%s)" % (
base_label, a_p50, comp_label, b_p50, (d or 0), trend)
else:
line = "无稳态 step 数据可比对"
if HAS_OPENPYXL:
out = write_xlsx(ma, mb, base_label, comp_label, args.output)
print("[OK] %s" % out)
else:
prefix = os.path.splitext(args.output)[0]
paths = write_csv_fallback(ma, mb, base_label, comp_label, prefix)
print("[WARN] openpyxl 未安装,降级为 CSV:%s" % ", ".join(paths))
print("[INFO] " + line)
return 0
if __name__ == "__main__":
raise SystemExit(main())Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 197k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |