Ascend/msagent / skills/profiler/ascend-prof-compare
ascend-prof-compare
VerifiedCompares two Prof datasets (GPU vs NPU / NPU vs NPU) by auto-installing and calling msprof-analyze compare, then parses the result xlsx into HTML report and Chinese xlsx. Invoke when user provides two prof dirs to compare, or has a compare result xlsx to analyze/translate.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 11.2 KB |
| references/analysis_methodology.md | markdown | 15.1 KB |
| references/compare_quickstart.md | markdown | 3.6 KB |
| scripts/csv_translator.py | script | 8.2 KB |
| scripts/html_generator.py | script | 38.7 KB |
| scripts/main_analyzer.py | script | 11.8 KB |
| scripts/parsers/__init__.py | script | 35 B |
| scripts/parsers/api_compare.py | script | 3.1 KB |
| scripts/parsers/common.py | script | 1.6 KB |
| scripts/parsers/communication.py | script | 2.5 KB |
| scripts/parsers/kernel.py | script | 13.3 KB |
| scripts/parsers/memory.py | script | 3.6 KB |
| scripts/parsers/module.py | script | 4.1 KB |
| scripts/parsers/operator.py | script | 4.0 KB |
| scripts/parsers/overall_metrics.py | script | 5.9 KB |
| scripts/run_compare.py | script | 9.3 KB |
scripts/html_generator.py
View on GitHub#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""HTML 报告生成器
将中间件 JSON 数据渲染为自包含的 HTML 报告。
"""
from datetime import datetime
def _fmt(val, suffix='', decimals=2):
"""格式化数值"""
if val is None:
return '-'
try:
if isinstance(val, float) and (val == float('inf') or val == float('-inf')):
return 'inf'
return f'{float(val):.{decimals}f}{suffix}'
except (ValueError, TypeError):
return str(val)
def _ratio_color(ratio):
"""根据比率返回颜色类名"""
if ratio is None:
return 'neutral'
try:
r = float(ratio)
if r > 1.1:
return 'bad'
elif r < 0.9:
return 'good'
else:
return 'neutral'
except (ValueError, TypeError):
return 'neutral'
def _diff_color(diff):
"""根据差异值返回颜色类名"""
if diff is None:
return 'neutral'
try:
d = float(diff)
if d > 0:
return 'bad'
elif d < 0:
return 'good'
else:
return 'neutral'
except (ValueError, TypeError):
return 'neutral'
def generate_html_report(results: dict, output_path: str):
"""生成 HTML 报告"""
meta = results.get('metadata', {})
summary = results.get('analysis_summary', {})
# 可用 Sheet 列表
available = summary.get('available_sheets', [])
unavailable = summary.get('unavailable_sheets', [])
# 总体性能
om = results.get('overall_metrics', {})
e2e = om.get('e2e', {})
dims = om.get('dimensions', [])
sub_cats = om.get('top_sub_categories', [])
# 算子统计
op_stat = results.get('operator_statistic', {})
# 模块统计
mod_stat = results.get('module_statistic', {})
# 通信
comm = results.get('communication', {})
# 内存
mem_stat = results.get('memory_statistic', {})
# Kernel
kc = results.get('kernel_compare', {})
# API
api = results.get('api_compare', {})
html = f"""<!DOCTYPE html>
<html lang="zh-CN">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>性能比对分析报告</title>
<style>
* {{ margin: 0; padding: 0; box-sizing: border-box; }}
body {{ font-family: -apple-system, 'Segoe UI', 'Noto Sans CJK SC', sans-serif; background: #f0f2f5; color: #1a1d23; line-height: 1.6; }}
.container {{ max-width: 1200px; margin: 0 auto; padding: 24px 24px 24px 200px; }}
.header {{ background: linear-gradient(135deg, #667eea 0%, #764ba2 100%); color: #fff; padding: 40px 32px; border-radius: 16px; margin-bottom: 24px; box-shadow: 0 4px 20px rgba(102,126,234,0.3); }}
.header h1 {{ font-size: 28px; margin-bottom: 8px; }}
.header .meta {{ font-size: 14px; opacity: 0.85; }}
.section {{ background: #fff; border-radius: 12px; padding: 24px; margin-bottom: 20px; box-shadow: 0 2px 8px rgba(0,0,0,0.06); }}
.section-title {{ font-size: 20px; font-weight: 700; margin-bottom: 16px; padding-bottom: 12px; border-bottom: 2px solid #e2e5ea; }}
.kpi-grid {{ display: grid; grid-template-columns: repeat(auto-fit, minmax(200px, 1fr)); gap: 16px; margin-bottom: 20px; }}
.kpi-card {{ background: #f8f9fb; border: 1px solid #e2e5ea; border-radius: 10px; padding: 16px; }}
.kpi-card .label {{ font-size: 13px; color: #6b7280; margin-bottom: 4px; }}
.kpi-card .value {{ font-size: 24px; font-weight: 700; }}
.kpi-card .sub {{ font-size: 12px; color: #6b7280; margin-top: 4px; }}
.bad {{ color: #dc2626; }}
.good {{ color: #16a34a; }}
.neutral {{ color: #6b7280; }}
table {{ width: 100%; border-collapse: collapse; font-size: 14px; }}
.table-wrap {{ overflow-x: auto; margin-bottom: 8px; }}
th {{ background: #f3f4f6; padding: 10px 12px; text-align: left; font-weight: 600; color: #374151; border-bottom: 2px solid #e2e5ea; white-space: nowrap; }}
td {{ padding: 10px 12px; border-bottom: 1px solid #f0f0f0; }}
tr:hover td {{ background: #f8f9fb; }}
.tag {{ display: inline-block; padding: 2px 8px; border-radius: 999px; font-size: 12px; font-weight: 500; }}
.tag-bad {{ background: #fee2e2; color: #dc2626; }}
.tag-good {{ background: #dcfce7; color: #16a34a; }}
.tag-neutral {{ background: #f3f4f6; color: #6b7280; }}
.findings {{ list-style: none; }}
.findings li {{ background: #fef3c7; border-left: 4px solid #f59e0b; padding: 12px 16px; margin-bottom: 8px; border-radius: 0 8px 8px 0; font-size: 14px; }}
.chart-container {{ position: relative; height: 300px; margin: 16px 0; }}
.unavailable {{ background: #f9fafb; color: #9ca3af; padding: 40px; text-align: center; border-radius: 8px; }}
.col-num {{ text-align: right; font-variant-numeric: tabular-nums; }}
.bar-container {{ display: flex; align-items: center; gap: 8px; }}
.bar {{ height: 8px; border-radius: 4px; background: #e5e7eb; flex: 1; overflow: hidden; }}
.bar-fill {{ height: 100%; border-radius: 4px; }}
.footer {{ text-align: center; padding: 24px; color: #9ca3af; font-size: 13px; }}
.dim-card {{ display: flex; align-items: center; justify-content: space-between; padding: 12px 16px; border-radius: 8px; margin-bottom: 8px; border: 1px solid #e2e5ea; }}
.dim-card.primary {{ border-color: #7c3aed; background: #f5f3ff; }}
/* 悬浮侧边导航 */
.nav-sidebar {{ position: fixed; top: 24px; left: 24px; width: 168px; max-height: calc(100vh - 48px); overflow-y: auto; background: #fff; border-radius: 12px; padding: 16px; box-shadow: 0 2px 8px rgba(0,0,0,0.08); font-size: 13px; }}
.nav-sidebar .nav-title {{ font-weight: 700; font-size: 14px; margin-bottom: 10px; color: #1a1d23; }}
.nav-sidebar a {{ display: block; padding: 4px 0 4px 12px; color: #374151; text-decoration: none; border-left: 2px solid #e2e5ea; margin-bottom: 2px; }}
.nav-sidebar a:hover {{ color: #7c3aed; border-left-color: #7c3aed; }}
.nav-sidebar a.nav-sub {{ font-size: 12px; color: #6b7280; padding-left: 24px; }}
.nav-sidebar .nav-sep {{ height: 1px; background: #e2e5ea; margin: 8px 0; }}
/* Excel 样式表 */
.excel-table {{ width: 100%; border-collapse: collapse; font-size: 12px; font-family: 'Segoe UI', sans-serif; }}
.excel-table th {{ background: #7c3aed; color: #fff; padding: 6px 10px; text-align: center; font-weight: 600; border: 1px solid #5b3fae; white-space: nowrap; font-size: 11px; }}
.excel-table td {{ padding: 5px 10px; border: 1px solid #e2e5ea; text-align: center; }}
.excel-table td.col-label {{ text-align: left; font-weight: 500; background: #f8f9fb; }}
.excel-table tr:nth-child(even) td {{ background: #fafafa; }}
.excel-table tr:nth-child(even) td.col-label {{ background: #f3f4f6; }}
.excel-table td.bad {{ color: #dc2626; font-weight: 600; }}
.excel-table td.good {{ color: #16a34a; font-weight: 600; }}
.excel-table td.indent {{ padding-left: 24px; }}
.excel-table td.indent2 {{ padding-left: 40px; }}
.excel-table tr.row-key td {{ background: #ede9fe; font-weight: 600; }}
.excel-table tr.row-key td.col-label-bold {{ background: #ddd6fe; font-weight: 700; color: #5b21b6; }}
</style>
</head>
<body>
<div class="container">
<!-- 悬浮侧边导航 -->
<div class="nav-sidebar">
<div class="nav-title">目录导航</div>
<a href="#overall">1. 总体性能比对</a>
<a href="#conclusion">2. 分析结论</a>
<div class="nav-sep"></div>
<a href="#operator">3. 算子统计</a>
<a href="#module">4. 模块统计</a>
<a href="#communication">5. 通信比对</a>
<a href="#memory">6. 内存统计</a>
<a href="#kernel">7. Kernel比对</a>
<a href="#kernel-data" class="nav-sub">- 两卡数据量比对</a>
<a href="#kernel-top10" class="nav-sub">- Top10计算算子</a>
<a href="#kernel-imbalance" class="nav-sub">- 负载不均预警</a>
<a href="#api">8. API比对</a>
<a href="#sheets">9. Sheet可用性</a>
</div>
<!-- 头部 -->
<div class="header">
<h1>性能比对分析报告</h1>
<div class="meta">
源文件: {meta.get('source_file', '-')} | 分析时间: {meta.get('analysis_time', '-')} |
Sheet 数: {meta.get('total_sheets', '-')} | 分析方法: analysis_methodology.md
</div>
</div>
<!-- 总体性能 -->
<div class="section" id="overall">
<div class="section-title">1. 总体性能比对 (OverallMetrics)</div>
{_render_overall_metrics(om, e2e, dims, sub_cats) if om.get('available') else '<div class="unavailable">该 Sheet 不存在或为空</div>'}
</div>
<!-- 分析结论 -->
<div class="section" id="conclusion">
<div class="section-title">2. 分析结论</div>
<div style="display: grid; grid-template-columns: 1fr 1fr; gap: 16px;">
<div>
<h4 style="color: #16a34a; margin-bottom: 8px;">[改善] 改善亮点</h4>
<ul class="findings" style="list-style: none;">
{''.join(f'<li style="background:#dcfce7;border-left-color:#16a34a;">{f}</li>' for f in summary.get('improvements', [])) if summary.get('improvements') else '<li style="background:#f3f4f6;border-left-color:#9ca3af;">暂无改善项</li>'}
</ul>
</div>
<div>
<h4 style="color: #dc2626; margin-bottom: 8px;">[风险] 劣化风险</h4>
<ul class="findings" style="list-style: none;">
{''.join(f'<li style="background:#fee2e2;border-left-color:#dc2626;">{f}</li>' for f in summary.get('risks', [])) if summary.get('risks') else '<li style="background:#f3f4f6;border-left-color:#9ca3af;">暂无劣化项</li>'}
</ul>
</div>
</div>
</div>
<!-- 算子统计 -->
<div class="section" id="operator">
<div class="section-title">3. 算子统计比对 (OperatorCompareStatistic)</div>
{_render_operator_statistic(op_stat) if op_stat.get('available') else '<div class="unavailable">该 Sheet 不存在或为空</div>'}
</div>
<!-- 模块统计 -->
<div class="section" id="module">
<div class="section-title">4. 模块统计比对 (ModuleCompareStatistic)</div>
{_render_module_statistic(mod_stat) if mod_stat.get('available') else '<div class="unavailable">该 Sheet 不存在或为空(需采集 Python Function 事件)</div>'}
</div>
<!-- 通信比对 -->
<div class="section" id="communication">
<div class="section-title">5. 通信比对 (CommunicationCompare)</div>
{_render_communication(comm) if comm.get('available') else '<div class="unavailable">该 Sheet 不存在或为空</div>'}
</div>
<!-- 内存统计 -->
<div class="section" id="memory">
<div class="section-title">6. 内存统计比对 (MemoryCompareStatistic)</div>
{_render_memory_statistic(mem_stat) if mem_stat.get('available') else '<div class="unavailable">该 Sheet 不存在或为空</div>'}
</div>
<!-- Kernel 比对 -->
<div class="section" id="kernel">
<div class="section-title">7. Kernel 比对 (KernelCompare)</div>
{_render_kernel_compare(kc) if kc.get('available') else '<div class="unavailable">该 Sheet 不存在或为空(仅 NPU vs NPU 场景)</div>'}
</div>
<!-- API 比对 -->
<div class="section" id="api">
<div class="section-title">8. API 比对 (ApiCompare)</div>
{_render_api_compare(api) if api.get('available') else '<div class="unavailable">该 Sheet 不存在或为空</div>'}
</div>
<!-- Sheet 可用性 -->
<div class="section" id="sheets">
<div class="section-title">9. Sheet 可用性总览</div>
<table>
<tr><th>状态</th><th>Sheet 名称</th></tr>
{''.join(f'<tr><td><span class="tag tag-good">可用</span></td><td>{s}</td></tr>' for s in available)}
{''.join(f'<tr><td><span class="tag tag-neutral">不可用</span></td><td>{s}</td></tr>' for s in unavailable)}
</table>
</div>
<div class="footer">
Generated by Compare Analyzer | {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}
</div>
</div>
</body>
</html>"""
with open(output_path, 'w', encoding='utf-8') as f:
f.write(html)
def _render_overall_metrics(om, e2e, dims, sub_cats):
"""渲染总体性能部分"""
parts = []
# E2E KPI + 趋势
trend = om.get('overall_trend', '')
trend_color = '#16a34a' if trend == '改善' else '#dc2626' if trend == '劣化' else '#6b7280'
if e2e:
parts.append(f"""
<div class="kpi-grid">
<div class="kpi-card">
<div class="label">E2E 总耗时(基准)</div>
<div class="value">{_fmt(e2e.get('base'), 'ms')}</div>
</div>
<div class="kpi-card">
<div class="label">E2E 总耗时(比对)</div>
<div class="value {_diff_color(e2e.get('diff'))}">{_fmt(e2e.get('comp'), 'ms')}</div>
</div>
<div class="kpi-card">
<div class="label">E2E 差异</div>
<div class="value {_diff_color(e2e.get('diff'))}">{_fmt(e2e.get('diff'), 'ms')}</div>
<div class="sub">比率: {_fmt(e2e.get('ratio'))}</div>
</div>
<div class="kpi-card">
<div class="label">整体趋势</div>
<div class="value" style="color:{trend_color};">{trend}</div>
</div>
</div>""")
# 维度拆解(区分改善/劣化)
if dims:
parts.append('<div style="margin-top:16px;"><h4 style="margin-bottom:4px;">维度拆解</h4>')
parts.append('<div style="font-size:12px;color:#6b7280;margin-bottom:12px;">比率 = 比对值/基准值(>1劣化,<1改善)。贡献占比仅对劣化维度计算,分母为所有劣化维度差异之和。</div>')
for d in dims:
is_imp = d.get('is_improved')
is_deg = d.get('is_degraded')
if is_imp:
tag = '<span class="tag tag-good" style="margin-left:8px;">改善</span>'
elif is_deg:
tag = '<span class="tag tag-bad" style="margin-left:8px;">劣化</span>'
else:
tag = ''
parts.append(f"""
<div class="dim-card">
<div>
<strong>{d['name']}</strong> ({d['label']})
{tag}
</div>
<div style="text-align:right;">
<span class="{_diff_color(d['diff_duration'])}">{_fmt(d['diff_duration'], 'ms')}</span>
<span class="neutral" style="margin-left:12px;">比率 {_fmt(d['diff_ratio'])}</span>
<span class="neutral" style="margin-left:12px;">劣化贡献 {d.get('contribution_str', '-')}</span>
</div>
</div>""")
parts.append('</div>')
# 子类别 Top
if sub_cats:
parts.append('<div style="margin-top:16px;"><h4 style="margin-bottom:12px;">子类别变化 Top 8</h4><table>')
parts.append('<tr><th>子类别</th><th class="col-num">基准(ms)</th><th class="col-num">比对(ms)</th><th class="col-num">差异(ms)</th><th class="col-num">比率</th><th>状态</th></tr>')
for s in sub_cats:
cls = _ratio_color(s.get('diff_ratio'))
if s.get('is_improved'):
tag = '<span class="tag tag-good">改善</span>'
elif s.get('is_degraded'):
tag = '<span class="tag tag-bad">劣化</span>'
else:
tag = '<span class="tag tag-neutral">持平</span>'
parts.append(f'<tr><td>{s["name"]}</td><td class="col-num">{_fmt(s.get("base_duration"))}</td><td class="col-num">{_fmt(s.get("comp_duration"))}</td><td class="col-num {_diff_color(s.get("diff_duration"))}">{_fmt(s.get("diff_duration"), "ms")}</td><td class="col-num {cls}">{_fmt(s.get("diff_ratio"))}</td><td>{tag}</td></tr>')
parts.append('</table></div>')
# Not minimal profiling 警告
if om.get('has_minimal_warning'):
parts.append('<div style="margin-top:12px;padding:12px;background:#fef3c7;border-radius:8px;color:#92400e;font-size:13px;">[警告] 检测到 Not minimal profiling,E2E 时间可能存在性能膨胀,影响通信和调度耗时判断。</div>')
# OverallMetrics 原始数据表(Excel 样式 + 中文翻译)
raw_metrics = om.get('raw_metrics', [])
if raw_metrics:
parts.append('<h4 style="margin:16px 0 8px;">总体性能原始数据</h4>')
parts.append('<div class="table-wrap"><table class="excel-table">')
parts.append('<tr><th>指标</th><th>基准耗时(ms)</th><th>基准占比</th><th>基准数量</th><th>比对耗时(ms)</th><th>比对占比</th><th>比对数量</th><th>差异(ms)</th><th>差异比率</th></tr>')
for m in raw_metrics:
idx = m['index']
# 缩进处理:一级 \t 用 indent,二级 \t\t 用 indent2
indent_cls = ''
clean_idx = idx
if idx.startswith('\t\t'):
indent_cls = 'indent2'
clean_idx = idx.strip()
elif idx.startswith('\t'):
indent_cls = 'indent'
clean_idx = idx.strip()
diff_cls = 'bad' if m['diff_duration'] > 0 else 'good' if m['diff_duration'] < 0 else ''
ratio_cls = 'bad' if m['diff_ratio'] > 1.05 else 'good' if m['diff_ratio'] < 0.95 else ''
# 四大核心维度行强调
key_metrics = ('Computing Time', 'Uncovered Communication Time', 'Free Time', 'E2E Time')
row_cls = ' class="row-key"' if clean_idx in key_metrics else ''
label_cls = 'col-label-bold' if clean_idx in key_metrics else 'col-label'
parts.append(f'<tr{row_cls}><td class="{label_cls} {indent_cls}">{clean_idx}</td><td>{_fmt(m["base_duration"])}</td><td>{_fmt(m["base_ratio"])}</td><td>{m["base_number"]}</td><td>{_fmt(m["comp_duration"])}</td><td>{_fmt(m["comp_ratio"])}</td><td>{m["comp_number"]}</td><td class="{diff_cls}">{_fmt(m["diff_duration"])}</td><td class="{ratio_cls}">{_fmt(m["diff_ratio"])}</td></tr>')
parts.append('</table></div>')
return ''.join(parts)
def _render_operator_statistic(op_stat):
"""渲染算子统计部分"""
parts = []
parts.append(f"""
<div class="kpi-grid">
<div class="kpi-card"><div class="label">参与比对算子总数</div><div class="value">{op_stat.get('total_operators', 0)}</div></div>
<div class="kpi-card"><div class="label">劣化算子数</div><div class="value bad">{op_stat.get('degraded_count', 0)}</div></div>
<div class="kpi-card"><div class="label">改善算子数</div><div class="value good">{op_stat.get('improved_count', 0)}</div></div>
<div class="kpi-card"><div class="label">Top10 集中度</div><div class="value">{_fmt(op_stat.get('concentration'), '%', 1)}</div><div class="sub">{'集中' if op_stat.get('is_concentrated') else '分散'}</div></div>
</div>""")
top10 = op_stat.get('top10', [])
if top10:
parts.append('<table><tr><th>排名</th><th>算子名称</th><th class="col-num">基准耗时(ms)</th><th class="col-num">比对耗时(ms)</th><th class="col-num">差异(ms)</th><th class="col-num">比率</th></tr>')
for i, op in enumerate(top10, 1):
parts.append(f'<tr><td>{i}</td><td style="font-family:monospace;">{op["name"]}</td><td class="col-num">{_fmt(op["base_duration"])}</td><td class="col-num">{_fmt(op["comp_duration"])}</td><td class="col-num {_diff_color(op["diff_duration"])}">{_fmt(op["diff_duration"])}</td><td class="col-num {_ratio_color(op["diff_ratio"])}">{_fmt(op["diff_ratio"])}</td></tr>')
parts.append('</table>')
return ''.join(parts)
def _render_module_statistic(mod_stat):
"""渲染模块统计部分"""
parts = []
parts.append(f"""
<div class="kpi-grid">
<div class="kpi-card"><div class="label">参与比对模块总数</div><div class="value">{mod_stat.get('total_modules', 0)}</div></div>
<div class="kpi-card"><div class="label">有调用栈</div><div class="value">{'是' if mod_stat.get('has_call_stack') else '否'}</div></div>
</div>""")
top10 = mod_stat.get('top10', [])
if top10:
parts.append('<table><tr><th>排名</th><th>模块类</th><th>层级</th><th>模块名</th><th class="col-num">基准总耗时(ms)</th><th class="col-num">比对总耗时(ms)</th><th class="col-num">差异(ms)</th><th class="col-num">比率</th></tr>')
for i, m in enumerate(top10, 1):
parts.append(f'<tr><td>{i}</td><td>{m["module_class"]}</td><td>{m["module_level"]}</td><td style="font-family:monospace;font-size:12px;">{m["module_name"]}</td><td class="col-num">{_fmt(m["base_total_time"])}</td><td class="col-num">{_fmt(m["comp_total_time"])}</td><td class="col-num {_diff_color(m["diff_total_time"])}">{_fmt(m["diff_total_time"])}</td><td class="col-num {_ratio_color(m["diff_total_ratio"])}">{_fmt(m["diff_total_ratio"])}</td></tr>')
parts.append('</table>')
return ''.join(parts)
def _render_communication(comm):
"""渲染通信部分"""
parts = []
parts.append(f"""
<div class="kpi-grid">
<div class="kpi-card"><div class="label">通信算子总数</div><div class="value">{comm.get('total_comm_ops', 0)}</div></div>
<div class="kpi-card"><div class="label">改善数</div><div class="value good">{comm.get('improved_count', 0)}</div></div>
<div class="kpi-card"><div class="label">劣化数</div><div class="value bad">{comm.get('degraded_count', 0)}</div></div>
<div class="kpi-card"><div class="label">极端改善</div><div class="value">{'是' if comm.get('has_extreme_improvement') else '否'}</div></div>
</div>""")
# 变化最大的算子(包含改善和劣化)
top_changed = comm.get('top_changed', [])
if top_changed:
parts.append('<h4 style="margin:12px 0 8px;">变化最大的通信算子</h4><table>')
parts.append('<tr><th>排名</th><th>通信算子名</th><th class="col-num">基准(us)</th><th class="col-num">比对(us)</th><th class="col-num">差异(us)</th><th class="col-num">比率</th><th>状态</th></tr>')
for i, c in enumerate(top_changed, 1):
if c['diff_duration'] < 0:
tag = '<span class="tag tag-good">改善</span>'
if c.get('is_extreme_improvement'):
tag += ' <span class="tag tag-good">极端</span>'
elif c['diff_duration'] > 0:
tag = '<span class="tag tag-bad">劣化</span>'
else:
tag = '<span class="tag tag-neutral">持平</span>'
parts.append(f'<tr><td>{i}</td><td style="font-family:monospace;">{c["comm_name"]}</td><td class="col-num">{_fmt(c["base_total"])}</td><td class="col-num">{_fmt(c["comp_total"])}</td><td class="col-num {_diff_color(c["diff_duration"])}">{_fmt(c["diff_duration"])}</td><td class="col-num {_ratio_color(c["diff_ratio"])}">{_fmt(c["diff_ratio"])}</td><td>{tag}</td></tr>')
parts.append('</table>')
return ''.join(parts)
def _render_memory_statistic(mem_stat):
"""渲染内存统计部分"""
parts = []
if mem_stat.get('all_zero_memory'):
parts.append(f"""
<div class="kpi-card" style="margin-bottom:12px;background:#f9fafb;border-color:#e5e7eb;">
<div class="label">参与比对算子数</div>
<div class="value">{mem_stat.get('total_operators', 0)}</div>
</div>
<div style="padding:16px;background:#fef3c7;border-radius:8px;color:#92400e;font-size:14px;">
[说明] {mem_stat.get('zero_memory_note', '所有算子内存无差异')}
</div>""")
return ''.join(parts)
parts.append(f"""
<div class="kpi-grid">
<div class="kpi-card"><div class="label">参与比对算子数</div><div class="value">{mem_stat.get('total_operators', 0)}</div></div>
<div class="kpi-card"><div class="label">总内存差异</div><div class="value {_diff_color(mem_stat.get('total_diff_memory'))}">{_fmt(mem_stat.get('total_diff_memory'), 'MB')}</div></div>
<div class="kpi-card"><div class="label">内存增长算子数</div><div class="value bad">{mem_stat.get('increased_count', 0)}</div></div>
<div class="kpi-card"><div class="label">内存减少算子数</div><div class="value good">{mem_stat.get('decreased_count', 0)}</div></div>
</div>""")
top10 = mem_stat.get('top10', [])
if top10:
parts.append('<table><tr><th>排名</th><th>算子名称</th><th class="col-num">基准内存(MB)</th><th class="col-num">比对内存(MB)</th><th class="col-num">差异(MB)</th><th class="col-num">比率</th></tr>')
for i, m in enumerate(top10, 1):
parts.append(f'<tr><td>{i}</td><td style="font-family:monospace;">{m["name"]}</td><td class="col-num">{_fmt(m["base_memory"])}</td><td class="col-num">{_fmt(m["comp_memory"])}</td><td class="col-num {_diff_color(m["diff_memory"])}">{_fmt(m["diff_memory"])}</td><td class="col-num {_ratio_color(m["diff_ratio"])}">{_fmt(m["diff_ratio"])}</td></tr>')
parts.append('</table>')
return ''.join(parts)
def _render_kernel_compare(kc):
"""渲染 Kernel 比对部分"""
parts = []
trend = kc.get('net_trend', '')
trend_color = '#16a34a' if trend == '改善' else '#dc2626' if trend == '劣化' else '#6b7280'
parts.append(f"""
<div class="kpi-grid">
<div class="kpi-card"><div class="label">Kernel 总数</div><div class="value">{kc.get('total_kernels', 0)}</div><div class="sub">计算 {kc.get('compute_kernel_count', 0)} / 通信 {kc.get('comm_kernel_count', 0)}</div></div>
<div class="kpi-card"><div class="label">劣化总量</div><div class="value bad">+{_fmt(kc.get('total_degraded_us', 0), 'us')}</div><div class="sub">{kc.get('degraded_count', 0)} 个</div></div>
<div class="kpi-card"><div class="label">改善总量</div><div class="value good">-{_fmt(kc.get('total_improved_us', 0), 'us')}</div><div class="sub">{kc.get('improved_count', 0)} 个</div></div>
<div class="kpi-card"><div class="label">净趋势</div><div class="value" style="color:{trend_color};">{trend}</div><div class="sub">显著劣化: {kc.get('significant_degraded_count', 0)} 个</div></div>
</div>""")
# 显著劣化 Kernel
sig_degraded = kc.get('significant_degraded', [])
if sig_degraded:
parts.append('<h4 style="margin:12px 0 8px;color:#dc2626;">显著劣化 Kernel(比率 > 1.05)</h4><table>')
parts.append('<tr><th>Kernel名称</th><th>输入Shape</th><th class="col-num">基准(us)</th><th class="col-num">比对(us)</th><th class="col-num">差异(us)</th><th class="col-num">比率</th><th>原因</th></tr>')
for k in sig_degraded:
parts.append(f'<tr><td style="font-family:monospace;font-size:12px;">{k["kernel"]}</td><td style="font-size:12px;">{k["input_shape"]}</td><td class="col-num">{_fmt(k["base_total"])}</td><td class="col-num">{_fmt(k["comp_total"])}</td><td class="col-num bad">+{_fmt(k["diff_total"])}</td><td class="col-num bad">{_fmt(k["total_ratio"])}</td><td><span class="tag tag-bad">{k.get("degradation_type", "-")}</span></td></tr>')
parts.append('</table>')
# === 两卡数据量比对分析(在 Top 10 变化最大 之前) ===
dvc = kc.get('data_volume_compare', {})
if dvc:
elements_changed = dvc.get('elements_changed', False)
calls_changed = dvc.get('calls_changed', False)
balance_tag = '<span class="tag tag-good">负载均衡</span>' if not elements_changed and not calls_changed else '<span class="tag tag-bad">负载变化</span>'
parts.append(f"""
<h4 id="kernel-data" style="margin:16px 0 8px;color:#2563eb;">两卡输入数据量比对分析 {balance_tag}</h4>
<div style="font-size:12px;color:#6b7280;margin-bottom:8px;">总数据量 = 各算子单次数据量 × calls 之和。若两卡总数据量和总 calls 完全一致,说明计算负载均衡;若不一致,说明模型结构或执行路径发生变化。</div>
<table>
<tr><th>指标</th><th class="col-num">基准卡</th><th class="col-num">比对卡</th><th class="col-num">差异</th></tr>
<tr><td>计算算子种类数</td><td class="col-num">{kc.get('compute_kernel_count', 0)}</td><td class="col-num">{kc.get('compute_kernel_count', 0)}</td><td class="col-num">-</td></tr>
<tr><td>总调用次数</td><td class="col-num">{dvc.get('base_total_calls', 0)}</td><td class="col-num">{dvc.get('comp_total_calls', 0)}</td><td class="col-num {_diff_color(dvc.get('comp_total_calls', 0) - dvc.get('base_total_calls', 0))}">{dvc.get('comp_total_calls', 0) - dvc.get('base_total_calls', 0):+d}</td></tr>
<tr><td>总输入数据量</td><td class="col-num">{dvc.get('base_total_elements_fmt', '-')}</td><td class="col-num">{dvc.get('comp_total_elements_fmt', '-')}</td><td class="col-num">{'一致' if not elements_changed else '变化'}</td></tr>
<tr><td>总耗时(us)</td><td class="col-num">{_fmt(dvc.get('base_total_us', 0))}</td><td class="col-num">{_fmt(dvc.get('comp_total_us', 0))}</td><td class="col-num {_diff_color(dvc.get('comp_total_us', 0) - dvc.get('base_total_us', 0))}">{_fmt(dvc.get('comp_total_us', 0) - dvc.get('base_total_us', 0))}</td></tr>
</table>""")
# Top10 计算算子的数据量明细对比
top10_data = kc.get('top10_data_compare', [])
if top10_data:
parts.append('<h4 id="kernel-top10" style="margin:12px 0 8px;color:#2563eb;">Top 10 计算算子数据量明细</h4>')
parts.append('<div class="table-wrap"><table>')
parts.append('<tr><th>排名</th><th>算子名</th><th>单次数据量</th><th class="col-num">calls(基准/比对)</th><th class="col-num">总数据量(基准/比对)</th><th class="col-num">耗时(us)(基准/比对)</th></tr>')
for i, k in enumerate(top10_data, 1):
calls_str = f'{k["base_calls"]} / {k["comp_calls"]}'
data_str = f'{k["base_total_elements_fmt"]} / {k["comp_total_elements_fmt"]}'
dur_str = f'{_fmt(k["base_total_us"])} / {_fmt(k["comp_total_us"])}'
parts.append(f'<tr><td>{i}</td><td style="font-family:monospace;font-size:12px;" title="{k["shape"]}">{k["kernel"]}</td><td class="col-num">{k["elements_per_call_fmt"]}</td><td class="col-num">{calls_str}</td><td class="col-num">{data_str}</td><td class="col-num">{dur_str}</td></tr>')
parts.append('</table></div>')
top10 = kc.get('top10', [])
if top10:
parts.append('<h4 style="margin:16px 0 8px;">Top 10 变化最大</h4><table>')
parts.append('<tr><th>排名</th><th>Kernel名称</th><th>输入Shape</th><th class="col-num">基准(us)</th><th class="col-num">比对(us)</th><th class="col-num">差异(us)</th><th class="col-num">比率</th><th>原因</th></tr>')
for i, k in enumerate(top10, 1):
parts.append(f'<tr><td>{i}</td><td style="font-family:monospace;font-size:12px;">{k["kernel"]}</td><td style="font-size:12px;">{k["input_shape"]}</td><td class="col-num">{_fmt(k["base_total"])}</td><td class="col-num">{_fmt(k["comp_total"])}</td><td class="col-num {_diff_color(k["diff_total"])}">{_fmt(k["diff_total"])}</td><td class="col-num {_ratio_color(k["total_ratio"])}">{_fmt(k["total_ratio"])}</td><td><span class="tag {_ratio_color(k["total_ratio"]).replace("bad","tag-bad").replace("good","tag-good").replace("neutral","tag-neutral")}">{k.get("degradation_type", "-")}</span></td></tr>')
parts.append('</table>')
# === Top 10 计算算子绝对耗时(排除通信算子和空Shape) ===
top10_compute = kc.get('top10_compute', [])
if top10_compute:
parts.append('<h4 style="margin:16px 0 4px;color:#7c3aed;">Top 10 计算算子绝对耗时(排除通信算子)</h4>')
parts.append('<div style="font-size:12px;color:#6b7280;margin-bottom:8px;">输入数据量 = Shape 各维度相乘后所有 tensor 求和。基准与比对共用同一 Shape,故数据量相同;关注 calls 变化和耗时方差判断负载均衡。</div>')
parts.append('<div class="table-wrap"><table>')
parts.append('<tr><th>排名</th><th>Kernel名称</th><th class="col-num">数据量</th><th class="col-num">calls(基准/比对)</th><th class="col-num">耗时(us) 总/均/最大/最小</th><th class="col-num">方差</th></tr>')
for i, k in enumerate(top10_compute, 1):
calls_changed = k.get('calls_changed', False)
calls_str = f'{k["base_calls"]} / {k["comp_calls"]}'
if calls_changed:
calls_str += ' <span class="tag tag-bad">变化</span>'
variance = k.get('variance_ratio', 0)
variance_str = f'{variance}x' if variance > 0 else '-'
variance_color = 'bad' if variance > 2 else 'neutral'
dur_str = f'{_fmt(k["comp_total"])} / {_fmt(k["comp_avg"])} / {_fmt(k["comp_max"])} / {_fmt(k["comp_min"])}'
parts.append(f'<tr><td>{i}</td><td style="font-family:monospace;font-size:12px;" title="{k["input_shape"]}">{k["kernel"]}</td><td class="col-num">{k.get("total_elements_fmt", "-")}</td><td class="col-num">{calls_str}</td><td class="col-num">{dur_str}</td><td class="col-num {variance_color}">{variance_str}</td></tr>')
parts.append('</table></div>')
# === 计算算子负载均衡分析(仅calls>10的计算算子) ===
imbalance = kc.get('load_imbalance', [])
if imbalance:
parts.append(f'<h4 id="kernel-imbalance" style="margin:16px 0 4px;color:#dc2626;">[预警] 计算算子负载不均预警(max/min > 2x,共 {kc.get("load_imbalance_count", 0)} 个)</h4>')
parts.append('<div style="font-size:12px;color:#6b7280;margin-bottom:8px;">负载不均 = 同一 Kernel 在多次调用中耗时差异大(最大值/最小值 > 2x)。通常由输入数据量不均匀(如 MoE 专家路由不均衡)或系统调度抖动导致。方差倍数越高,说明部分调用的耗时远超平均水平。</div>')
parts.append('<div class="table-wrap"><table>')
parts.append('<tr><th>Kernel名称</th><th class="col-num">输入数据量</th><th class="col-num">平均(us)</th><th class="col-num">最大(us)</th><th class="col-num">最小(us)</th><th class="col-num">方差倍数</th><th class="col-num">调用次数</th><th>严重程度</th></tr>')
for k in imbalance:
sev = k.get('imbalance_severity', '')
sev_tag = f'<span class="tag tag-bad">{sev}</span>' if sev in ('严重', '显著') else f'<span class="tag tag-neutral">{sev}</span>'
parts.append(f'<tr><td style="font-family:monospace;font-size:12px;" title="{k["input_shape"]}">{k["kernel"]}</td><td class="col-num">{k.get("total_elements_fmt", "-")}</td><td class="col-num">{_fmt(k["comp_avg"])}</td><td class="col-num bad">{_fmt(k["comp_max"])}</td><td class="col-num good">{_fmt(k["comp_min"])}</td><td class="col-num bad">{k.get("variance_ratio", 0)}x</td><td class="col-num">{k["comp_calls"]}</td><td>{sev_tag}</td></tr>')
parts.append('</table></div>')
else:
parts.append('<div style="margin:16px 0;padding:12px;background:#dcfce7;border-radius:8px;color:#16a34a;font-size:14px;">[正常] 计算算子未检测到负载不均(所有 calls>10 的计算 Kernel 的 max/min 比值 < 2x)</div>')
# === 多 Shape 计算算子分布 ===
multi_shape = kc.get('multi_shape_kernels', {})
if multi_shape:
parts.append(f'<h4 style="margin:16px 0 4px;color:#2563eb;">同名计算算子多 Shape 分布(共 {kc.get("multi_shape_kernel_count", 0)} 个)</h4>')
parts.append('<div style="font-size:12px;color:#6b7280;margin-bottom:8px;">同一 Kernel 名称对应不同输入 Shape,反映模型中不同层/专家的数据量分配差异。数据量差异大说明各层负载不均。</div>')
for name, entries in multi_shape.items():
parts.append(f'<div style="margin-bottom:12px;"><div style="font-weight:600;font-family:monospace;font-size:13px;margin-bottom:4px;">{name}</div><table>')
parts.append('<tr><th>Shape</th><th class="col-num">输入数据量</th><th class="col-num">总耗时(us)</th><th class="col-num">平均(us)</th><th class="col-num">调用次数</th></tr>')
for e in entries:
parts.append(f'<tr><td style="font-size:11px;">{e["shape"][:40]}</td><td class="col-num">{e.get("total_elements_fmt", "-")}</td><td class="col-num">{_fmt(e["comp_total"])}</td><td class="col-num">{_fmt(e["comp_avg"])}</td><td class="col-num">{e["comp_calls"]}</td></tr>')
parts.append('</table></div>')
# === 通信算子概览(单独区域,不分析Shape和方差) ===
comm_kernels = kc.get('comm_kernels', [])
if comm_kernels:
parts.append('<details style="margin-top:16px;"><summary style="cursor:pointer;font-weight:600;color:#6b7280;font-size:14px;margin-bottom:8px;">通信算子耗时概览(点击展开)</summary><table>')
parts.append('<tr><th>通信算子</th><th class="col-num">基准(us)</th><th class="col-num">比对(us)</th><th class="col-num">差异(us)</th><th class="col-num">比率</th><th class="col-num">平均(us)</th><th class="col-num">最大(us)</th><th class="col-num">最小(us)</th><th class="col-num">调用次数</th></tr>')
for k in comm_kernels:
parts.append(f'<tr><td style="font-family:monospace;font-size:12px;">{k["kernel"]}</td><td class="col-num">{_fmt(k.get("base_total", 0))}</td><td class="col-num">{_fmt(k["comp_total"])}</td><td class="col-num {_diff_color(k["diff_total"])}">{_fmt(k["diff_total"])}</td><td class="col-num {_ratio_color(k["total_ratio"])}">{_fmt(k["total_ratio"])}</td><td class="col-num">{_fmt(k["comp_avg"])}</td><td class="col-num">{_fmt(k["comp_max"])}</td><td class="col-num">{_fmt(k["comp_min"])}</td><td class="col-num">{k["comp_calls"]}</td></tr>')
parts.append('</table></details>')
return ''.join(parts)
def _render_api_compare(api):
"""渲染 API 比对部分"""
parts = []
trend = api.get('net_trend', '')
trend_color = '#16a34a' if trend == '改善' else '#dc2626' if trend == '劣化' else '#6b7280'
parts.append(f"""
<div class="kpi-grid">
<div class="kpi-card"><div class="label">API 总数</div><div class="value">{api.get('total_apis', 0)}</div></div>
<div class="kpi-card"><div class="label">劣化总量</div><div class="value bad">+{_fmt(api.get('total_degraded_ms', 0), 'ms')}</div><div class="sub">{api.get('degraded_count', 0)} 个</div></div>
<div class="kpi-card"><div class="label">改善总量</div><div class="value good">-{_fmt(api.get('total_improved_ms', 0), 'ms')}</div><div class="sub">{api.get('improved_count', 0)} 个</div></div>
<div class="kpi-card"><div class="label">净趋势</div><div class="value" style="color:{trend_color};">{trend}</div><div class="sub">纯耗时变化: {api.get('pure_duration_change_count', 0)} / 调用次数变化: {api.get('call_count_change_count', 0)}</div></div>
</div>""")
top10 = api.get('top10', [])
if top10:
parts.append('<table><tr><th>排名</th><th>API名称</th><th class="col-num">基准(ms)</th><th class="col-num">比对(ms)</th><th class="col-num">差异(ms)</th><th class="col-num">Self比率</th><th class="col-num">次数比率</th><th>原因</th></tr>')
for i, a in enumerate(top10, 1):
parts.append(f'<tr><td>{i}</td><td style="font-family:monospace;font-size:12px;">{a["name"]}</td><td class="col-num">{_fmt(a["base_total"])}</td><td class="col-num">{_fmt(a["comp_total"])}</td><td class="col-num {_diff_color(a["diff_total"])}">{_fmt(a["diff_total"])}</td><td class="col-num {_ratio_color(a["self_ratio"])}">{_fmt(a["self_ratio"])}</td><td class="col-num {_ratio_color(a["calls_ratio"])}">{_fmt(a["calls_ratio"])}</td><td><span class="tag tag-neutral">{a.get("degradation_type", "-")}</span></td></tr>')
parts.append('</table>')
return ''.join(parts)
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| canvas-designanthropics/skillsCreate beautiful visual art in .png and .pdf documents using design philosophy. You should use this skill when the user asks to create a poster, piece of art, design, or other static piece. Create original visual designs, never copying existing artists' work to avoid copyright violations. | Stars: 197k | Updated | Claude CodeCodex | Verified | Install |
| docxanthropics/skillsUse this skill whenever the user wants to create, read, edit, or manipulate Word documents (.docx files) or Word templates (.dotx files). Triggers include: any mention of 'Word doc', 'word document', '.docx', '.dotx', or requests to produce professional documents with formatting like tables of contents, headings, page numbers, or letterheads. Also use when extracting or reorganizing content from .docx or .dotx files, inserting or replacing images in documents, performing find-and-replace in Word files, working with tracked changes or comments, or converting content into a polished Word document. If the user asks for a 'report', 'memo', 'letter', 'template', or similar deliverable as a Word or .docx file, use this skill. Do NOT use for PDFs, spreadsheets, Google Docs, or general coding tasks unrelated to document generation. | Stars: 197k | Updated | Claude CodeCodex | Verified | Install |
| pdfanthropics/skillsUse this skill whenever the user wants to do anything with PDF files. This includes reading or extracting text/tables from PDFs, combining or merging multiple PDFs into one, splitting PDFs apart, rotating pages, adding watermarks, creating new PDFs, filling PDF forms, encrypting/decrypting PDFs, extracting images, and OCR on scanned PDFs to make them searchable. If the user mentions a .pdf file or asks to produce one, use this skill. | Stars: 197k | Updated | Claude CodeCodex | Verified | Install |
| pptxanthropics/skillsUse this skill any time a .pptx or .potx file is involved in any way — as input, output, or both. This includes: creating slide decks, pitch decks, or presentations; reading, parsing, or extracting text from any .pptx or .potx file (even if the extracted content will be used elsewhere, like in an email or summary); editing, modifying, or updating existing presentations; combining or splitting slide files; working with templates (.potx), layouts, speaker notes, or comments. Trigger whenever the user mentions "deck," "slides," "presentation," or references a .pptx or .potx filename, regardless of what they plan to do with the content afterward. If a .pptx or .potx file needs to be opened, created, or touched, use this skill. | Stars: 197k | Updated | Claude CodeCodex | Verified | Install |