ascend-profiling-analysis
Automates performance profiling analysis for models trained on Ascend NPUs.
Install
mkdir -p .claude/skills/ascend-profiling-analysis && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/15606" && unzip -o skill.zip -d .claude/skills/ascend-profiling-analysis && rm skill.zipInstalls to .claude/skills/ascend-profiling-analysis
Activation
This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.
Analyze Ascend NPU profiling data to identify training performance bottlenecks. Breaks down step-level time into compute, unoverlapped communication, and freetime; within compute, analyzes compute vs memory-bound ratios and cube vs vector utilization to summarize the model's performance bottleneck.Key capabilities
- →Analyze Ascend NPU profiling data
- →Identify training performance bottlenecks
- →Break down step-level time into compute, communication, and freetime
- →Analyze compute vs memory-bound ratios
- →Summarize cube vs vector utilization
How it works
This skill analyzes Ascend NPU profiling data by executing a Python script that breaks down step-level time into compute, communication, and freetime. It further analyzes compute vs memory-bound ratios and cube vs vector utilization to identify training performance bottlenecks.
Inputs & outputs
When to use ascend-profiling-analysis
- →Identify training bottlenecks
- →Analyze operator runtime
- →Optimize npu performance
- →Check compute vs memory bound
About this skill
模型 Profiling 数据分析
对基于 Ascend NPU 训练的模型，通过 torch.profile 采集的 profiling 数据进行自动化分析，拆解模型性能瓶颈，统计 cube/vector 算子耗时比例，计算和访存比例。
触发条件
当用户提供 profiling 数据目录路径并要求以下任一分析时调用此 skill：
- 分析模型训练耗时瓶颈
- 分析 step trace 中计算/通信/空闲占比
- 分析 kernel 算子耗时分布
- 分析计算 vs 访存带宽比例
执行步骤
1. 确认路径
用户需提供 profiling 数据根目录（包含 *_ascend_pt 子目录的路径），例如：
D:\profile\gr00t-n1.6
如果用户未提供路径，使用 AskUserQuestion 询问。
2. 运行分析脚本
执行以下命令（使用 miniconda3 的 Python 环境）：
& "$env:USERPROFILE\miniconda3\python.exe" "D:\code\DrivingSDK\mx_driving\tools\profile_analyse\profile_analyse.py" "<用户提供的路径>"
3. 解读结果
脚本会输出以下分析内容，将结果整理后呈现给用户：
Section 1 — Step Trace 分析：
- 各卡耗时明细表（Computing / Comm / FreeTime in us）
- 数据分析表（耗时占比 + 耗时波动 min~max）
Section 2 — Kernel Details 分析：
- 每种算子计算耗时与访存带宽耗时占比（Top 20，含计算类型 cube/vector）
Section 3 — 整网计算 vs 访存比例：
- 计算:访存百分比比值（如 43.20% : 56.80%）
- Cube:Vector 百分比比值（如 64.85% : 35.15%）
CSV 输出文件（保存在 mx_driving\tools\profile_analyse\ 目录下，所有数值保留两位小数）：
| 文件 | 说明 |
|---|---|
step_trace_analysis.csv | Step 1 — 各卡耗时明细 + Summary（耗时占比与波动） |
top20_ops_compute_memory_breakdown.csv | Step 2 — Top 20 算子计算/访存耗时占比 |
computing_summary.csv | Step 3 — cube/vector 大类汇总 + 计算:访存 / Cube:Vector 比例 |
注意事项
- Kernel Details 分析只取第一张卡的 kernel_details.csv 以加速计算
- 若 step_trace_time.csv 包含多个 Step，脚本自动取最后一个 Step 进行分析
依赖
- Python 环境：miniconda3（路径
$env:USERPROFILE\miniconda3\python.exe） - 三方库：pandas, numpy
When not to use it
- →The task does not involve Ascend NPU profiling data
- →The task does not require analysis of training performance bottlenecks
Prerequisites
Limitations
- →Kernel Details analysis only takes the first card's kernel_details.csv
- →If step_trace_time.csv contains multiple Steps, the script automatically takes the last Step for analysis
How it compares
This skill automates the analysis of complex Ascend NPU profiling data, providing a structured breakdown of performance bottlenecks and detailed reports, which is more efficient than manual interpretation of raw profiling logs.
Compared to similar skills
ascend-profiling-analysis side by side with the closest alternatives in the catalog.
| Skill | Installs | Updated | Safety | Difficulty |
|---|---|---|---|---|
| ascend-profiling-analysis (this skill) | 0 | 2mo | No flags | Advanced |
| google-analytics | 43 | 7mo | Review | Intermediate |
| backtesting-trading-strategies | 10 | 1mo | Review | Intermediate |
| model-usage | 5 | 2mo | Review | Beginner |
Try saying
Example prompts that trigger this skill in your AI assistant.
More by Ascend
View all by Ascend →You might also like
google-analytics
davila7
Analyze Google Analytics data, review website performance metrics, identify traffic patterns, and suggest data-driven improvements. Use when the user asks about analytics, website metrics, traffic analysis, conversion rates, user behavior, or performance optimization.
backtesting-trading-strategies
jeremylongshore
Backtest crypto and traditional trading strategies against historical data. Calculates performance metrics (Sharpe, Sortino, max drawdown), generates equity curves, and optimizes strategy parameters. Use when user wants to test a trading strategy, validate signals, or compare approaches. Trigger with phrases like "backtest strategy", "test trading strategy", "historical performance", "simulate trades", "optimize parameters", or "validate signals".
model-usage
openclaw
Use CodexBar CLI local cost usage to summarize per-model usage for Codex or Claude, including the current (most recent) model or a full model breakdown. Trigger when asked for model-level usage/cost data from codexbar, or when you need a scriptable per-model summary from codexbar cost JSON.
tracking-crypto-derivatives
jeremylongshore
Track cryptocurrency futures, options, and perpetual swaps with funding rates, open interest, liquidations, and comprehensive derivatives market analysis. Use when monitoring derivatives markets, analyzing funding rates, tracking open interest, finding liquidation levels, or researching options flow. Trigger with phrases like "funding rate", "open interest", "perpetual swap", "futures basis", "liquidation levels", "options flow", "put call ratio", "derivatives analysis", or "BTC perps".
spark-optimization
wshobson
Optimize Apache Spark jobs with partitioning, caching, shuffle optimization, and memory tuning. Use when improving Spark performance, debugging slow jobs, or scaling data processing pipelines.
weights-and-biases
davila7
Track ML experiments with automatic logging, visualize training in real-time, optimize hyperparameters with sweeps, and manage model registry with W&B - collaborative MLOps platform