🚀 ACLNN Supported Dashboard

English | 中文

Dashboard: https://fzilan.github.io/aclnn-dashboard/

Automatically generates and visualizes ACLNN support coverage across Torch-NPU and MindSpore, including:

• Daily automated scanning and data refresh
• Coverage comparison and 7-day onboarding speed trend

Repository Structure

aclnn-dashboard/
├─ .github/workflows/
│  └─ refresh-aclnn-dashboard.yml     # Daily data refresh workflow
├─ data/
│  ├─ raw/
│  │  └─ aclnn-aa.generated.md        # Full ACLNN list (after crawl + cleanup)
│  └─ reports/
│     ├─ aclnn_to_torch_npu.csv/.md   # Torch-NPU scan report
│     ├─ aclnn_to_mindspore.csv/.md   # MindSpore scan report
│     └─ aclnn_to_all.csv/.md         # Merged comparison report
├─ scripts/
│  ├─ crawl/
│  │  ├─ aclnn_scrape_op_api_list.py
│  │  └─ clean_aclnn_md_table.py
│  ├─ scan/
│  │  ├─ aclnn_to_torch_npu_report.py
│  │  ├─ aclnn_to_mindspore_report.py
│  │  └─ aclnn_merge_report.py
│  └─ build/
│     ├─ build_dashboard_data.py
│     └─ update_coverage_history.py
├─ index.html                         # Dashboard page
├─ data.json                          # Frontend-consumable data
├─ coverage_history.json              # Daily coverage history
└─ run_pipeline.sh                    # One-command pipeline

What It Does

1. Crawls the full ACLNN API list and normalizes it
2. Scans codebases independently:
   • op-plugin (master) for Torch-NPU coverage
   • mindspore (master) for MindSpore coverage
3. Merges both sides into a unified comparison report
4. Builds data.json (metrics + operators + history)
5. Maintains coverage_history.json and computes 7-day speed (pp/day)

Requirements

• Python 3.9+
• Python packages: pyyaml, playwright
• Playwright browser: chromium

Install example:

python3 -m pip install -U pip pyyaml playwright
python3 -m playwright install chromium

Local Usage

Setup

git clone https://github.com/Fzilan/aclnn-dashboard.git
cd aclnn-dashboard

Choose One Workflow (3 options)

1. Scan + Merge + Build (recommended for daily updates)

bash run_pipeline.sh \
  --skip-scrape \
  --op-plugin-root /path/to/op-plugin \
  --mindspore-root /path/to/mindspore

2. Full pipeline including ACLNN crawl (only when ACLNN upstream list changes)

bash run_pipeline.sh \
  --op-plugin-root /path/to/op-plugin \
  --mindspore-root /path/to/mindspore

3. Rebuild frontend data only (no scan)

python3 scripts/build/build_dashboard_data.py \
  --history-file coverage_history.json \
  --output data.json

Preview locally

python3 -m http.server 8000

Open: http://localhost:8000

Automation (CI)

Workflow: .github/workflows/refresh-aclnn-dashboard.yml

• Schedule: daily at 08:30 Beijing Time (UTC 00:30)
• Pulls automatically:
  • https://gitcode.com/Ascend/op-plugin at master
  • https://gitcode.com/mindspore/mindspore at master
• Runs run_pipeline.sh --skip-scrape (skip ACLNN crawling by default)
• Auto commits and pushes when artifacts change
• After workflow succeeds, view the dashboard directly on GitHub Pages: https://fzilan.github.io/aclnn-dashboard/

Frontend Data Contract

• data.json.metrics: global metrics + 7-day speed
• data.json.operators: per-operator details (Torch/MindSpore evidence)
• data.json.history.daily_coverage: historical daily series
• coverage_history.json: source of daily snapshots (deduplicated by date)