plc-llm-bench / scripts /make_notebook.py
iulio's picture
Upload folder using huggingface_hub
3bccd69 verified
Raw History Blame Contribute Delete
12.3 kB
"""Build the self-contained Colab T4 notebook PLC_LLM_Bench_Colab.ipynb.
Usage: python scripts/make_notebook.py
"""
from __future__ import annotations
import json
import pathlib
ROOT = pathlib.Path(__file__).resolve().parents[1]
def source_files() -> list[pathlib.Path]:
files = [*sorted((ROOT / "plc_bench").glob("*.py")), *sorted((ROOT / "scripts").glob("*.py"))]
return [*files, *sorted((ROOT / "tests").glob("*.py")), ROOT / "pyproject.toml"]
def _chunk_text(text: str, max_bytes: int = 26_000) -> list[str]:
"""Split text along line boundaries into UTF-8 chunks under ``max_bytes``."""
chunks: list[str] = []
current: list[str] = []
current_bytes = 0
for line in text.splitlines(keepends=True):
line_bytes = len(line.encode("utf-8"))
if current and current_bytes + line_bytes > max_bytes:
chunks.append("".join(current))
current = [line]
current_bytes = line_bytes
else:
current.append(line)
current_bytes += line_bytes
if current:
chunks.append("".join(current))
return chunks
def cells() -> list[tuple[str, str]]:
out: list[tuple[str, str]] = [
(
"markdown",
"# PLC-LLM-Bench v0.2: Offline Industrial Automation LLM Benchmark (Colab & Kaggle)\n\n"
"Evaluates **13 open-weight LLMs** (Tier 1: <=12 GB **RTX 5070 Ti Mobile** fit; Tier 2: 12–16 GB / MoE partial offload) "
"on **30 custom Siemens TIA Portal SCL** tasks (5 categories) and **6 academic IEC 61131-3 ST** tasks.\n\n"
"Pipeline: `llama.cpp` (`llama-server` CUDA `sm_75`, `q8_0` KV cache) -> SCL Dialect Linter -> "
"SCL-to-IEC Normalizer -> `MatIEC iec2c` + `gcc` -> Multi-Scan PLC Simulator (`scl_interp.py`) -> Pass@2 Compiler Self-Repair.\n\n"
"Works automatically on **Google Colab** (`/content` + optional Drive) and **Kaggle** (`/kaggle/working`).",
),
("code", "!nvidia-smi --query-gpu=name,memory.total,memory.free --format=csv\n!df -h /content /kaggle/working 2>/dev/null || df -h ."),
(
"code",
"import os, pathlib\n"
"IS_KAGGLE = os.path.isdir('/kaggle/working')\n"
"WORK_ROOT = pathlib.Path('/kaggle/working' if IS_KAGGLE else '/content')\n"
"PROJECT_DIR = WORK_ROOT / 'plc-llm-bench'\n"
"MATIEC_DIR = WORK_ROOT / 'matiec'\n"
"LLAMA_DIR = WORK_ROOT / 'llama_bin'\n"
"GGUF_CACHE = WORK_ROOT / 'gguf_cache'\n"
"AUTOPLC_DIR = WORK_ROOT / 'AutoPLC'\n"
"AGENTS4PLC_DIR = WORK_ROOT / 'Agents4PLC_release'\n"
"USE_GDRIVE = (not IS_KAGGLE) # Set False on Colab to run solely in /content\n"
"RESULTS_DIR = PROJECT_DIR / 'results'\n"
"BIN_CACHE = WORK_ROOT / 'bin_cache'\n"
"if USE_GDRIVE:\n"
" try:\n"
" from google.colab import drive\n"
" drive.mount('/content/drive')\n"
" RESULTS_DIR = pathlib.Path('/content/drive/MyDrive/plc-llm-bench/results')\n"
" BIN_CACHE = pathlib.Path('/content/drive/MyDrive/plc-llm-bench/bin_t4')\n"
" except Exception as exc:\n"
" print('Drive mount skipped:', exc)\n"
"RESULTS_DIR.mkdir(parents=True, exist_ok=True)\n"
"BIN_CACHE.mkdir(parents=True, exist_ok=True)\n"
"for d in ('plc_bench', 'scripts', 'tests'):\n"
" (PROJECT_DIR / d).mkdir(parents=True, exist_ok=True)\n"
"os.chdir(PROJECT_DIR)\n"
"print(f'Workspace: {PROJECT_DIR} (Kaggle={IS_KAGGLE})')\n"
"print(f'Results: {RESULTS_DIR}')\n"
"!pip -q install -U huggingface_hub requests numpy matplotlib pytest",
),
(
"markdown",
"## Build / Restore MatIEC (`iec2c`) & `llama.cpp` (`llama-server` for T4 `sm_75`)",
),
(
"code",
"%%bash\n"
"set -e\n"
"WORK_ROOT=/content; [ -d /kaggle/working ] && WORK_ROOT=/kaggle/working\n"
"apt-get -qq update && apt-get -qq install -y build-essential git autoconf automake libtool bison flex cmake ninja-build >/dev/null\n"
"if [ ! -x \"$WORK_ROOT/matiec/iec2c\" ]; then\n"
" git clone --depth=1 https://github.com/beremiz/matiec.git \"$WORK_ROOT/matiec\"\n"
" cd \"$WORK_ROOT/matiec\" && autoreconf -i >/dev/null && ./configure >/dev/null && make -j$(nproc) >/dev/null\n"
"fi\n"
"ln -sf \"$WORK_ROOT/matiec/iec2c\" /usr/local/bin/iec2c\n"
"echo 'MATIEC:' && /usr/local/bin/iec2c -h | head -n 3\n",
),
(
"code",
"import json, urllib.request, subprocess, pathlib, os\n"
"target_server = pathlib.Path('/usr/local/bin/llama-server')\n"
"if not target_server.is_file():\n"
" req = urllib.request.Request(\n"
" 'https://api.github.com/repos/ggml-org/llama.cpp/releases?per_page=10',\n"
" headers={'User-Agent': 'plc-llm-bench'},\n"
" )\n"
" with urllib.request.urlopen(req, timeout=20) as resp:\n"
" rels = json.loads(resp.read().decode())\n"
" chosen = next(\n"
" r for r in rels\n"
" if any(a['name'].startswith('llama-') and 'bin-ubuntu-cuda-12.8-x64' in a['name'] for a in r.get('assets', []))\n"
" and any(a['name'].startswith('cudart-') and 'bin-ubuntu-cuda-12.8-x64' in a['name'] for a in r.get('assets', []))\n"
" )\n"
" llama_url = next(a['browser_download_url'] for a in chosen['assets'] if a['name'].startswith('llama-') and 'bin-ubuntu-cuda-12.8-x64' in a['name'])\n"
" cudart_url = next(a['browser_download_url'] for a in chosen['assets'] if a['name'].startswith('cudart-') and 'bin-ubuntu-cuda-12.8-x64' in a['name'])\n"
" LLAMA_DIR.mkdir(parents=True, exist_ok=True)\n"
" subprocess.run(f\"curl -fsSL -L '{llama_url}' -o /tmp/l.tar.gz && tar -xzf /tmp/l.tar.gz -C '{LLAMA_DIR}' && rm /tmp/l.tar.gz\", shell=True, check=True)\n"
" subprocess.run(f\"curl -fsSL -L '{cudart_url}' -o /tmp/c.tar.gz && tar -xzf /tmp/c.tar.gz -C '{LLAMA_DIR}' && rm /tmp/c.tar.gz\", shell=True, check=True)\n"
" srv = next(LLAMA_DIR.rglob('llama-server'))\n"
" cudart_dir = next(LLAMA_DIR.rglob('libcudart.so.12')).parent\n"
" for so in cudart_dir.glob('*.so*'):\n"
" dst = srv.parent / so.name\n"
" if not dst.exists():\n"
" dst.symlink_to(so)\n"
" target_server.write_text(f'#!/bin/bash\\nexport LD_LIBRARY_PATH={srv.parent}:${{LD_LIBRARY_PATH}}\\nexec {srv} \"$@\"\\n')\n"
" target_server.chmod(0o755)\n"
"!/usr/local/bin/llama-server --list-devices",
),
("markdown", "## Project Sources\n\nWrite `plc_bench`, `scripts`, and `tests` files."),
]
for f in source_files():
if f.name == "make_notebook.py":
continue
rel = f.relative_to(ROOT).as_posix()
txt = f.read_text(encoding="utf-8")
if len(txt.encode("utf-8")) > 26_000:
# Split oversized files across multiple cells to stay under Colab/Jupyter payload limits
for idx, chunk in enumerate(_chunk_text(txt, max_bytes=26_000)):
flag = "" if idx == 0 else "-a "
out.append(("code", f"%%writefile {flag}{rel}\n" + chunk))
continue
out.append(("code", f"%%writefile {rel}\n" + txt))
out += [
("markdown", "## Pre-Flight Verification (`pytest`)"),
("code", "!python -m pytest -q"),
("markdown", "## Fetch Academic PLC Benchmarks (`AutoPLC` & `Agents4PLC_release`)"),
(
"code",
"%%bash\n"
"set -e\n"
"WORK_ROOT=/content; [ -d /kaggle/working ] && WORK_ROOT=/kaggle/working\n"
"[ -d \"$WORK_ROOT/AutoPLC\" ] || git clone --depth=1 https://github.com/cangkui/AutoPLC.git \"$WORK_ROOT/AutoPLC\"\n"
"[ -d \"$WORK_ROOT/Agents4PLC_release\" ] || git clone --depth=1 https://github.com/Luoji-zju/Agents4PLC_release.git \"$WORK_ROOT/Agents4PLC_release\"\n",
),
(
"markdown",
"## Run Benchmark\n\n"
"Choose `MODE`: `'smoke'` (1 fast model, 3 tasks), `'tier1'` (8 models <=12 GB), `'tier2'` (5 models 12–16 GB / MoE), or `'all'`.",
),
(
"code",
"import os, pathlib, subprocess, time\n\n"
"MODE = 'smoke' # Change to 'tier1', 'tier2', or 'all'\n"
"os.environ['MATIEC_INCLUDE'] = str(MATIEC_DIR / 'lib')\n"
"os.environ['PYTHONPATH'] = str(PROJECT_DIR) + ':' + os.environ.get('PYTHONPATH', '')\n"
"ckpt = RESULTS_DIR / 'checkpoint.jsonl'\n"
"cmd = [\n"
" 'python', '-u', 'scripts/run_bench.py',\n"
" '--checkpoint', str(ckpt),\n"
" '--out-dir', str(RESULTS_DIR),\n"
" '--llama-server', '/usr/local/bin/llama-server',\n"
" '--cache-dir', str(GGUF_CACHE),\n"
" '--autoplc-dir', str(AUTOPLC_DIR),\n"
" '--agents4plc-dir', str(AGENTS4PLC_DIR),\n"
"]\n"
"if MODE == 'smoke':\n"
" cmd += ['--smoke', '--models', 'qwen3.5-9b']\n"
"elif MODE == 'tier1':\n"
" cmd += ['--tier', '1']\n"
"elif MODE == 'tier2':\n"
" cmd += ['--tier', '2']\n\n"
"log_file = PROJECT_DIR / 'bench.log'\n"
"with open(log_file, 'w', encoding='utf-8') as lf:\n"
" proc = subprocess.Popen(cmd, stdout=lf, stderr=subprocess.STDOUT, text=True, cwd=str(PROJECT_DIR))\n\n"
"print(f\"Launched benchmark (PID {proc.pid}) in background.\")\n"
"print(f\"Streaming {log_file} (polling up to 120s)...\\n\")\n"
"t0 = time.time()\n"
"last_pos = 0\n"
"while time.time() - t0 < 120:\n"
" if log_file.is_file():\n"
" with open(log_file, 'r', encoding='utf-8', errors='ignore') as f:\n"
" f.seek(last_pos)\n"
" new_chunk = f.read()\n"
" if new_chunk:\n"
" print(new_chunk, end='', flush=True)\n"
" last_pos = f.tell()\n"
" if proc.poll() is not None:\n"
" print(f\"\\n[DONE] Process finished with return code {proc.returncode}\")\n"
" break\n"
" time.sleep(3)",
),
("markdown", "## Leaderboard & Pareto Charts"),
(
"code",
"from IPython.display import Image, Markdown, display\n"
"md_file = RESULTS_DIR / 'leaderboard.md'\n"
"png_file = RESULTS_DIR / 'benchmark_charts.png'\n"
"if md_file.is_file():\n"
" display(Markdown(md_file.read_text(encoding='utf-8')))\n"
"if png_file.is_file():\n"
" display(Image(filename=str(png_file)))",
),
]
return out
def main() -> None:
nb = {
"nbformat": 4,
"nbformat_minor": 5,
"metadata": {
"accelerator": "GPU",
"colab": {"gpuType": "T4", "provenance": []},
"kernelspec": {"display_name": "Python 3", "name": "python3"},
"language_info": {"name": "python"},
},
"cells": [
{
"cell_type": kind,
"id": f"cell-{i:02d}",
"metadata": {},
"source": src.splitlines(keepends=True),
**({"outputs": [], "execution_count": None} if kind == "code" else {}),
}
for i, (kind, src) in enumerate(cells())
],
}
out_nb = ROOT / "PLC_LLM_Bench_Colab.ipynb"
out_nb.write_text(json.dumps(nb, indent=1), encoding="utf-8")
print(f"Wrote {out_nb} ({len(nb['cells'])} cells)")
if __name__ == "__main__":
main()