菜单
Run `uv sync` and the four CI gate checks (ruff check, ruff format, pyright, pytest unit). First command after cloning. Auto-handles non-CUDA hosts via `--no-sources`.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | install-and-verify |
| description | Run `uv sync` and the four CI gate checks (ruff check, ruff format, pyright, pytest unit). First command after cloning. Auto-handles non-CUDA hosts via `--no-sources`. |
git clone, before any other skill.git pull if the user asks for it.Run every step in the background so the user isn't blocked — in Claude Code, set run_in_background=true on each Bash call and report status as each completes.
Detect the host. The torch wheel pin in pyproject.toml is CUDA-only:
command -v nvidia-smi >/dev/null 2>&1 && nvidia-smi -L >/dev/null 2>&1 && echo CUDA || echo NO_CUDA
On NO_CUDA, print this warning to the user verbatim before proceeding, and append --no-sources to every uv run and uv sync invocation in this skill (the flag is per-command, not sticky):
No CUDA GPU detected. Using
--no-sourcesto fetch torch from PyPI instead of the CUDA wheel index. Sufficient for the dev loop (lint, type-check, CPU unit tests); training and GPU tests still need a Linux GPU box. Note:--no-sourcesalso bypasses any other[tool.uv.sources]entries the project may add later.
Baseline check (use the form matching your host):
uv run python scripts/check_env.py # CUDA
uv run --no-sources python scripts/check_env.py # non-CUDA
Non-zero exit: print stdout verbatim and stop. Common: install uv (curl -LsSf https://astral.sh/uv/install.sh | sh) or cd into the KempnerForge checkout.
Lint scope: kempnerforge/ tests/ scripts/ Type-check scope: kempnerforge/ Dev group: ruff, pyright[nodejs], pytest, pytest-timeout, vulture GPU test paths (NOT here): tests/integration/, tests/distributed/, tests/e2e/ Non-CUDA fallback: uv sync --no-sources (bypasses [tool.uv.sources] torch CUDA pin)
On non-CUDA hosts, append --no-sources to every uv command below.
Sync:
uv sync
Expected: ~30 s cached, 2–5 min cold.
Confirm Python ≥ 3.12:
uv run python -c "import sys; assert sys.version_info >= (3, 12), sys.version_info; print(sys.version_info[:3])"
Lint:
uv run ruff check kempnerforge/ tests/ scripts/
Format check:
uv run ruff format --check kempnerforge/ tests/ scripts/
Type check:
uv run pyright kempnerforge/
Unit tests (CPU-only):
uv run pytest tests/unit/ -v --timeout=60
Expected: ~890 tests in 10–60 s.
Report each step OK / FAIL. Next:
/kempnerforge:cluster-config/kempnerforge:smoke-test--no-sources is per-command. Forgetting it on one call → that call fails with the torch wheel error..venv from a different Python or torch flavor → rm -rf .venv && uv sync. Symptom: tests run on the wrong Python or pyright diverges from CI.tests/integration/ / tests/distributed/ / tests/e2e/ are GPU-only; skipped here. Use /kempnerforge:smoke-test for the GPU path..github/workflows/ci.yml. Local-vs-CI divergence: check uv --version against astral-sh/setup-uv@v6./kempnerforge:cluster-config — first cluster onboarding step./kempnerforge:smoke-test — short single-GPU training to validate end-to-end.Add a new optimizer to KempnerForge. Covers implementation, registry hook, config fields, tests, and a preset TOML.
First-run setup. Detect SLURM account/partition/QoS from env and write configs/cluster/local.toml so every other skill can preflight.
Per-subsystem status report. What is implemented, what is tested, what is planned but unwired, and known limitations.
Walk through KempnerForge's subsystems in the order a forward pass encounters them. Starting point for anyone new to the codebase.
Submit a training job via sbatch. Wraps singlenode.sh for one node and multinode.sh for multiple. Injects account, partition, QoS, and time overrides from local.toml.
End-to-end one-GPU sanity check. Runs a short training loop to confirm torch, CUDA, NCCL, uv, and the dataloader all work before committing to longer runs.