configuration-ui
Guide JSON config editing, smoke configs, CLI override precedence, stage command construction, metrics JSONL inspection, and the Streamlit control panel/job manager.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- configuration-ui
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Guide JSON config editing, smoke configs, CLI override precedence, stage command construction, metrics JSONL inspection, and the Streamlit control panel/job manager.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/configuration.md7.4 KB
- 参考文档references/control-panel-ui.md6.7 KB
- 参考文档references/troubleshooting.md7.7 KB
- 脚本scripts/inspect_metrics_jsonl.py5.9 KB
- 脚本scripts/print_config_summary.py6.3 KB
03 SKILL.md 正文
configuration-ui
Use this sub-skill when a future Researcher needs to edit or inspect this repo's JSON stage configs, build safe training/UI command lines, use smoke configs, inspect metrics JSONL, or operate the Streamlit control panel.
Read first
- For config precedence, dataclass fields, smoke/full JSON behavior, and CLI override rules, read references/configuration.md.
- For Streamlit pages, job registry behavior, GPU-busy guarding, logs, and UI metrics, read references/control-panel-ui.md.
- For common config/UI/metrics failures, read references/troubleshooting.md.
Safe bundled helpers
- scripts/print_config_summary.py summarizes one or more JSON config files and compares explicit keys against a base JSON without importing the repo.
- scripts/inspect_metrics_jsonl.py summarizes metrics JSONL columns, last row, numeric ranges, and malformed-line counts without pandas or repo imports.
Both helpers are read-only by default, have --help, and include --demo tiny
fixture modes.
Route elsewhere
- Algorithm choices, loss interpretation, RLHF stage sequencing, KL/reward tuning, and training-stage failure analysis: use ../post-training-rlhf/SKILL.md.
- Base architecture, legacy pretraining, DDP/bf16 pretraining, checkpoint structure, and memory planning: use ../model-pretraining/SKILL.md.
- Dataset schemas, packed HDF5 validation, preference JSONL validation, and RL prompt validation: use ../data-preparation/SKILL.md.
- GSM8K evaluation details, direct chat/raw inference, checkpoint loading, and sampling controls: use ../evaluation-chat/SKILL.md.
Operating guardrails
- Prefer
configs/smoke/*.jsonfor command/config smoke checks; those use a siblingconfigs/smoke/base.jsonand are CPU-oriented. - Use
--print-configbefore launching long jobs when changing JSON or CLI overrides. - Treat UI launches as process-spawning operations: check the UI job registry status and the GPU-busy guard before starting GPU jobs.
- Treat Weights & Biases as optional. The repo always writes metrics JSONL under the configured log directory even when W&B is disabled or unavailable.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).