training-and-experiments
Run and debug H2O LLM Studio training experiments, distributed launchers, backend choices, and experiment artifacts.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- training-and-experiments
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Run and debug H2O LLM Studio training experiments, distributed launchers, backend choices, and experiment artifacts.
- disable-model-invocation
- true
02 资源文件 7
- 参考文档references/backend-and-performance.md5.4 KB
- 参考文档references/experiment-artifacts.md5.1 KB
- 参考文档references/training-cli.md5.6 KB
- 参考文档references/troubleshooting.md6.3 KB
- 脚本scripts/check_training_environment.py10.3 KB
- 脚本scripts/distributed_train_wrapper.sh5.3 KB
- 脚本scripts/make_minimal_config.py8.8 KB
03 SKILL.md 正文
Training And Experiments
Use this sub-skill when the task is to launch, monitor, or debug an H2O LLM Studio experiment after a dataset/config has already been chosen. It owns direct CLI training, GUI-launched training behavior, multi-GPU launch mechanics, backend/performance settings, and the files written to an experiment output directory.
Read These First
- Training CLI: direct
llm_studio/train.py -Yruns, deprecated-C, dynamic--section.fieldoverrides, GUI launch behavior, and multi-GPU commands. - Experiment artifacts: output directory contents, status flags, logs, charts, checkpoints, predictions, and how to decide whether a run finished.
- Backend and performance: CUDA, DeepSpeed, DDP, LoRA/DoRA/RSLoRA, mixed precision, optimizer/scheduler, and memory/runtime trade-offs.
- Troubleshooting: symptoms, likely causes, recovery checks, and routes to sibling sub-skills for data/model/export issues.
Bundled Helpers
scripts/make_minimal_config.py: create a tiny CSV plus CPU-like YAML config and print a non-executing training command.scripts/check_training_environment.py: inspect Python, package, PyTorch/CUDA,CUDA_HOME,nvcc, DeepSpeed importability, and a YAML config without starting training.scripts/distributed_train_wrapper.sh: build a validatedtorchrunordeepspeedmulti-GPU command; dry-run by default and executes only with--execute.
Route Away
- Dataset columns, schema, YAML construction, config class details, and CSV/Parquet validation belong to
../configuration-and-data/SKILL.md. - Model wrapper internals, metrics, losses, prediction semantics, and evaluation internals belong to
../modeling-and-evaluation/SKILL.md. - Interactive prompting, model-card/export preflights, and Hugging Face publishing belong to
../export-and-prompt/SKILL.md. - App startup, work directories, H2O Wave settings, and server lifecycle belong to
../app-and-ui/SKILL.md.
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).