sweep

name	sweep
description	Batch-run one or more ModernTSF experiment configs (concurrently) via `tool/tsf.py run`. Use when the user wants to launch several TOML run configs in one go, train across multiple sweeps, or run experiments on specific GPUs.

When to use / what to ask

Use this skill when the user wants to run multiple experiment configs (or a single config) through the unified runner. Before running, confirm:

1. Config paths — one or more TOML files under configs/runs/. Defaults to configs/runs/run_single_data.toml when none are given.
2. Concurrency — --jobs N runs configs in parallel (default 1). Only raise it when the configs fit together (e.g. one per GPU); heavy trainings on one GPU should stay at --jobs 1.
3. GPUs — --gpus 0,1 round-robins CUDA_VISIBLE_DEVICES across the jobs.

Command

# Single config
uv run python tool/tsf.py run configs/runs/run_single_data.toml

# Two configs in parallel, one per GPU
uv run python tool/tsf.py run configs/runs/sweep_data.toml configs/runs/sweep_model.toml --jobs 2 --gpus 0,1

# Several configs, all on GPU 1, sequentially
uv run python tool/tsf.py run configs/runs/a.toml configs/runs/b.toml --gpus 1

tsf run reports OK / FAIL per config and exits non-zero if any config fails.

Notes

• Each config runs as CUDA_VISIBLE_DEVICES=<gpu> modern-tsf --config <config>.
• To preview what a sweep config expands to (datasets × models × pred_lens) before launching, run:

  uv run python tool/tsf.py inspect --config <config>
  

• Results land in work_dirs/<dataset>/<model>/performance.csv after each run.

Reference

See docs/en/scripts.md for the full tsf command reference and docs/en/configs.md for config structure, sweep syntax, and extends chains.