AGENT SKILL

benchmarking-and-results

Use this sub-skill when a user needs to run bounded timm benchmarks, compare inference/train/profile timing modes, interpret bundled result CSVs, or look up ImageNet-1k model metadata.

License 许可
Apache-2.0
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
benchmarking-and-results
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Use this sub-skill when a user needs to run bounded timm benchmarks, compare inference/train/profile timing modes, interpret bundled result CSVs, or look up ImageNet-1k model metadata.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

Benchmarking and Results

Use this sub-skill for safe, reproducible performance checks with timm benchmark tooling and for interpreting bundled result tables.

Route Here When

  • The user asks how to run benchmark.py for inference, train-step, profile, or combined timing.
  • The user wants a bounded benchmark command for one model, a wildcard subset, or a short model-list file.
  • The user asks what benchmark CSV columns mean or how to compare rows across devices, precision, layouts, or PyTorch versions.
  • The user needs ImageNet result fields or model_metadata-in1k pretraining labels explained.
  • The user requests all-model benchmarking and needs safe scoping before launching long bulk work.

Start Safely

  1. Pick the smallest representative scope first: one model, one device, low batch size, and low iteration counts.
  2. Use the bundled command builder to create bounded commands instead of hand-assembling broad runs: python scripts/timm_benchmark_command_builder.py --model resnet50 --bench inference --device cuda --batch-size 32 --num-warm-iter 2 --num-bench-iter 5 --amp
  3. Run debug timing before publication-quality timing, then increase --num-warm-iter, --num-bench-iter, and batch size only after memory and precision choices are stable.
  4. Treat bulk or all-model requests as potentially expensive; require an explicit model pattern, model-list file, or --allow-bulk acknowledgement.

References

  • references/benchmarking.md: benchmark modes, important flags, safe command patterns, and profiling choices.
  • references/results-tables.md: benchmark, validation, and metadata CSV columns plus interpretation cautions.
  • references/troubleshooting.md: OOM, timing comparability, warmup, torch.compile, optional profiler, AMP/layout, and stale metadata issues.
  • scripts/timm_benchmark_command_builder.py: guarded command generator for bounded single-model and optional bulk benchmark commands.

Boundary Notes

  • This sub-skill covers benchmark execution and result interpretation only.
  • Use the cli-workflows sub-skill for full train.py, validate.py, or dataset-driven CLI workflows.
  • Do not present bundled benchmark CSV values as universal hardware truth; they are snapshots from specific accelerators, software versions, precision modes, and layouts.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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).

★ 351