AgentSkillsCN

kebab-gemm-config-tuning

在 Kebab 中使用 config.yaml 和优化文档调整 GEMM 实验设置。适用于被要求调整矩阵尺寸、精度、布局模式、分块大小、内核版本，或设计可重复的性能实验时使用。

SKILL.md
--- frontmatter
name: kebab-gemm-config-tuning
description: Tune GEMM experiment settings in Kebab using config.yaml and optimization docs. Use when asked to adjust matrix sizes, precision, layout modes, tile sizes, kernel versions, or to design reproducible performance experiments.

Kebab GEMM Config Tuning

When to Use This Skill

  • •User asks to change GEMM benchmark parameters
  • •User asks to compare layout modes (RR, RC, CR, CC)
  • •User asks to test precision (float16, bfloat16, float32) or kernel version lists
  • •User asks for tile-size and occupancy-oriented tuning plans

Key Configuration Surface

Main file: config.yaml

Important keys under operators.gemm:

  • •impl
  • •versions
  • •matrix_sizes
  • •modes
  • •precisions
  • •tile_sizes
  • •init_method

Step-by-Step Workflow

  1. •Edit config.yaml with one controlled change at a time.
  2. •Build and run benchmark:
    • •make build
    • •make bench-gemm
  3. •For deeper analysis, run:
    • •make tune-gemm-cute (or cuda/ref)
  4. •Save and compare outputs from bench_results/, profiling/, and reports/.

Tuning Rules from Repository Context

  • •Hopper/WGMMA paths generally expect sm_90a.
  • •The documented tile constraint notes K dimension compatibility requirements for GMMA pathways.
  • •Keep experiment dimensions and precision explicit to ensure reproducibility.

Troubleshooting

IssueMitigation
Unexpected perf regressionRevert to previous versions/tile_sizes baseline and isolate one variable
Precision mismatchVerify selected kernel path supports requested precision
Layout-specific anomalyTest all modes and compare independently

References

  • •config.yaml
  • •docs/optimization_plan.md
  • •docs/TILE_SIZE_CONFIGURATION.md
  • •Makefile GEMM benchmark/profile targets