soar-autoresearch

name	soar-autoresearch
description	For autonomous SOAR inference optimization loops. Trigger when user asks for iterative tuning, overnight experiments, benchmark-driven optimization, or automated keep/discard decisions.

SOAR Strategy Skill

Trigger Conditions

• 用户提到: 自动调优 / 自动实验 / overnight 跑实验 / 提速竞赛 / SOAR。
• 有明确评估指标: S1/S8/Smax + accuracy。

Core Rules

1. Accuracy gate first (normalized_acc >= 97), otherwise fail.
2. Keep only if estimated final score improves.
3. One primary hypothesis per round.
4. Always append results.tsv.

Strategy Ladder

1. args_tuning_low_risk
2. quantization_medium_risk
3. kernel_backend_medium_high_risk
4. speculative_high_risk
5. stacked_combo_when_stable

Suggested Experiment Templates

• args tuning:
  • switch backend
  • tune chunked-prefill-size
  • toggle cuda graph/radix cache/warmup
• quantization:
  • GPTQ W4A16 baseline
  • group_size search
  • selective high-precision fallback
• kernels:
  • trim redundant cast/copy
  • sparse topk path micro-optimizations
• speculative:
  • minimal viable MiniCPM compatibility path

Stop / Pivot Rules

• 3 consecutive crashes in one lane -> downgrade risk lane.
• no gain in 8 rounds -> pivot to next strategy phase.