65de33fc创建于 4月17日历史提交

文件	最后提交记录	最后更新时间
README.md	Update Skill Evolution paths to use user home directory - Changed default output paths for SKILL.md files to `~/.akg/evolved_skills/{dsl}/` from `op/resources/skills/{dsl}/`. - Updated documentation and scripts to reflect the new directory structure for evolved skills. - Ensured consistency across various components regarding the new paths for skill management.	2 个月前
ab_test_utils.py	Enhance KernelGen and Skill Evolution with A/B Testing Support - Updated KernelGen to implement a layered stage-based skill selection mechanism, allowing for dynamic skill loading based on generation stages (initial, debug, optimize). - Introduced `exclude_skill_names` and `force_skill_names` parameters for precise A/B testing of evolved skills, enabling users to control skill inclusion during generation. - Enhanced documentation to reflect changes in skill selection processes and A/B testing capabilities. - Added symlink deployment instructions for evolved skills to ensure automatic discovery by KernelGen. - Refactored related scripts and utilities to support the new skill management features and improve overall usability.	2 个月前
run_ab_test.py	update logs path	1 个月前
run_skill_evolution.py	Refactor DSL Handling and Skill Metadata Management - Moved the `dsl_to_dir_key` function from `core.utils` to `core_v2.skill.metadata` for better organization and clarity. - Updated various modules to utilize the new location of `dsl_to_dir_key`, ensuring consistent DSL naming conventions across the codebase. - Enhanced the `SPEC.md` documentation to reflect changes in skill metadata, including a more detailed description for the `dsl` field and the overall skill directory structure. - Added unit tests for skill selection logic and DSL conversion to validate functionality and ensure robustness in skill management processes.	2 个月前
tracking.md	tracking and single run	2 个月前
verify_evolved_skill.py	Update Skill Evolution paths to use user home directory - Changed default output paths for SKILL.md files to `~/.akg/evolved_skills/{dsl}/` from `op/resources/skills/{dsl}/`. - Updated documentation and scripts to reflect the new directory structure for evolved skills. - Ensured consistency across various components regarding the new paths for skill management.	2 个月前

Skill Evolution 工具集

本目录包含 Skill 自进化系统的 CLI 工具和验证脚本。

文件说明

文件	功能
`run_skill_evolution.py`	四模式 Skill 生成 CLI（不依赖 Agent 框架）
`verify_evolved_skill.py`	验证 evolved skill 对算子生成效果的 A/B 对比脚本
`run_ab_test.py`	批量 A/B 测试运行器（多 group 多算子）
`ab_test_utils.py`	A/B 测试工具函数（运行管理、日志解析、结果收集）
`tracking.md`	实验结果跟踪文档

Skill Evolution 模式速查

模式	场景	输入
`search_log`	adaptive_search 跑完后，从搜索日志提取进化链 diff 生成优化经验	节点 logs 目录
`expert_tuning`	用户交互式调优后，从对话历史提取"建议→代码→性能"因果链	对话目录
`error_fix`	从失败→成功的修复记录中提取调试经验，持续追加到同一 skill	节点 logs 目录
`organize`	`~/.akg/evolved_skills/{dsl}/` 下 skill 过多时，fix 拆分 + improvement 合并去重	DSL 名称

快速上手

1. 生成 Skill

cd akg_agents

# 从搜索日志生成
python examples/kernel_related/skill_evolution/run_skill_evolution.py search_log /path/to/logs relu

# 从人工调优生成
python examples/kernel_related/skill_evolution/run_skill_evolution.py expert_tuning ~/.akg/conversations/cli_xxx relu

# 从错误修复生成
python examples/kernel_related/skill_evolution/run_skill_evolution.py error_fix /path/to/logs matmul

# 合并已有 skill
python examples/kernel_related/skill_evolution/run_skill_evolution.py organize triton_cuda

2. 验证 Skill 效果

verify_evolved_skill.py 支持通过 --skill-paths 指定要验证的 skill（文件、目录或多个混合），默认 DSL 为 triton_ascend。

# 验证单个 skill 文件
python examples/kernel_related/skill_evolution/verify_evolved_skill.py \
    --task-file /path/to/op_task.py \
    --skill-paths ~/.akg/evolved_skills/triton-cuda/triton-cuda-error-fix/SKILL.md
    --dsl triton_cuda --backend cuda --arch a100

# 混合多个路径
python examples/kernel_related/skill_evolution/verify_evolved_skill.py \
    --task-file /path/to/op_task.py \
    --skill-paths /path/to/error-fix/SKILL.md /path/to/exp-skill/

# 仅跑 B 组
python examples/kernel_related/skill_evolution/verify_evolved_skill.py \
    --task-file /path/to/op_task.py \
    --skill-paths /path/to/SKILL.md \
    --mode B

默认使用 --task-type profile 同时验证正确性和性能（Speedup / GenTime），如只需正确性验证可传 --task-type precision_only。

验证原理：使用 kernelgen_only_workflow，A 组仅加载 guides/ 下原始 skill，B 组通过 KernelGen.extra_skills 在 LLM 精筛完成后追加指定 skill（不参与粗筛和精筛，确保一定被选中），对比两组的 Speedup 和 GenTime。

3. 批量 A/B 测试

# 运行 Group 1 的批量对比
python examples/kernel_related/skill_evolution/run_ab_test.py --group 1 --mode both --device 0

详细文档参见 docs/v2/CN/SkillEvolution.md。