| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
【collaborative-agent-kernel-evolution】新增 CAKE2 进化式 AscendC 算子生成 Team(仅 Vector 算子) Co-authored-by: Junyifannnn<fan.junyi@huawei.com> # message auto-generated for no-merge-commit merge: !313 merge feat/add-cake-plugin into master 【collaborative-agent-kernel-evolution】新增 CAKE2 进化式 AscendC 算子生成 Team(仅 Vector 算子) Created-by: gcw_P2Mh5H1S Commit-by: Junyifannnn Merged-by: cann-robot Description: ## 描述 **新特性**: 【collaborative-agent-kernel-evolution】新增 plugins-community/collaborative-agent-kernel-evolution —— 进化式 AscendC 算子生成 Team(仅 Vector 算子)。 含 3 个 Agent 与 19 个内置 Skill,覆盖「算子描述生成 → PyTorch 参考 → functional 转换 → AscendDSL baseline → DSL lowering 到 AscendC → 编译评估 →(可选)进化式多变体性能优化」全流程: - cake(primary):标准单算子生成主流程 - cake-evo(primary):并行多变体进化优化 - cake-partial(subagent):被 cake-evo 调度的半程生成子任务 自包含单插件(目录 plugins-community/collaborative-agent-kernel-evolution/,对齐 ops-easyasc-dsl 自包含形态) ## 关联的Issue 关联 Issue [#230](https://gitcode.com/cann/cannbot-skills/issues/230) ## 测试 - **L1 结构门禁 skill_validator.py**:19 skills + 3 agents **0 error**(warn-only:6× S-CON-05、2× S-CON-06「缺错误处理章节」,不阻断) - **tests/run-tests.sh --fast**:10 项全过(初次 test-line-endings 报 6 个 CRLF 文件,均为插件内 vendored 外部文档,已全部转 LF,复跑 PASS) ## 文档更新 - 新增 plugins-community/collaborative-agent-kernel-evolution/(含 README / docs / tests) - 根 .claude-plugin/marketplace.json 注册 cake(source ./plugins-community/collaborative-agent-kernel-evolution,category development) - 更新根 CHANGELOG.md(2026-06-02 条目) ## 类型标签 - [x] 🌟 新特性 ## AscendC 算子评测结果 | Operator | Job ID | Cases | Score | Avg speedup | Rel path | |---|---|---:|---:|---:|---|---| | apply_rotary_pos_emb | job_940a45830a2f | 20/20 | 63.339075 | 0.355671 | level2/apply_rotary_pos_emb | | softmax | job_5ec935012955 | 20/20 | 68.015671 | 0.661441 | level2/softmax | | topk | job_bc7b660e83a6 | 20/20 | 54.233631 | 0.119558 | level3/top_k | ## apply_rotary_pos_emb - Job: job_940a45830a2f; operator: ApplyRotaryPosEmb; cases: 20/20; score: 63.339075; avg_speedup: 0.355671. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level2/apply_rotary_pos_emb_1 | 182.06 | 89.56 | 35.02 | 0.491926 | 0.192354 | 0.270563 | | 2 | level2/apply_rotary_pos_emb_2 | 779.98 | 113.24 | 59.22 | 0.145183 | 0.075925 | 0.069723 | | 3 | level2/apply_rotary_pos_emb_3 | 125.34 | 78.24 | 14.56 | 0.624222 | 0.116164 | 0.365012 | | 4 | level2/apply_rotary_pos_emb_4 | 140.26 | 246.2 | 32.77 | 1.755312 | 0.233638 | 0.665057 | | 5 | level2/apply_rotary_pos_emb_5 | 3001.06 | 391.07 | 140.36 | 0.130311 | 0.04677 | 0.080578 | | 6 | level2/apply_rotary_pos_emb_6 | 654.54 | 347.8 | 34.78 | 0.531366 | 0.053137 | 0.335578 | | 7 | level2/apply_rotary_pos_emb_7 | 146.94 | 73.08 | 23.1 | 0.497346 | 0.157207 | 0.287539 | | 8 | level2/apply_rotary_pos_emb_8 | 360.54 | 74.4 | 26.64 | 0.206357 | 0.073889 | 0.125138 | | 9 | level2/apply_rotary_pos_emb_9 | 178.12 | 198.1 | 19.81 | 1.112172 | 0.111217 | 0.529679 | | 10 | level2/apply_rotary_pos_emb_10 | 274.06 | 163 | 16.3 | 0.59476 | 0.059476 | 0.362706 | | 11 | level2/apply_rotary_pos_emb_11 | 1398.00 | 186.7 | 98.62 | 0.133548 | 0.070544 | 0.063483 | | 12 | level2/apply_rotary_pos_emb_12 | 896.26 | 48 | 4.8 | 0.053556 | 0.005356 | 0.04622 | | 13 | level2/apply_rotary_pos_emb_13 | 595.38 | 425.07 | 100 | 0.713947 | 0.16796 | 0.396209 | | 14 | level2/apply_rotary_pos_emb_14 | 27.18 | 10 | 1 | 0.367918 | 0.036792 | 0.255827 | | 15 | level2/apply_rotary_pos_emb_15 | 86.22 | 21.7 | 2.17 | 0.251682 | 0.025168 | 0.18855 | | 16 | level2/apply_rotary_pos_emb_16 | 82.48 | 41.34 | 4.17 | 0.501212 | 0.050558 | 0.321874 | | 17 | level2/apply_rotary_pos_emb_17 | 5265.48 | 2621.60 | 262.16 | 0.497884 | 0.049788 | 0.320456 | | 18 | level2/apply_rotary_pos_emb_18 | 97.2 | 43.9 | 4.39 | 0.451646 | 0.045165 | 0.298594 | | 19 | level2/apply_rotary_pos_emb_19 | 27.12 | 10 | 1 | 0.368732 | 0.036873 | 0.256264 | | 20 | level2/apply_rotary_pos_emb_20 | 1260.70 | 183.04 | 54.04 | 0.145189 | 0.042865 | 0.096581 | ## softmax - Job: job_5ec935012955; operator: Softmax; cases: 20/20; score: 68.015671; avg_speedup: 0.661441. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level2/softmax_1 | 19.42 | 15.7 | 1.57 | 0.808445 | 0.080844 | 0.441839 | | 2 | level2/softmax_2 | 30.96 | 37.98 | 8.74 | 1.226744 | 0.2823 | 0.568208 | | 3 | level2/softmax_3 | 104.2 | 91.76 | 25.17 | 0.880614 | 0.241555 | 0.457286 | | 4 | level2/softmax_4 | 2175.34 | 1006.60 | 100.66 | 0.462732 | 0.046273 | 0.303943 | | 5 | level2/softmax_5 | 458.2 | 422.73 | 139.81 | 0.922588 | 0.305129 | 0.470506 | | 6 | level2/softmax_6 | 1907.08 | 786.04 | 101.69 | 0.412169 | 0.053322 | 0.274868 | | 7 | level2/softmax_7 | 22.42 | 20.04 | 3.14 | 0.893845 | 0.140054 | 0.467109 | | 8 | level2/softmax_8 | 61.92 | 58.12 | 17.49 | 0.93863 | 0.282461 | 0.477663 | | 9 | level2/softmax_9 | 1044.48 | 412.79 | 50.08 | 0.395211 | 0.047947 | 0.267266 | | 10 | level2/softmax_10 | 21.82 | 14.2 | 1.55 | 0.650779 | 0.071036 | 0.384265 | | 11 | level2/softmax_11 | 2669.88 | 1063.10 | 108.09 | 0.398183 | 0.040485 | 0.271557 | | 12 | level2/softmax_12 | 61.3 | 62.58 | 14.59 | 1.020881 | 0.23801 | 0.506758 | | 13 | level2/softmax_13 | 2246.58 | 30 | 3 | 0.013354 | 0.001335 | 0.011891 | | 14 | level2/softmax_14 | 238.52 | 227.3 | 22.73 | 0.95296 | 0.095296 | 0.486654 | | 15 | level2/softmax_15 | 30.6 | 30.46 | 4.56 | 0.995425 | 0.14902 | 0.498652 | | 16 | level2/softmax_16 | 115.28 | 15.7 | 1.57 | 0.13619 | 0.013619 | 0.110529 | | 17 | level2/softmax_17 | 19.98 | 18.64 | 4.35 | 0.932933 | 0.217718 | 0.477607 | | 18 | level2/softmax_18 | 24.9 | 20.12 | 3.14 | 0.808032 | 0.126104 | 0.438307 | | 19 | level2/softmax_19 | 119.64 | 31.3 | 3.13 | 0.261618 | 0.026162 | 0.194706 | | 20 | level2/softmax_20 | 1870.84 | 219.8 | 21.98 | 0.117487 | 0.011749 | 0.096654 | ## topk - Job: job_bc7b660e83a6; operator: TopK; cases: 20/20; score: 54.233631; avg_speedup: 0.119558. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level3/top_k_1 | 2431.10 | 10.9 | 1.09 | 0.004484 | 0.000448 | 0.004021 | | 2 | level3/top_k_2 | 279.02 | 21.8 | 2.18 | 0.078131 | 0.007813 | 0.066181 | | 3 | level3/top_k_3 | 995.88 | 174.8 | 17.48 | 0.175523 | 0.017552 | 0.13852 | | 4 | level3/top_k_4 | 281395.82 | 1398.10 | 139.81 | 0.004968 | 0.000497 | 0.004454 | | 5 | level3/top_k_5 | 2508.46 | 1398.10 | 139.81 | 0.557354 | 0.055735 | 0.346929 | | 6 | level3/top_k_6 | 6939698.16 | 1398.10 | 139.81 | 0.000201 | 2.015e-05 | 0.000181 | | 7 | level3/top_k_7 | 74.9 | 10.9 | 1.09 | 0.145527 | 0.014553 | 0.117316 | | 8 | level3/top_k_8 | 3213.32 | 21.5 | 2.15 | 0.006691 | 0.000669 | 0.00599 | | 9 | level3/top_k_9 | 97.66 | 13.8 | 1.38 | 0.141307 | 0.014131 | 0.114259 | | 10 | level3/top_k_10 | 95309.40 | 1035.20 | 103.52 | 0.010861 | 0.001086 | 0.009691 | | 11 | level3/top_k_11 | 104.16 | 10.9 | 1.09 | 0.104647 | 0.010465 | 0.086906 | | 12 | level3/top_k_12 | 111.12 | 22.8 | 2.28 | 0.205184 | 0.020518 | 0.158627 | | 13 | level3/top_k_13 | 4755.26 | 14.3 | 1.43 | 0.003007 | 0.000301 | 0.0027 | | 14 | level3/top_k_14 | 2448.26 | 20.8 | 2.08 | 0.008496 | 0.00085 | 0.007595 | | 15 | level3/top_k_15 | 19245.22 | 170.5 | 17.05 | 0.008859 | 0.000886 | 0.007917 | | 16 | level3/top_k_16 | 211.92 | 126.8 | 12.68 | 0.598339 | 0.059834 | 0.364182 | | 17 | level3/top_k_17 | 80.84 | 21.9 | 2.19 | 0.270905 | 0.027091 | 0.200386 | | 18 | level3/top_k_18 | 6133.36 | 24.4 | 2.44 | 0.003978 | 0.000398 | 0.003569 | | 19 | level3/top_k_19 | 184.64 | 10.9 | 1.09 | 0.059034 | 0.005903 | 0.050734 | | 20 | level3/top_k_20 | 5940.70 | 21.8 | 2.18 | 0.00367 | 0.000367 | 0.003293 | ## 未来三个月的计划 ### 第 1 个月 —— 完善 补全 reduce类,数学函数类,排序类算子的生成与优化skill和harness ### 第 2 个月 —— 拓展CV融合算子 增加对CV类融合算子如quant_batch_matmul,grouped_matmul_swiglu_quant等的生成适配,保障生成精度 ### 第 3 个月 —— 拓展attention算子 增加对attention类算子的适配 See merge request: cann/cannbot-skills!313 | 10 天前 | |
【collaborative-agent-kernel-evolution】新增 CAKE2 进化式 AscendC 算子生成 Team(仅 Vector 算子) Co-authored-by: Junyifannnn<fan.junyi@huawei.com> # message auto-generated for no-merge-commit merge: !313 merge feat/add-cake-plugin into master 【collaborative-agent-kernel-evolution】新增 CAKE2 进化式 AscendC 算子生成 Team(仅 Vector 算子) Created-by: gcw_P2Mh5H1S Commit-by: Junyifannnn Merged-by: cann-robot Description: ## 描述 **新特性**: 【collaborative-agent-kernel-evolution】新增 plugins-community/collaborative-agent-kernel-evolution —— 进化式 AscendC 算子生成 Team(仅 Vector 算子)。 含 3 个 Agent 与 19 个内置 Skill,覆盖「算子描述生成 → PyTorch 参考 → functional 转换 → AscendDSL baseline → DSL lowering 到 AscendC → 编译评估 →(可选)进化式多变体性能优化」全流程: - cake(primary):标准单算子生成主流程 - cake-evo(primary):并行多变体进化优化 - cake-partial(subagent):被 cake-evo 调度的半程生成子任务 自包含单插件(目录 plugins-community/collaborative-agent-kernel-evolution/,对齐 ops-easyasc-dsl 自包含形态) ## 关联的Issue 关联 Issue [#230](https://gitcode.com/cann/cannbot-skills/issues/230) ## 测试 - **L1 结构门禁 skill_validator.py**:19 skills + 3 agents **0 error**(warn-only:6× S-CON-05、2× S-CON-06「缺错误处理章节」,不阻断) - **tests/run-tests.sh --fast**:10 项全过(初次 test-line-endings 报 6 个 CRLF 文件,均为插件内 vendored 外部文档,已全部转 LF,复跑 PASS) ## 文档更新 - 新增 plugins-community/collaborative-agent-kernel-evolution/(含 README / docs / tests) - 根 .claude-plugin/marketplace.json 注册 cake(source ./plugins-community/collaborative-agent-kernel-evolution,category development) - 更新根 CHANGELOG.md(2026-06-02 条目) ## 类型标签 - [x] 🌟 新特性 ## AscendC 算子评测结果 | Operator | Job ID | Cases | Score | Avg speedup | Rel path | |---|---|---:|---:|---:|---|---| | apply_rotary_pos_emb | job_940a45830a2f | 20/20 | 63.339075 | 0.355671 | level2/apply_rotary_pos_emb | | softmax | job_5ec935012955 | 20/20 | 68.015671 | 0.661441 | level2/softmax | | topk | job_bc7b660e83a6 | 20/20 | 54.233631 | 0.119558 | level3/top_k | ## apply_rotary_pos_emb - Job: job_940a45830a2f; operator: ApplyRotaryPosEmb; cases: 20/20; score: 63.339075; avg_speedup: 0.355671. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level2/apply_rotary_pos_emb_1 | 182.06 | 89.56 | 35.02 | 0.491926 | 0.192354 | 0.270563 | | 2 | level2/apply_rotary_pos_emb_2 | 779.98 | 113.24 | 59.22 | 0.145183 | 0.075925 | 0.069723 | | 3 | level2/apply_rotary_pos_emb_3 | 125.34 | 78.24 | 14.56 | 0.624222 | 0.116164 | 0.365012 | | 4 | level2/apply_rotary_pos_emb_4 | 140.26 | 246.2 | 32.77 | 1.755312 | 0.233638 | 0.665057 | | 5 | level2/apply_rotary_pos_emb_5 | 3001.06 | 391.07 | 140.36 | 0.130311 | 0.04677 | 0.080578 | | 6 | level2/apply_rotary_pos_emb_6 | 654.54 | 347.8 | 34.78 | 0.531366 | 0.053137 | 0.335578 | | 7 | level2/apply_rotary_pos_emb_7 | 146.94 | 73.08 | 23.1 | 0.497346 | 0.157207 | 0.287539 | | 8 | level2/apply_rotary_pos_emb_8 | 360.54 | 74.4 | 26.64 | 0.206357 | 0.073889 | 0.125138 | | 9 | level2/apply_rotary_pos_emb_9 | 178.12 | 198.1 | 19.81 | 1.112172 | 0.111217 | 0.529679 | | 10 | level2/apply_rotary_pos_emb_10 | 274.06 | 163 | 16.3 | 0.59476 | 0.059476 | 0.362706 | | 11 | level2/apply_rotary_pos_emb_11 | 1398.00 | 186.7 | 98.62 | 0.133548 | 0.070544 | 0.063483 | | 12 | level2/apply_rotary_pos_emb_12 | 896.26 | 48 | 4.8 | 0.053556 | 0.005356 | 0.04622 | | 13 | level2/apply_rotary_pos_emb_13 | 595.38 | 425.07 | 100 | 0.713947 | 0.16796 | 0.396209 | | 14 | level2/apply_rotary_pos_emb_14 | 27.18 | 10 | 1 | 0.367918 | 0.036792 | 0.255827 | | 15 | level2/apply_rotary_pos_emb_15 | 86.22 | 21.7 | 2.17 | 0.251682 | 0.025168 | 0.18855 | | 16 | level2/apply_rotary_pos_emb_16 | 82.48 | 41.34 | 4.17 | 0.501212 | 0.050558 | 0.321874 | | 17 | level2/apply_rotary_pos_emb_17 | 5265.48 | 2621.60 | 262.16 | 0.497884 | 0.049788 | 0.320456 | | 18 | level2/apply_rotary_pos_emb_18 | 97.2 | 43.9 | 4.39 | 0.451646 | 0.045165 | 0.298594 | | 19 | level2/apply_rotary_pos_emb_19 | 27.12 | 10 | 1 | 0.368732 | 0.036873 | 0.256264 | | 20 | level2/apply_rotary_pos_emb_20 | 1260.70 | 183.04 | 54.04 | 0.145189 | 0.042865 | 0.096581 | ## softmax - Job: job_5ec935012955; operator: Softmax; cases: 20/20; score: 68.015671; avg_speedup: 0.661441. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level2/softmax_1 | 19.42 | 15.7 | 1.57 | 0.808445 | 0.080844 | 0.441839 | | 2 | level2/softmax_2 | 30.96 | 37.98 | 8.74 | 1.226744 | 0.2823 | 0.568208 | | 3 | level2/softmax_3 | 104.2 | 91.76 | 25.17 | 0.880614 | 0.241555 | 0.457286 | | 4 | level2/softmax_4 | 2175.34 | 1006.60 | 100.66 | 0.462732 | 0.046273 | 0.303943 | | 5 | level2/softmax_5 | 458.2 | 422.73 | 139.81 | 0.922588 | 0.305129 | 0.470506 | | 6 | level2/softmax_6 | 1907.08 | 786.04 | 101.69 | 0.412169 | 0.053322 | 0.274868 | | 7 | level2/softmax_7 | 22.42 | 20.04 | 3.14 | 0.893845 | 0.140054 | 0.467109 | | 8 | level2/softmax_8 | 61.92 | 58.12 | 17.49 | 0.93863 | 0.282461 | 0.477663 | | 9 | level2/softmax_9 | 1044.48 | 412.79 | 50.08 | 0.395211 | 0.047947 | 0.267266 | | 10 | level2/softmax_10 | 21.82 | 14.2 | 1.55 | 0.650779 | 0.071036 | 0.384265 | | 11 | level2/softmax_11 | 2669.88 | 1063.10 | 108.09 | 0.398183 | 0.040485 | 0.271557 | | 12 | level2/softmax_12 | 61.3 | 62.58 | 14.59 | 1.020881 | 0.23801 | 0.506758 | | 13 | level2/softmax_13 | 2246.58 | 30 | 3 | 0.013354 | 0.001335 | 0.011891 | | 14 | level2/softmax_14 | 238.52 | 227.3 | 22.73 | 0.95296 | 0.095296 | 0.486654 | | 15 | level2/softmax_15 | 30.6 | 30.46 | 4.56 | 0.995425 | 0.14902 | 0.498652 | | 16 | level2/softmax_16 | 115.28 | 15.7 | 1.57 | 0.13619 | 0.013619 | 0.110529 | | 17 | level2/softmax_17 | 19.98 | 18.64 | 4.35 | 0.932933 | 0.217718 | 0.477607 | | 18 | level2/softmax_18 | 24.9 | 20.12 | 3.14 | 0.808032 | 0.126104 | 0.438307 | | 19 | level2/softmax_19 | 119.64 | 31.3 | 3.13 | 0.261618 | 0.026162 | 0.194706 | | 20 | level2/softmax_20 | 1870.84 | 219.8 | 21.98 | 0.117487 | 0.011749 | 0.096654 | ## topk - Job: job_bc7b660e83a6; operator: TopK; cases: 20/20; score: 54.233631; avg_speedup: 0.119558. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level3/top_k_1 | 2431.10 | 10.9 | 1.09 | 0.004484 | 0.000448 | 0.004021 | | 2 | level3/top_k_2 | 279.02 | 21.8 | 2.18 | 0.078131 | 0.007813 | 0.066181 | | 3 | level3/top_k_3 | 995.88 | 174.8 | 17.48 | 0.175523 | 0.017552 | 0.13852 | | 4 | level3/top_k_4 | 281395.82 | 1398.10 | 139.81 | 0.004968 | 0.000497 | 0.004454 | | 5 | level3/top_k_5 | 2508.46 | 1398.10 | 139.81 | 0.557354 | 0.055735 | 0.346929 | | 6 | level3/top_k_6 | 6939698.16 | 1398.10 | 139.81 | 0.000201 | 2.015e-05 | 0.000181 | | 7 | level3/top_k_7 | 74.9 | 10.9 | 1.09 | 0.145527 | 0.014553 | 0.117316 | | 8 | level3/top_k_8 | 3213.32 | 21.5 | 2.15 | 0.006691 | 0.000669 | 0.00599 | | 9 | level3/top_k_9 | 97.66 | 13.8 | 1.38 | 0.141307 | 0.014131 | 0.114259 | | 10 | level3/top_k_10 | 95309.40 | 1035.20 | 103.52 | 0.010861 | 0.001086 | 0.009691 | | 11 | level3/top_k_11 | 104.16 | 10.9 | 1.09 | 0.104647 | 0.010465 | 0.086906 | | 12 | level3/top_k_12 | 111.12 | 22.8 | 2.28 | 0.205184 | 0.020518 | 0.158627 | | 13 | level3/top_k_13 | 4755.26 | 14.3 | 1.43 | 0.003007 | 0.000301 | 0.0027 | | 14 | level3/top_k_14 | 2448.26 | 20.8 | 2.08 | 0.008496 | 0.00085 | 0.007595 | | 15 | level3/top_k_15 | 19245.22 | 170.5 | 17.05 | 0.008859 | 0.000886 | 0.007917 | | 16 | level3/top_k_16 | 211.92 | 126.8 | 12.68 | 0.598339 | 0.059834 | 0.364182 | | 17 | level3/top_k_17 | 80.84 | 21.9 | 2.19 | 0.270905 | 0.027091 | 0.200386 | | 18 | level3/top_k_18 | 6133.36 | 24.4 | 2.44 | 0.003978 | 0.000398 | 0.003569 | | 19 | level3/top_k_19 | 184.64 | 10.9 | 1.09 | 0.059034 | 0.005903 | 0.050734 | | 20 | level3/top_k_20 | 5940.70 | 21.8 | 2.18 | 0.00367 | 0.000367 | 0.003293 | ## 未来三个月的计划 ### 第 1 个月 —— 完善 补全 reduce类,数学函数类,排序类算子的生成与优化skill和harness ### 第 2 个月 —— 拓展CV融合算子 增加对CV类融合算子如quant_batch_matmul,grouped_matmul_swiglu_quant等的生成适配,保障生成精度 ### 第 3 个月 —— 拓展attention算子 增加对attention类算子的适配 See merge request: cann/cannbot-skills!313 | 10 天前 | |
【collaborative-agent-kernel-evolution】新增 CAKE2 进化式 AscendC 算子生成 Team(仅 Vector 算子) Co-authored-by: Junyifannnn<fan.junyi@huawei.com> # message auto-generated for no-merge-commit merge: !313 merge feat/add-cake-plugin into master 【collaborative-agent-kernel-evolution】新增 CAKE2 进化式 AscendC 算子生成 Team(仅 Vector 算子) Created-by: gcw_P2Mh5H1S Commit-by: Junyifannnn Merged-by: cann-robot Description: ## 描述 **新特性**: 【collaborative-agent-kernel-evolution】新增 plugins-community/collaborative-agent-kernel-evolution —— 进化式 AscendC 算子生成 Team(仅 Vector 算子)。 含 3 个 Agent 与 19 个内置 Skill,覆盖「算子描述生成 → PyTorch 参考 → functional 转换 → AscendDSL baseline → DSL lowering 到 AscendC → 编译评估 →(可选)进化式多变体性能优化」全流程: - cake(primary):标准单算子生成主流程 - cake-evo(primary):并行多变体进化优化 - cake-partial(subagent):被 cake-evo 调度的半程生成子任务 自包含单插件(目录 plugins-community/collaborative-agent-kernel-evolution/,对齐 ops-easyasc-dsl 自包含形态) ## 关联的Issue 关联 Issue [#230](https://gitcode.com/cann/cannbot-skills/issues/230) ## 测试 - **L1 结构门禁 skill_validator.py**:19 skills + 3 agents **0 error**(warn-only:6× S-CON-05、2× S-CON-06「缺错误处理章节」,不阻断) - **tests/run-tests.sh --fast**:10 项全过(初次 test-line-endings 报 6 个 CRLF 文件,均为插件内 vendored 外部文档,已全部转 LF,复跑 PASS) ## 文档更新 - 新增 plugins-community/collaborative-agent-kernel-evolution/(含 README / docs / tests) - 根 .claude-plugin/marketplace.json 注册 cake(source ./plugins-community/collaborative-agent-kernel-evolution,category development) - 更新根 CHANGELOG.md(2026-06-02 条目) ## 类型标签 - [x] 🌟 新特性 ## AscendC 算子评测结果 | Operator | Job ID | Cases | Score | Avg speedup | Rel path | |---|---|---:|---:|---:|---|---| | apply_rotary_pos_emb | job_940a45830a2f | 20/20 | 63.339075 | 0.355671 | level2/apply_rotary_pos_emb | | softmax | job_5ec935012955 | 20/20 | 68.015671 | 0.661441 | level2/softmax | | topk | job_bc7b660e83a6 | 20/20 | 54.233631 | 0.119558 | level3/top_k | ## apply_rotary_pos_emb - Job: job_940a45830a2f; operator: ApplyRotaryPosEmb; cases: 20/20; score: 63.339075; avg_speedup: 0.355671. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level2/apply_rotary_pos_emb_1 | 182.06 | 89.56 | 35.02 | 0.491926 | 0.192354 | 0.270563 | | 2 | level2/apply_rotary_pos_emb_2 | 779.98 | 113.24 | 59.22 | 0.145183 | 0.075925 | 0.069723 | | 3 | level2/apply_rotary_pos_emb_3 | 125.34 | 78.24 | 14.56 | 0.624222 | 0.116164 | 0.365012 | | 4 | level2/apply_rotary_pos_emb_4 | 140.26 | 246.2 | 32.77 | 1.755312 | 0.233638 | 0.665057 | | 5 | level2/apply_rotary_pos_emb_5 | 3001.06 | 391.07 | 140.36 | 0.130311 | 0.04677 | 0.080578 | | 6 | level2/apply_rotary_pos_emb_6 | 654.54 | 347.8 | 34.78 | 0.531366 | 0.053137 | 0.335578 | | 7 | level2/apply_rotary_pos_emb_7 | 146.94 | 73.08 | 23.1 | 0.497346 | 0.157207 | 0.287539 | | 8 | level2/apply_rotary_pos_emb_8 | 360.54 | 74.4 | 26.64 | 0.206357 | 0.073889 | 0.125138 | | 9 | level2/apply_rotary_pos_emb_9 | 178.12 | 198.1 | 19.81 | 1.112172 | 0.111217 | 0.529679 | | 10 | level2/apply_rotary_pos_emb_10 | 274.06 | 163 | 16.3 | 0.59476 | 0.059476 | 0.362706 | | 11 | level2/apply_rotary_pos_emb_11 | 1398.00 | 186.7 | 98.62 | 0.133548 | 0.070544 | 0.063483 | | 12 | level2/apply_rotary_pos_emb_12 | 896.26 | 48 | 4.8 | 0.053556 | 0.005356 | 0.04622 | | 13 | level2/apply_rotary_pos_emb_13 | 595.38 | 425.07 | 100 | 0.713947 | 0.16796 | 0.396209 | | 14 | level2/apply_rotary_pos_emb_14 | 27.18 | 10 | 1 | 0.367918 | 0.036792 | 0.255827 | | 15 | level2/apply_rotary_pos_emb_15 | 86.22 | 21.7 | 2.17 | 0.251682 | 0.025168 | 0.18855 | | 16 | level2/apply_rotary_pos_emb_16 | 82.48 | 41.34 | 4.17 | 0.501212 | 0.050558 | 0.321874 | | 17 | level2/apply_rotary_pos_emb_17 | 5265.48 | 2621.60 | 262.16 | 0.497884 | 0.049788 | 0.320456 | | 18 | level2/apply_rotary_pos_emb_18 | 97.2 | 43.9 | 4.39 | 0.451646 | 0.045165 | 0.298594 | | 19 | level2/apply_rotary_pos_emb_19 | 27.12 | 10 | 1 | 0.368732 | 0.036873 | 0.256264 | | 20 | level2/apply_rotary_pos_emb_20 | 1260.70 | 183.04 | 54.04 | 0.145189 | 0.042865 | 0.096581 | ## softmax - Job: job_5ec935012955; operator: Softmax; cases: 20/20; score: 68.015671; avg_speedup: 0.661441. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level2/softmax_1 | 19.42 | 15.7 | 1.57 | 0.808445 | 0.080844 | 0.441839 | | 2 | level2/softmax_2 | 30.96 | 37.98 | 8.74 | 1.226744 | 0.2823 | 0.568208 | | 3 | level2/softmax_3 | 104.2 | 91.76 | 25.17 | 0.880614 | 0.241555 | 0.457286 | | 4 | level2/softmax_4 | 2175.34 | 1006.60 | 100.66 | 0.462732 | 0.046273 | 0.303943 | | 5 | level2/softmax_5 | 458.2 | 422.73 | 139.81 | 0.922588 | 0.305129 | 0.470506 | | 6 | level2/softmax_6 | 1907.08 | 786.04 | 101.69 | 0.412169 | 0.053322 | 0.274868 | | 7 | level2/softmax_7 | 22.42 | 20.04 | 3.14 | 0.893845 | 0.140054 | 0.467109 | | 8 | level2/softmax_8 | 61.92 | 58.12 | 17.49 | 0.93863 | 0.282461 | 0.477663 | | 9 | level2/softmax_9 | 1044.48 | 412.79 | 50.08 | 0.395211 | 0.047947 | 0.267266 | | 10 | level2/softmax_10 | 21.82 | 14.2 | 1.55 | 0.650779 | 0.071036 | 0.384265 | | 11 | level2/softmax_11 | 2669.88 | 1063.10 | 108.09 | 0.398183 | 0.040485 | 0.271557 | | 12 | level2/softmax_12 | 61.3 | 62.58 | 14.59 | 1.020881 | 0.23801 | 0.506758 | | 13 | level2/softmax_13 | 2246.58 | 30 | 3 | 0.013354 | 0.001335 | 0.011891 | | 14 | level2/softmax_14 | 238.52 | 227.3 | 22.73 | 0.95296 | 0.095296 | 0.486654 | | 15 | level2/softmax_15 | 30.6 | 30.46 | 4.56 | 0.995425 | 0.14902 | 0.498652 | | 16 | level2/softmax_16 | 115.28 | 15.7 | 1.57 | 0.13619 | 0.013619 | 0.110529 | | 17 | level2/softmax_17 | 19.98 | 18.64 | 4.35 | 0.932933 | 0.217718 | 0.477607 | | 18 | level2/softmax_18 | 24.9 | 20.12 | 3.14 | 0.808032 | 0.126104 | 0.438307 | | 19 | level2/softmax_19 | 119.64 | 31.3 | 3.13 | 0.261618 | 0.026162 | 0.194706 | | 20 | level2/softmax_20 | 1870.84 | 219.8 | 21.98 | 0.117487 | 0.011749 | 0.096654 | ## topk - Job: job_bc7b660e83a6; operator: TopK; cases: 20/20; score: 54.233631; avg_speedup: 0.119558. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level3/top_k_1 | 2431.10 | 10.9 | 1.09 | 0.004484 | 0.000448 | 0.004021 | | 2 | level3/top_k_2 | 279.02 | 21.8 | 2.18 | 0.078131 | 0.007813 | 0.066181 | | 3 | level3/top_k_3 | 995.88 | 174.8 | 17.48 | 0.175523 | 0.017552 | 0.13852 | | 4 | level3/top_k_4 | 281395.82 | 1398.10 | 139.81 | 0.004968 | 0.000497 | 0.004454 | | 5 | level3/top_k_5 | 2508.46 | 1398.10 | 139.81 | 0.557354 | 0.055735 | 0.346929 | | 6 | level3/top_k_6 | 6939698.16 | 1398.10 | 139.81 | 0.000201 | 2.015e-05 | 0.000181 | | 7 | level3/top_k_7 | 74.9 | 10.9 | 1.09 | 0.145527 | 0.014553 | 0.117316 | | 8 | level3/top_k_8 | 3213.32 | 21.5 | 2.15 | 0.006691 | 0.000669 | 0.00599 | | 9 | level3/top_k_9 | 97.66 | 13.8 | 1.38 | 0.141307 | 0.014131 | 0.114259 | | 10 | level3/top_k_10 | 95309.40 | 1035.20 | 103.52 | 0.010861 | 0.001086 | 0.009691 | | 11 | level3/top_k_11 | 104.16 | 10.9 | 1.09 | 0.104647 | 0.010465 | 0.086906 | | 12 | level3/top_k_12 | 111.12 | 22.8 | 2.28 | 0.205184 | 0.020518 | 0.158627 | | 13 | level3/top_k_13 | 4755.26 | 14.3 | 1.43 | 0.003007 | 0.000301 | 0.0027 | | 14 | level3/top_k_14 | 2448.26 | 20.8 | 2.08 | 0.008496 | 0.00085 | 0.007595 | | 15 | level3/top_k_15 | 19245.22 | 170.5 | 17.05 | 0.008859 | 0.000886 | 0.007917 | | 16 | level3/top_k_16 | 211.92 | 126.8 | 12.68 | 0.598339 | 0.059834 | 0.364182 | | 17 | level3/top_k_17 | 80.84 | 21.9 | 2.19 | 0.270905 | 0.027091 | 0.200386 | | 18 | level3/top_k_18 | 6133.36 | 24.4 | 2.44 | 0.003978 | 0.000398 | 0.003569 | | 19 | level3/top_k_19 | 184.64 | 10.9 | 1.09 | 0.059034 | 0.005903 | 0.050734 | | 20 | level3/top_k_20 | 5940.70 | 21.8 | 2.18 | 0.00367 | 0.000367 | 0.003293 | ## 未来三个月的计划 ### 第 1 个月 —— 完善 补全 reduce类,数学函数类,排序类算子的生成与优化skill和harness ### 第 2 个月 —— 拓展CV融合算子 增加对CV类融合算子如quant_batch_matmul,grouped_matmul_swiglu_quant等的生成适配,保障生成精度 ### 第 3 个月 —— 拓展attention算子 增加对attention类算子的适配 See merge request: cann/cannbot-skills!313 | 10 天前 | |
【collaborative-agent-kernel-evolution】新增 CAKE2 进化式 AscendC 算子生成 Team(仅 Vector 算子) Co-authored-by: Junyifannnn<fan.junyi@huawei.com> # message auto-generated for no-merge-commit merge: !313 merge feat/add-cake-plugin into master 【collaborative-agent-kernel-evolution】新增 CAKE2 进化式 AscendC 算子生成 Team(仅 Vector 算子) Created-by: gcw_P2Mh5H1S Commit-by: Junyifannnn Merged-by: cann-robot Description: ## 描述 **新特性**: 【collaborative-agent-kernel-evolution】新增 plugins-community/collaborative-agent-kernel-evolution —— 进化式 AscendC 算子生成 Team(仅 Vector 算子)。 含 3 个 Agent 与 19 个内置 Skill,覆盖「算子描述生成 → PyTorch 参考 → functional 转换 → AscendDSL baseline → DSL lowering 到 AscendC → 编译评估 →(可选)进化式多变体性能优化」全流程: - cake(primary):标准单算子生成主流程 - cake-evo(primary):并行多变体进化优化 - cake-partial(subagent):被 cake-evo 调度的半程生成子任务 自包含单插件(目录 plugins-community/collaborative-agent-kernel-evolution/,对齐 ops-easyasc-dsl 自包含形态) ## 关联的Issue 关联 Issue [#230](https://gitcode.com/cann/cannbot-skills/issues/230) ## 测试 - **L1 结构门禁 skill_validator.py**:19 skills + 3 agents **0 error**(warn-only:6× S-CON-05、2× S-CON-06「缺错误处理章节」,不阻断) - **tests/run-tests.sh --fast**:10 项全过(初次 test-line-endings 报 6 个 CRLF 文件,均为插件内 vendored 外部文档,已全部转 LF,复跑 PASS) ## 文档更新 - 新增 plugins-community/collaborative-agent-kernel-evolution/(含 README / docs / tests) - 根 .claude-plugin/marketplace.json 注册 cake(source ./plugins-community/collaborative-agent-kernel-evolution,category development) - 更新根 CHANGELOG.md(2026-06-02 条目) ## 类型标签 - [x] 🌟 新特性 ## AscendC 算子评测结果 | Operator | Job ID | Cases | Score | Avg speedup | Rel path | |---|---|---:|---:|---:|---|---| | apply_rotary_pos_emb | job_940a45830a2f | 20/20 | 63.339075 | 0.355671 | level2/apply_rotary_pos_emb | | softmax | job_5ec935012955 | 20/20 | 68.015671 | 0.661441 | level2/softmax | | topk | job_bc7b660e83a6 | 20/20 | 54.233631 | 0.119558 | level3/top_k | ## apply_rotary_pos_emb - Job: job_940a45830a2f; operator: ApplyRotaryPosEmb; cases: 20/20; score: 63.339075; avg_speedup: 0.355671. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level2/apply_rotary_pos_emb_1 | 182.06 | 89.56 | 35.02 | 0.491926 | 0.192354 | 0.270563 | | 2 | level2/apply_rotary_pos_emb_2 | 779.98 | 113.24 | 59.22 | 0.145183 | 0.075925 | 0.069723 | | 3 | level2/apply_rotary_pos_emb_3 | 125.34 | 78.24 | 14.56 | 0.624222 | 0.116164 | 0.365012 | | 4 | level2/apply_rotary_pos_emb_4 | 140.26 | 246.2 | 32.77 | 1.755312 | 0.233638 | 0.665057 | | 5 | level2/apply_rotary_pos_emb_5 | 3001.06 | 391.07 | 140.36 | 0.130311 | 0.04677 | 0.080578 | | 6 | level2/apply_rotary_pos_emb_6 | 654.54 | 347.8 | 34.78 | 0.531366 | 0.053137 | 0.335578 | | 7 | level2/apply_rotary_pos_emb_7 | 146.94 | 73.08 | 23.1 | 0.497346 | 0.157207 | 0.287539 | | 8 | level2/apply_rotary_pos_emb_8 | 360.54 | 74.4 | 26.64 | 0.206357 | 0.073889 | 0.125138 | | 9 | level2/apply_rotary_pos_emb_9 | 178.12 | 198.1 | 19.81 | 1.112172 | 0.111217 | 0.529679 | | 10 | level2/apply_rotary_pos_emb_10 | 274.06 | 163 | 16.3 | 0.59476 | 0.059476 | 0.362706 | | 11 | level2/apply_rotary_pos_emb_11 | 1398.00 | 186.7 | 98.62 | 0.133548 | 0.070544 | 0.063483 | | 12 | level2/apply_rotary_pos_emb_12 | 896.26 | 48 | 4.8 | 0.053556 | 0.005356 | 0.04622 | | 13 | level2/apply_rotary_pos_emb_13 | 595.38 | 425.07 | 100 | 0.713947 | 0.16796 | 0.396209 | | 14 | level2/apply_rotary_pos_emb_14 | 27.18 | 10 | 1 | 0.367918 | 0.036792 | 0.255827 | | 15 | level2/apply_rotary_pos_emb_15 | 86.22 | 21.7 | 2.17 | 0.251682 | 0.025168 | 0.18855 | | 16 | level2/apply_rotary_pos_emb_16 | 82.48 | 41.34 | 4.17 | 0.501212 | 0.050558 | 0.321874 | | 17 | level2/apply_rotary_pos_emb_17 | 5265.48 | 2621.60 | 262.16 | 0.497884 | 0.049788 | 0.320456 | | 18 | level2/apply_rotary_pos_emb_18 | 97.2 | 43.9 | 4.39 | 0.451646 | 0.045165 | 0.298594 | | 19 | level2/apply_rotary_pos_emb_19 | 27.12 | 10 | 1 | 0.368732 | 0.036873 | 0.256264 | | 20 | level2/apply_rotary_pos_emb_20 | 1260.70 | 183.04 | 54.04 | 0.145189 | 0.042865 | 0.096581 | ## softmax - Job: job_5ec935012955; operator: Softmax; cases: 20/20; score: 68.015671; avg_speedup: 0.661441. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level2/softmax_1 | 19.42 | 15.7 | 1.57 | 0.808445 | 0.080844 | 0.441839 | | 2 | level2/softmax_2 | 30.96 | 37.98 | 8.74 | 1.226744 | 0.2823 | 0.568208 | | 3 | level2/softmax_3 | 104.2 | 91.76 | 25.17 | 0.880614 | 0.241555 | 0.457286 | | 4 | level2/softmax_4 | 2175.34 | 1006.60 | 100.66 | 0.462732 | 0.046273 | 0.303943 | | 5 | level2/softmax_5 | 458.2 | 422.73 | 139.81 | 0.922588 | 0.305129 | 0.470506 | | 6 | level2/softmax_6 | 1907.08 | 786.04 | 101.69 | 0.412169 | 0.053322 | 0.274868 | | 7 | level2/softmax_7 | 22.42 | 20.04 | 3.14 | 0.893845 | 0.140054 | 0.467109 | | 8 | level2/softmax_8 | 61.92 | 58.12 | 17.49 | 0.93863 | 0.282461 | 0.477663 | | 9 | level2/softmax_9 | 1044.48 | 412.79 | 50.08 | 0.395211 | 0.047947 | 0.267266 | | 10 | level2/softmax_10 | 21.82 | 14.2 | 1.55 | 0.650779 | 0.071036 | 0.384265 | | 11 | level2/softmax_11 | 2669.88 | 1063.10 | 108.09 | 0.398183 | 0.040485 | 0.271557 | | 12 | level2/softmax_12 | 61.3 | 62.58 | 14.59 | 1.020881 | 0.23801 | 0.506758 | | 13 | level2/softmax_13 | 2246.58 | 30 | 3 | 0.013354 | 0.001335 | 0.011891 | | 14 | level2/softmax_14 | 238.52 | 227.3 | 22.73 | 0.95296 | 0.095296 | 0.486654 | | 15 | level2/softmax_15 | 30.6 | 30.46 | 4.56 | 0.995425 | 0.14902 | 0.498652 | | 16 | level2/softmax_16 | 115.28 | 15.7 | 1.57 | 0.13619 | 0.013619 | 0.110529 | | 17 | level2/softmax_17 | 19.98 | 18.64 | 4.35 | 0.932933 | 0.217718 | 0.477607 | | 18 | level2/softmax_18 | 24.9 | 20.12 | 3.14 | 0.808032 | 0.126104 | 0.438307 | | 19 | level2/softmax_19 | 119.64 | 31.3 | 3.13 | 0.261618 | 0.026162 | 0.194706 | | 20 | level2/softmax_20 | 1870.84 | 219.8 | 21.98 | 0.117487 | 0.011749 | 0.096654 | ## topk - Job: job_bc7b660e83a6; operator: TopK; cases: 20/20; score: 54.233631; avg_speedup: 0.119558. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level3/top_k_1 | 2431.10 | 10.9 | 1.09 | 0.004484 | 0.000448 | 0.004021 | | 2 | level3/top_k_2 | 279.02 | 21.8 | 2.18 | 0.078131 | 0.007813 | 0.066181 | | 3 | level3/top_k_3 | 995.88 | 174.8 | 17.48 | 0.175523 | 0.017552 | 0.13852 | | 4 | level3/top_k_4 | 281395.82 | 1398.10 | 139.81 | 0.004968 | 0.000497 | 0.004454 | | 5 | level3/top_k_5 | 2508.46 | 1398.10 | 139.81 | 0.557354 | 0.055735 | 0.346929 | | 6 | level3/top_k_6 | 6939698.16 | 1398.10 | 139.81 | 0.000201 | 2.015e-05 | 0.000181 | | 7 | level3/top_k_7 | 74.9 | 10.9 | 1.09 | 0.145527 | 0.014553 | 0.117316 | | 8 | level3/top_k_8 | 3213.32 | 21.5 | 2.15 | 0.006691 | 0.000669 | 0.00599 | | 9 | level3/top_k_9 | 97.66 | 13.8 | 1.38 | 0.141307 | 0.014131 | 0.114259 | | 10 | level3/top_k_10 | 95309.40 | 1035.20 | 103.52 | 0.010861 | 0.001086 | 0.009691 | | 11 | level3/top_k_11 | 104.16 | 10.9 | 1.09 | 0.104647 | 0.010465 | 0.086906 | | 12 | level3/top_k_12 | 111.12 | 22.8 | 2.28 | 0.205184 | 0.020518 | 0.158627 | | 13 | level3/top_k_13 | 4755.26 | 14.3 | 1.43 | 0.003007 | 0.000301 | 0.0027 | | 14 | level3/top_k_14 | 2448.26 | 20.8 | 2.08 | 0.008496 | 0.00085 | 0.007595 | | 15 | level3/top_k_15 | 19245.22 | 170.5 | 17.05 | 0.008859 | 0.000886 | 0.007917 | | 16 | level3/top_k_16 | 211.92 | 126.8 | 12.68 | 0.598339 | 0.059834 | 0.364182 | | 17 | level3/top_k_17 | 80.84 | 21.9 | 2.19 | 0.270905 | 0.027091 | 0.200386 | | 18 | level3/top_k_18 | 6133.36 | 24.4 | 2.44 | 0.003978 | 0.000398 | 0.003569 | | 19 | level3/top_k_19 | 184.64 | 10.9 | 1.09 | 0.059034 | 0.005903 | 0.050734 | | 20 | level3/top_k_20 | 5940.70 | 21.8 | 2.18 | 0.00367 | 0.000367 | 0.003293 | ## 未来三个月的计划 ### 第 1 个月 —— 完善 补全 reduce类,数学函数类,排序类算子的生成与优化skill和harness ### 第 2 个月 —— 拓展CV融合算子 增加对CV类融合算子如quant_batch_matmul,grouped_matmul_swiglu_quant等的生成适配,保障生成精度 ### 第 3 个月 —— 拓展attention算子 增加对attention类算子的适配 See merge request: cann/cannbot-skills!313 | 10 天前 | |
【collaborative-agent-kernel-evolution】新增 CAKE2 进化式 AscendC 算子生成 Team(仅 Vector 算子) Co-authored-by: Junyifannnn<fan.junyi@huawei.com> # message auto-generated for no-merge-commit merge: !313 merge feat/add-cake-plugin into master 【collaborative-agent-kernel-evolution】新增 CAKE2 进化式 AscendC 算子生成 Team(仅 Vector 算子) Created-by: gcw_P2Mh5H1S Commit-by: Junyifannnn Merged-by: cann-robot Description: ## 描述 **新特性**: 【collaborative-agent-kernel-evolution】新增 plugins-community/collaborative-agent-kernel-evolution —— 进化式 AscendC 算子生成 Team(仅 Vector 算子)。 含 3 个 Agent 与 19 个内置 Skill,覆盖「算子描述生成 → PyTorch 参考 → functional 转换 → AscendDSL baseline → DSL lowering 到 AscendC → 编译评估 →(可选)进化式多变体性能优化」全流程: - cake(primary):标准单算子生成主流程 - cake-evo(primary):并行多变体进化优化 - cake-partial(subagent):被 cake-evo 调度的半程生成子任务 自包含单插件(目录 plugins-community/collaborative-agent-kernel-evolution/,对齐 ops-easyasc-dsl 自包含形态) ## 关联的Issue 关联 Issue [#230](https://gitcode.com/cann/cannbot-skills/issues/230) ## 测试 - **L1 结构门禁 skill_validator.py**:19 skills + 3 agents **0 error**(warn-only:6× S-CON-05、2× S-CON-06「缺错误处理章节」,不阻断) - **tests/run-tests.sh --fast**:10 项全过(初次 test-line-endings 报 6 个 CRLF 文件,均为插件内 vendored 外部文档,已全部转 LF,复跑 PASS) ## 文档更新 - 新增 plugins-community/collaborative-agent-kernel-evolution/(含 README / docs / tests) - 根 .claude-plugin/marketplace.json 注册 cake(source ./plugins-community/collaborative-agent-kernel-evolution,category development) - 更新根 CHANGELOG.md(2026-06-02 条目) ## 类型标签 - [x] 🌟 新特性 ## AscendC 算子评测结果 | Operator | Job ID | Cases | Score | Avg speedup | Rel path | |---|---|---:|---:|---:|---|---| | apply_rotary_pos_emb | job_940a45830a2f | 20/20 | 63.339075 | 0.355671 | level2/apply_rotary_pos_emb | | softmax | job_5ec935012955 | 20/20 | 68.015671 | 0.661441 | level2/softmax | | topk | job_bc7b660e83a6 | 20/20 | 54.233631 | 0.119558 | level3/top_k | ## apply_rotary_pos_emb - Job: job_940a45830a2f; operator: ApplyRotaryPosEmb; cases: 20/20; score: 63.339075; avg_speedup: 0.355671. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level2/apply_rotary_pos_emb_1 | 182.06 | 89.56 | 35.02 | 0.491926 | 0.192354 | 0.270563 | | 2 | level2/apply_rotary_pos_emb_2 | 779.98 | 113.24 | 59.22 | 0.145183 | 0.075925 | 0.069723 | | 3 | level2/apply_rotary_pos_emb_3 | 125.34 | 78.24 | 14.56 | 0.624222 | 0.116164 | 0.365012 | | 4 | level2/apply_rotary_pos_emb_4 | 140.26 | 246.2 | 32.77 | 1.755312 | 0.233638 | 0.665057 | | 5 | level2/apply_rotary_pos_emb_5 | 3001.06 | 391.07 | 140.36 | 0.130311 | 0.04677 | 0.080578 | | 6 | level2/apply_rotary_pos_emb_6 | 654.54 | 347.8 | 34.78 | 0.531366 | 0.053137 | 0.335578 | | 7 | level2/apply_rotary_pos_emb_7 | 146.94 | 73.08 | 23.1 | 0.497346 | 0.157207 | 0.287539 | | 8 | level2/apply_rotary_pos_emb_8 | 360.54 | 74.4 | 26.64 | 0.206357 | 0.073889 | 0.125138 | | 9 | level2/apply_rotary_pos_emb_9 | 178.12 | 198.1 | 19.81 | 1.112172 | 0.111217 | 0.529679 | | 10 | level2/apply_rotary_pos_emb_10 | 274.06 | 163 | 16.3 | 0.59476 | 0.059476 | 0.362706 | | 11 | level2/apply_rotary_pos_emb_11 | 1398.00 | 186.7 | 98.62 | 0.133548 | 0.070544 | 0.063483 | | 12 | level2/apply_rotary_pos_emb_12 | 896.26 | 48 | 4.8 | 0.053556 | 0.005356 | 0.04622 | | 13 | level2/apply_rotary_pos_emb_13 | 595.38 | 425.07 | 100 | 0.713947 | 0.16796 | 0.396209 | | 14 | level2/apply_rotary_pos_emb_14 | 27.18 | 10 | 1 | 0.367918 | 0.036792 | 0.255827 | | 15 | level2/apply_rotary_pos_emb_15 | 86.22 | 21.7 | 2.17 | 0.251682 | 0.025168 | 0.18855 | | 16 | level2/apply_rotary_pos_emb_16 | 82.48 | 41.34 | 4.17 | 0.501212 | 0.050558 | 0.321874 | | 17 | level2/apply_rotary_pos_emb_17 | 5265.48 | 2621.60 | 262.16 | 0.497884 | 0.049788 | 0.320456 | | 18 | level2/apply_rotary_pos_emb_18 | 97.2 | 43.9 | 4.39 | 0.451646 | 0.045165 | 0.298594 | | 19 | level2/apply_rotary_pos_emb_19 | 27.12 | 10 | 1 | 0.368732 | 0.036873 | 0.256264 | | 20 | level2/apply_rotary_pos_emb_20 | 1260.70 | 183.04 | 54.04 | 0.145189 | 0.042865 | 0.096581 | ## softmax - Job: job_5ec935012955; operator: Softmax; cases: 20/20; score: 68.015671; avg_speedup: 0.661441. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level2/softmax_1 | 19.42 | 15.7 | 1.57 | 0.808445 | 0.080844 | 0.441839 | | 2 | level2/softmax_2 | 30.96 | 37.98 | 8.74 | 1.226744 | 0.2823 | 0.568208 | | 3 | level2/softmax_3 | 104.2 | 91.76 | 25.17 | 0.880614 | 0.241555 | 0.457286 | | 4 | level2/softmax_4 | 2175.34 | 1006.60 | 100.66 | 0.462732 | 0.046273 | 0.303943 | | 5 | level2/softmax_5 | 458.2 | 422.73 | 139.81 | 0.922588 | 0.305129 | 0.470506 | | 6 | level2/softmax_6 | 1907.08 | 786.04 | 101.69 | 0.412169 | 0.053322 | 0.274868 | | 7 | level2/softmax_7 | 22.42 | 20.04 | 3.14 | 0.893845 | 0.140054 | 0.467109 | | 8 | level2/softmax_8 | 61.92 | 58.12 | 17.49 | 0.93863 | 0.282461 | 0.477663 | | 9 | level2/softmax_9 | 1044.48 | 412.79 | 50.08 | 0.395211 | 0.047947 | 0.267266 | | 10 | level2/softmax_10 | 21.82 | 14.2 | 1.55 | 0.650779 | 0.071036 | 0.384265 | | 11 | level2/softmax_11 | 2669.88 | 1063.10 | 108.09 | 0.398183 | 0.040485 | 0.271557 | | 12 | level2/softmax_12 | 61.3 | 62.58 | 14.59 | 1.020881 | 0.23801 | 0.506758 | | 13 | level2/softmax_13 | 2246.58 | 30 | 3 | 0.013354 | 0.001335 | 0.011891 | | 14 | level2/softmax_14 | 238.52 | 227.3 | 22.73 | 0.95296 | 0.095296 | 0.486654 | | 15 | level2/softmax_15 | 30.6 | 30.46 | 4.56 | 0.995425 | 0.14902 | 0.498652 | | 16 | level2/softmax_16 | 115.28 | 15.7 | 1.57 | 0.13619 | 0.013619 | 0.110529 | | 17 | level2/softmax_17 | 19.98 | 18.64 | 4.35 | 0.932933 | 0.217718 | 0.477607 | | 18 | level2/softmax_18 | 24.9 | 20.12 | 3.14 | 0.808032 | 0.126104 | 0.438307 | | 19 | level2/softmax_19 | 119.64 | 31.3 | 3.13 | 0.261618 | 0.026162 | 0.194706 | | 20 | level2/softmax_20 | 1870.84 | 219.8 | 21.98 | 0.117487 | 0.011749 | 0.096654 | ## topk - Job: job_bc7b660e83a6; operator: TopK; cases: 20/20; score: 54.233631; avg_speedup: 0.119558. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level3/top_k_1 | 2431.10 | 10.9 | 1.09 | 0.004484 | 0.000448 | 0.004021 | | 2 | level3/top_k_2 | 279.02 | 21.8 | 2.18 | 0.078131 | 0.007813 | 0.066181 | | 3 | level3/top_k_3 | 995.88 | 174.8 | 17.48 | 0.175523 | 0.017552 | 0.13852 | | 4 | level3/top_k_4 | 281395.82 | 1398.10 | 139.81 | 0.004968 | 0.000497 | 0.004454 | | 5 | level3/top_k_5 | 2508.46 | 1398.10 | 139.81 | 0.557354 | 0.055735 | 0.346929 | | 6 | level3/top_k_6 | 6939698.16 | 1398.10 | 139.81 | 0.000201 | 2.015e-05 | 0.000181 | | 7 | level3/top_k_7 | 74.9 | 10.9 | 1.09 | 0.145527 | 0.014553 | 0.117316 | | 8 | level3/top_k_8 | 3213.32 | 21.5 | 2.15 | 0.006691 | 0.000669 | 0.00599 | | 9 | level3/top_k_9 | 97.66 | 13.8 | 1.38 | 0.141307 | 0.014131 | 0.114259 | | 10 | level3/top_k_10 | 95309.40 | 1035.20 | 103.52 | 0.010861 | 0.001086 | 0.009691 | | 11 | level3/top_k_11 | 104.16 | 10.9 | 1.09 | 0.104647 | 0.010465 | 0.086906 | | 12 | level3/top_k_12 | 111.12 | 22.8 | 2.28 | 0.205184 | 0.020518 | 0.158627 | | 13 | level3/top_k_13 | 4755.26 | 14.3 | 1.43 | 0.003007 | 0.000301 | 0.0027 | | 14 | level3/top_k_14 | 2448.26 | 20.8 | 2.08 | 0.008496 | 0.00085 | 0.007595 | | 15 | level3/top_k_15 | 19245.22 | 170.5 | 17.05 | 0.008859 | 0.000886 | 0.007917 | | 16 | level3/top_k_16 | 211.92 | 126.8 | 12.68 | 0.598339 | 0.059834 | 0.364182 | | 17 | level3/top_k_17 | 80.84 | 21.9 | 2.19 | 0.270905 | 0.027091 | 0.200386 | | 18 | level3/top_k_18 | 6133.36 | 24.4 | 2.44 | 0.003978 | 0.000398 | 0.003569 | | 19 | level3/top_k_19 | 184.64 | 10.9 | 1.09 | 0.059034 | 0.005903 | 0.050734 | | 20 | level3/top_k_20 | 5940.70 | 21.8 | 2.18 | 0.00367 | 0.000367 | 0.003293 | ## 未来三个月的计划 ### 第 1 个月 —— 完善 补全 reduce类,数学函数类,排序类算子的生成与优化skill和harness ### 第 2 个月 —— 拓展CV融合算子 增加对CV类融合算子如quant_batch_matmul,grouped_matmul_swiglu_quant等的生成适配,保障生成精度 ### 第 3 个月 —— 拓展attention算子 增加对attention类算子的适配 See merge request: cann/cannbot-skills!313 | 10 天前 | |
【collaborative-agent-kernel-evolution】新增 CAKE2 进化式 AscendC 算子生成 Team(仅 Vector 算子) Co-authored-by: Junyifannnn<fan.junyi@huawei.com> # message auto-generated for no-merge-commit merge: !313 merge feat/add-cake-plugin into master 【collaborative-agent-kernel-evolution】新增 CAKE2 进化式 AscendC 算子生成 Team(仅 Vector 算子) Created-by: gcw_P2Mh5H1S Commit-by: Junyifannnn Merged-by: cann-robot Description: ## 描述 **新特性**: 【collaborative-agent-kernel-evolution】新增 plugins-community/collaborative-agent-kernel-evolution —— 进化式 AscendC 算子生成 Team(仅 Vector 算子)。 含 3 个 Agent 与 19 个内置 Skill,覆盖「算子描述生成 → PyTorch 参考 → functional 转换 → AscendDSL baseline → DSL lowering 到 AscendC → 编译评估 →(可选)进化式多变体性能优化」全流程: - cake(primary):标准单算子生成主流程 - cake-evo(primary):并行多变体进化优化 - cake-partial(subagent):被 cake-evo 调度的半程生成子任务 自包含单插件(目录 plugins-community/collaborative-agent-kernel-evolution/,对齐 ops-easyasc-dsl 自包含形态) ## 关联的Issue 关联 Issue [#230](https://gitcode.com/cann/cannbot-skills/issues/230) ## 测试 - **L1 结构门禁 skill_validator.py**:19 skills + 3 agents **0 error**(warn-only:6× S-CON-05、2× S-CON-06「缺错误处理章节」,不阻断) - **tests/run-tests.sh --fast**:10 项全过(初次 test-line-endings 报 6 个 CRLF 文件,均为插件内 vendored 外部文档,已全部转 LF,复跑 PASS) ## 文档更新 - 新增 plugins-community/collaborative-agent-kernel-evolution/(含 README / docs / tests) - 根 .claude-plugin/marketplace.json 注册 cake(source ./plugins-community/collaborative-agent-kernel-evolution,category development) - 更新根 CHANGELOG.md(2026-06-02 条目) ## 类型标签 - [x] 🌟 新特性 ## AscendC 算子评测结果 | Operator | Job ID | Cases | Score | Avg speedup | Rel path | |---|---|---:|---:|---:|---|---| | apply_rotary_pos_emb | job_940a45830a2f | 20/20 | 63.339075 | 0.355671 | level2/apply_rotary_pos_emb | | softmax | job_5ec935012955 | 20/20 | 68.015671 | 0.661441 | level2/softmax | | topk | job_bc7b660e83a6 | 20/20 | 54.233631 | 0.119558 | level3/top_k | ## apply_rotary_pos_emb - Job: job_940a45830a2f; operator: ApplyRotaryPosEmb; cases: 20/20; score: 63.339075; avg_speedup: 0.355671. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level2/apply_rotary_pos_emb_1 | 182.06 | 89.56 | 35.02 | 0.491926 | 0.192354 | 0.270563 | | 2 | level2/apply_rotary_pos_emb_2 | 779.98 | 113.24 | 59.22 | 0.145183 | 0.075925 | 0.069723 | | 3 | level2/apply_rotary_pos_emb_3 | 125.34 | 78.24 | 14.56 | 0.624222 | 0.116164 | 0.365012 | | 4 | level2/apply_rotary_pos_emb_4 | 140.26 | 246.2 | 32.77 | 1.755312 | 0.233638 | 0.665057 | | 5 | level2/apply_rotary_pos_emb_5 | 3001.06 | 391.07 | 140.36 | 0.130311 | 0.04677 | 0.080578 | | 6 | level2/apply_rotary_pos_emb_6 | 654.54 | 347.8 | 34.78 | 0.531366 | 0.053137 | 0.335578 | | 7 | level2/apply_rotary_pos_emb_7 | 146.94 | 73.08 | 23.1 | 0.497346 | 0.157207 | 0.287539 | | 8 | level2/apply_rotary_pos_emb_8 | 360.54 | 74.4 | 26.64 | 0.206357 | 0.073889 | 0.125138 | | 9 | level2/apply_rotary_pos_emb_9 | 178.12 | 198.1 | 19.81 | 1.112172 | 0.111217 | 0.529679 | | 10 | level2/apply_rotary_pos_emb_10 | 274.06 | 163 | 16.3 | 0.59476 | 0.059476 | 0.362706 | | 11 | level2/apply_rotary_pos_emb_11 | 1398.00 | 186.7 | 98.62 | 0.133548 | 0.070544 | 0.063483 | | 12 | level2/apply_rotary_pos_emb_12 | 896.26 | 48 | 4.8 | 0.053556 | 0.005356 | 0.04622 | | 13 | level2/apply_rotary_pos_emb_13 | 595.38 | 425.07 | 100 | 0.713947 | 0.16796 | 0.396209 | | 14 | level2/apply_rotary_pos_emb_14 | 27.18 | 10 | 1 | 0.367918 | 0.036792 | 0.255827 | | 15 | level2/apply_rotary_pos_emb_15 | 86.22 | 21.7 | 2.17 | 0.251682 | 0.025168 | 0.18855 | | 16 | level2/apply_rotary_pos_emb_16 | 82.48 | 41.34 | 4.17 | 0.501212 | 0.050558 | 0.321874 | | 17 | level2/apply_rotary_pos_emb_17 | 5265.48 | 2621.60 | 262.16 | 0.497884 | 0.049788 | 0.320456 | | 18 | level2/apply_rotary_pos_emb_18 | 97.2 | 43.9 | 4.39 | 0.451646 | 0.045165 | 0.298594 | | 19 | level2/apply_rotary_pos_emb_19 | 27.12 | 10 | 1 | 0.368732 | 0.036873 | 0.256264 | | 20 | level2/apply_rotary_pos_emb_20 | 1260.70 | 183.04 | 54.04 | 0.145189 | 0.042865 | 0.096581 | ## softmax - Job: job_5ec935012955; operator: Softmax; cases: 20/20; score: 68.015671; avg_speedup: 0.661441. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level2/softmax_1 | 19.42 | 15.7 | 1.57 | 0.808445 | 0.080844 | 0.441839 | | 2 | level2/softmax_2 | 30.96 | 37.98 | 8.74 | 1.226744 | 0.2823 | 0.568208 | | 3 | level2/softmax_3 | 104.2 | 91.76 | 25.17 | 0.880614 | 0.241555 | 0.457286 | | 4 | level2/softmax_4 | 2175.34 | 1006.60 | 100.66 | 0.462732 | 0.046273 | 0.303943 | | 5 | level2/softmax_5 | 458.2 | 422.73 | 139.81 | 0.922588 | 0.305129 | 0.470506 | | 6 | level2/softmax_6 | 1907.08 | 786.04 | 101.69 | 0.412169 | 0.053322 | 0.274868 | | 7 | level2/softmax_7 | 22.42 | 20.04 | 3.14 | 0.893845 | 0.140054 | 0.467109 | | 8 | level2/softmax_8 | 61.92 | 58.12 | 17.49 | 0.93863 | 0.282461 | 0.477663 | | 9 | level2/softmax_9 | 1044.48 | 412.79 | 50.08 | 0.395211 | 0.047947 | 0.267266 | | 10 | level2/softmax_10 | 21.82 | 14.2 | 1.55 | 0.650779 | 0.071036 | 0.384265 | | 11 | level2/softmax_11 | 2669.88 | 1063.10 | 108.09 | 0.398183 | 0.040485 | 0.271557 | | 12 | level2/softmax_12 | 61.3 | 62.58 | 14.59 | 1.020881 | 0.23801 | 0.506758 | | 13 | level2/softmax_13 | 2246.58 | 30 | 3 | 0.013354 | 0.001335 | 0.011891 | | 14 | level2/softmax_14 | 238.52 | 227.3 | 22.73 | 0.95296 | 0.095296 | 0.486654 | | 15 | level2/softmax_15 | 30.6 | 30.46 | 4.56 | 0.995425 | 0.14902 | 0.498652 | | 16 | level2/softmax_16 | 115.28 | 15.7 | 1.57 | 0.13619 | 0.013619 | 0.110529 | | 17 | level2/softmax_17 | 19.98 | 18.64 | 4.35 | 0.932933 | 0.217718 | 0.477607 | | 18 | level2/softmax_18 | 24.9 | 20.12 | 3.14 | 0.808032 | 0.126104 | 0.438307 | | 19 | level2/softmax_19 | 119.64 | 31.3 | 3.13 | 0.261618 | 0.026162 | 0.194706 | | 20 | level2/softmax_20 | 1870.84 | 219.8 | 21.98 | 0.117487 | 0.011749 | 0.096654 | ## topk - Job: job_bc7b660e83a6; operator: TopK; cases: 20/20; score: 54.233631; avg_speedup: 0.119558. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level3/top_k_1 | 2431.10 | 10.9 | 1.09 | 0.004484 | 0.000448 | 0.004021 | | 2 | level3/top_k_2 | 279.02 | 21.8 | 2.18 | 0.078131 | 0.007813 | 0.066181 | | 3 | level3/top_k_3 | 995.88 | 174.8 | 17.48 | 0.175523 | 0.017552 | 0.13852 | | 4 | level3/top_k_4 | 281395.82 | 1398.10 | 139.81 | 0.004968 | 0.000497 | 0.004454 | | 5 | level3/top_k_5 | 2508.46 | 1398.10 | 139.81 | 0.557354 | 0.055735 | 0.346929 | | 6 | level3/top_k_6 | 6939698.16 | 1398.10 | 139.81 | 0.000201 | 2.015e-05 | 0.000181 | | 7 | level3/top_k_7 | 74.9 | 10.9 | 1.09 | 0.145527 | 0.014553 | 0.117316 | | 8 | level3/top_k_8 | 3213.32 | 21.5 | 2.15 | 0.006691 | 0.000669 | 0.00599 | | 9 | level3/top_k_9 | 97.66 | 13.8 | 1.38 | 0.141307 | 0.014131 | 0.114259 | | 10 | level3/top_k_10 | 95309.40 | 1035.20 | 103.52 | 0.010861 | 0.001086 | 0.009691 | | 11 | level3/top_k_11 | 104.16 | 10.9 | 1.09 | 0.104647 | 0.010465 | 0.086906 | | 12 | level3/top_k_12 | 111.12 | 22.8 | 2.28 | 0.205184 | 0.020518 | 0.158627 | | 13 | level3/top_k_13 | 4755.26 | 14.3 | 1.43 | 0.003007 | 0.000301 | 0.0027 | | 14 | level3/top_k_14 | 2448.26 | 20.8 | 2.08 | 0.008496 | 0.00085 | 0.007595 | | 15 | level3/top_k_15 | 19245.22 | 170.5 | 17.05 | 0.008859 | 0.000886 | 0.007917 | | 16 | level3/top_k_16 | 211.92 | 126.8 | 12.68 | 0.598339 | 0.059834 | 0.364182 | | 17 | level3/top_k_17 | 80.84 | 21.9 | 2.19 | 0.270905 | 0.027091 | 0.200386 | | 18 | level3/top_k_18 | 6133.36 | 24.4 | 2.44 | 0.003978 | 0.000398 | 0.003569 | | 19 | level3/top_k_19 | 184.64 | 10.9 | 1.09 | 0.059034 | 0.005903 | 0.050734 | | 20 | level3/top_k_20 | 5940.70 | 21.8 | 2.18 | 0.00367 | 0.000367 | 0.003293 | ## 未来三个月的计划 ### 第 1 个月 —— 完善 补全 reduce类,数学函数类,排序类算子的生成与优化skill和harness ### 第 2 个月 —— 拓展CV融合算子 增加对CV类融合算子如quant_batch_matmul,grouped_matmul_swiglu_quant等的生成适配,保障生成精度 ### 第 3 个月 —— 拓展attention算子 增加对attention类算子的适配 See merge request: cann/cannbot-skills!313 | 10 天前 | |
【collaborative-agent-kernel-evolution】新增 CAKE2 进化式 AscendC 算子生成 Team(仅 Vector 算子) Co-authored-by: Junyifannnn<fan.junyi@huawei.com> # message auto-generated for no-merge-commit merge: !313 merge feat/add-cake-plugin into master 【collaborative-agent-kernel-evolution】新增 CAKE2 进化式 AscendC 算子生成 Team(仅 Vector 算子) Created-by: gcw_P2Mh5H1S Commit-by: Junyifannnn Merged-by: cann-robot Description: ## 描述 **新特性**: 【collaborative-agent-kernel-evolution】新增 plugins-community/collaborative-agent-kernel-evolution —— 进化式 AscendC 算子生成 Team(仅 Vector 算子)。 含 3 个 Agent 与 19 个内置 Skill,覆盖「算子描述生成 → PyTorch 参考 → functional 转换 → AscendDSL baseline → DSL lowering 到 AscendC → 编译评估 →(可选)进化式多变体性能优化」全流程: - cake(primary):标准单算子生成主流程 - cake-evo(primary):并行多变体进化优化 - cake-partial(subagent):被 cake-evo 调度的半程生成子任务 自包含单插件(目录 plugins-community/collaborative-agent-kernel-evolution/,对齐 ops-easyasc-dsl 自包含形态) ## 关联的Issue 关联 Issue [#230](https://gitcode.com/cann/cannbot-skills/issues/230) ## 测试 - **L1 结构门禁 skill_validator.py**:19 skills + 3 agents **0 error**(warn-only:6× S-CON-05、2× S-CON-06「缺错误处理章节」,不阻断) - **tests/run-tests.sh --fast**:10 项全过(初次 test-line-endings 报 6 个 CRLF 文件,均为插件内 vendored 外部文档,已全部转 LF,复跑 PASS) ## 文档更新 - 新增 plugins-community/collaborative-agent-kernel-evolution/(含 README / docs / tests) - 根 .claude-plugin/marketplace.json 注册 cake(source ./plugins-community/collaborative-agent-kernel-evolution,category development) - 更新根 CHANGELOG.md(2026-06-02 条目) ## 类型标签 - [x] 🌟 新特性 ## AscendC 算子评测结果 | Operator | Job ID | Cases | Score | Avg speedup | Rel path | |---|---|---:|---:|---:|---|---| | apply_rotary_pos_emb | job_940a45830a2f | 20/20 | 63.339075 | 0.355671 | level2/apply_rotary_pos_emb | | softmax | job_5ec935012955 | 20/20 | 68.015671 | 0.661441 | level2/softmax | | topk | job_bc7b660e83a6 | 20/20 | 54.233631 | 0.119558 | level3/top_k | ## apply_rotary_pos_emb - Job: job_940a45830a2f; operator: ApplyRotaryPosEmb; cases: 20/20; score: 63.339075; avg_speedup: 0.355671. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level2/apply_rotary_pos_emb_1 | 182.06 | 89.56 | 35.02 | 0.491926 | 0.192354 | 0.270563 | | 2 | level2/apply_rotary_pos_emb_2 | 779.98 | 113.24 | 59.22 | 0.145183 | 0.075925 | 0.069723 | | 3 | level2/apply_rotary_pos_emb_3 | 125.34 | 78.24 | 14.56 | 0.624222 | 0.116164 | 0.365012 | | 4 | level2/apply_rotary_pos_emb_4 | 140.26 | 246.2 | 32.77 | 1.755312 | 0.233638 | 0.665057 | | 5 | level2/apply_rotary_pos_emb_5 | 3001.06 | 391.07 | 140.36 | 0.130311 | 0.04677 | 0.080578 | | 6 | level2/apply_rotary_pos_emb_6 | 654.54 | 347.8 | 34.78 | 0.531366 | 0.053137 | 0.335578 | | 7 | level2/apply_rotary_pos_emb_7 | 146.94 | 73.08 | 23.1 | 0.497346 | 0.157207 | 0.287539 | | 8 | level2/apply_rotary_pos_emb_8 | 360.54 | 74.4 | 26.64 | 0.206357 | 0.073889 | 0.125138 | | 9 | level2/apply_rotary_pos_emb_9 | 178.12 | 198.1 | 19.81 | 1.112172 | 0.111217 | 0.529679 | | 10 | level2/apply_rotary_pos_emb_10 | 274.06 | 163 | 16.3 | 0.59476 | 0.059476 | 0.362706 | | 11 | level2/apply_rotary_pos_emb_11 | 1398.00 | 186.7 | 98.62 | 0.133548 | 0.070544 | 0.063483 | | 12 | level2/apply_rotary_pos_emb_12 | 896.26 | 48 | 4.8 | 0.053556 | 0.005356 | 0.04622 | | 13 | level2/apply_rotary_pos_emb_13 | 595.38 | 425.07 | 100 | 0.713947 | 0.16796 | 0.396209 | | 14 | level2/apply_rotary_pos_emb_14 | 27.18 | 10 | 1 | 0.367918 | 0.036792 | 0.255827 | | 15 | level2/apply_rotary_pos_emb_15 | 86.22 | 21.7 | 2.17 | 0.251682 | 0.025168 | 0.18855 | | 16 | level2/apply_rotary_pos_emb_16 | 82.48 | 41.34 | 4.17 | 0.501212 | 0.050558 | 0.321874 | | 17 | level2/apply_rotary_pos_emb_17 | 5265.48 | 2621.60 | 262.16 | 0.497884 | 0.049788 | 0.320456 | | 18 | level2/apply_rotary_pos_emb_18 | 97.2 | 43.9 | 4.39 | 0.451646 | 0.045165 | 0.298594 | | 19 | level2/apply_rotary_pos_emb_19 | 27.12 | 10 | 1 | 0.368732 | 0.036873 | 0.256264 | | 20 | level2/apply_rotary_pos_emb_20 | 1260.70 | 183.04 | 54.04 | 0.145189 | 0.042865 | 0.096581 | ## softmax - Job: job_5ec935012955; operator: Softmax; cases: 20/20; score: 68.015671; avg_speedup: 0.661441. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level2/softmax_1 | 19.42 | 15.7 | 1.57 | 0.808445 | 0.080844 | 0.441839 | | 2 | level2/softmax_2 | 30.96 | 37.98 | 8.74 | 1.226744 | 0.2823 | 0.568208 | | 3 | level2/softmax_3 | 104.2 | 91.76 | 25.17 | 0.880614 | 0.241555 | 0.457286 | | 4 | level2/softmax_4 | 2175.34 | 1006.60 | 100.66 | 0.462732 | 0.046273 | 0.303943 | | 5 | level2/softmax_5 | 458.2 | 422.73 | 139.81 | 0.922588 | 0.305129 | 0.470506 | | 6 | level2/softmax_6 | 1907.08 | 786.04 | 101.69 | 0.412169 | 0.053322 | 0.274868 | | 7 | level2/softmax_7 | 22.42 | 20.04 | 3.14 | 0.893845 | 0.140054 | 0.467109 | | 8 | level2/softmax_8 | 61.92 | 58.12 | 17.49 | 0.93863 | 0.282461 | 0.477663 | | 9 | level2/softmax_9 | 1044.48 | 412.79 | 50.08 | 0.395211 | 0.047947 | 0.267266 | | 10 | level2/softmax_10 | 21.82 | 14.2 | 1.55 | 0.650779 | 0.071036 | 0.384265 | | 11 | level2/softmax_11 | 2669.88 | 1063.10 | 108.09 | 0.398183 | 0.040485 | 0.271557 | | 12 | level2/softmax_12 | 61.3 | 62.58 | 14.59 | 1.020881 | 0.23801 | 0.506758 | | 13 | level2/softmax_13 | 2246.58 | 30 | 3 | 0.013354 | 0.001335 | 0.011891 | | 14 | level2/softmax_14 | 238.52 | 227.3 | 22.73 | 0.95296 | 0.095296 | 0.486654 | | 15 | level2/softmax_15 | 30.6 | 30.46 | 4.56 | 0.995425 | 0.14902 | 0.498652 | | 16 | level2/softmax_16 | 115.28 | 15.7 | 1.57 | 0.13619 | 0.013619 | 0.110529 | | 17 | level2/softmax_17 | 19.98 | 18.64 | 4.35 | 0.932933 | 0.217718 | 0.477607 | | 18 | level2/softmax_18 | 24.9 | 20.12 | 3.14 | 0.808032 | 0.126104 | 0.438307 | | 19 | level2/softmax_19 | 119.64 | 31.3 | 3.13 | 0.261618 | 0.026162 | 0.194706 | | 20 | level2/softmax_20 | 1870.84 | 219.8 | 21.98 | 0.117487 | 0.011749 | 0.096654 | ## topk - Job: job_bc7b660e83a6; operator: TopK; cases: 20/20; score: 54.233631; avg_speedup: 0.119558. | # | Case ID | Elapsed us | Baseline us | HW limit us | Speedup | Speedup vs HW | Perf score | |---:|---|---:|---:|---:|---:|---:|---:| | 1 | level3/top_k_1 | 2431.10 | 10.9 | 1.09 | 0.004484 | 0.000448 | 0.004021 | | 2 | level3/top_k_2 | 279.02 | 21.8 | 2.18 | 0.078131 | 0.007813 | 0.066181 | | 3 | level3/top_k_3 | 995.88 | 174.8 | 17.48 | 0.175523 | 0.017552 | 0.13852 | | 4 | level3/top_k_4 | 281395.82 | 1398.10 | 139.81 | 0.004968 | 0.000497 | 0.004454 | | 5 | level3/top_k_5 | 2508.46 | 1398.10 | 139.81 | 0.557354 | 0.055735 | 0.346929 | | 6 | level3/top_k_6 | 6939698.16 | 1398.10 | 139.81 | 0.000201 | 2.015e-05 | 0.000181 | | 7 | level3/top_k_7 | 74.9 | 10.9 | 1.09 | 0.145527 | 0.014553 | 0.117316 | | 8 | level3/top_k_8 | 3213.32 | 21.5 | 2.15 | 0.006691 | 0.000669 | 0.00599 | | 9 | level3/top_k_9 | 97.66 | 13.8 | 1.38 | 0.141307 | 0.014131 | 0.114259 | | 10 | level3/top_k_10 | 95309.40 | 1035.20 | 103.52 | 0.010861 | 0.001086 | 0.009691 | | 11 | level3/top_k_11 | 104.16 | 10.9 | 1.09 | 0.104647 | 0.010465 | 0.086906 | | 12 | level3/top_k_12 | 111.12 | 22.8 | 2.28 | 0.205184 | 0.020518 | 0.158627 | | 13 | level3/top_k_13 | 4755.26 | 14.3 | 1.43 | 0.003007 | 0.000301 | 0.0027 | | 14 | level3/top_k_14 | 2448.26 | 20.8 | 2.08 | 0.008496 | 0.00085 | 0.007595 | | 15 | level3/top_k_15 | 19245.22 | 170.5 | 17.05 | 0.008859 | 0.000886 | 0.007917 | | 16 | level3/top_k_16 | 211.92 | 126.8 | 12.68 | 0.598339 | 0.059834 | 0.364182 | | 17 | level3/top_k_17 | 80.84 | 21.9 | 2.19 | 0.270905 | 0.027091 | 0.200386 | | 18 | level3/top_k_18 | 6133.36 | 24.4 | 2.44 | 0.003978 | 0.000398 | 0.003569 | | 19 | level3/top_k_19 | 184.64 | 10.9 | 1.09 | 0.059034 | 0.005903 | 0.050734 | | 20 | level3/top_k_20 | 5940.70 | 21.8 | 2.18 | 0.00367 | 0.000367 | 0.003293 | ## 未来三个月的计划 ### 第 1 个月 —— 完善 补全 reduce类,数学函数类,排序类算子的生成与优化skill和harness ### 第 2 个月 —— 拓展CV融合算子 增加对CV类融合算子如quant_batch_matmul,grouped_matmul_swiglu_quant等的生成适配,保障生成精度 ### 第 3 个月 —— 拓展attention算子 增加对attention类算子的适配 See merge request: cann/cannbot-skills!313 | 10 天前 |
CAKE2 Documentation
This directory contains comprehensive documentation for CAKE2 features and capabilities.
Main Documentation Files
PROJECT_OVERVIEW.md - Project Introduction
High-level overview of CAKE2 project, including:
- Project purpose and capabilities
- Architecture overview (local mode)
- Supported operator types
- Technology stack and workflow
Use this when: You're new to CAKE2 and want to understand what it does.
GETTING_STARTED.md - Installation and First Steps
Quick start guide for local CANN environment setup, including:
- Prerequisites and installation
- First operator generation
- Output structure understanding
- Troubleshooting common issues
Use this when: You want to install CAKE2 and generate your first operator.
ARCHITECTURE.md - System Architecture
Deep technical overview of CAKE2's architecture, including:
- Core components and pipeline stages
- Local compilation architecture
- Skill system organization
- Performance optimizations and design patterns
Use this when: You want to understand how CAKE2 works internally.
EVO.md - Evolutionary Kernel Generation
Complete guide to the evolutionary kernel generation system, including:
- Quick start guide
- How to use the cake_evo agent
- How inspirations and diversity work
- Subagent execution and result evaluation
- Troubleshooting evolutionary workflows
Use this when: You want to generate multiple kernel variants and optimize through evolutionary search.
RETURN_CODE_EXPLANATIONS.md - Return Codes
Reference for CAKE2 return codes and error handling.
Quick Navigation
For Users
I want to understand what CAKE2 is:
- Start with PROJECT_OVERVIEW.md
I want to generate my first kernel:
- Follow GETTING_STARTED.md
- Use the
cakeagent in Claude Code
I want to generate multiple kernel variants and optimize:
- Use the
cake_evoagent in Claude Code - See EVO.md for complete guide
For Developers
I want to understand the system architecture:
- Read ARCHITECTURE.md
I want to understand the evolutionary algorithm:
- See EVO.md → How Inspirations and Diversity Work
I want to modify the evolutionary system:
- See EVO.md and study the evolution/ module
Document Structure
PROJECT_OVERVIEW.md
- Project introduction and purpose
- Key features and capabilities
- Technology stack overview
- Use cases and limitations
GETTING_STARTED.md
- Installation prerequisites
- CANN environment setup
- First operator generation
- Troubleshooting guide
ARCHITECTURE.md
- System architecture overview
- Core components breakdown
- Local compilation pipeline
- Design patterns and optimizations
EVO.md
- Evolutionary generation workflow
- Tiered sampling and inspirations
- Subagent execution model
- Result evaluation and classification
Additional Resources
- Evolution strategies:
skills/cake-evo/references/meta_prompts/— optimization strategy index and details
Getting Started
1. Install CANN Environment
Follow GETTING_STARTED.md for CANN 8.3+ setup.
2. Load CAKE2 Plugin
claude --plugin-dir /path/to/cake2
3. Generate Kernels
Type /agents to select an agent:
cake— Generate a single high-quality kernelcake-evo— Evolutionary multi-variant optimization
See EVO.md for evolution guide.
Common Questions
Q: What's the difference between cake and cake_evo?
A:
cake: Generates a single high-quality kernelcake_evo: Generates multiple variants and optimizes through evolution
Q: Do I need NPU hardware locally?
A: Yes, CAKE2 requires local CANN environment and NPU hardware for compilation.
Q: How does the evolutionary system improve across rounds?
A: It uses tiered sampling to select inspirations from successful implementations, then generates new variants with different optimization strategies. See EVO.md → How Inspirations and Diversity Work.
Q: How are results evaluated?
A: Results are evaluated on compilation success, correctness (vs PyTorch reference), and performance (speedup). See EVO.md → Result Evaluation.
Contributing
When adding new documentation:
- Update the appropriate existing file or create a new one
- Update this README with navigation links
Support
For issues or questions:
- Check the relevant documentation file
- See main project README for contact information
Last Updated: 2026-03-16 Documentation Version: Local Mode Only