cake

Generate AscendC project scaffold (pybind, CMake, host code, kernel skeleton) from functional PyTorch for pure Vector operators. Use after functional-conversion, before dsl-baseline-generation.

Multi-case operator evaluation with precision testing and performance profiling, use when you want to evaluate the performance and correctness of your AscendC operator implementation.

Unified AscendC operator runtime debug skill. Diagnose precision errors, runtime crashes, hangs, and multicore inconsistencies using a hypothesis-driven protocol with 3-layer evidence (code review → log analysis → tools). Covers hypothesis patterns for Vector operators (cache line conflict, workspace sizing, cross-tile accumulation, einsum semantics, TBuf compiler merge, etc.) plus msSanitizer, msDebug, and msaicerr tool escalation. MUST be invoked before any fix attempt when operator produces wrong output, crashes, hangs, or shows non-deterministic multicore results. Do NOT self-fix without running this skill first. 触发：算子产生错误输出、崩溃、挂死或多核结果不一致时。

Review AscendC kernel code for structural red-line violations (P0–P3) and algorithm correctness. Use after dsl-lowering, before ascendc-evaluation. Outputs rectification report; guides fix → recompile → unit test → system test.

Ascend C 开发资源检索技能。通过本地 API 文档索引、示例代码映射和在线文档兜底搜索定位开发资料，优先查本地、缺失时再查在线。当需要查询 API 用法、示例代码、兼容性信息、官方资料入口或定位文档来源时使用。本技能（含 scripts/ 脚本）源自 cannbot-skills 仓库（gitcode.com/cann/cannbot-skills）的 ops/ascendc-docs-search，重命名而来，为自研代码。

Evolutionary AscendC operator generation — spawn parallel variants with different optimization strategies and select the best. Use as the top-level orchestrator for multi-round kernel optimization. 触发：需要多轮并行进化以优化内核性能时。

Comprehensive review of the entire kernel generation process after all stages complete. Analyzes problems encountered during generation, compile/environment issues, and skill/agent document quality. Produces a REVIEW.md in the kernel output directory. Use when a formal review of the kernel generation process is needed, or when manually triggered by the user.

Diagnose AscendC kernel performance bottlenecks from profiling data (msprof). Matches expert rules first, falls back to LLM analysis. Use when speedup is below target after correctness passes.

Generate initial AscendDSL code (class structure, compute, tiling) from functional PyTorch for pure Vector operators. Use after ascend-call-generation creates the project scaffold.

Translate the operator DSL into AscendC code through multiple passes. Also used when diagnosing compilation errors.

Iteratively optimize AscendDSL code for performance — tiling tuning, vectorization, pipeline adjustments. Use after evaluation shows correctness passes but speedup is below target.

Convert PyTorch nn.Module reference implementation to stateless functional API (module_fn + get_inputs). Use after reference-generation, before ascend-call-generation.

Git 版本管理 - 算子工作区初始化、逐阶段提交、worktree 并行隔离与审计追踪

Generate self-contained interactive HTML dashboard (4 tabs: algo flow, UB tiling, precision, performance) from an AscendC operator output directory. Use when asked to visualize or report operator results. Works for both precision-pass and precision-fail states.

Generate operator description JSON (shapes, dtypes, attributes) from API description or user specification. First step in the cake agent pipeline.

Generate reference PyTorch implementation (nn.Module with get_init_inputs/get_random_inputs) from operator description JSON. Faithfully reproduces the Golden definition when provided. Use after op-desc-generation.

Unified remote NPU development — sync, exec, and test across multiple NPU backends (docker containers, hdspace cloud) via a single Python CLI. Use when: (1) syncing code to any remote NPU server (2) executing build/test commands on remote NPU (3) probing NPU platform info (4) managing multiple remote targets (A2/910B + A3/910C) (5) setting up hdspace tunnel or Docker SSH connections (6) troubleshooting remote compile/run issues (507033, ctypes, rsync)

Track skill invocations, durations, and outcomes during operator generation. Records which skills were called, their inputs/outputs, and correlates with final results. MUST use when: (1) Starting any skill step, (2) Completing any skill step, (3) Final summary to correlate skills with outcomes.

Track and manage task progress via a PROGRESS.md. MUST use when: (1) Starting a new task or operator workflow, (2) Entering or completing any step/stage/substage, (3) After context compaction or reset — read PROGRESS.md first to recover state. Also call when the user asks about current progress or status.