WangWindow
Community@WangWindow · China
Life beyond the screen matters.
Agent Skills by WangWindow
Showing 99 vetted skills indexed across 1 GitHub repositories.
WangWindowevolution-knowledge
Provides domain knowledge for AscendC kernel optimization on Ascend 910B hardware.
WangWindowtilelang-env-check
Validates TileLang-Ascend environment configuration and auto-repairs submodules, builds, and environment variables.
WangWindowdsl-baseline-generation
Generate initial AscendDSL kernel code from functional PyTorch for NPU vector operators.
WangWindowascendc-st-design
Generates L0/L1/L2 system test cases for Ascend C operators from aclnn interface documentation.
WangWindowtorch-custom-ops-guide
Guides registration of custom PyTorch operators for npugraph_ex graph mode compilation.
WangWindowtriton-op-coding
Generates and iteratively fixes Triton Ascend NPU kernel code from operator task descriptions.
WangWindowaog-prior-art-verify
Scan, stage, build, and classify prior-art arch35 operator candidates with digest-bound provenance.
WangWindowascendc-crash-debug
Diagnose Ascend C operator hangs, crashes, and memory errors on NPU hardware.
WangWindowpypto-precision-compare
Diagnose PyPTO operator precision issues using tensor graph verification, Pass checks, and binary search.
WangWindowpypto-precision-debug
Diagnose PyPTO operator precision failures through syntax checks and prioritized workaround attempts.
WangWindowtriton-npu-convert
Convert PyTorch operators into Triton Ascend NPU kernel-backed operators with validated correctness.
WangWindowgitcode-issue-gen
Creates GitCode Issues from PR changes or manual descriptions with template selection and bidirectional linking.
WangWindowops-evaluation
Builds, installs, and benchmarks AscendC operators from ops repositories against baseline versions.
WangWindowknowledge-issue-report
Generates and validates GitCode issue submission materials for cannbot-knowledge knowledge base feedback.
WangWindowtilelang-programming-model-guide
Guides selection and configuration of TileLang Ascend Developer and Expert programming modes.
WangWindowop-dashboard
Generate self-contained interactive HTML dashboards from AscendC operator output directories.
WangWindowcatlass-op-develop
Generate CATLASS kernel code for Ascend NPU operators from design selections.
WangWindowskill-trace
Records skill invocations, durations, and outcomes in JSON trace files during operator generation.
WangWindowcannbot-skill-reviewer
Reviews CANNBot SKILL.md submissions against repository gates and nine-dimension quality scoring.
WangWindowascendc-performance-best-practices
Provides performance optimization guides and template code for Ascend C operator families.
WangWindowmodel-infer-prefetch
Adds torch_npu.npu_prefetch weight prefetching to NPU models to overlap memory-bound MatMul weight transfers with computation.
WangWindowtriton-latency-optimizer
Optimizes Triton kernel latency on Ascend NPU through sequential single-point optimization passes.
WangWindowmodel-train-oom-analysis
Diagnose NPU out-of-memory failures in PyTorch training via log classification, static estimation, and snapshot analysis.
WangWindowascendc-blaze-best-practice
Guides development of MatMul-class operators on Ascend 950 NPU using Blaze and tensor_api.
Frequently Asked Questions About WangWindow
FAQPage SchemaWhat tasks can I accomplish using WangWindow's skills?▼
You can develop AscendC custom operators end-to-end (spec, design, Tiling, coding, UT/ST testing), convert PyTorch operators to Triton/TileLang/PyPTO kernels, adapt LLM inference with KVCache, quantization, graph mode, and TP/EP/DP parallelism, plus debug precision, crash, and performance issues on Ascend NPUs.
Who are these skills designed for?▼
They target NPU kernel engineers, AscendC/CANN operator developers, and PyTorch model inference engineers working on Huawei Ascend chips (910B/910C/950, arch22/arch35), including those handling SHMEM communication operators, CATLASS matmul kernels, and model deployment baselines.
How do the operator development skills run in practice?▼
Skills are invoked by trigger keywords or explicit slash commands (e.g., /aog-perf-eval {output_dir}), orchestrated through staged checkpoints (CP0-CP5) covering environment check, requirement confirmation, design, testing, performance acceptance, and code review, with PROGRESS.md tracking state across sessions.
Are WangWindow's skills open source and what do they cost?▼
Several skills declare the CANN-2.0 license (e.g., ge-stream-log-analysis, gitcode handlers, ascendc-sync-audit), indicating open-source availability at no cost. The registry itself is publicly hosted under the WangWindow account with 73 public repositories.
What prerequisites and dependencies do these skills require?▼
Most skills require an Ascend NPU environment with CANN Toolkit installed, torch_npu for PyTorch integration, and tools like npu-smi, msprof, msnpureport, and msaicerr for diagnostics. Some support simulator-only runs (npusim) without physical NPU hardware, and remote development via Docker or hdspace backends.