Loading...
Loading...
Found 905 Skills
Generate Triton operator requirement documents suitable for Ascend NPU. Used when users need to design new Triton operators, write operator requirement documents, or perform operator performance optimization design.
Debugging and Root Cause Localization for AscendC Operator Precision Issues. Used when operator precision tests fail (such as allclose failure, result deviation, all-zero/NaN output, etc.). Process: Error Distribution Analysis → Code Error-Prone Point Review → Experimental Isolation → printf/DumpTensor Instrumentation → Fix Verification. Keywords: precision debugging, precision issue, result inconsistency, error localization, allclose failure, output deviation, NaN, all-zero, precision debug.
HCCL (Huawei Collective Communication Library) performance testing for Ascend NPU clusters. Use for testing distributed communication bandwidth, verifying HCCL functionality, and benchmarking collective operations like AllReduce, AllGather. Covers MPI installation, multi-node pre-flight checks (SSH/CANN version/NPU health), and production testing workflows.
Create Docker containers for Huawei Ascend NPU development with proper device mappings and volume mounts. Use when setting up Ascend development environments in Docker, running CANN applications in containers, or creating isolated NPU development workspaces. Supports privileged mode (default), basic mode, and full mode with profiling/logging. Auto-detects available NPU devices.
Provides installation guidance for CANN on Ascend NPU. Call this skill when users need to install CANN, configure the Ascend environment, or resolve installation issues.
Analyze Huawei Ascend NPU profiling data to discover hidden performance anomalies and produce a detailed model architecture report reverse-engineered from profiling. Trigger on Ascend profiling traces, NPU bottlenecks, device idle gaps, host-device issues, kernel_details.csv / trace_view.json / op_summary / communication.json. Also trigger on "profiling", "step time", "device bubble", "underfeed", "host bound", "device bound", "AICPU", "wait anchor", "kernel gap", "Ascend performance", "model architecture", "layer structure", "forward pass", "model structure". Runs anomaly discovery (bubble detection, wait-anchor, AICPU exposure) alongside model architecture analysis (layer classification, per-layer sub-structure, communication pipeline). Outputs a separate Markdown architecture report alongside anomaly analysis.
Troubleshoot and optimize the performance of Ascend C operators. This skill is applicable when users develop, review or optimize Ascend C kernel operators, or triggered when users mention keywords such as Ascend C performance optimization, operator optimization, tiling, pipeline, data copy, memory optimization, NPU/Ascend.
Optimize the performance of Triton operators optimized for Ascend NPU. This guide is for users who need to optimize the performance of Triton operators on Ascend NPU, resolve UB overflow, improve Cube unit utilization, and design Tiling strategies.
Maintain JSONL-only profiler performance test cases under csrc/ops/<op>/test in ascend-kernel. Collect data using torch_npu.profiler (with fixed warmup=5 and active=5), aggregate the Total Time(us) from ASCEND_PROFILER_OUTPUT/op_statistic.csv, and output a unified Markdown comparison report (custom operator vs baseline) that includes a DType column. Do not generate perf_cases.json or *_profiler_results.json. Refer to examples/layer_norm_profiler_reference/ for the reference implementation.
Complete toolkit for Huawei Ascend NPU model conversion and end-to-end inference adaptation. Workflow 1 auto-discovers input shapes and parameters from user source code. Workflow 2 exports PyTorch models to ONNX. Workflow 3 converts ONNX to .om via ATC with multi-CANN version support. Workflow 4 adapts the user's full inference pipeline (preprocessing + model + postprocessing) to run end-to-end on NPU. Workflow 5 verifies precision between ONNX and OM outputs. Workflow 6 generates a reproducible README. Supports any standard PyTorch/ONNX model. Use when converting, testing, or deploying models on Ascend AI processors.
Initialize AscendC operator project and create compilable operator skeleton. Trigger scenarios: (1) User requests to create a new operator; (2) Keywords: ascendc operator, new operator, operator directory, operator initialization; (3) Need to quickly implement based on ascend-kernel template. This skill not only creates directories, but also outputs standard files and checklists for "continuous development".
Generate Triton kernel code for Ascend NPU based on operator design documents. Used when users need to implement Triton operator kernels and convert requirement documents into executable code. Core capabilities: (1) Parse requirement documents to confirm computing logic (2) Design tiling partitioning strategy (3) Generate high-performance kernel code (4) Generate test code to verify correctness.