operator-dev

Follow the cannbot-skills chain routed by the skill map: triton-task-extractor, triton-op-designer, triton-op-coding, then triton-latency-optimizer for performance tuning. Apply MindIE-SD-specific guidance in parallel, such as warm/cold benchmark measurement and kernel diff methodology.

Use catlass for CV fusion containing matmul, such as quantized matmul plus activation epilogue, where cube unit gains exist. Use triton for pure vector elementwise fusion like swiglu or gating without matmul, since catlass provides no benefit there.

Custom operator kernel objects are cached by tiling-key hash, so source changes without key changes reuse old binaries. Perform a full clean rebuild, verify with md5 checksums, and use the sentinel method of adding an observable semantic change to confirm the new kernel runs.

Import mindiesd before initializing the NPU or creating tensors, since custom operator registration fails after GE initialization. If the error persists, the operator package has not entered the runtime CANN environment, which is a deployment issue rather than a parameter problem.

Run scripts/install_cannbot.sh on Linux or install_cannbot.ps1 on Windows to clone it to ~/.cannbot-skills. If unavailable, the Skill skips external loading as an availability fallback while repository-specific experience still applies, and the gap is logged for review.

A pattern hit does not guarantee benefit; check kernel form first, since triton codegen memory access may be worse than hand-tuned aclnn kernels. Follow the kernel diff to per-pass AB testing to R1-R5 root cause workflow in the benefit rootcause guide.