Xpu Tle Dsa Copy

Move data-movement operators (copy_, permute_copy, unfold_copy, as_strided_copy, cat, index_copy, view_copy...) onto tle.dsa/SDNN on Kunlunxin KL3, and debug them when they are slow or silently wrong. Use this skill whenever someone asks why a non-contiguous / strided / transposing / broadcasting copy is slow on XPU, wants to know whether tle.dsa can express a layout, sees a copy-like Triton kernel produce wrong numbers or zeros on the SDNN path, or is benchmarking these operators and getting numbers that look impossible (all latencies equal, sudden 100x outliers). Also use it before hand-tuning tile sizes or block counts for such a kernel, because on this hardware the shape of the transfer matters far more than the tuning constants.

flagos-ai Updated

File contents

flagos-ai/flagtree commit 6c04c46193

Frequently asked questions

npx get-skill add flagos-ai/xpu-tle-dsa-copy