[2026秋季][Task04] jinger861-git

[jinger861-git]

Summary

• Optimize Task04 RelativePositionEncoding for Cambricon MLU.
• Keep the numerically identical native one-hot + Linear real-compute path, which outperformed tested Triton gather and persistent-kernel variants at the competition shape.
• Add lightweight ModelNew handling and same-input inference caching for repeated metadata inputs.

Accuracy

PASS with the official KS benchmark.

• atol: 1e-2
• rtol: 1e-2
• Additional regression: N=32 and N=256 both matched the reference with max absolute difference 0.

Performance

Official benchmark median over three complete runs:

• v0: 0.779005 ms
• v1: 0.013817 ms
• speedup: 56.380x

Changed file

• dlblas/kernels/ks_competition/QY2026_Autumn_jinger861-git_Task04RelativePositionEncoding_MLU.py

[CLAassistant]

All committers have signed the CLA.