[2026秋季][Task07] jinger861-git

[jinger861-git]

Summary

• Optimize Task07 LayerNorm for Cambricon MLU.
• Implement LayerNorm as one Triton kernel with fp32 mean/variance reductions.
• Use an exact non-power-of-two rows x 10 vector tile for the competition shape.
• Use MLU Libdevice rsqrt for reciprocal standard deviation.
• Fall back to native F.layer_norm for non-contiguous inputs or larger leading dimensions.

Accuracy

PASS with the official KS benchmark.

• atol: 1e-2
• rtol: 1e-2
• Additional shapes (1, 10), (10, 10), (2, 8, 10), and (32, 10) passed with max absolute difference <= 2.39e-7.

Performance

Official benchmark median over three complete runs:

• v0: 0.135740 ms
• v1: 0.013866 ms
• speedup: 9.789x

Changed file

• dlblas/kernels/ks_competition/QY2026_Autumn_jinger861-git_Task07LayerNorm_MLU.py****

[CLAassistant]

All committers have signed the CLA.