[2026秋季][Task05] jinger861-git

[jinger861-git]

Summary

• Optimize Task05 TriAttentionFallback for Cambricon MLU.
• Fuse QK^T, fp32 stable softmax, and PV into one Triton kernel.
• Read Q/K/V directly from [B, N, S, H, D] strides, avoiding v0 permutation/reshape copies.
• Tune MLU Union1 execution with num_warps=4 and num_stages=4.

Accuracy

PASS with the official KS benchmark.

• atol: 1e-2
• rtol: 1e-2
• Additional shape regression passed for S=64/100/128; worst max absolute difference was 1.44e-3.

Performance

Official benchmark median over three complete runs:

• v0: 0.141894 ms
• v1: 0.013620 ms
• speedup: 10.418x

Changed file

• dlblas/kernels/ks_competition/QY2026_Autumn_jinger861-git_Task05TriAttentionFallback_MLU.py

[CLAassistant]

All committers have signed the CLA.