[2026秋季][Task01] Foxxiaohuli

[Foxxiaohuli]

使用自己写的测试脚本，对比原始参考实现（PyTorch）输出结果

--- Correctness---

  [PASS] Correctness(N=128, B=64)  max_abs_diff=1.490116e-07
  [PASS] Correctness(N=512, B=128)  max_abs_diff=1.490116e-07

--- Performance ---

   N_TOKEN  NO_BINS     Triton(ms)    PyTorch(ms)    Speedup
        64       32         0.0290         0.2161      7.45x
        64       64         0.0293         0.2215      7.57x
       128       64         0.0322         0.2378      7.38x
       256       64         0.0417         0.2994      7.19x
       256      128         0.0529         0.3713      7.02x
       512       64         0.0768         0.5463      7.11x
       512      128         0.1188         0.8601      7.24x
      1024       64         0.2151         1.5335      7.13x