Why is it much slower than Tilelang?

[zzyplaybasketball]

Hello, may I ask why the operator for the DSA's gemm calculation is much slower than the DSA gemm operator implemented by Tilelang? In a single-card environment, when measuring all the sizes, the speed will be reduced by more than 20%.Logically speaking, the implementation of Cutlass that you have developed should obviously be faster than that of Tilelang, whether it is in terms of load balancing or the use of TMA, etc.

The code of Tilelang： https://github.com/tile-ai/tilelang/blob/main/examples/deepseek_v32/sparse_mla_fwd_pipelined.py

thank you very much.

[zzyplaybasketball]

The device we are using is the single-card H20.

[IISuperluminaLII]

What quantization?