Compare Append Kernel's Results with Xformers about flashinfer

[LiuXiaoxuanPKU]

In the example above, if I deliberately make prefix length '0'
prefix_lens = [random.randint(0, 0) for _ in range(batch_size)]
and I find that the output(output) of flashInfer are all 0s, this is not expected because the append kernel should degrade to pure prefill kernel in that case, and running prefill should output non-zero values.

[LiuXiaoxuanPKU]

Fixed! I did not pass the parameters correctly, the output is similar to xformers.