Closed zhanglx13 closed 10 months ago
Make V tensor k-major so that we don't need to suffer from the transposition issue when ds_reading from LDS for V. This PR adds a new version of fused-attention tutorial so that the original one is not "polluted".
ds_read
Performance: T: causal=True F: causal=False
Make V tensor k-major so that we don't need to suffer from the transposition issue when
ds_read
ing from LDS for V.This PR adds a new version of fused-attention tutorial so that the original one is not "polluted".
Performance: T: causal=True F: causal=False