Releases: lucidrains/native-sparse-attention-pytorch
Releases · lucidrains/native-sparse-attention-pytorch
0.0.77
fixes for triton pathway
0.0.76
causal flag for the transformer and setting correct flag for flex att…
0.0.75
some progress towards non-causal variant
0.0.73
fix
0.0.72
initial forward needs to return cache with rotated keys
0.0.71
update NSA inference so rotated queries and keys are cached
0.0.70
fix some padding issues for gating with importance score
0.0.69
move the gating back onto the selected keys for improved differentiab…
0.0.68
some intermittent issue with flex attention on sample, just disable a…
0.0.66
fix intermittent issue with triton nsa dk