mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-29 03:07:02 +00:00
Use partition_index parameter for all get_x_indices_from_distributed_indices() calls
This commit is contained in:
@@ -438,7 +438,9 @@ struct HstuAttentionWithSoftmaxFwdPipelineQRKSVS
|
||||
sweep_tile_span(p_spans[number<0>{}], [&](auto idx0) {
|
||||
sweep_tile_span(p_spans[number<1>{}], [&](auto idx1) {
|
||||
const auto tile_idx = get_x_indices_from_distributed_indices(
|
||||
pcomp_tile.get_tile_distribution(), make_tuple(idx0, idx1));
|
||||
pcomp_tile.get_tile_distribution(),
|
||||
make_tuple(idx0, idx1),
|
||||
partition_index);
|
||||
|
||||
const auto col = seqlen_k_curr + tile_idx.at(number<1>{});
|
||||
constexpr auto i_j_idx = make_tuple(idx0, idx1);
|
||||
|
||||
@@ -447,7 +447,9 @@ struct HstuAttentionWithSoftmaxFwdPipelineQRKSVSTrLoad
|
||||
sweep_tile_span(p_spans[number<0>{}], [&](auto idx0) {
|
||||
sweep_tile_span(p_spans[number<1>{}], [&](auto idx1) {
|
||||
const auto tile_idx = get_x_indices_from_distributed_indices(
|
||||
pcomp_tile.get_tile_distribution(), make_tuple(idx0, idx1));
|
||||
pcomp_tile.get_tile_distribution(),
|
||||
make_tuple(idx0, idx1),
|
||||
partition_index);
|
||||
|
||||
const auto col = seqlen_k_curr + tile_idx.at(number<1>{});
|
||||
constexpr auto i_j_idx = make_tuple(idx0, idx1);
|
||||
|
||||
Reference in New Issue
Block a user