Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[XLA:SPMD] Fix scatter index-parallel partitioning issues. #17243

Merged
merged 1 commit into from
Sep 17, 2024
Merged

[XLA:SPMD] Fix scatter index-parallel partitioning issues. #17243

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 11 additions & 9 deletions xla/hlo/utils/hlo_sharding_util.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2319,7 +2319,6 @@ std::optional<GatherScatterParallelDims> GetGatherScatterBatchParallelDims(
index_parallel_in_dim[i] = -1;
}
}
absl::c_sort(indices_parallel_dims);
if (!indices_parallel_dims.empty()) {
return GatherScatterParallelDims{
indices_parallel_dims, operand_parallel_dims, index_parallel_in_dim};
Expand Down Expand Up @@ -2362,15 +2361,18 @@ GetGatherOutputOrScatterUpdateParallelDims(
int64_t index_vector_dim, absl::Span<const int64_t> offset_or_window_dims) {
absl::InlinedVector<int64_t, 1> output_parallel_dims;
auto indices_parallel_dims = parallel_dim.indices_parallel_dims;
for (int i = 0, idx_dim = 0; i < shape.dimensions_size(); ++i) {
if (absl::c_linear_search(offset_or_window_dims, i)) {
continue;
}
const int index_dim = idx_dim < index_vector_dim ? idx_dim : idx_dim + 1;
if (absl::c_binary_search(indices_parallel_dims, index_dim)) {
output_parallel_dims.push_back(i);
for (int64_t indices_parallel_dim : indices_parallel_dims) {
for (int i = 0, idx_dim = 0; i < shape.dimensions_size(); ++i) {
if (absl::c_linear_search(offset_or_window_dims, i)) {
continue;
}
const int index_dim = idx_dim < index_vector_dim ? idx_dim : idx_dim + 1;
if (indices_parallel_dim == index_dim) {
output_parallel_dims.push_back(i);
break;
}
++idx_dim;
}
++idx_dim;
}
return output_parallel_dims;
}
Expand Down
20 changes: 12 additions & 8 deletions xla/service/spmd/gather_scatter_handler.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -684,11 +684,13 @@ absl::StatusOr<HloInstruction*> PartitionGatherIndexParallelDimensions(
hlo_sharding_util::GroupShardingOnDims(new_indices.sharding(),
indices_parallel_dims);
const GroupedSharding operand_grouped =
hlo_sharding_util::GroupShardingOnDims(operand.sharding(),
operand_parallel_dims);
AlignGroupsWith(hlo_sharding_util::GroupShardingOnDims(
operand.sharding(), operand_parallel_dims),
new_indices_grouped);
const GroupedSharding output_grouped =
hlo_sharding_util::GroupShardingOnDims(gather_output_sharding,
output_parallel_dims);
AlignGroupsWith(hlo_sharding_util::GroupShardingOnDims(
gather_output_sharding, output_parallel_dims),
new_indices_grouped);
PartitionedHlo per_group_operand =
PerGroupPartitionedHlo(operand, operand_grouped, b, clean_ups);
PartitionedHlo per_group_new_indices = PerGroupPartitionedHlo(
Expand Down Expand Up @@ -1130,11 +1132,13 @@ absl::StatusOr<HloInstruction*> PartitionScatterIndexParallelDimensions(
hlo_sharding_util::GroupShardingOnDims(new_indices.sharding(),
indices_parallel_dims);
const GroupedSharding operand_grouped =
hlo_sharding_util::GroupShardingOnDims(operands[0].sharding(),
operand_parallel_dims);
AlignGroupsWith(hlo_sharding_util::GroupShardingOnDims(
operands[0].sharding(), operand_parallel_dims),
new_indices_grouped);
const GroupedSharding update_grouped =
hlo_sharding_util::GroupShardingOnDims(updates[0].sharding(),
update_parallel_dims);
AlignGroupsWith(hlo_sharding_util::GroupShardingOnDims(
updates[0].sharding(), update_parallel_dims),
new_indices_grouped);
const GroupedSharding& output_grouped = operand_grouped;
std::vector<PartitionedHlo> per_group_operands =
PerGroupPartitionedHlos(operands, operand_grouped, b, clean_ups);
Expand Down
Loading