Skip to content

[KVCache] Fix the aux data syncing order of paged KV cache #16988

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

Merged
tqchen merged 1 commit into from
May 13, 2024
Merged

[KVCache] Fix the aux data syncing order of paged KV cache #16988

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
21 changes: 11 additions & 10 deletions src/runtime/relax_vm/paged_kv_cache.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1709,24 +1709,28 @@ class PagedAttentionKVCacheObj : public AttentionKVCacheObj {
// - Reset the copy.
aux_data_manager_->ResetCopy();

// 1. qo_indptr_on_depths
// 1. q_rope_position_map
// q_rope_position_map has to be synced first so that it has a 0 byte offset
ICHECK_EQ(q_rope_position_map_host_.size(), total_append_length);
q_rope_position_map_view_ = aux_data_manager_->CopyQRoPEPosMapAsync(&q_rope_position_map_host_);
// 2. qo_indptr_on_depths
for (int d = 0; d < num_depths_; ++d) {
qo_indptr_on_depths_view_[d] =
aux_data_manager_->CopyQOIndptrOnDepthAsync(&qo_indptr_on_depths_host_[d], d);
}
// 2. page_indptr_on_depths
// 3. page_indptr_on_depths
for (int d = 0; d < num_depths_; ++d) {
ICHECK_EQ(page_indptr_on_depths_host_[d].size(), qo_indptr_on_depths_host_[d].size());
page_indptr_on_depths_view_[d] =
aux_data_manager_->CopyPageIndptrOnDepthAsync(&page_indptr_on_depths_host_[d], d);
}
// 3. page_indices_on_depths
// 4. page_indices_on_depths
for (int d = 0; d < num_depths_; ++d) {
ICHECK_EQ(page_indices_on_depths_host_[d].size(), page_indptr_on_depths_host_[d].back());
page_indices_on_depths_view_[d] =
aux_data_manager_->CopyPageIndicesOnDepthAsync(&page_indices_on_depths_host_[d], d);
}
// 4. length_info_on_depths
// 5. length_info_on_depths
// last_page_len_on_depths_host_;
// sliding_window_offset_on_depths_host_;
// sink_size_on_depths_host_;
Expand All @@ -1746,23 +1750,20 @@ class PagedAttentionKVCacheObj : public AttentionKVCacheObj {
&sink_size_on_depths_host_[d], d);
}
}
// 5. k_rope_pos_offset_on_depths
// 6. k_rope_pos_offset_on_depths
for (int d = 0; d < num_depths_; ++d) {
ICHECK_EQ(k_rope_pos_offset_on_depths_host_[d].size() + 1,
qo_indptr_on_depths_host_[d].size());
k_rope_pos_offset_view_[d] = aux_data_manager_->CopyKRoPEPosOffsetOnDepthAsync(
&k_rope_pos_offset_on_depths_host_[d], d);
}
// 6. cur_append_lengths_indptr
// 7. cur_append_lengths_indptr
cur_append_length_indptr_view_ =
aux_data_manager_->CopyCurAppendLengthIndptrAsync(&cur_append_lengths_indptr_host_);
// 7. k_ragged_rope_pos_offset
// 8. k_ragged_rope_pos_offset
ICHECK_EQ(k_ragged_rope_pos_offset_host_.size(), num_sequences);
k_ragged_rope_pos_offset_view_ =
aux_data_manager_->CopyKRaggedRoPEPosOffsetAsync(&k_ragged_rope_pos_offset_host_);
// 8. q_rope_position_map
ICHECK_EQ(q_rope_position_map_host_.size(), total_append_length);
q_rope_position_map_view_ = aux_data_manager_->CopyQRoPEPosMapAsync(&q_rope_position_map_host_);
// 9. append_position_map
append_position_map_view_ =
aux_data_manager_->CopyAppendPositionMapAsync(&append_position_map_host_);
Expand Down