Skip to content

Extending memory allocator coverage to PDX's maintenance path #20

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
lkuffo merged 1 commit into from
Oct 16, 2026
Merged

Extending memory allocator coverage to PDX's maintenance path #20

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
10 changes: 6 additions & 4 deletions AGENTS.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -182,10 +182,12 @@ Performance-critical — weigh every copy/allocation.
SuperKMeans' k-means) and to new clusters; each `Cluster` keeps it for `Grow`.
`LoadPDXIndex`/`LoadPDXIndexFromStream` take it as a parameter, since loading allocates the clusters.
Routed: cluster `data`/`indices`, the pruner's rotation (an `AllocatedBuffer` behind an `Eigen::Map`),
`IVF::centroid_norms`, `NormalizeAndRotate`, the population gathers, split/merge/reassign scratch,
`ReadRotationMatrix`, `FlatIndex::GetEmbeddings`. On the heap: centroids (SuperKMeans' `Train`
vector), `RowIdClusterMapping`, tombstones, `Flat` storage, per-query searcher buffers, the
single-embedding transform in `Append`, d-sized arrays, `MmapFile` and Eigen's QR internals.
`IVF::centroid_norms`, `NormalizeAndRotate`, the population gathers, split/merge/reassign scratch
(the split's A/B embeddings and ids, sized up front by `CountRowsToSteal`), `BulkAppend`'s row-id
arrays, `ReadRotationMatrix`, `FlatIndex::GetEmbeddings`. On the heap: centroids (SuperKMeans'
`Train` vector), `RowIdClusterMapping`, tombstones, `Flat` storage, per-query searcher buffers, the
single-embedding transform in `Append`, d-sized and per-cluster (k-sized) arrays, the split's
`SplitPartition` positions and k=2 assignments, `MmapFile` and Eigen's QR internals.
`IPDXIndex::GetAllocatedSizeInBytes()` counts the routed buffers an index keeps (cluster data, norms,
an owned pruner's rotation); `GetHeapSizeInBytes()` is the rest of `GetInMemorySizeInBytes()`, for a
caller whose allocator already charges the routed bytes.
Expand Down
Loading
Loading