Repository navigation

Add Asymmetric Scalar Hashing (ASH) quantization - #796

Draft

tlwillke wants to merge 254 commits into

mainfrom

ash-dev

Draft

Add Asymmetric Scalar Hashing (ASH) quantization#796
tlwillke wants to merge 254 commits into
mainfrom
ash-dev

This pull request is big! We're only showing the most recent 250 commits

Commits on Dec 27, 2025
- First implementation with fast scalar and SIMD projection encoding.
  tlwillke
  committed
  e91a842
Commits on Dec 30, 2025
Commits on Dec 31, 2025
- New distance calculation benchmark for PQ.
  tlwillke
  committed
  d194319
Commits on Jan 1, 2026
Commits on Jan 3, 2026
Commits on Jan 4, 2026
- Added multi-landmark support to the scalar scorer. Updated DistancesASH to acceptt landmark count.
  tlwillke
  committed
  ae8bf04
Commits on Jan 5, 2026
- Created a pooled landmark-specific query buffer. Still using the non-pooled version for performance purposes though.
  tlwillke
  committed
  92bf98d
- Added debug instrumentation to kmeans++ clusterer.
  tlwillke
  committed
  d903eb6
Commits on Jan 6, 2026
- Bug fixes in SIMD scoreRange for copying queries into contiguous scratch and mixed landmarks.
  tlwillke
  committed
  db5bfbd
Commits on Jan 7, 2026
Commits on Jan 8, 2026
Commits on Jan 12, 2026
- Replaced QR with Thin SVD in runWithinTraining and temporarily overriding payload alignment exception for recall benchmarking purposes.
  tlwillke
  committed
  d8097e1
Commits on Jan 13, 2026
- Changed computeNativeV(X) to computeNativeVt(X) to fix ITQ. Other small fixes. ITQ and multi-landmarks work.
  tlwillke
  committed
  2fad117
Commits on Jan 14, 2026
- Partial implementation of ASH single vector scoring logic SIMD support. Does not work yet.
  tlwillke
  committed
  74b0642
- Initial ASH integration into graph index. Scalar scoring with separated vectors. Recall and perf look acceptable.
  tlwillke
  committed
  a4d35b9
Commits on Feb 14, 2026
- Batch kmeans clustering. Renamed various scorers. Introduced new packed layout with header interleaving. Introduced ashMaskedAdd_512.
  tlwillke
  committed
  e9ee895
Commits on Feb 15, 2026
- Merge latest main into branch.
  tlwillke
  committed
  73a4014
- Additional cleanup for the merge of latest main into branch.
  tlwillke
  committed
  2f4e02d
Commits on Feb 23, 2026
- Merge latest main.
  tlwillke
  committed
  46d6d45
Commits on Mar 3, 2026
- Merged latest main, including logging support. Hierarchical indexing does not work.
  tlwillke
  committed
  6d4cf9e
- Merging latest main into branch.
  tlwillke
  committed
  c80194e
Commits on Mar 5, 2026
- Extend csv logging to include build and search compressor information.
  tlwillke
  committed
  a71359b
- Merge remote-tracking branch 'origin/logging-improvements' into ash-dev
  tlwillke
  committed
  6a1a1b7
Commits on Mar 24, 2026
- Keep ITQ in flat column-major buffers and compute Xbin in row blocks to reduce conversions and memory traffic without changing training behavior. Iterations are ~2.6X faster.
  tlwillke
  committed
  df7cb7e
Commits on Apr 14, 2026
- Latest work on SIMD masked adds for ASH.
  tlwillke
  committed
  7ba8e7e
- Merge latest main into branch.
  tlwillke
  committed
  6cea37c
- Refactored YAML benchmark configs post-rebase. Added an ASH compressionRatio.
  tlwillke
  committed
  e2d9579
Commits on Jun 10, 2026
- Initial implementation of multibit ASH trainer.
  tlwillke
  committed
  b81c149
- Removed Landing algo from DistancesASH.
  tlwillke
  committed
  b03b39b
- Remove Landing trainer from Asymmetric Hashing.
  tlwillke
  committed
  63c113d
- Removed Landing algo selection from Compression.
  tlwillke
  committed
  5ba29d9
Commits on Jun 11, 2026
- Initial implementation of scalar/fallback multibit ASH support.
  tlwillke
  committed
  d5aa241
Commits on Jun 12, 2026
- Initial implementation of Fused ASH.
  tlwillke
  committed
  118ff38
Commits on Sep 11, 2026
Commits on Oct 28, 2026
- Merge latest main into dev branch.
  tlwillke
  committed
  4096cbe
- Stub off ASH compaction for now.
  tlwillke
  committed
  129d4ba
- Minor dataset config change.
  tlwillke
  committed
  49b74c2
- Initial commit for adaptive termination algo.
  tlwillke
  committed
  fd53a1f
- Implementation of Distance Adaptive Beam Search.
  tlwillke
  committed
  f067fb6
- Integrate adaptive termination for ASH graph search
  tlwillke
  committed
  91ebf35
- Enable adaptive termination for ASH construction scores
  tlwillke
  committed
  5da7910
- Parallelize ASH ITQ training encoding
  tlwillke
  committed
  a9d4064
- Use divide-and-conquer SVD for ASH ITQ polar steps
  tlwillke
  committed
  d2d9168
- Set ASH adaptive termination default gamma to 0.01
  tlwillke
  committed
  990c5aa
- Resize ASH quantization workspace for each vector shape
  tlwillke
  committed
  d34afca
- Use 20x ASH ITQ training sample with early stopping
  tlwillke
  committed
  7a08308
- Align ASH benchmark configs at 32x and retain PQ alternatives
  tlwillke
  committed
  a88033a
- Fix ASH SIMD projection gather for partial blocks
  tlwillke
  committed
  3f52d28
- Validate ASH exhaustive recall across scalar and SIMD scorers
  tlwillke
  committed
  1ed906c
- Validate 64-landmark ASH fused graph scoring
  tlwillke
  committed
  32ff4bd
- Align ASH and PQ single-thread throughput benchmarks
  tlwillke
  committed
  f650cdb
- Share ASH query projection in fused graph scorer
  tlwillke
  committed
  fa5b1a3
- Write fused ASH neighborhoods in bulk
  tlwillke
  committed
  3695c3f
- Complete symmetric ASH construction scoring for packed codes
  tlwillke
  committed
  3dba09f
- Benchmark prepared symmetric ASH queries in DistancesASH
  tlwillke
  committed
  406774d
- Compare ASH construction scoring policies and setup costs
  tlwillke
  committed
  fc48335
- Optimize symmetric ASH scoring from encoded vectors
  tlwillke
  committed
  b48f725
- Measure ASH scoring with correct query setup and isolated modes
  tlwillke
  committed
  430819e
- Prevent boxed SIMD allocations during ASH graph pruning
  tlwillke
  committed
  76ff343
- Measure graph throughput over sustained query batches
  tlwillke
  committed
  ffa463d
- Expose symmetric ASH scoring through the diversity API
  tlwillke
  committed
  869052c
- Document symmetric ASH correctness and performance study
  tlwillke
  committed
  ffa6608
- Control graph benchmark query concurrency and verify result parity
  tlwillke
  committed
  c0b4dfb
- Document completed ASH versus PQ Figure 9 graph study
  tlwillke
  committed
  cf8eaa8
- Optimize fused ASH packing with 8x8 byte transposition
  tlwillke
  committed
  04bd54c
- Handle fresh Meson build directories
  tlwillke
  committed
  889eb5a
- Store heap-backed native float vectors directly
  tlwillke
  committed
  87d0b80
- Share prepared ASH neighborhood scoring with graph decoding
  tlwillke
  committed
  2b40957
- Add Highway scoring for complete ASH neighborhoods
  tlwillke
  committed
  1f57e5c
- Benchmark isolated ASH neighborhood full scans
  tlwillke
  committed
  7f6cf1c
- Measure full-scan ASH recall through neighborhood scoring
  tlwillke
  committed
  47d58a0
- Document ASH neighborhood scoring and benchmark contracts
  tlwillke
  committed
  cb02aef
- Measure full-scan PQ recall at exact code byte budgets
  tlwillke
  committed
  e7a4e0f
- Document trained ASH versus PQ recall at equal code sizes
  tlwillke
  committed
  7d088b1
- Clean up ASH rebase dependency and writer integration
  tlwillke
  committed
  2b25871
- Register fused ASH in the version 7 format extension
  tlwillke
  committed
  ca1d485
Commits on Oct 29, 2026
- Normalize graph format extension imports
  tlwillke
  committed
  ae9f956
- Report inclusive and construction-only BenchYAML measurements
  tlwillke
  committed
  f81d2c0
- Revert "Report inclusive and construction-only BenchYAML measurements"
  tlwillke
  committed
  139a8ab
- Use validated divide-and-conquer PCA for ASH training
  tlwillke
  committed
  93a6f98
- Document native PQ versus trained ASH graph results
  tlwillke
  committed
  5c3ae32
- Project dense ASH queries with one Highway dispatch
  tlwillke
  committed
  9ca91ac
- Score symmetric ASH construction codes with Highway
  tlwillke
  committed
  93997f3
- Cover native ASH preparation, exact scoring and compatibility
  tlwillke
  committed
  a10f4f8
- Record BenchYAML results for native ASH preparation
  tlwillke
  committed
  37eb9ae
- Complete native ASH results for both MPNet and OpenAI compression budgets
  tlwillke
  committed
  d7605a9
Commits on Oct 31, 2026
- test: expect dot product in Ada dataset metadata
  tlwillke
  committed
  cc1d119
- docs: escape comparison notation in API comments
  tlwillke
  committed
  d83a23d
- native: add optional sequential oneDNN FP16 projection engine
  tlwillke
  committed
  8c6325b
- ash: default native query projection to FP16 with FP32 fallback
  tlwillke
  committed
  514d45d
- docs: document portable FP16 query projection and validation
  tlwillke
  committed
  800bdd0
- native: add portable FP32 batch projection for ASH
  tlwillke
  committed
  5a8a78c
- ash: use native FP32 projection for single and bulk encoding
  tlwillke
  committed
  cb9164b
Commits on Nov 1, 2026
- docs: document FP32 ASH encoding and encodeAll measurements
  tlwillke
  committed
  7d46fe6
- ash: streamline residual normalization during encoding
  tlwillke
  committed
  e2f9384
- native: prefer oneDNN FP32 for ASH batch projection
  tlwillke
  committed
  1b84ae2
- docs: recommend oneDNN for native ASH installation
  tlwillke
  committed
  6eaec4c
- ash: prepare single-landmark training from one shared sample
  tlwillke
  committed
  a35367b
- docs: fix PQ entropy diagnostic paragraph markup
  tlwillke
  committed
  0851582
Commits on Nov 2, 2026
- ash: reuse FP64 ITQ workspaces and quantize contiguous rows
  tlwillke
  committed
  322e179
- ash: use guarded covariance PCA for training
  tlwillke
  committed
  6436b6e
- ash: accelerate ITQ polar steps and share quantization workers
  tlwillke
  committed
  ac624ab
- native: include oneDNN in default ASH builds
  tlwillke
  committed
  7e99a0c
- Optimize native ASH direct scoring and reuse query scratch
  tlwillke
  committed
  4a9efd1
Commits on Nov 3, 2026
- Fix native ASH projection ownership accounting
  tlwillke
  committed
  ea7357a
- Account retained ASH model, vector, and cache storage
  tlwillke
  committed
  dac1c61
- Add true native FP16 batch projection with portable fallback
  tlwillke
  committed
  d4403b7
- Combine concurrent ASH query projections on caller threads
  tlwillke
  committed
  5d0c7bc
- Add reusable prepared ASH queries for graph search
  tlwillke
  committed
  3553ea4
- Add bounded execution with batched request preparation
  tlwillke
  committed
  d342b86
- Add queued query execution mode to BenchYAML
  tlwillke
  committed
  6aefea3
- Document queued ASH POC and unresolved latency regression
  tlwillke
  committed
  5c95726
Commits on Nov 4, 2026
Commits on Nov 5, 2026
- Add standard MPNet-10M ASH index configuration
  tlwillke
  committed
  e9b1991
- Include construction timings in queued BenchYAML reports
  tlwillke
  committed
  8582de5
- Keep benchmark diagnostics separate from result tables
  tlwillke
  committed
  8cdf18b
- Reuse ASH encoding corrections during construction setup
  tlwillke
  committed
  ed55e1f
- Reuse input-scan means for NVQ reranking
  tlwillke
  committed
  4d71ec8
Commits on Nov 6, 2026
- Add opt-in oneDNN INT8 ASH query projection
  tlwillke
  committed
  31b5d7e
- Default ASH construction to symmetric-all and document configuration
  tlwillke
  committed
  9e7f639
- Expose ASH projection precision in BenchYAML run YAML
  tlwillke
  committed
  2e683e6
- Move BenchYAML ASH precision into index search parameters
  tlwillke
  committed
  4e36c9b
- Add introductory fused ASH graph examples for INT8 and FP16
  tlwillke
  committed
  9f3ebaa
- Demonstrate synchronous and queued ASH searches in hello examples
  tlwillke
  committed
  9aebbef
- Preserve incremental native builds across Maven phases
  tlwillke
  committed
  090faaa
Commits on Nov 7, 2026
- Merge main into ash-dev with modular Grid integration
  tlwillke
  committed
  775ef9e
- Add BenchYAML configuration guide
  tlwillke
  committed
  bac007e
Commits on Nov 8, 2026
- Add BenchYAML schema 3 construction and search options
  tlwillke
  committed
  3f42a42
- Use the JVector API in HelloVectorWorld
  tlwillke
  committed
  3671a37
- Avoid per-call heap segment allocation in symmetric ASH pair kernel
  tlwillke
  committed
  88df291
- Add DistancesFP scoring and allocation benchmark
  tlwillke
  committed
  a2cedea
Commits on Nov 9, 2026
- Group ASH benchmark index recipes and leave saved-index reuse off by default
  tlwillke
  committed
  08c9776
- Organize index parameters into dataset defaults and example pipelines
  tlwillke
  committed
  ba0a9f2
- Use lock-free paged storage for dense graph node maps
  tlwillke
  committed
  8ebde41
- Join pool workers before the TestNeighbors leak check
  tlwillke
  committed
  fddd5f0
- Store ASH codes in compact pages and score them in place
  tlwillke
  committed
  beb274e
- Keep the tuned ASH projection kernel within short query arrays
  tlwillke
  committed
  24ebb57
- Report graph build, flush and reopen times for on-disk builds
  tlwillke
  committed
  2478da8
- Name ESMC recipes after their catalog datasets
  tlwillke
  committed
  8e1aa6d
- Log ASH training through SLF4J instead of standard output
  tlwillke
  committed
  ff4d30d
- Set construction and search workers in the index YAML
  tlwillke
  committed
  28396dd
- Add a build-and-run launcher and IntelliJ run configurations
  tlwillke
  committed
  fe80b2d
- Remove benchmark run records and machine paths
  tlwillke
  committed
  a09f65d
- Default BenchYAML to a few public datasets and key output
  tlwillke
  committed
  955039f
- Remove one-off PQ analysis tools and kernel tuning harnesses
  tlwillke
  committed
  3549f97
- Make the ASH vs PQ study runner portable
  tlwillke
  committed
  4821179
- Summarize the ASH vs PQ results and make reproduction portable
  tlwillke
  committed
  e994c72
- Add an ASH performance page with the ASH vs PQ results
  tlwillke
  committed
  5f2cb46
- Align the ASH and benchmark docs and fix broken links
  tlwillke
  committed
  7582603
Commits on Nov 10, 2026
- Use ASH with main's graph settings in the dataset defaults
  tlwillke
  committed
  dbe74d1
- Keep low-dimensional defaults unquantized and drop plot_output.py references
  tlwillke
  committed
  5bb6a9e
- Keep uncompressed defaults scoring with source vectors
  tlwillke
  committed
  cab432e
- Reject any reranker on full-precision search
  tlwillke
  committed
  a248c9b
- Let the library set BLAS threads for ASH training
  tlwillke
  committed
  918d1ae
- Add one-JVM recipes for every ASH vs PQ case
  tlwillke
  committed
  2ed3205
- Resolve the ASH TODOs
  tlwillke
  committed
  a2d020d
- Test every selectable format version, including 6 with fused PQ and NVQ
  tlwillke
  committed
  564e2d5
- Document on-disk format version 7 in UPGRADING.md
  tlwillke
  committed
  b8525fc
- Write the 4.1.0 release notes for ASH as PR 796 entries
  tlwillke
  committed
  261b1ba
- Accept schema 1 index YAML in BenchYAML
  tlwillke
  committed
  6f5e6b4
- Drop license comments from docs excluded from the license check
  tlwillke
  committed
  1982296
- Explain landmarks and shared projection in the ASH release note
  tlwillke
  committed
  2ec06af
- Declare netlib BLAS and LAPACK where the published jvector artifact inherits them
  tlwillke
  committed
  66c0ef8
- Rewrite the README around getting started, quantizer choice and BenchYAML
  tlwillke
  committed
  c3d4c2c
- Install CMake for oneDNN in Linux CI
  tlwillke
  committed
  1587e76
- Guard the ASH native subprocess test on hosts without the library
  tlwillke
  committed
  b16ffc1
- Include KleidiAI objects in the ARM oneDNN add-on
  tlwillke
  committed
  f29789f
- Avoid the masked gather in Panama ASH projection scoring
  tlwillke
  committed
  3507c3c
- Mark the key ASH APIs as experimental
  tlwillke
  committed
  9b5819a
- Use scalar ASH kernels on JDK 20 and drop masked memory ops from projection
  tlwillke
  committed
  ade1ec9
- Keep the 16-lane ASH neighborhood path off scalable SVE targets
  tlwillke
  committed
  9418c19
- Exclude Meson's subproject lock and package cache from RAT
  tlwillke
  committed
  4fd3af0
- Restore the regression-tests dataset section used by CI benchmarks
  tlwillke
  committed
  14e82bf
- Split the oneDNN add-on into its own Meson project and fetch Highway as a wrap
  tlwillke
  committed
  0681448
- Quote the oneDNN Maven flag so PowerShell passes it intact on Windows
  tlwillke
  committed
  cfb9cdf
- Let the oneDNN FP32 ASH query view report batch projection
  tlwillke
  committed
  2281ff8
- Key cached benchmark graphs by everything that shapes the index file
  tlwillke
  committed
  e3a5ded
Commits on Nov 11, 2026
Commits on Nov 12, 2026
Commits on Nov 13, 2026

Commits

Commits on Dec 27, 2025

Commits on Dec 30, 2025

Commits on Dec 31, 2025

Commits on Jan 1, 2026

Commits on Jan 3, 2026

Commits on Jan 4, 2026

Commits on Jan 5, 2026

Commits on Jan 6, 2026

Commits on Jan 7, 2026

Commits on Jan 8, 2026

Commits on Jan 12, 2026

Commits on Jan 13, 2026

Commits on Jan 14, 2026

Commits on Feb 14, 2026

Commits on Feb 15, 2026

Commits on Feb 23, 2026

Commits on Mar 3, 2026

Commits on Mar 5, 2026

Commits on Mar 24, 2026

Commits on Apr 14, 2026

Commits on Jun 10, 2026

Commits on Jun 11, 2026

Commits on Jun 12, 2026

Commits on Sep 11, 2026

Commits on Oct 28, 2026

Commits on Oct 29, 2026

Commits on Oct 31, 2026

Commits on Nov 1, 2026