Track conceptual kernel deduplication PRs and binary size savings #2888

Residual closure for the SOL comparison

The frozen integration of all 40 open/draft cuVS swarm heads on the bb8251f5282f7e3755b42de4aaa1389cd076bd54 review basis retained 93 duplicate specializations per profile. Two additional signed draft PRs close that ownership gap while preserving one concrete owning instantiation per generated TU:

• #3014: k-means fit and centroid helpers, 24 generated owners; removes the remaining 92 k-means duplicate specializations.
• #3015: shared IVF-flat/IVF-binary build-index launchers, nine generated owners; removes the remaining shared build-index specialization.

With both heads integrated, the freshly linked CUDA 13.3.73 library has 4,581 unique kernels and 4,581 total copies on every one of the seven native profiles and compute_120 PTX: zero duplicate copies. Required kernel identities and public cuVS exports are preserved. Every new owner module matches the linked library, and all SM90a kernel text/per-kernel constant payload sets match the pre-fix library.

All 70 focused native H200 CPU-oracle cases pass: 48 SG/MG k-means cases and 22 IVF/refinement cases. MG uses NCCL_P2P_DISABLE=1 after a standalone direct-P2P NCCL sum hung on this host. A cross-index-width shared-clique sequence also hangs in the pre-fix library; the remaining combinations pass with fresh cliques/processes. These are bounded native checks, not universal backend or runtime-path coverage.

No standalone compile-time or size saving is assigned to these new PRs. The matched baseline-versus-SOL study remains pending; its intended variable is grouping the same already-deduplicated instantiations. The older measurements and 96-residual census below retain their historical frozen basis and do not describe this new zero-duplicate library.

The swarm now uses one concrete owning instantiation per generated CUDA translation unit, through ordinary JSON matrices and the existing generate_inst_matrix. The revised signed graph has been published: 33 layout PRs updated, with 119 new signed commits, followed by signed review fixes for #2896 and #2937 (121 new signed commits total). The original 42 branch refs advanced atomically; signed validation fixes and the affected dependent branch followed. Existing PR dependencies are preserved.

Launcher headers provide declarations. Host definitions and real device kernels compile in CUDA owners. Shared lower kernel instantiations have separate owners. Includes stay at the top of files. No unity/GROUP build, .cu inclusion, RDC, new production CMake/Python module, or local RAFT patch/pin is used. RAFT and CUB remain separate concepts.

The combined validation source has 1,215 generated owners plus 43 ordinary CUDA owners. All 184 native templates are registered; each generated TU contains exactly one concrete owning instantiation. Actual project CMake configurations pass with the original full architecture lists for CUDA 12.9 and 13.3: 3,264 and 3,295 library objects respectively. Every exact PR commit tree passes native generation checks. The complete SM90 owner census finds 2,595 unique kernel names with zero duplicate copies. Full-library and runtime results are reported below.

Revised complete-library measurements

Paired full-architecture builds on the frozen 825e7438174fce3e998a9f0fb33b6a40dfcc7108 compile-time-experiment source basis, with identical Release/GCC13/compression settings. These controls differ from the older accepted per-PR measurement basis; the figures must not be mixed. Sizes are decimal MB after strip --strip-unneeded. The revised aggregate includes the approved header changes from RAFT #3189, #3187 and #3193 and the #2896 IVF-binary histogram review follow-up. The preceding v12 source and artifacts remain immutable.

Compared with the grouped-owner implementation on this same basis (includes the new owner/closure fixes as well as the TU split):

The grouped frozen control has 16 unresolved cuVS helper references. The revised set also completes those owners; this net comparison does not isolate CUDA-container overhead.

Complete-library native/PTX census (distinct from the 1,258-owner SM90 proof):

The owner set has zero duplicate copies. Any remaining full-library families are reported explicitly above; no claim is made that the entire library has zero duplicates. The per-PR remeasurements for the 34 changed rows remain pending; the combined delta cannot be allocated by summing owner objects.

The single remaining cuRAND table-bearing module is the existing IVF-RaBitQ random-scaling owner, which uses XORWOW. It carries eight arrays totaling 218,048 uncompressed native bytes per architecture profile. All 32 split RNG dispatch owners remain free of XORWOW/MRG implementation tables.

The #2896 review follow-up removes 33 IVF-binary histogram duplicate families from the preceding aggregate, which retained 129. The 96 residual families are 92 k-means helper families, one IVF-binary build-index family, and three PQ half-map families. All were already present in the grouped control. The published #2913 also addresses those three PQ map families; the accepted frozen caller context retains them. This is not evidence that the complete published library has zero duplicates. No new duplicate family is introduced by the revised owners.

The full-architecture owner census has 2,601 unique kernels per profile and zero duplicate copies. The standalone SM90 owner census has 2,595: the architecture lists select five extra half-radix fallback kernels and the MST fallback contributes one net extra kernel. The difference is explained by dispatch selection, with no unexplained kernel loss.

Both toolkit libraries link with -Wl,-z,defs, have zero unresolved internal references, and retain every strong export and declared public cuVS API function. Removed dynamic exports are classified weak private implementation, header-only RAFT, or standard-library helpers.

Per-PR measurement basis

The previous grouped-owner aggregate (209.495904 MB saved on CUDA 12.9; 116.015744 MB on CUDA 13.3) is historical and does not measure this new layout. The new paired aggregate is reported above. The changed PRs’ old individual deltas are retired pending fresh per-PR measurements. PR deltas overlap and must not be summed.

The measured rows below retain exact accepted frozen paired reports. All figures are decimal MB of stripped complete libcuvs.so, with identical Release/GCC13/architecture/compression settings on each side. CUDA 12.9 includes SM70/75/80/86/90a/100f/120a/120 and compute_120 PTX; CUDA 13.3 excludes SM70. They are not measurements of current nightly packages.

#2896 has two generated histogram instantiations. Its older isolated savings of 5.085520 / 2.550192 MB (CUDA 12.9 / 13.3) are historical: that accepted frozen source has no IVF-binary caller, so it cannot measure the review follow-up. The current full-PR standalone delta is pending. The freshly linked combined-library increment from this follow-up is 0.336960 / 0.189472 MB, included in the aggregate above. #2913 has 171 generated map instantiations and saves −5.336648 / −5.014024 MB: real standalone growth. Its extra container/host metadata cost is included. One newer-main bitwise-decode owner uses its exact signed current dependency context; the other 170 use the accepted frozen context.

Measured PRs, descending CUDA 13.3 savings

Status: 🟣 Merged; 🟢 Open for review; 📝 Draft. The owner column reports generated and retained static CUDA sources separately.

Changed PRs awaiting per-instantiation remeasurement

These rows show the published diffs and real prerequisites. Pending entries are not estimates; the former grouped-owner deltas no longer apply. They will be sorted by descending savings when remeasured.

RAFT work

• 🟣 Merged — RAFT#3189: narrow RNG headers to prevent unused cuRAND table leaks. The accepted isolated frozen result, measured after cuVS Narrow k-means and ScaNN serialization includes #2934, saves 42.401792 / 29.995008 MB. That baseline is not a Git dependency. No production RAFT TU is added.
• 📝 Draft — RAFT#3193: inline reduction and linewise launch hooks used by cuVS [WIP] Share RAFT norm, reduction and PCA helper kernels #2898 and [WIP] Share RAFT matrix-broadcast kernels #2914. Three existing RAFT headers change; RAFT stays header-only. cuVS owns the generated CUDA TUs. Its effect is included only in the revised combined analysis, not assigned an unmeasured standalone delta.
• 📝 Draft — RAFT#3187: independent PCA/TSVD header structuring. Its exact approved header changes are included in the revised aggregate validation context, alongside #3189 and #3193. Earlier standalone grouped-owner figures remain historical; no new isolated delta is assigned.

Review and validation scope

Both outstanding swarm reviews are addressed in signed follow-ups: #2896 routes IVF-binary histogram calls through the existing shared owner, and #2937 adds the forced device-sort trustworthiness regression. Both pass pre-commit. The histogram caller and owners freshly compile for every supported architecture/PTX on both toolkits; four public IVF-binary tests pass per toolkit against the complete follow-up library. An actual compiler-dependency audit shows this header affects only the rebuilt IVF-binary caller in the aggregate. The trustworthiness test and exact PR dispatch objects compile and pass on both toolkits; kernel traces confirm CUB segmented radix sorting with zero block-sort launches. No generation or build-system input changes, new dependencies, or stack rebases are required.

The pre-follow-up v12 aggregate passed runtime verification on H200/SM90 for both toolkits: 35 bounded groups: 209 gtests with zero failures/errors/skips, plus all 48 brute-force CPU-oracle cases (float/half, three metrics, all row/column input layouts) and six MST device/managed/pinned cases. All 11 fresh strict executable links pass. The retained C API wrapper came from a different C++ ABI snapshot; rebuilding its 19 objects per toolkit against the final headers resolves that fixture mismatch. Both matching wrappers pass -z,defs, and loader-path checks confirm the final wrappers and libcuvs.so are used. No GPU test failed; the final DSO size/mtime remains unchanged throughout the run.

The current review batch remains #2896, #2937, and #2913. Reviews prioritize prerequisites that unblock the largest savings. Revised #2889, #2928 and #2907 need fresh sizes before their former ranking can be asserted.

The 11 test include-extension errors in #2911 are fixed; its full include checker and all applicable pre-commit hooks pass. No library input changes in that follow-up.

GPU runtime checks use H200/SM90; compilation covers every supported architecture. All32 split RNG dispatch owners have been checked on both toolkits across every native profile and compute_120 PTX: no XORWOW/MRG implementation tables were reintroduced. The complete-library census is reported above, separately from the owner-only census.

Only real dependencies form stacks. The cuVS graph retains 16 open roots targeting main plus the merged #2934 root; multi-parent concepts use the union of their actual providers. The linear GitHub stacks remain #2937 → #2906 → #2927 → #2911 and #2913 → #2914.

Closed redundant host-dispatch relocations: #2903, #2912. The upstream CUB experiment remains closed. No CUB changes are proposed upstream.