[mc8max]

Fixes #2856

Summary

CAGRA’s MULTI_CTA and MULTI_KERNEL search plans passed hashmap_size in bytes to hashmap.resize(), which expects an element count. Both plans now pass hashmap_size / sizeof(INDEX_T). For uint32_t indices, this reduces the hashmap allocation to one quarter of its previous size.

Validation

Built with CUDA 12.9 and ran 17 existing CAGRA tests on an NVIDIA RTX PRO 6000 Blackwell Server Edition. All passed with zero failures or skips.

• NEIGHBORS_ANN_CAGRA_BUGS_TEST — 8 passed, including hashmap sizing and fallback cases.
• NEIGHBORS_ANN_CAGRA_FLOAT_UINT32_TEST — 5 passed.
• NEIGHBORS_ANN_CAGRA_HALF_UINT32_TEST — 4 passed.

The float and half cases checked recall and distances for both search algorithms. The allocation reduction follows from the size conversion.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[coderabbitai]

No actionable comments were generated in the recent review.

Summary

Summary by CodeRabbit

• Bug Fixes
  • Corrected memory sizing during multi-CTA search setup. This helps ensure the search hashmap is allocated to the intended size and supports more reliable search behavior.

The multi-CTA and multi-kernel search paths now divide the byte-sized hashmap size by sizeof(INDEX_T) before passing it to hashmap.resize.

Changes: Hashmap allocation sizing

Layer / File(s)	Summary
Convert hashmap size before resize cpp/src/neighbors/detail/cagra/search_multi_cta.cuh, cpp/src/neighbors/detail/cagra/search_multi_kernel.cuh	Both search paths pass the hashmap size as an INDEX_T element count to hashmap.resize, instead of passing the byte count.

Estimated code review effort: 2 (Simple) | ~8 minutes

Change: Bug fix · Severity of issue fixed: Low

Suggested reviewers: achirkin, divyegala, bdice

Merge Risk: Minimal · up to e2660

This change reduces the hashmap allocation in multi-CTA and multi-kernel CAGRA search to the intended size. No concrete merge-blocking risk was identified. The label checker failure only needs the missing labels added.

Architecture risk: Low · up to e2660

The change affects 1 system.

Changed systems: cpp

Architecture concerns
No architecture-level concerns identified.

Review details

Systems and components

• observed — cpp (service) was modified; 2 changed files map to changed impact.

Before / after behavior

• observed — Modified behavior in cpp/src/neighbors/detail/cagra/search_multi_cta.cuh: The initial hashmap allocation now passes hashmap_size / sizeof(INDEX_T) elements to resize instead of passing the byte count directly.
• observed — Modified behavior in cpp/src/neighbors/detail/cagra/search_multi_kernel.cuh: search::set_params now passes hashmap_size / sizeof(INDEX_T) to hashmap.resize instead of passing hashmap_size directly.

Pre-merge checks: 4

Check name	Status	Explanation
Title check		The title names the fix for issue #2856 and the affected search plans (MULTI_CTA and MULTI_KERNEL) and the over-allocated hashmap size. The changeset divides hashmap_size by sizeof(INDEX_T) in both se…
Description check		The description states the bug: MULTI_CTA and MULTI_KERNEL pass a byte count to hashmap.resize(), which expects an element count. It explains the fix (dividing by sizeof(INDEX_T)), the effect on uint3…
Linked Issues check		Issue #2856 requires the multi-CTA and multi-kernel search plans to use element-count units for hashmap sizing, matching the single-CTA plan. At head e2660ce, search_multi_cta.cuh (set_params) calls h…
Out of Scope Changes check		The whole-PR diff (merge base 918c8a3 to head e2660ce) changes only two files: search_multi_cta.cuh and search_multi_kernel.cuh. Each change is a single resize-argument conversion with a comment, and …

Full details: Title check

Explanation

The title names the fix for issue #2856 and the affected search plans (MULTI_CTA and MULTI_KERNEL) and the over-allocated hashmap size. The changeset divides hashmap_size by sizeof(INDEX_T) in both search plans, so the title matches the main change. It is somewhat long and includes the issue reference, but it is specific and accurate.

Full details: Description check

Explanation

The description states the bug: MULTI_CTA and MULTI_KERNEL pass a byte count to hashmap.resize(), which expects an element count. It explains the fix (dividing by sizeof(INDEX_T)), the effect on uint32_t allocation, and the tests run. This directly describes the changeset.

Full details: Linked Issues check

Explanation

Issue #2856 requires the multi-CTA and multi-kernel search plans to use element-count units for hashmap sizing, matching the single-CTA plan. At head e2660ce, search_multi_cta.cuh (set_params) calls hashmap.resize(hashmap_size / sizeof(INDEX_T), ...), and search_multi_kernel.cuh (set_params) makes the same change. Both now convert the byte count stored by calc_hashmap_params to an element count before resize(). The division is exact because hashmap_size is computed as sizeof(INDEX_T) * max_queries * hashmap::get_size(hash_bitlen). The single-CTA plan in search_single_cta.cuh already uses hashmap_element_count(), so the three plans now use consistent units. The PR adds no automated test. The author reports existing CAGRA tests passing, but the issue does not request a new test, so this does not count as a failure.

Full details: Out of Scope Changes check

Explanation

The whole-PR diff (merge base 918c8a3 to head e2660ce) changes only two files: search_multi_cta.cuh and search_multi_kernel.cuh. Each change is a single resize-argument conversion with a comment, and each directly implements the requirement in #2856. No unrelated changes, declaration changes, or refactoring appear in the diff.

Finishing Touches

Generate unit tests (beta)

• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Recent review info

Run configuration

• Source mode: Auto
• Repository: NVIDIA/cuvs
• Branch: main
• File: .coderabbit.yaml
• Commit: 918c8a3
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Enterprise
• Run ID: 82696eed-351c-4d2b-8137-d95066587484

Commits

Reviewing files that changed from the base of the PR and between 918c8a3 and e2660ce.

Files selected for processing: 2

• cpp/src/neighbors/detail/cagra/search_multi_cta.cuh
• cpp/src/neighbors/detail/cagra/search_multi_kernel.cuh

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.

[mc8max]

@irina-resh-nvda can you help check this PR pls?

[irina-resh-nvda]

LGTM

[irina-resh-nvda]

@achirkin Can you take a look and approve if it looks good?