[BUG] cuVS 25.12 failed to build index with ITERATIVE_CAGRA_SEARCH and small intermediate_graph_degree

[lanstonchu]

Describe the bug
We have tried tup upgraded the cuVS version from v25.10 to v25.12, but we saw lots of segmentation fault when we tried to build index that we haven't seen before in v25.10. For example, if we used below build parameters, we saw an seg-fault error at the 4th iteration (see attached error message).

We also tried larger intermediate_graph_degree, saying 32 and 64, but still see error, and only won't see error when intermediate_graph_degree is increased to 128. We also see error in other algorithm, but we saw errors more often at ITERATIVE_CAGRA_SEARCH then other algorithms.

Steps/Code to reproduce bug

cuDA version: 13.1
cuVS version: 25.10
metric: IP
build_algo: ITERATIVE_CAGRA_SEARCH
intermediate_graph_degree: 16
graph_degree: 16

Expected behavior

• index build without error in cuVS v25.10, so we expect cuVS v25.12 to behave same

Environment details (please complete the following information):

• cuDA version: 13.1
• cuVS version: 25.10
• We are installing cuDA/cuVS using Conda;
• The implementation mainly relies on cuVS’s C API (e.g., cuvsCagraBuild, cuvsCagraSearch, cuvsResourcesCreate, etc.), using opaque handle types such as cuvsCagraIndex_t. However, it also bypasses the C API abstraction by casting the internal handle (index->addr) back to the underlying C++ type (cuvs::neighbors::cagra::index<float, uint32_t>) and invoking a C++ member function (size()). Because it uses C++-specific constructs such as reinterpret_cast, namespaces (cuvs::neighbors::cagra), templates, and auto, the file must be compiled with a C++ compiler, even though most of its cuVS interactions go through the C API.

Additional context
error_message.txt

[tfeher]

Thank you @lanstonchu for reporting! Could you share the dataset shape (rows, dims) and data type so that we can reproduce?

Tagging @irina-resh-nvda to fix this. The error is likely related to internal alignment requirement not satisfied by the dataset dimension. The search kernel would run into the error and the subsequent memcpy op will report the error. We should check for this and add padding if necessary.

We also see error in other algorithm

It is not expected that other algorithms run into the same problem. @lanstonchu could you share the error message that. you get whet you use other algorithms?

[lanstonchu]

rows = 10000 and dims = 1024

We have already uninstalled the erratic version so I don't have error messages of other algorithms on hand. But their problems are not exactly same as this one (some of them are not seg-fault, but just returning invalid neighbours) so they are not same as the issue of this ticket.