Skip to content

armgit / LEANN

Public Imported from GitHub
0

[MLsys2026 Best Paper]: https://arxiv.org/abs/2506.08276. RAG on Everything with LEANN. Enjoy 97% storage savings while running a fast, accurate, and 100% private RAG application on your personal device.

fix(diskann): warn when dataset has fewer vectors than PQ centroids (#300)

Octopusd83d8a6c5651
diff --git a/packages/leann-backend-diskann/leann_backend_diskann/diskann_backend.py b/packages/leann-backend-diskann/leann_backend_diskann/diskann_backend.py
index 6de63ef..8e97609 100644
--- a/packages/leann-backend-diskann/leann_backend_diskann/diskann_backend.py
+++ b/packages/leann-backend-diskann/leann_backend_diskann/diskann_backend.py
@@ -218,6 +218,21 @@ class DiskannBuilder(LeannBackendBuilderInterface):
logger.warning(f"Converting data to float32, shape: {data.shape}")
data = data.astype(np.float32)
+ # DiskANN uses 256 PQ centroids by default for product quantization.
+ # When num_vectors < 256, MKL's cblas_sgemm receives degenerate matrix dimensions
+ # and emits non-fatal "Parameter N was incorrect" errors to stderr on Windows.
+ # The index still builds correctly, but the noise can be confusing.
+ # See: https://github.com/yichuan-w/LEANN/issues/280
+ _NUM_PQ_CENTROIDS = 256
+ if data.shape[0] < _NUM_PQ_CENTROIDS:
+ logger.warning(
+ f"DiskANN dataset has only {data.shape[0]} vector(s), which is fewer than "
+ f"the default number of PQ centroids ({_NUM_PQ_CENTROIDS}). "
+ f"On Windows with Intel MKL this produces non-fatal 'cblas_sgemm parameter' "
+ f"errors in the output — the index will still build correctly. "
+ f"Consider using '--backend-name hnsw' for small datasets to avoid these messages."
+ )
+
data_filename = f"{index_prefix}_data.bin"
_write_vectors_to_bin(data, index_dir / data_filename)