finalfusion / reductive Goto Github PK

It seems that cluster centroids are not randomized when multiple clustering attempts are performed:

This was probably introduced when switching to rayon for parallel quantization.

Currently, we are computing the distances between each instance and each cluster centroid using a matrix multiplication of both matrices. This requires n_instances x n_centroids memory. When more than one thread is used, we have multiple such matrices in flight, resulting in large memory use.

Alternatively, we could compute the centroid distances per instance. This requires far less memory, but is more CPU intensive.

The reductive API should offer both approaches.

The unit tests currently test many bits and pieces of a quantizer, but a full quantizer (are the bits and pieces used together correctly). Add tests for each quantizer to check whether they work end-to-end.

Due to this lack of coverage, I only found a panic in PR #5 when testing with finalfusion-utils.