CPU performance and backend selection

Squeeze uses different execution paths for different algorithms. A Rust extension being installed does not mean an entire UMAP fit executes in Rust.

UMAP neighbor search

UMAP(use_hnsw=False) selects PyNNDescent for the approximate neighbor path. use_hnsw=True requests the Rust HNSW wrapper. Unsupported metrics or a missing extension can fall back to PyNNDescent with a warning. Small-data and precomputed paths may bypass approximate search entirely.

hnsw_prune_strategy="simple" is the default; hnsw_prune_strategy="robust" selects RobustPrune, with hnsw_alpha=1.2 by default. Measure graph quality and end-to-end fit time before changing these settings. UMAP's subsequent layout is Python/Numba code.

Rust reducers

The Rust reducers use CPU distance, linear algebra, and optimization routines. The new graph methods parallelize exact neighbor construction with Rayon and use runtime-selected SIMD squared-distance kernels where supported. They retain only the nearest edges, but still compare all pairs: graph construction time is quadratic in the sample count. They are not an approximate HNSW search mode.

PCA diagonalizes a feature covariance matrix. MDS, Isomap, LLE, PHATE and the current t-SNE affinity construction can allocate dense matrices. Barnes-Hut repulsion in t-SNE does not eliminate its dense affinity construction.

Measure the complete operation

Use release builds and a controlled thread budget. Record fit time, quality, input shape, backend, and hardware. A distance-kernel speedup may be a small fraction of total runtime. The current benchmarks publish measured results rather than a universal speed multiplier.