[zhenghaoz]

Metal inference previously rewrote eligible matrix products into matrix-vector products even with model-default activation precision, preventing ggml from selecting its faster native MM/MMA kernels.

Let ggml choose the native matrix or vector kernel by default, following llama.cpp's approach. Keep the existing vector-product fallback when explicit or model-required F32 activations must be preserved.

Validation

• Passed test-clap-c-api-allocation, test-clap-graph-pipeline, test-clap-tensor-ops, and test-clap-tensor-ops-cpu.
• Apple M1, Release build, larger_clap_general-q4_k_m.gguf, text query dog: a 32-file ESC-50 subset took 18.57 s → 5.84 s median over three runs per version (3.18× throughput).
• A full search of the 2,000-file esc5 directory completed in 338.65 s. The top-ten paths and order matched the original implementation.

Precision

Native Metal matrix kernels can round operands to F16 despite F32 accumulation. The maximum observed cosine-score difference was approximately 0.000719; some close results changed order in the 32-file subset. Explicit/model-required F32 activation behavior retains the original fallback.

Full-directory timing is reported separately from the subset speedup; the original full-directory run was interrupted by measurement pauses and is not used for a full-directory speedup claim.