CPU Optimizations - Prefill, Tokenization, and Token Generation - #27032

Draft

lexasub wants to merge 9 commits into

ggml-org:masterfrom

Commits on Aug 14, 2026

improve kv cache performance
lexasub
committed
7e4cb24
Revert "z"
lexasub
committed
6f1152d
tokenizer: very fast tokenizing

lexasub
authored and
lexasub
committed
50ff3e3
improove set_input_kq_mask_impl
lexasub
committed
60ed106
add cache in unicode_regex_split_stl
lexasub
committed
dcadaba
add hardware fp16_to_fp32
lexasub
committed
4461db0
unroll dequantize_row_q4_K
lexasub
committed
e0c6a72
opt1
lexasub
committed
79c2cb9
opt2
lexasub
committed
76c4751