The bs32 variants existed because pre-fix the pipeline required
kBlockN <= page_size, so page_size=32 forced a kBlockN=32 kernel
family. The multi-page-tile fix (commit 473869aba) lifted that
constraint and made kBlockN compile-time-independent of the runtime
page size, so the bs32 family is now redundant: every non-bs32 variant
is correct for any page_size.
This was validated in advance by forcing use_bs32=false in the
dispatcher and running the full correctness suite -- 236/240, identical
to baseline (the 4 remaining failures are the pre-existing int32-
overflow case, orthogonal).
Removes:
* 16 instances/unified_attention_*_bs32_*.cpp files
* unified_attention_decode_bs32_kernel_traits in unified_attention_impl.hpp
* 3 _BS32 dispatch macros in unified_attention.cpp
* 3 _p32 entries from the KernelVariant enum
* 3 dispatch_*_p32 helper functions and their switch cases
* the page_blk_size branch in select_config (now a pure tile-tier ladder)
Net: 12 fewer compile units (build time -6s on JIT), 78 fewer dispatcher
lines, and "which kernel runs?" is now driven purely by Q-tile shape.
Co-authored-by: Cursor <cursoragent@cursor.com>
5bd8f73a28