silu(gate(x)) * up(x) for M <= 4 rows in one kernel when gate and up are coded
at different k. Same lanes, accumulation order and bf16 SwiGLU boundaries as
fusedGateUpSwiglu; each projection keeps its own word layout. exactBits
applies at the qualified single-row bf16 geometry, where the decoder builds
its value from exact float32 bits: layers pass the choice they resolved at
construction (qualifiedRepresentations), benches and tests the one they compare.
silu(gate(x)) * up(x) for M <= 4 rows in one kernel when gate and up are coded at different k. Same lanes, accumulation order and bf16 SwiGLU boundaries as fusedGateUpSwiglu; each projection keeps its own word layout.
exactBitsapplies at the qualified single-row bf16 geometry, where the decoder builds its value from exact float32 bits: layers pass the choice they resolved at construction (qualifiedRepresentations), benches and tests the one they compare.