=== Transformer layer pipeline: unimem vs standard ===
dim=512 heads=8 seq=128 ffn=1376
>>> STANDARD (Vec + separate allocs) <<<
alloc: 793.875µs
CPU rmsnorm: 34.667µs
AMX qkv: 612.875µs
CPU rope: 30.833µs
AMX attn: 14.458µs
CPU softmax: 6.458µs
AMX attn@V: 7.875µs
AMX ffn up: 669.125µs
CPU silu: 78.583µs
AMX ffn down: 316.875µs
ANE compile: 41.440791ms
ANE run: 231.584µs
TOTAL: 46.979541ms
pass: true
>>> CYB-MEM (tape, zero-copy) <<<
alloc (tape): 908.625µs (14.7 MB used / 15.7 MB cap)
CPU rmsnorm: 11.333µs
AMX qkv: 307.416µs
CPU rope: 17.167µs
AMX attn: 7.833µs
CPU softmax: 6.75µs
AMX attn@V: 5.042µs
AMX ffn up: 423.166µs
CPU silu: 81.167µs
AMX ffn down: 281.709µs
ANE compile: 17.151209ms
ANE run: 180.833µs
TOTAL: 19.985084ms
pass: true
=== Comparison ===
Stage Standard unimem Speedup
----- -------- ------- -------
alloc + fill 793.9µs 908.6µs 0.9x
CPU: rmsnorm 34.7µs 11.3µs 3.1x
AMX: qkv projection 612.9µs 307.4µs 2.0x
CPU: rope 30.8µs 17.2µs 1.8x
AMX: attention matmul 14.5µs 7.8µs 1.8x
CPU: softmax 6.5µs 6.8µs 1.0x
AMX: attn @ V 7.9µs 5.0µs 1.6x
AMX: ffn up+gate 669.1µs 423.2µs 1.6x
CPU: silu 78.6µs 81.2µs 1.0x
AMX: ffn down 316.9µs 281.7µs 1.1x
ANE: compile+load 41.4ms 17.2ms 2.4x
ANE: matmul run 231.6µs 180.8µs 1.3x
TOTAL 47.0ms 20.0ms 2.4x
reference verified: true
unimem verified: true