Add Metal 4 M5 prefill optimizations#15

Draft

ivanfioravanti wants to merge 16 commits into

antirez:mainfrom

ivanfioravanti:codex/metal4-m5-scaffold

Commits on May 13, 2026

Add Metal 4 M5 scaffold
ivanfioravanti
committed
Improve Metal MPP diagnostics and safe defaults
ivanfioravanti
committed
Tune Metal MPP defaults and thinking checkpoints
ivanfioravanti
committed
Improve Metal MPP prefill throughput
ivanfioravanti
committed
Add low-power Metal MPP Q8 profile
ivanfioravanti
committed
Add M5 Max drift-patch macro plumbing and --dump-logits tooling

ivanfioravanti
and
claude
committed
Stabilize HC mixer sigmoid behind DS4_METAL_HC_STABLE (default on)

ivanfioravanti
and
claude
committed
Unify RMSNorm scale formula behind DS4_METAL_NORM_RSQRT_DISABLE (default on)

ivanfioravanti
and
claude
committed
Add diagnostic DS4_METAL_KV_RAW_F32 to skip FP16 KV round-trip

ivanfioravanti
and
claude
committed
Add diagnostic DS4_METAL_ROPE_EXP2_LOG2 RoPE angle path

ivanfioravanti
and
claude
committed
Fix DS4_METAL_TENSOR_MATMUL_DISABLE host dispatch

ivanfioravanti
and
claude
committed
Default Metal Tensor Q8_0 matmul OFF on M5 Max

ivanfioravanti
and
claude
committed
Add DS4_METAL_MATH_SAFE diagnostic to pin shader library to IEEE-754

ivanfioravanti
and
claude
committed
Fix: F16 compressor Tensor matmul incorrectly coupled to Q8 default

ivanfioravanti
and
claude
committed
Fix Q8 MPP kernel test: reference must take the legacy path

ivanfioravanti
and
claude
committed
Update README to match new M5 Tensor defaults and refreshed drift numbers

ivanfioravanti
and
claude
committed