Ferrox v0.4.1

Highlights — simdgroup-MMA flash attention at d=128

The d=64 MMA prefill attention kernel from v0.4.0 now covers head_dim
128 as well: its body is parameterised over the head dim, so one macro
emits both gqa_prefill_fa_ext_mma_d64 and _d128. That closes the
last group of dense Metal prefill rows.

Interleaved A/B on an Apple M2 Pro, -p 512 -n 0 -r 3 --ngl 99, two
reps per arm (FERROX_METAL_FA_MMA off vs on):

Model	FA-vec	MMA	speedup	llama.cpp	gap
Qwen3-0.6B Q8_0	1917	3277	1.71×	3510	1.81× → 1.07×
Phi-4-mini Q4_K_M	450	540	1.20×	562	1.24× → 1.04×
Llama-3.2-3B Q4_K_M	523	630	1.20×	563	1.08× → 0.89×
Mistral-7B-v0.2 Q4_K_M	235	266	1.13×	257	1.10× → 0.97×

Gap = llama.cpp / ferrox; under 1 means ferrox is faster. Greedy
output is byte-identical with the kernel on and off.

Caveat on the numbers. These are interleaved A/B measurements, valid
as relative comparisons. The full benchmarks/RESULTS.md ledger was not
regenerated for this release — the benchmark host was above its load bar
throughout — so that table still shows the pre-v0.4.1 Metal figures.

FERROX_METAL_FA_MMA=0 restores the previous kernel (the scalar
fa_ext at d=64, FA-vec at d=128).

Install

curl -fsSL https://raw.githubusercontent.com/antonellof/ferrox/main/scripts/install.sh | bash

Or pin a version:

FERROX_VERSION=v0.4.1 \
  curl -fsSL https://raw.githubusercontent.com/antonellof/ferrox/main/scripts/install.sh | bash

Prebuilt archives

Asset	Notes
ferrox-v0.4.1-darwin-arm64.tar.gz	macOS Apple Silicon, Metal
ferrox-v0.4.1-linux-x86_64.tar.gz	Linux x86_64, CPU

CUDA and other hosts: build from source (--features cuda).

Full Changelog: v0.4.0...v0.4.1