metal : per-device tuned (Q, NE) for flash-attn vec (replay) - #26570

Merged

ggerganov merged 14 commits into

stack/metal-fa-vec

Aug 24, 2026

Commits on Aug 24, 2026

metal : per-device tuned (Q, NE) for flash-attn vec (#25750)

forforever73
authored and
ggerganov
committed
d73917f
add device id -> token lookup for the offline tuning tool

forforever73
authored and
ggerganov
committed
88f905f
add ggml-metal-tuning skeleton

forforever73
authored and
ggerganov
committed
1f9a360
add op-agnostic perf cell + median timing for the tuner

forforever73
authored and
ggerganov
committed
0a540a5
add FA-vec graph build + tensor init to the tuner

forforever73
authored and
ggerganov
committed
f09bef3
tools : add FA-vec (Q,NE) sweep, compression and table emit

forforever73
authored and
ggerganov
committed
9c99dbc
cool down and re-measure the dirty window on thermal drift

forforever73
authored and
ggerganov
committed
48c64c6
test-backend-ops : replace the FA vec tune mode with a bounded (Q,NE) slice

forforever73
authored and
ggerganov
committed
6c65b7d
tools : document the Metal tuner, point the table comment at it

forforever73
authored and
ggerganov
committed
ea337d0
abort on unknown KV type, single-source fa_vec_legal_ne

forforever73
authored and
ggerganov
committed
f801d5d
cleanup

forforever73
authored and
ggerganov
committed
26bf53d
honor -o in the FA vec (Q,NE) slice

forforever73
authored and
ggerganov
committed
545c66e
retune FA-vec (Q, NE) under a pointwise no-harm gate

forforever73
authored and
ggerganov
committed
7658a24
cont : add fa-vec tunings for M1 Pro, M2 Ultra, M5 Max
ggerganov
committed
efeda76