ggml: add PTQ1_0, ternary at group 128 (1.75 bpw, lossless vs PQ2_0) - #148

Open

bri-prism wants to merge 14 commits into

feat/tq1_0-g128

Commits on Sep 3, 2026

ggml: add PTQ1_0, Prism ternary at group 128 (1.75 bpw)
bri-prism
committed
e198192
ggml: wire PTQ1_0 into quantize dispatch, ftype map and row validation
bri-prism
committed
dce9a4b
metal: PTQ1_0 kernels — correct, but 4.2x slower decode than PQ2_0
bri-prism
committed
e5bb8d3
metal: PTQ1_0 decode 29.8 -> 165.3 tok/s, now 0.89x of PQ2_0
bri-prism
committed
745bf55
cuda: PTQ1_0 dequantize and MMVQ paths — NOT COMPILED, no toolchain here
bri-prism
committed
e2ecfa6
metal: tune PTQ1_0 to N_SG=1, and record what optimising it ruled out
bri-prism
committed
9208cdc
metal, cpu: close PTQ1_0 type-list gaps, add the missing mul_mv_ext r1_5
bri-prism
committed
f844a75
tests: verify the CUDA MMVQ dot numerically on the host
bri-prism
committed
6385749
cuda: move ptq1_0_trit to common.cuh so vecdotq.cuh can see it — BUILDS NOW
bri-prism
committed
10462d0
vulkan: PTQ1_0 support, mirrored from Q1_0 not Q2_0 — UNTESTED as committed
bri-prism
committed
4d583a9
cuda: PTQ1_0 MMVQ takes the whole block — measured, and still 6-9x slower
bri-prism
committed
1e29f20
vulkan: fix three build breaks — shared ptq1_0.glsl, skip coopmat2, drop cm2 pipelines
bri-prism
committed
599d0b8
vulkan: make PTQ1_0 reachable, and guard the coopmat2 path
bri-prism
committed
705d9cc
cuda: accelerate PTQ1_0 MMVQ with packed DP4A
bri-prism
committed
d213a38