mfritsche 8adda8448c rknpu2: guard OMP parallel regions with if(M>=8) for decode ...

Experiment: decode runs M=1; opening+joining an OpenMP parallel region per row (A-prep and C-collect loops) is pure fork/join overhead. Guard both pragmas with if(M>=8) so decode runs serially while prefill (M>=300) stays parallel. Byte-identical output. tg128 2.57 -> 2.73-2.80 (+6-9%), tg256 2.65; pp unchanged (35.8/41.6/40.6 for 300/512/700). Zero risk (same math, threading policy only).

2026-07-09 01:25:39 +02:00

..

cmake

ggml: backend-agnostic tensor parallelism (experimental) (#19378)

2026-04-09 16:42:19 +02:00

include

CUDA: lower-case PCI bus id, standardize for ggml (#22820)

2026-05-08 10:09:38 +02:00

src

rknpu2: guard OMP parallel regions with if(M>=8) for decode

2026-07-09 01:25:39 +02:00

.gitignore

vulkan : cmake integration (#8119)

2024-07-13 18:12:39 +02:00

CMakeLists.txt

Merge tag 'b9101' into dspark

2026-07-08 21:00:48 +00:00