LoRA: Optimise LoKr at runtime - #1233

Merged

leejet merged 16 commits into

leejet:masterfrom

stduhpf:lokr-forward

Feb 7, 2026

Commits on Feb 1, 2026

LoRA: Optimise LoKr at runtime
stduhpf
committed
b4db4be
lokr: fix convs
stduhpf
committed
d608b37
lokr: fix lienar forward for CUDA/HIP and CPU backends
stduhpf
committed
b486097
lokr: disable "optimization" for convolutions
stduhpf
committed
8553862
LoKR: re-implement conv
stduhpf
committed
2430989
lokr: fix conv bypass implementation
stduhpf
committed
fbf401b
lokr: cleanup linear path code
stduhpf
committed
04f9b1f

Commits on Feb 2, 2026

reshape to 2d before mat_mul
stduhpf
committed
5b67c4b
maxComputeWorkGroupCount workaround for vulkan
stduhpf
committed
f7d53b6
Avoid too large tensors dims in matmul for smaller vk workgroups
stduhpf
committed
244480e
make it vk only
stduhpf
committed
1ab9ed2
remove unncesary casts for non-conv weights
stduhpf
committed
c7629d9
fix wrong flag (oops)
stduhpf
committed
30051a2

Commits on Feb 7, 2026

format code
leejet
committed
da67c5b
eliminate the compilation warnings on MSVC
leejet
committed
26d4ddd
with_lora -> with_lora_and_lokr
leejet
committed
de98ba9