ggml-meta : split a host-resident KV cache by head - #5

Open

Piggidragon wants to merge 11 commits into

mgpu/host-kv-split-by-head

Commits on Sep 14, 2026

ggml : keep the source name of a scheduler copy
Piggidragon
committed
205cb8d
ggml-meta : split a host-resident KV cache by head
Piggidragon
committed
e1c5abe
llama : split a V tensor on its own head size
Piggidragon
committed
fa56aca
ggml : read back a strided row split from a meta buffer
Piggidragon
committed
5719cbf
llama : keep the recurrent state on device under split mode tensor
Piggidragon
committed
5cdde60
tests : cover a host-resident KV cache split by tensor
Piggidragon
committed
346b38b
ggml-meta : write every stream of a head-split host KV cache
Piggidragon
committed
1f90680
ggml-meta : close the last subgraph on a host-resident cache
Piggidragon
committed
35c6243
ggml-meta : identify a scheduler copy by its name, not by its flags
Piggidragon
committed
20b0891
sched : write a multi-stream window to a meta backend in one copy
Piggidragon
committed
50b379a

Commits on Sep 17, 2026

tests : align the parallel check with the host KV rows
Piggidragon
committed
9faffa9