]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
overview / pkg / ggml / sources / llama.cpp / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: d413dca)
model : avoid ggml_cont_3d for fused QKV weights (#15662)
authorGeorgi Gerganov <redacted>
Mon, 8 Sep 2025 07:25:33 +0000 (10:25 +0300)
committerGitHub <redacted>
Mon, 8 Sep 2025 07:25:33 +0000 (10:25 +0300)
commitcf0e3ba1500bd23635b444f64ba23cbdb56c92ef
tree320fe018579a7a1b85a5eebe06fb608adaf2b759tree
parentd413dca00360a7e4cb71441dacecfa32556fcc31commit | diff
model : avoid ggml_cont_3d for fused QKV weights (#15662)

* model : avoid ggml_cont_3d for fused QKV weights

ggml-ci

* kv-cache : make cpy_k and cpy_v implementation more readable

ggml-ci

* cont : add comments

ggml-ci

* cont : minor fix [no ci]

* cont : one more fix

* cont : clarity

ggml-ci

* kv-cache : require contiguous heads of k_cur and v_cur

ggml-ci
src/llama-kv-cache.cpp diff | blob | history
src/llama-kv-cache.h diff | blob | history
src/llama-model.cpp diff | blob | history
Packaging of ggml-org/llama.cpp