]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
overview / pkg / ggml / sources / llama.cpp / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: 082b326)
llama : make all KQ masks f16 if FA is used, remove zero attention bias, remove raw_k...
authorfairydreaming <redacted>
Fri, 10 Jul 2026 07:06:58 +0000 (09:06 +0200)
committerGitHub <redacted>
Fri, 10 Jul 2026 07:06:58 +0000 (09:06 +0200)
commit2ed3c1abbb8e155226b0b2cbeb9e9efad77fbb02
tree6177a70fec2363c9ea1bf806657f9f1210398698tree
parent082b326fc76f6e9bbb835b3920a3022bfdb6691ccommit | diff
llama : make all KQ masks f16 if FA is used, remove zero attention bias, remove raw_k repeats in DeepSeek V4 (#25370)

* llama : make all KQ masks (except the lightning indexer one) f16 if FA is used and remove zero attention bias in DeepSeek V4

* llama : remove dead code that repeats unified raw_k cache for each stream in DeepSeek V4 - no longer needed as raw_k is always non-unified.

---------

Co-authored-by: Stanisław Szymczyk <redacted>
src/llama-graph.cpp diff | blob | history
src/models/deepseek4.cpp diff | blob | history
Packaging of ggml-org/llama.cpp