]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
models : dedup Kimi Linear delta net implementation (#19668)
authorymcki <redacted>
Thu, 19 Feb 2026 06:15:17 +0000 (14:15 +0800)
committerGitHub <redacted>
Thu, 19 Feb 2026 06:15:17 +0000 (08:15 +0200)
commitad9f692f8ffd552648bff68e1d8cf30512677ffe
tree21e89f8d0759e521f92e1d6f0adb25b3f2d8bcbd
parent8a70973557ff2605dada63f0dd2c08aa1cfcceb2
models : dedup Kimi Linear delta net implementation (#19668)

* models : add llm_build_delta_net_base

* cont : keep qwen35 and qwen35moe graphs intact

* cont : add comments [no ci]

* add kimi linear to delta-net-base

* removed unnecessary ggml_cont from g_exp_t

* removed ggml_cont from g_diff_exp_t. moved ggml_cont for o to kimi-linear.cpp

* removed unnecessary diag mask

* cont : simplify

* cont : avoid graph splits

* scale q after mul instead of beginning

* scale q after mul instead of beginning

* identical ppl

* cont : fix scale and decay mask

* minor : remove TODO

---------

Co-authored-by: Georgi Gerganov <redacted>
src/models/delta-net-base.cpp
src/models/kimi-linear.cpp
src/models/models.h
src/models/qwen3next.cpp