]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
sycl: fuse the gated-delta-net state writeback cpy (#26643)
authorTitaniumtown <redacted>
Fri, 14 Aug 2026 06:00:36 +0000 (23:00 -0700)
committerGitHub <redacted>
Fri, 14 Aug 2026 06:00:36 +0000 (09:00 +0300)
commit3d93885352a0049c8388a0da0698ec1a69e60d90
tree035e8e09bb78aa44338f93f10641aa939c8fbb90
parent2bacf9ea5cf9e191430d4c786202d5984df7c193
sycl: fuse the gated-delta-net state writeback cpy (#26643)

Port of https://github.com/ggml-org/llama.cpp/pull/23940.

Arc Pro B70, Qwen 3.6 27B Q4_K - Medium (48 of its 64 blocks run
gated_delta_net), -ngl 99 -fa 1 -ctk f16 -ctv f16 -b 2048 -ub 2048,
interleaved A/B passes of r=3:

  tg128           23.91 / 23.90 / 23.90 -> 24.19 / 24.17 / 24.20   +1.2%
  tg128 (rebuild) 23.81 / 23.81         -> 24.09 / 24.10           +1.2%
  pp2048        1050.8  / 1053.9        -> 1053.8 / 1054.5         flat
  2 seqs, tg128    32.73 / 32.75        ->  33.11 / 33.10          +1.1%
ggml/src/ggml-sycl/gated_delta_net.cpp
ggml/src/ggml-sycl/gated_delta_net.hpp
ggml/src/ggml-sycl/ggml-sycl.cpp