]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
sycl: fuse UNARY(silu|sigmoid|softplus) + MUL (#26411)
authorTitaniumtown <redacted>
Thu, 13 Aug 2026 08:41:38 +0000 (01:41 -0700)
committerGitHub <redacted>
Thu, 13 Aug 2026 08:41:38 +0000 (11:41 +0300)
commit1ee1cd9bc65a56ab50e2ed19a48709dc42d1dd9d
tree28269b0cd5ee9cdfb0d004fa083e3f6a06c3faea
parent8efbf65dbd55bc2387d9632ce2f3cd66d33b3813
sycl: fuse UNARY(silu|sigmoid|softplus) + MUL (#26411)

Measured on Arc Pro B70 (Battlemage), Qwen3.6-27B Q4_K_M, -fa on, f16 KV,
-b 2048 -ub 2048, llama-bench -r 3, three interleaved A/B rounds:

  pp2048        1014.70 -> 1018.56 t/s   (+0.38%, within run-to-run spread)
  tg128         23.73 -> 23.86 t/s       (+0.57%)
  tg128 @ d4096 22.71 -> 22.86 t/s       (+0.62%)
ggml/src/ggml-sycl/element_wise.cpp
ggml/src/ggml-sycl/element_wise.hpp
ggml/src/ggml-sycl/fusion.cpp
ggml/src/ggml-sycl/fusion.hpp
ggml/src/ggml-sycl/ggml-sycl.cpp
tests/test-backend-ops.cpp