]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
overview / pkg / ggml / sources / llama.cpp / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: bcfd198)
SYCL: improve MoE prefill throughput (#23142)
authorAlexey Kopytko <redacted>
Fri, 22 May 2026 12:50:17 +0000 (21:50 +0900)
committerGitHub <redacted>
Fri, 22 May 2026 12:50:17 +0000 (15:50 +0300)
commitcc9e331213b6a9cb186aabe01a4ec6a61419dd80
treee17cea5e2aa002616af24115778f00575958c1b7tree
parentbcfd1989e9a90af74669d94057ff2468682c3f4acommit | diff
SYCL: improve MoE prefill throughput (#23142)

- change `k_copy_src1_to_contiguous` so that uses a precomputed contiguous mapping where all rows "owned" by an expert are in one slice with a know starts and ends
- switch the `O(n_as * n_routed_rows)` contraption to a counting sort-based procedure with `O(n_as + n_routed_rows)` complexity
ggml/src/ggml-sycl/ggml-sycl.cpp diff | blob | history
Packaging of ggml-org/llama.cpp