]> git.djapps.eu Git - pkg/ggml/sources/ggml/commit
Fix the sub group size of Intel (llama/8106)
authorluoyu-intel <redacted>
Tue, 2 Jul 2024 02:16:00 +0000 (02:16 +0000)
committerGeorgi Gerganov <redacted>
Mon, 8 Jul 2024 10:03:28 +0000 (13:03 +0300)
commitc6b5e9abc2ffed028de703f8d82144e500576a7c
tree2440c58c29c09cbb1e948c671d04382487825d09
parentd1bbf97fcf4d5e93e2d6933a6778ed8acac28fed
Fix the sub group size of Intel (llama/8106)

* use warp_size macro for all sycl kernels

* fix mask of permute_sub_group_by_xor

* fix rms_norm with correct warp number

* fix rms_norm_f32/group_norm_f32

* move norm to norm.cpp file

* fix quantize bug

* fix mmvq's batch size
src/CMakeLists.txt
src/ggml-sycl.cpp
src/ggml-sycl/backend.hpp
src/ggml-sycl/common.hpp
src/ggml-sycl/dmmv.cpp
src/ggml-sycl/mmvq.cpp
src/ggml-sycl/norm.cpp [new file with mode: 0644]
src/ggml-sycl/norm.hpp [new file with mode: 0644]
src/ggml-sycl/presets.hpp