]> git.djapps.eu Git - pkg/ggml/sources/whisper.cpp/commit
OpenCL: add initial FA support (llama/14987)
authorrmatif <redacted>
Sat, 16 Aug 2025 08:05:55 +0000 (10:05 +0200)
committerGeorgi Gerganov <redacted>
Mon, 18 Aug 2025 17:30:45 +0000 (20:30 +0300)
commite27f4f205d9e60368ef95893462cb5b7a7cd4fff
tree2af5352921b2485e84ff8c69e5a7181d17e1045d
parent77771b2711cb8fb7af9fdc7b6fb98a04cd55f3e0
OpenCL: add initial FA support (llama/14987)

* add F16/F16 fa support

* fix kernel init

* use mad instead of fma

* use inline function

* mark FA with sinks as unsupported for now

* add pragma unroll to loops
ggml/src/ggml-opencl/CMakeLists.txt
ggml/src/ggml-opencl/ggml-opencl.cpp
ggml/src/ggml-opencl/kernels/flash_attn_f16.cl [new file with mode: 0644]
ggml/src/ggml-opencl/kernels/flash_attn_f32.cl [new file with mode: 0644]
ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl [new file with mode: 0644]