Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ggml: add ggml_rope_set_offset (+ metal support) (#27120)

Browse Source 
* add params

* cpu kernel

* metal kernel

* add test backend ops

* gate other backends

* ggml: (cuda) support ggml_rope_set_offset (#27121)

* rm cuda supports_op guard, fix webgpu clang-format

* ggml: support ggml_rope_set_offset on vulkan (#27344)

* ggml: support ggml_rope_set_offset on vulkan

* remove inplace optimization
This commit is contained in:
Xuan-Son Nguyen
2026-08-19 14:04:57 +02:00
committed by GitHub
parent 77acca437f
commit fe8156f789
19 changed files with 272 additions and 116 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-et/ggml-et.cpp
+3 -1
View File
@@ -1061,9 +1061,11 @@ static bool ggml_backend_et_device_supports_op(ggml_backend_dev_t dev, const ggm
const bool zero_view_offset = op->src[0]->view_src == nullptr || op->src[0]->view_offs == 0;
const bool has_sections = ggml_get_op_params_i32(op, 11) > 0 || ggml_get_op_params_i32(op, 12) > 0 ||
ggml_get_op_params_i32(op, 13) > 0;
// FIXME: support ggml_rope_set_offset
const bool zero_rot_offset = ggml_get_op_params_i32(op, 15) == 0;
supported =
zero_view_offset && ndims <= 512 &&
zero_view_offset && zero_rot_offset && ndims <= 512 &&
(is_normal || (is_neox && ndims % 16 == 0) || (is_imrope && ndims % 16 == 0 && has_sections));
} else {
supported = false;