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-vulkan/ggml-vulkan.cpp
+7 -1
View File
@@ -1645,6 +1645,7 @@ struct vk_op_rope_push_constants {
uint32_t rope_mode;
uint32_t nrows;
uint32_t n_dims;
uint32_t n_offs;
float freq_scale;
float freq_base;
float ext_factor;
@@ -13144,6 +13145,7 @@ static uint32_t ggml_vk_rms_partials_size(ggml_backend_vk_context * ctx, const g
static vk_op_rope_push_constants ggml_vk_make_rope_constants(const ggml_tensor *dst, const ggml_tensor *src0, const bool has_ff, bool backprop, const uint32_t set_rows_stride) {
const int n_dims = ((const int32_t *) dst->op_params)[1];
const int mode = ((const int32_t *) dst->op_params)[2];
const int n_offs = ((const int32_t *) dst->op_params)[15];
// const int n_ctx = ((const int32_t *) dst->op_params)[3];
const int n_ctx_orig = ((const int32_t *) dst->op_params)[4];
const float freq_base = ((const float *) dst->op_params)[5];
@@ -13173,7 +13175,7 @@ static vk_op_rope_push_constants ggml_vk_make_rope_constants(const ggml_tensor *
uint32_t nb13 = dst->nb[3] / ggml_type_size(dst->type);
vk_op_rope_push_constants rope {
(uint32_t)mode, (uint32_t)ggml_nrows(src0), (uint32_t)n_dims, freq_scale,
(uint32_t)mode, (uint32_t)ggml_nrows(src0), (uint32_t)n_dims, (uint32_t)n_offs, freq_scale,
freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale, has_ff,
{ sections[0], sections[1], sections[2], sections[3] }, is_imrope, backprop, set_rows_stride,
@@ -19219,6 +19221,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
tensor_clone = ggml_rope_ext_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
}
}
const int n_offs = ((int32_t *) tensor->op_params)[15];
if (n_offs != 0) {
tensor_clone = ggml_rope_set_offset(tensor_clone, n_offs);
}
} else if (tensor->op == GGML_OP_UNARY) {
switch (ggml_get_unary_op(tensor)) {
case GGML_UNARY_OP_EXP: