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

* 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
+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:
@@ -50,19 +50,21 @@ void rope_norm(const uint i0, const uint i1, const uint i2, const uint i3, rope_
}
idst += p.d_offset;
if (i0 >= p.n_dims) {
if (i0 < p.n_offs || i0 >= p.n_offs + p.n_dims) {
rope_data_d[idst + 0] = ROPE_D_TYPE(rope_data_a[ix + 0]);
rope_data_d[idst + 1] = ROPE_D_TYPE(rope_data_a[ix + 1]);
return;
}
const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, i0/2.0f);
const uint iw = i0 - p.n_offs; // relative idx
const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, iw/2.0f);
const float freq_factor = p.has_ff != 0 ? rope_data_ff[iw/2] : 1.0f;
float cos_theta, sin_theta;
rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
rope_yarn(theta_base / freq_factor, iw, cos_theta, sin_theta, p);
const float x0 = float(rope_data_a[ix + 0]);
const float x1 = float(rope_data_a[ix + 1]);
@@ -87,25 +89,28 @@ void rope_neox(const uint i0, const uint i1, const uint i2, const uint i3, rope_
}
idst += p.d_offset;
if (i0 >= p.n_dims) {
if (i0 < p.n_offs || i0 >= p.n_offs + p.n_dims) {
rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
rope_data_d[idst + i0/2 + 1] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 1]);
return;
}
const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, i0/2.0f);
const uint iw = i0 - p.n_offs; // relative idx
const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, iw/2.0f);
const float freq_factor = p.has_ff != 0 ? rope_data_ff[iw/2] : 1.0f;
float cos_theta, sin_theta;
rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
rope_yarn(theta_base / freq_factor, iw, cos_theta, sin_theta, p);
const float x0 = float(rope_data_a[ix + 0]);
const float x1 = float(rope_data_a[ix + p.n_dims/2]);
// idst/ix point at channel i0/2; the first channel of the rotated pair is p.n_offs + iw/2 = i0/2 + p.n_offs/2
const float x0 = float(rope_data_a[ix + p.n_offs/2 + 0]);
const float x1 = float(rope_data_a[ix + p.n_offs/2 + p.n_dims/2]);
rope_data_d[idst + 0] = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
rope_data_d[idst + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
rope_data_d[idst + p.n_offs/2 + 0] = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
rope_data_d[idst + p.n_offs/2 + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
}
@@ -125,53 +130,56 @@ void rope_multi(const uint i0, const uint i1, const uint i2, const uint i3, rope
}
idst += p.d_offset;
if (i0 >= p.n_dims) {
if (i0 < p.n_offs || i0 >= p.n_offs + p.n_dims) {
rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
rope_data_d[idst + i0/2 + 1] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 1]);
return;
}
const uint iw = i0 - p.n_offs; // relative idx
const int sect_dims = p.sections[0] + p.sections[1] + p.sections[2] + p.sections[3];
const int sec_w = p.sections[1] + p.sections[0];
const uint sector = (i0 / 2) % sect_dims;
const uint sector = (iw / 2) % sect_dims;
float theta_base = 0.0;
if (p.is_imrope != 0) {
if (sector % 3 == 1 && sector < 3 * p.sections[1]) {
theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, i0/2.0f);
theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, iw/2.0f);
} else if (sector % 3 == 2 && sector < 3 * p.sections[2]) {
theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, i0/2.0f);
theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, iw/2.0f);
} else if (sector % 3 == 0 && sector < 3 * p.sections[0]) {
theta_base = rope_data_pos[i2]*pow(p.theta_scale, i0/2.0f);
theta_base = rope_data_pos[i2]*pow(p.theta_scale, iw/2.0f);
} else {
theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, i0/2.0f);
theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, iw/2.0f);
}
} else {
if (sector < p.sections[0]) {
theta_base = rope_data_pos[i2]*pow(p.theta_scale, i0/2.0f);
theta_base = rope_data_pos[i2]*pow(p.theta_scale, iw/2.0f);
}
else if (sector >= p.sections[0] && sector < sec_w) {
theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, i0/2.0f);
theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, iw/2.0f);
}
else if (sector >= sec_w && sector < sec_w + p.sections[2]) {
theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, i0/2.0f);
theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, iw/2.0f);
}
else if (sector >= sec_w + p.sections[2]) {
theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, i0/2.0f);
theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, iw/2.0f);
}
}
const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
const float freq_factor = p.has_ff != 0 ? rope_data_ff[iw/2] : 1.0f;
float cos_theta, sin_theta;
rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
rope_yarn(theta_base / freq_factor, iw, cos_theta, sin_theta, p);
const float x0 = float(rope_data_a[ix + 0]);
const float x1 = float(rope_data_a[ix + p.n_dims/2]);
// idst/ix point at channel i0/2; the first channel of the rotated pair is p.n_offs + iw/2 = i0/2 + p.n_offs/2
const float x0 = float(rope_data_a[ix + p.n_offs/2 + 0]);
const float x1 = float(rope_data_a[ix + p.n_offs/2 + p.n_dims/2]);
rope_data_d[idst + 0] = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
rope_data_d[idst + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
rope_data_d[idst + p.n_offs/2 + 0] = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
rope_data_d[idst + p.n_offs/2 + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
}
void rope_vision(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
@@ -5,6 +5,7 @@ struct rope_params {
uint rope_mode;
uint nrows;
uint n_dims;
uint n_offs;
float freq_scale;
float freq_base;
float ext_factor;