Skip to content

ggml : implement GLU for split up/gate #14181

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 6 commits into from
Jun 18, 2025
Merged

ggml : implement GLU for split up/gate #14181

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
94361fd
implement GLU for split up/gate
CISC Jun 14, 2025
ac3194d
add tests for ggml_glu_split
CISC Jun 14, 2025
be3f78c
Vulkan: Implement glu_split logic and shader support
0cc4m Jun 15, 2025
06362b0
add split to logging [no ci]
CISC Jun 15, 2025
42c2870
SYCL: refactor element_size ops and add split up and gate support to …
qnixsynapse Jun 15, 2025
832efa9
SYCL: switch GEGLU to use tanh approximation
qnixsynapse Jun 17, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
23 changes: 23 additions & 0 deletions ggml/include/ggml.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1131,6 +1131,29 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);

// A: n columns, r rows,
// B: n columns, r rows,
GGML_API struct ggml_tensor * ggml_glu_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
enum ggml_glu_op op);

GGML_API struct ggml_tensor * ggml_reglu_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);

GGML_API struct ggml_tensor * ggml_geglu_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);

GGML_API struct ggml_tensor * ggml_swiglu_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);

// normalize along rows
GGML_API struct ggml_tensor * ggml_norm(
struct ggml_context * ctx,
Expand Down
150 changes: 120 additions & 30 deletions ggml/src/ggml-cpu/ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3201,14 +3201,24 @@ static void ggml_compute_forward_reglu_f32(
ggml_tensor * dst) {

const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
char * src0_d = (char *) src0->data;
char * src1_d = (char *) (src1 ? src1->data : src0->data);
const size_t src0_o = src0->nb[1];
const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];

GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));

if (src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src0->type == src1->type);
}

const int ith = params->ith;
const int nth = params->nth;

const int nc = src0->ne[0] / 2;
const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
const int nr = ggml_nrows(src0);

GGML_ASSERT(dst->ne[0] == nc);
Expand All @@ -3224,10 +3234,15 @@ static void ggml_compute_forward_reglu_f32(
const int ir1 = MIN(ir0 + dr, nr);

for (int i1 = ir0; i1 < ir1; i1++) {
ggml_vec_reglu_f32(nc,
(float *) ((char *) dst->data + i1*( dst->nb[1])),
(float *) ((char *) src0->data + i1*(src0->nb[1])) + (swapped ? nc : 0),
(float *) ((char *) src0->data + i1*(src0->nb[1])) + (swapped ? 0 : nc));
float * src0_p = (float *) (src0_d + i1*src0_o);
float * src1_p = (float *) (src1_d + i1*src1_o);

if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}

ggml_vec_reglu_f32(nc, (float *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);

#ifndef NDEBUG
for (int k = 0; k < nc; k++) {
Expand All @@ -3245,14 +3260,24 @@ static void ggml_compute_forward_reglu_f16(
ggml_tensor * dst) {

const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
char * src0_d = (char *) src0->data;
char * src1_d = (char *) (src1 ? src1->data : src0->data);
const size_t src0_o = src0->nb[1];
const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];

GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));

if (src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src0->type == src1->type);
}

const int ith = params->ith;
const int nth = params->nth;

const int nc = src0->ne[0] / 2;
const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
const int nr = ggml_nrows(src0);

GGML_ASSERT(dst->ne[0] == nc);
Expand All @@ -3268,10 +3293,15 @@ static void ggml_compute_forward_reglu_f16(
const int ir1 = MIN(ir0 + dr, nr);

for (int i1 = ir0; i1 < ir1; i1++) {
ggml_vec_reglu_f16(nc,
(ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])),
(ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])) + (swapped ? nc : 0),
(ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])) + (swapped ? 0 : nc));
ggml_fp16_t * src0_p = (ggml_fp16_t *) (src0_d + i1*src0_o);
ggml_fp16_t * src1_p = (ggml_fp16_t *) (src1_d + i1*src1_o);

if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}

ggml_vec_reglu_f16(nc, (ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);

#ifndef NDEBUG
for (int k = 0; k < nc; k++) {
Expand Down Expand Up @@ -3314,14 +3344,24 @@ static void ggml_compute_forward_geglu_f32(
ggml_tensor * dst) {

const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
char * src0_d = (char *) src0->data;
char * src1_d = (char *) (src1 ? src1->data : src0->data);
const size_t src0_o = src0->nb[1];
const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];

GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));

if (src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src0->type == src1->type);
}

const int ith = params->ith;
const int nth = params->nth;

const int nc = src0->ne[0] / 2;
const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
const int nr = ggml_nrows(src0);

GGML_ASSERT(dst->ne[0] == nc);
Expand All @@ -3337,10 +3377,15 @@ static void ggml_compute_forward_geglu_f32(
const int ir1 = MIN(ir0 + dr, nr);

for (int i1 = ir0; i1 < ir1; i1++) {
ggml_vec_geglu_f32(nc,
(float *) ((char *) dst->data + i1*( dst->nb[1])),
(float *) ((char *) src0->data + i1*(src0->nb[1])) + (swapped ? nc : 0),
(float *) ((char *) src0->data + i1*(src0->nb[1])) + (swapped ? 0 : nc));
float * src0_p = (float *) (src0_d + i1*src0_o);
float * src1_p = (float *) (src1_d + i1*src1_o);

if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}

ggml_vec_geglu_f32(nc, (float *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);

#ifndef NDEBUG
for (int k = 0; k < nc; k++) {
Expand All @@ -3358,14 +3403,24 @@ static void ggml_compute_forward_geglu_f16(
ggml_tensor * dst) {

const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
char * src0_d = (char *) src0->data;
char * src1_d = (char *) (src1 ? src1->data : src0->data);
const size_t src0_o = src0->nb[1];
const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];

GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));

if (src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src0->type == src1->type);
}

const int ith = params->ith;
const int nth = params->nth;

const int nc = src0->ne[0] / 2;
const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
const int nr = ggml_nrows(src0);

GGML_ASSERT(dst->ne[0] == nc);
Expand All @@ -3381,10 +3436,15 @@ static void ggml_compute_forward_geglu_f16(
const int ir1 = MIN(ir0 + dr, nr);

for (int i1 = ir0; i1 < ir1; i1++) {
ggml_vec_geglu_f16(nc,
(ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])),
(ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])) + (swapped ? nc : 0),
(ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])) + (swapped ? 0 : nc));
ggml_fp16_t * src0_p = (ggml_fp16_t *) (src0_d + i1*src0_o);
ggml_fp16_t * src1_p = (ggml_fp16_t *) (src1_d + i1*src1_o);

if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}

ggml_vec_geglu_f16(nc, (ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);

#ifndef NDEBUG
for (int k = 0; k < nc; k++) {
Expand Down Expand Up @@ -3427,14 +3487,24 @@ static void ggml_compute_forward_swiglu_f32(
ggml_tensor * dst) {

const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
char * src0_d = (char *) src0->data;
char * src1_d = (char *) (src1 ? src1->data : src0->data);
const size_t src0_o = src0->nb[1];
const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];

GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));

if (src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src0->type == src1->type);
}

const int ith = params->ith;
const int nth = params->nth;

const int nc = src0->ne[0] / 2;
const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
const int nr = ggml_nrows(src0);

GGML_ASSERT(dst->ne[0] == nc);
Expand All @@ -3450,10 +3520,15 @@ static void ggml_compute_forward_swiglu_f32(
const int ir1 = MIN(ir0 + dr, nr);

for (int i1 = ir0; i1 < ir1; i1++) {
ggml_vec_swiglu_f32(nc,
(float *) ((char *) dst->data + i1*( dst->nb[1])),
(float *) ((char *) src0->data + i1*(src0->nb[1])) + (swapped ? nc : 0),
(float *) ((char *) src0->data + i1*(src0->nb[1])) + (swapped ? 0 : nc));
float * src0_p = (float *) (src0_d + i1*src0_o);
float * src1_p = (float *) (src1_d + i1*src1_o);

if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}

ggml_vec_swiglu_f32(nc, (float *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);

#ifndef NDEBUG
for (int k = 0; k < nc; k++) {
Expand All @@ -3471,14 +3546,24 @@ static void ggml_compute_forward_swiglu_f16(
ggml_tensor * dst) {

const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
char * src0_d = (char *) src0->data;
char * src1_d = (char *) (src1 ? src1->data : src0->data);
const size_t src0_o = src0->nb[1];
const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];

GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));

if (src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src0->type == src1->type);
}

const int ith = params->ith;
const int nth = params->nth;

const int nc = src0->ne[0] / 2;
const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
const int nr = ggml_nrows(src0);

GGML_ASSERT(dst->ne[0] == nc);
Expand All @@ -3494,10 +3579,15 @@ static void ggml_compute_forward_swiglu_f16(
const int ir1 = MIN(ir0 + dr, nr);

for (int i1 = ir0; i1 < ir1; i1++) {
ggml_vec_swiglu_f16(nc,
(ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])),
(ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])) + (swapped ? nc : 0),
(ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])) + (swapped ? 0 : nc));
ggml_fp16_t * src0_p = (ggml_fp16_t *) (src0_d + i1*src0_o);
ggml_fp16_t * src1_p = (ggml_fp16_t *) (src1_d + i1*src1_o);

if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}

ggml_vec_swiglu_f16(nc, (ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);

#ifndef NDEBUG
for (int k = 0; k < nc; k++) {
Expand Down
Loading
Loading