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

ggml: add SWIGLU_CLAMP (#27930)

Browse Source 
* ggml: add SWIGLU_CLAMP

* add vulkan shader
This commit is contained in:
Aman Gupta
2026-08-30 23:00:02 +08:00
committed by GitHub
parent 2578138397
commit 0190529ec4
40 changed files with 768 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/ggml-cpu.c
+1
View File
@@ -2311,6 +2311,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_GLU_OP_SWIGLU_OAI:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
case GGML_GLU_OP_SWIGLU_CLAMP:
{
n_tasks = n_threads;
} break;
ggml/src/ggml-cpu/ops.cpp
+137
View File
@@ -3403,6 +3403,139 @@ static void ggml_compute_forward_swiglu_oai(
}
}
// ggml_compute_forward_swiglu_clamp
static void ggml_compute_forward_swiglu_clamp_f32(const ggml_compute_params * params, 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 = src1 ? src0->ne[0] : src0->ne[0] / 2;
const int nr = ggml_nrows(src0);
GGML_ASSERT(dst->ne[0] == nc);
GGML_ASSERT(ggml_nrows(dst) == nr);
const int32_t swapped = ggml_get_op_params_i32(dst, 1);
const float limit = ggml_get_op_params_f32(dst, 3);
const int dr = (nr + nth - 1) / nth;
const int ir0 = dr * ith;
const int ir1 = MIN(ir0 + dr, nr);
for (int i1 = ir0; i1 < ir1; i1++) {
float * src0_p = (float *) (src0_d + i1 * src0_o);
float * src1_p = (float *) (src1_d + i1 * src1_o);
float * dst_p = (float *) ((char *) dst->data + i1 * (dst->nb[1]));
if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}
for (int k = 0; k < nc; k++) {
const float gate = std::min(src0_p[k], limit);
const float up = std::clamp(src1_p[k], -limit, limit);
dst_p[k] = gate / (1.f + expf(-gate)) * up;
}
#ifndef NDEBUG
for (int k = 0; k < nc; k++) {
const float x = dst_p[k];
GGML_UNUSED(x);
assert(!isnan(x));
assert(!isinf(x));
}
#endif // NDEBUG
}
}
static void ggml_compute_forward_swiglu_clamp_f16(const ggml_compute_params * params, 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 = src1 ? src0->ne[0] : src0->ne[0] / 2;
const int nr = ggml_nrows(src0);
GGML_ASSERT(dst->ne[0] == nc);
GGML_ASSERT(ggml_nrows(dst) == nr);
const int32_t swapped = ggml_get_op_params_i32(dst, 1);
const float limit = ggml_get_op_params_f32(dst, 3);
const int dr = (nr + nth - 1) / nth;
const int ir0 = dr * ith;
const int ir1 = MIN(ir0 + dr, nr);
for (int i1 = ir0; i1 < ir1; i1++) {
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);
ggml_fp16_t * dst_p = (ggml_fp16_t *) ((char *) dst->data + i1 * (dst->nb[1]));
if (!src1) {
src0_p += swapped ? nc : 0;
src1_p += swapped ? 0 : nc;
}
for (int k = 0; k < nc; k++) {
const float gate = std::min(GGML_FP16_TO_FP32(src0_p[k]), limit);
const float up = std::clamp(GGML_FP16_TO_FP32(src1_p[k]), -limit, limit);
dst_p[k] = GGML_FP32_TO_FP16(gate / (1.f + expf(-gate)) * up);
}
#ifndef NDEBUG
for (int k = 0; k < nc; k++) {
const float x = GGML_FP16_TO_FP32(dst_p[k]);
GGML_UNUSED(x);
assert(!isnan(x));
assert(!isinf(x));
}
#endif // NDEBUG
}
}
static void ggml_compute_forward_swiglu_clamp(const ggml_compute_params * params, ggml_tensor * dst) {
switch (dst->src[0]->type) {
case GGML_TYPE_F32:
ggml_compute_forward_swiglu_clamp_f32(params, dst);
break;
case GGML_TYPE_F16:
ggml_compute_forward_swiglu_clamp_f16(params, dst);
break;
default:
GGML_ABORT("fatal error");
}
}
// ggml_compute_forward_geglu_erf
static void ggml_compute_forward_geglu_erf_f32(
@@ -10136,6 +10269,10 @@ void ggml_compute_forward_glu(
{
ggml_compute_forward_geglu_quick(params, dst);
} break;
case GGML_GLU_OP_SWIGLU_CLAMP:
{
ggml_compute_forward_swiglu_clamp(params, dst);
} break;
default:
{
GGML_ABORT("fatal error");