EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

ggml : implement GEGLU_ERF and GEGLU_QUICK ops (#14445)

Browse Source
This commit is contained in:
Sigbjørn Skjæret
2025-07-03 23:07:22 +02:00
committed by GitHub
parent bee28421be
commit 28657a8229
20 changed files with 789 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
ggml/src/ggml-metal/ggml-metal.m
View File

@@ -530,6 +530,8 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_REGLU,
GGML_METAL_KERNEL_TYPE_GEGLU,
GGML_METAL_KERNEL_TYPE_SWIGLU,
GGML_METAL_KERNEL_TYPE_GEGLU_ERF,
GGML_METAL_KERNEL_TYPE_GEGLU_QUICK,
GGML_METAL_KERNEL_TYPE_SUM_ROWS,
GGML_METAL_KERNEL_TYPE_MEAN,
GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
@@ -1510,6 +1512,8 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REGLU, reglu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU, geglu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SWIGLU, swiglu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU_ERF, geglu_erf, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU_QUICK, geglu_quick, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MEAN, mean, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX, argmax, true);
@@ -1693,6 +1697,8 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
case GGML_GLU_OP_REGLU:
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_SWIGLU:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
return ggml_is_contiguous_1(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
default:
return false;
@@ -2456,6 +2462,12 @@ static bool ggml_metal_encode_node(
case GGML_GLU_OP_SWIGLU:
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SWIGLU].pipeline;
break;
case GGML_GLU_OP_GEGLU_ERF:
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GEGLU_ERF].pipeline;
break;
case GGML_GLU_OP_GEGLU_QUICK:
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GEGLU_QUICK].pipeline;
break;
default:
GGML_ABORT("fatal error");
}

44
ggml/src/ggml-metal/ggml-metal.metal
View File

@@ -1258,6 +1258,50 @@ kernel void kernel_swiglu(
}
}
kernel void kernel_geglu_erf(
device const char * src0,
device const char * src1,
device char * dst,
constant ggml_metal_kargs_glu & args,
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
uint ntg[[threads_per_threadgroup]]) {
device const float * src0_row = (device const float *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
device const float * src1_row = (device const float *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
device float * dst_row = (device float *) ((device char *) dst + tgpig*args.nb1);
for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
const float x0 = src0_row[i0];
const float x1 = src1_row[i0];
const float gelu_erf = 0.5f*x0*(1.0f+erf_approx<float>(x0*SQRT_2_INV));
dst_row[i0] = gelu_erf*x1;
}
}
kernel void kernel_geglu_quick(
device const char * src0,
device const char * src1,
device char * dst,
constant ggml_metal_kargs_glu & args,
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
uint ntg[[threads_per_threadgroup]]) {
device const float * src0_row = (device const float *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
device const float * src1_row = (device const float *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
device float * dst_row = (device float *) ((device char *) dst + tgpig*args.nb1);
for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
const float x0 = src0_row[i0];
const float x1 = src1_row[i0];
const float gelu_quick = x0*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x0)));
dst_row[i0] = gelu_quick*x1;
}
}
template <bool norm>
kernel void kernel_sum_rows(
constant ggml_metal_kargs_sum_rows & args,