ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

ggml : add ggml_gelu_erf() (#13667)

Browse source 
* ggml : add ggml_gelu_na (not approximated)

* fix naming order

* rename na --> erf

* apply review suggesions

* revert naming order
This commit is contained in:
Xuan-Son Nguyen 2025-05-21 16:26:33 +02:00 • committed by GitHub
parent 0d5c742161
commit cf4cb59e64
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 213 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

24
ggml/src/ggml-metal/ggml-metal.m
View file

@ -149,6 +149,8 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_SIGMOID,
GGML_METAL_KERNEL_TYPE_GELU,
GGML_METAL_KERNEL_TYPE_GELU_4,
GGML_METAL_KERNEL_TYPE_GELU_ERF,
GGML_METAL_KERNEL_TYPE_GELU_ERF_4,
GGML_METAL_KERNEL_TYPE_GELU_QUICK,
GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,
GGML_METAL_KERNEL_TYPE_SILU,
@ -1103,6 +1105,8 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIGMOID, sigmoid, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU, gelu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4, gelu_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_ERF, gelu_erf, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_ERF_4, gelu_erf_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK, gelu_quick, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, gelu_quick_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU, silu, true);
@ -1613,6 +1617,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
case GGML_UNARY_OP_RELU:
case GGML_UNARY_OP_SIGMOID:
case GGML_UNARY_OP_GELU:
case GGML_UNARY_OP_GELU_ERF:
case GGML_UNARY_OP_GELU_QUICK:
case GGML_UNARY_OP_SILU:
case GGML_UNARY_OP_ELU:
@ -2251,6 +2256,25 @@ static bool ggml_metal_encode_node(
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_GELU_ERF:
{
int64_t n = ggml_nelements(dst);
id<MTLComputePipelineState> pipeline = nil;
if (n % 4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_ERF_4].pipeline;
n /= 4;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_ERF].pipeline;
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_GELU_QUICK:
{
int64_t n = ggml_nelements(dst);

37
ggml/src/ggml-metal/ggml-metal.metal
View file

@ -856,6 +856,7 @@ kernel void kernel_tanh(
constant float GELU_COEF_A = 0.044715f;
constant float GELU_QUICK_COEF = -1.702f;
constant float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
constant float SQRT_2_INV = 0.70710678118654752440084436210484f;
kernel void kernel_gelu(
device const float * src0,
@ -897,6 +898,42 @@ kernel void kernel_gelu_quick_4(
dst[tpig] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
}
// based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
// ref: https://www.johndcook.com/blog/python_erf/
constant float p_erf = 0.3275911f;
constant float a1_erf = 0.254829592f;
constant float a2_erf = -0.284496736f;
constant float a3_erf = 1.421413741f;
constant float a4_erf = -1.453152027f;
constant float a5_erf = 1.061405429f;
template<typename T>
T erf_approx(T x) {
T sign_x = sign(x);
x = fabs(x);
T t = 1.0f / (1.0f + p_erf * x);
T y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
return sign_x * y;
}
kernel void kernel_gelu_erf(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
device const float & x = src0[tpig];
dst[tpig] = 0.5f*x*(1.0f+erf_approx<float>(x*SQRT_2_INV));
}
kernel void kernel_gelu_erf_4(
device const float4 * src0,
device float4 * dst,
uint tpig[[thread_position_in_grid]]) {
device const float4 & x = src0[tpig];
dst[tpig] = 0.5f*x*(1.0f+erf_approx<float4>(x*SQRT_2_INV));
}
kernel void kernel_silu(
device const float * src0,
device float * dst,