CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-16 11:27:03 +00:00
Code Issues Packages 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

8
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@@ -2314,6 +2314,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
case GGML_GLU_OP_SWIGLU:
ggml_cuda_op_swiglu(ctx, dst);
break;
case GGML_GLU_OP_GEGLU_ERF:
ggml_cuda_op_geglu_erf(ctx, dst);
break;
case GGML_GLU_OP_GEGLU_QUICK:
ggml_cuda_op_geglu_quick(ctx, dst);
break;
default:
return false;
}
@@ -3116,6 +3122,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
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]);
default:
return false;

8
ggml/src/ggml-cuda/unary.cu
View File

@@ -285,6 +285,14 @@ void ggml_cuda_op_swiglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_unary_gated<op_silu>(ctx, dst);
}
void ggml_cuda_op_geglu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_unary_gated<op_gelu_erf>(ctx, dst);
}
void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_unary_gated<op_gelu_quick>(ctx, dst);
}
/* silu_back */
static __device__ __forceinline__ float op_silu_back(float grad, float x) {

4
ggml/src/ggml-cuda/unary.cuh
View File

@@ -64,3 +64,7 @@ void ggml_cuda_op_reglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_geglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_swiglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_geglu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst);