vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 09:34:37 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

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
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
20 changed files with 789 additions and 32 deletions
Download patch file Download diff file Expand all files Collapse all files

16
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file

@ -456,6 +456,8 @@ struct vk_device_struct {
vk_pipeline pipeline_geglu[2];
vk_pipeline pipeline_reglu[2];
vk_pipeline pipeline_swiglu[2];
vk_pipeline pipeline_geglu_erf[2];
vk_pipeline pipeline_geglu_quick[2];
vk_pipeline pipeline_leaky_relu_f32;
vk_pipeline pipeline_silu_back_f32;
@ -2821,6 +2823,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
CREATE_GLU(geglu)
CREATE_GLU(reglu)
CREATE_GLU(swiglu)
CREATE_GLU(geglu_erf)
CREATE_GLU(geglu_quick)
#undef CREATE_GLU
ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
@ -6575,6 +6579,10 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
return ctx->device->pipeline_reglu[dst->type == GGML_TYPE_F16];
case GGML_GLU_OP_SWIGLU:
return ctx->device->pipeline_swiglu[dst->type == GGML_TYPE_F16];
case GGML_GLU_OP_GEGLU_ERF:
return ctx->device->pipeline_geglu_erf[dst->type == GGML_TYPE_F16];
case GGML_GLU_OP_GEGLU_QUICK:
return ctx->device->pipeline_geglu_quick[dst->type == GGML_TYPE_F16];
default:
break;
}
@ -8919,6 +8927,8 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_REGLU:
case GGML_GLU_OP_SWIGLU:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
break;
default:
return false;
@ -9166,6 +9176,8 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_REGLU:
case GGML_GLU_OP_SWIGLU:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
ggml_vk_glu(ctx, compute_ctx, src0, src1, node, dryrun);
break;
default:
@ -9384,6 +9396,8 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_REGLU:
case GGML_GLU_OP_SWIGLU:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
buf = tensor->buffer;
break;
default:
@ -10194,6 +10208,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_GLU_OP_GEGLU:
case GGML_GLU_OP_REGLU:
case GGML_GLU_OP_SWIGLU:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
return ggml_is_contiguous(op->src[0]) &&
(op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
(op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&