CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-28 08:31:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

CANN: Implement GLU ops (#14884)

Browse Source 
Implement REGLU, GEGLU, SWIGLU ops according to #14158
This commit is contained in:
hipudding
2025-07-26 17:56:18 +08:00
committed by GitHub
parent 9b8f3c6c77
commit 11dd5a44eb
4 changed files with 194 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
ggml/src/ggml-cann/aclnn_ops.cpp
View File

@@ -99,7 +99,7 @@ void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst, aclT
}
}
void ggml_cann_unary_op(
void ggml_cann_op_unary(
std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src = dst->src[0];
@@ -111,6 +111,42 @@ void ggml_cann_unary_op(
ggml_cann_release_resources(ctx, acl_src, acl_dst);
}
void ggml_cann_op_unary_gated(
std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src0 = dst->src[0];
ggml_tensor* src1 = dst->src[1];
GGML_ASSERT(ggml_is_contiguous_1(src0));
GGML_ASSERT(ggml_is_contiguous_1(dst));
const int32_t swapped = ggml_get_op_params_i32(dst, 1);
aclTensor* acl_dst = ggml_cann_create_tensor(dst);
aclTensor *acl_src0 = nullptr, *acl_src1 = nullptr;
if(src1) {
GGML_ASSERT(ggml_is_contiguous_1(src1));
GGML_ASSERT(src0->type == src1->type);
acl_src0 = ggml_cann_create_tensor(src0);
acl_src1 = ggml_cann_create_tensor(src1);
} else {
int64_t ne[] = {src0->ne[0] / 2, src0->ne[1], src0->ne[2], src0->ne[3]};
size_t nb[] = {src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]};
acl_src0 = ggml_cann_create_tensor(src0, ne, nb, GGML_MAX_DIMS, ACL_FORMAT_ND, 0);
acl_src1 = ggml_cann_create_tensor(src0, ne, nb, GGML_MAX_DIMS, ACL_FORMAT_ND, ne[0] * ggml_element_size(src0));
if (swapped) {
std::swap(acl_src0, acl_src1);
}
}
unary_op(ctx, acl_src0, acl_dst);
GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, acl_dst, acl_src1);
ggml_cann_release_resources(ctx, acl_src0, acl_dst);
if(src1)
ggml_cann_release_resources(ctx, acl_src1);
}
/**
* @brief Repeats elements of a tensor along each dimension according to the
* specified repeat array.