DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-11 11:26:34 +00:00
Code Issues Packages Projects Releases Wiki Activity

sync: update ggml (#134)

Browse Source
This commit is contained in:
leejet authored and GitHub committed 2024-01-05 23:18:41 +08:00
1 parent db382348cc
commit 2b6ec97fe2
9 files changed
+21 -65
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

clip.hpp
+1 -8
View File
@@ -443,8 +443,6 @@ struct ResidualAttentionBlock {
struct ggml_tensor* ln2_w; // [hidden_size, ]
struct ggml_tensor* ln2_b; // [hidden_size, ]
struct ggml_tensor* attn_scale; // [hidden_size, ]
size_t calculate_mem_size(ggml_type wtype) {
double mem_size = 0;
mem_size += 4 * hidden_size * hidden_size * ggml_type_sizef(wtype); // q_w/k_w/v_w/out_w
@@ -452,7 +450,6 @@ struct ResidualAttentionBlock {
mem_size += 2 * hidden_size * intermediate_size * ggml_type_sizef(wtype); // fc1_w/fc2_w
mem_size += intermediate_size * ggml_type_sizef(GGML_TYPE_F32); // fc1_b
mem_size += hidden_size * ggml_type_sizef(GGML_TYPE_F32); // fc2_b
mem_size += ggml_type_sizef(GGML_TYPE_F32); // attn_scale
return static_cast<size_t>(mem_size);
}
@@ -479,10 +476,6 @@ struct ResidualAttentionBlock {
ln2_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
ln2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
attn_scale = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
ggml_allocr_alloc(alloc, attn_scale);
float scale = 1.0f / sqrt((float)d_model);
ggml_backend_tensor_set(attn_scale, &scale, 0, sizeof(scale));
}
void map_by_name(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
@@ -521,7 +514,7 @@ struct ResidualAttentionBlock {
// self-attention
{
struct ggml_tensor* q = ggml_nn_linear(ctx, x, q_w, q_b);
q = ggml_scale_inplace(ctx, q, attn_scale);
q = ggml_scale_inplace(ctx, q, 1.0f / sqrt((float)d_model));
q = ggml_reshape_4d(ctx, q, d_model, n_head, n_token, N); // [N, n_token, n_head, d_model]
q = ggml_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3)); // [N, n_head, n_token, d_model]
q = ggml_reshape_3d(ctx, q, d_model, n_token, n_head * N); // [N * n_head, n_token, d_model]