vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-12 18:09:42 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	.github/workflows/docker.yml
#	CMakeLists.txt
#	Makefile
#	README.md
#	flake.lock
#	tests/test-backend-ops.cpp
This commit is contained in:
Concedo 2024-05-28 21:57:19 +08:00
commit 4ed9ba7352
58 changed files with 2404 additions and 361 deletions
Download patch file Download diff file Expand all files Collapse all files

6
otherarch/sdcpp/clip.hpp
View file

@ -617,7 +617,7 @@ public:
class_embedding = ggml_repeat(ctx, class_embed_weight, class_embedding); // [N, embed_dim]
class_embedding = ggml_reshape_4d(ctx, class_embedding, 1, embed_dim, 1, N); // [N, 1, embed_dim, 1]
struct ggml_tensor* x = ggml_concat(ctx, class_embedding, patch_embedding); // [N, num_positions, embed_dim, 1]
struct ggml_tensor* x = ggml_concat(ctx, class_embedding, patch_embedding, 2); // [N, num_positions, embed_dim, 1]
x = ggml_reshape_3d(ctx, x, embed_dim, num_positions, N); // [N, num_positions, embed_dim]
x = ggml_add(ctx, x, position_embed_weight);
return x; // [N, num_positions, embed_dim]
@ -948,7 +948,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public GGMLModule {
hidden_states2->ne[3]);
hidden_states2 = ggml_cont(ctx, ggml_permute(ctx, hidden_states2, 2, 0, 1, 3));
hidden_states = ggml_concat(ctx, hidden_states, hidden_states2); // [N, n_token, hidden_size + hidden_size2]
hidden_states = ggml_concat(ctx, hidden_states, hidden_states2, 2); // [N, n_token, hidden_size + hidden_size2]
hidden_states = ggml_cont(ctx, ggml_permute(ctx, hidden_states, 1, 2, 0, 3));
}
@ -980,7 +980,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public GGMLModule {
auto token_embed_weight = text_model.get_token_embed_weight();
token_embed_weight = ggml_reshape_3d(compute_ctx, token_embed_weight, token_embed_weight->ne[0], 1, token_embed_weight->ne[1]);
// concatenate custom embeddings
embeddings = ggml_concat(compute_ctx, token_embed_weight, custom_embeddings);
embeddings = ggml_concat(compute_ctx, token_embed_weight, custom_embeddings, 2);
embeddings = ggml_reshape_2d(compute_ctx, embeddings, embeddings->ne[0], embeddings->ne[2]);
}

8
otherarch/sdcpp/esrgan.hpp
View file

@ -42,13 +42,13 @@ public:
auto conv5 = std::dynamic_pointer_cast<Conv2d>(blocks["conv5"]);
auto x1 = lrelu(ctx, conv1->forward(ctx, x));
auto x_cat = ggml_concat(ctx, x, x1);
auto x_cat = ggml_concat(ctx, x, x1, 2);
auto x2 = lrelu(ctx, conv2->forward(ctx, x_cat));
x_cat = ggml_concat(ctx, x_cat, x2);
x_cat = ggml_concat(ctx, x_cat, x2, 2);
auto x3 = lrelu(ctx, conv3->forward(ctx, x_cat));
x_cat = ggml_concat(ctx, x_cat, x3);
x_cat = ggml_concat(ctx, x_cat, x3, 2);
auto x4 = lrelu(ctx, conv4->forward(ctx, x_cat));
x_cat = ggml_concat(ctx, x_cat, x4);
x_cat = ggml_concat(ctx, x_cat, x4, 2);
auto x5 = conv5->forward(ctx, x_cat);
x5 = ggml_add(ctx, ggml_scale(ctx, x5, 0.2f), x);

4
otherarch/sdcpp/unet.hpp
View file

@ -396,7 +396,7 @@ public:
if (c_concat->ne[3] != x->ne[3]) {
c_concat = ggml_repeat(ctx, c_concat, x);
}
x = ggml_concat(ctx, x, c_concat);
x = ggml_concat(ctx, x, c_concat, 2);
}
if (y != NULL) {
@ -491,7 +491,7 @@ public:
control_offset--;
}
h = ggml_concat(ctx, h, h_skip);
h = ggml_concat(ctx, h, h_skip, 2);
std::string name = "output_blocks." + std::to_string(output_block_idx) + ".0";