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

fixed a bug in ggml's conv_2d, and now > 1 input images working

Browse Source
This commit is contained in:
bssrdf committed 2024-02-18 22:18:53 -05:00
1 parent 6064b0e28b
commit d38b85d87e
4 files changed
+164 -55
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

clip.hpp
+8 -4
View File
@@ -975,6 +975,10 @@ struct CLIPVisionModel {
// patch_embeddings_f16 = ggml_cast(ctx0, patch_embeddings_f16, GGML_TYPE_F16);
// patch_embeddings_f16 = ggml_reshape_4d(ctx0, patch_embeddings_f16, ne[0], ne[1], ne[2], ne[3]);
struct ggml_tensor *patch_embeddings_f16 = ggml_cast(ctx0, patch_embeddings, GGML_TYPE_F16);
ggml_set_name(patch_embeddings_f16, "patch_embeddings_f16");
print_ggml_tensor(patch_embeddings_f16, true, "patch_embeddings_f16");
struct ggml_tensor * inp = ggml_conv_2d(ctx0, patch_embeddings_f16, x, patch_size, patch_size, 0, 0, 1, 1);
ggml_set_name(inp, "inp_conv_2d");
@@ -1023,7 +1027,7 @@ struct CLIPVisionModel {
embeddings =
// ggml_add(ctx0, embeddings, ggml_repeat(ctx0, ggml_get_rows(ctx0, position_embeddings, positions), embeddings));
ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, position_embeddings, positions));
ggml_set_name(embeddings, "embeddings_after_add");
ggml_set_name(embeddings, "embeddings_to_transformer");
print_ggml_tensor(embeddings, true, "embeddings_to_transformer");
// pre-layernorm
@@ -1035,12 +1039,12 @@ struct CLIPVisionModel {
embeddings = resblocks[i].forward(ctx0, embeddings); // [N, n_token, hidden_size]
}
ggml_set_name(embeddings, "embeddings_after_transformer");
// print_ggml_tensor(embeddings, true, "embeddings_after_encoder");
print_ggml_tensor(embeddings, true, "embeddings_after_transformer");
// get the output of cls token, e.g., 0th index
embeddings = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, embeddings, hidden_size, num_positions * batch_size), cls);
ggml_set_name(embeddings, "embeddings_after_cls_token");
// print_ggml_tensor(embeddings, true, "embeddings_pooled_after_encoder");
ggml_set_name(embeddings, "embeddings_pooled_after_encoder");
print_ggml_tensor(embeddings, true, "embeddings_pooled_after_encoder");
// post-layernorm
embeddings = ggml_nn_layer_norm(ctx0, embeddings, post_ln_w, post_ln_b);