happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'sf/deepseek-ocr' of github.com:sfallah/llama.cpp into sf/deepseek-ocr

This commit is contained in:
bluebread 2025-11-22 02:22:48 +00:00
parent effe66958e 86f111f8b7
commit f8f66a151b
1 changed files with 7 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
tools/mtmd/clip.cpp
View File

@ -820,6 +820,7 @@ struct clip_graph {
// TODO: better implementation
layer = ggml_permute(ctx0, ggml_norm(ctx0, ggml_cont(ctx0, ggml_permute(ctx0, layer, 1, 2, 0, 3)), eps), 2, 0,
1, 3);
layer = ggml_cont(ctx0, layer);
layer =
ggml_add(ctx0, ggml_mul(ctx0, ggml_repeat(ctx0, ggml_reshape_3d(ctx0, w, 1, 1, n_channels), layer), layer),
@ -1538,8 +1539,7 @@ struct clip_graph {
GGML_ASSERT(model.position_embeddings != nullptr);
const int n_pos = n_patches + 1;
ggml_tensor * inp = ggml_permute(ctx0, patch_embeds,2,1,0,3);
inp = ggml_cont(ctx0, inp);
ggml_tensor * inp = ggml_cont(ctx0,ggml_permute(ctx0, patch_embeds,2,1,0,3));
inp = ggml_reshape_2d(ctx0, inp, n_embd, n_patches);
@ -1551,7 +1551,7 @@ struct clip_graph {
norm_type norm_t = NORM_TYPE_NORMAL;
// for selecting learned pos embd, used by ViT
struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
cb(positions, "positions", -1);
ggml_tensor * learned_pos_embd = ggml_get_rows(ctx0, model.position_embeddings, positions);
@ -5237,7 +5237,10 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
} break;
case PROJECTOR_TYPE_DEEPSEEKOCR:
{
n_patches += 2;
int x_patch = img->nx / (params.patch_size);
n_patches += x_patch + 1;
} break;
default:
GGML_ABORT("unsupported projector type");