Skip to content

clip : remove boi/eoi embeddings for GLM-edge model (⚠️ breaking change) #13081

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Apr 24, 2025
Merged

clip : remove boi/eoi embeddings for GLM-edge model (⚠️ breaking change) #13081

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 0 additions & 2 deletions examples/llava/clip-impl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -90,8 +90,6 @@
#define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s"
#define TN_GLM_ADAPTER_GATE "adapter.linear.gate.%s"
#define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
#define TN_GLM_BOI_W "adapter.boi"
#define TN_GLM_EOI_W "adapter.eoi"

enum projector_type {
PROJECTOR_TYPE_MLP,
Expand Down
17 changes: 1 addition & 16 deletions examples/llava/clip.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -244,8 +244,6 @@ struct clip_vision_model {
//GLMV-Edge projection
struct ggml_tensor * mm_model_adapter_conv_w = nullptr;
struct ggml_tensor * mm_model_adapter_conv_b = nullptr;
struct ggml_tensor * boi_w = nullptr;
struct ggml_tensor * eoi_w = nullptr;

// MobileVLM projection
struct ggml_tensor * mm_model_mlp_1_w = nullptr;
Expand Down Expand Up @@ -1697,8 +1695,6 @@ struct clip_model_loader {
vision_model.mm_model_mlp_1_w = get_tensor(string_format(TN_GLM_ADAPTER_D_H_2_4H,"weight"));
vision_model.mm_model_mlp_2_w = get_tensor(string_format(TN_GLM_ADAPTER_GATE,"weight"));
vision_model.mm_model_mlp_3_w = get_tensor(string_format(TN_GLM_ADAPTER_D_4H_2_H,"weight"));
vision_model.boi_w = get_tensor(TN_GLM_BOI_W);
vision_model.eoi_w = get_tensor(TN_GLM_EOI_W);
} break;
case PROJECTOR_TYPE_MERGER:
{
Expand Down Expand Up @@ -2593,8 +2589,7 @@ void clip_free(clip_ctx * ctx) {
}

size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
int extra_tokens = ctx->has_glm_projector ? 2 : 0;
return (clip_n_patches(ctx) + extra_tokens) * clip_n_mmproj_embd(ctx) * sizeof(float);
return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
}

size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w) {
Expand Down Expand Up @@ -2790,9 +2785,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
}
if (ctx->has_glm_projector) {
GGML_ASSERT(batch_size == 1);
ggml_tensor * boi = ctx->vision_model.boi_w;
ggml_backend_tensor_get(boi,vec,0,ggml_nbytes(boi));
vec = (float*)(vec+ggml_nelements(boi)); //offset for boi
}

// build the inference graph
Expand Down Expand Up @@ -3001,13 +2993,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
// copy the embeddings to the location passed by the user
ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));

if (ctx->has_glm_projector) {
//eoi
ggml_tensor * eoi = ctx->vision_model.eoi_w;
int offset = ggml_nelements(embeddings);
ggml_backend_tensor_get(eoi, vec+offset, 0, ggml_nbytes(eoi));
}

return true;
}

Expand Down
5 changes: 5 additions & 0 deletions examples/llava/mtmd.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -186,6 +186,11 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
marker_modified = "<start_of_image>" + ctx->image_marker + "<end_of_image>";
string_replace_all(prompt_modified, ctx->image_marker, marker_modified);

} else if (proj_type == PROJECTOR_TYPE_GLM_EDGE) {
// <|begin_of_image|> ... (image embeddings) ... <|end_of_image|>
marker_modified = "<|begin_of_image|>" + ctx->image_marker + "<|end_of_image|>";
string_replace_all(prompt_modified, ctx->image_marker, marker_modified);

} else if (proj_type == PROJECTOR_TYPE_IDEFICS3) {
// https://github.com/huggingface/transformers/blob/a42ba80fa520c784c8f11a973ca9034e5f859b79/src/transformers/models/idefics3/processing_idefics3.py#L192-L215
marker_modified = "<fake_token_around_image><global-img>" + ctx->image_marker + "<fake_token_around_image>";
Expand Down
Loading