tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-30 12:55:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

mtmd : support SmolVLM (version 1 and 2) (#13050)

Browse Source 
* mtmd : support SmolVLM (version 1 and 2)

* correct chat template

* fix n_patches

* scale_factor is an int

* add more models to test
This commit is contained in:
Xuan-Son Nguyen
2025-04-22 16:24:54 +02:00
committed by GitHub
parent ab47dec3d3
commit dc39a5e7a8
10 changed files with 279 additions and 65 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
examples/llava/mtmd.cpp
View File

@ -176,6 +176,8 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
std::string prompt_modified(text.text);
std::string marker_modified(ctx->image_marker);
projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
// a bit hacky here, but works for now
// for some models, we need to add prefix and suffix to the image embeddings
if (clip_is_gemma3(ctx->ctx_clip)) {
@ -183,6 +185,11 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
// <start_of_image> ... (image embeddings) ... <end_of_image>
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_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>";
string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
}
// llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix