CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-30 08:42:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

mtmd : fix glm-edge redundant token count (#13139)

Browse Source 
* mtmd : fix glm-edge redundant token count

* fix chat template

* temporary disable GLMEdge test chat tmpl
This commit is contained in:
Xuan-Son Nguyen
2025-04-28 16:12:56 +02:00
committed by GitHub
parent fb0471d175
commit 4e87962e34
3 changed files with 11 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
examples/llava/mtmd.cpp
View File

@@ -203,9 +203,6 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
}
// llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix
// for glm-edge, we don't need to add because the tokens are already in the returned embeddings
// TODO @ngxson : glm-edge : remove BOI / EOI tokens embeddings, decode them as normal tokens
std::vector<std::string> parts = string_split_str(prompt_modified, ctx->image_marker);
output.clear();
@@ -246,7 +243,7 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
};
for (const auto & part : parts) {
//printf("tokenizing part: %s\n", part.c_str());
// printf("tokenizing part: %s\n", part.c_str());
bool add_bos = &parts.front() == &part;
auto tokens = mtmd_tokenize_text_internal(vocab, part, text.add_special && add_bos, text.parse_special);
if (tokens.empty()) {
@@ -338,11 +335,6 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
LOG_DBG("image_tokens->ny = %d\n", image_tokens->ny);
LOG_DBG("batch_f32 size = %d\n", (int)image_tokens->batch_f32.entries.size());
if (clip_is_glm(ctx->ctx_clip)) {
// glm-edge
image_tokens->nx += 2; // add 2 for the begin_of_image and end_of_image token embeddings
}
mtmd_input_chunk chunk{
MTMD_INPUT_CHUNK_TYPE_IMAGE,
{},