Skip to content

mtmd : fix glm-edge redundant token count #13139

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 5 commits into from
Apr 28, 2025
Merged

mtmd : fix glm-edge redundant token count #13139

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
b33c312
mtmd : fix glm-edge redundant token count
ngxson Apr 27, 2025
971aa25
Merge branch 'master' into xsn/mtmd_rm_glm_eoi_boi
ngxson Apr 27, 2025
4d01373
fix chat template
ngxson Apr 27, 2025
ed3a496
temporary disable GLMEdge test chat tmpl
ngxson Apr 28, 2025
cdcf561
Merge branch 'master' into xsn/mtmd_rm_glm_eoi_boi
ngxson Apr 28, 2025
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
10 changes: 1 addition & 9 deletions examples/llava/mtmd.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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();
Expand Down Expand Up @@ -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()) {
Expand Down Expand Up @@ -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,
{},
Expand Down
10 changes: 1 addition & 9 deletions src/llama-chat.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -447,7 +447,7 @@ int32_t llm_chat_apply_template(
if (add_ass) {
ss << "<|assistant|>";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4) {
} else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4 || tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
ss << "[gMASK]" << "<sop>";
for (auto message : chat) {
std::string role(message->role);
Expand All @@ -456,14 +456,6 @@ int32_t llm_chat_apply_template(
if (add_ass) {
ss << "<|assistant|>";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
for (auto message : chat) {
std::string role(message->role);
ss << "<|" << role << "|>" << "\n" << message->content;
}
if (add_ass) {
ss << "<|assistant|>";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_MINICPM) {
// MiniCPM-3B-OpenHermes-2.5-v2-GGUF
for (auto message : chat) {
Expand Down
17 changes: 9 additions & 8 deletions tests/test-chat-template.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -187,14 +187,15 @@ int main(void) {
/* .bos_token= */ "",
/* .eos_token= */ "",
},
{
/* .name= */ "GLMEdge",
/* .template_str= */ "{% for item in messages %}{% if item['role'] == 'system' %}<|system|>\n{{ item['content'] }}{% elif item['role'] == 'user' %}<|user|>\n{{ item['content'] }}{% elif item['role'] == 'assistant' %}<|assistant|>\n{{ item['content'] }}{% endif %}{% endfor %}<|assistant|>",
/* .expected_output= */ "<|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n I am an assistant <|user|>\nAnother question<|assistant|>",
/* .expected_output_jinja= */ "<|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n I am an assistant <|user|>\nAnother question<|assistant|>",
/* .bos_token= */ "",
/* .eos_token= */ "",
},
// TODO @ngxson : GLMEdge produces poor result without `[gMASK]<sop>`, so we're temporarily using GLM4 template for it. We should fix this in the future.
// {
// /* .name= */ "GLMEdge",
// /* .template_str= */ "{% for item in messages %}{% if item['role'] == 'system' %}<|system|>\n{{ item['content'] }}{% elif item['role'] == 'user' %}<|user|>\n{{ item['content'] }}{% elif item['role'] == 'assistant' %}<|assistant|>\n{{ item['content'] }}{% endif %}{% endfor %}<|assistant|>",
// /* .expected_output= */ "<|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n I am an assistant <|user|>\nAnother question<|assistant|>",
// /* .expected_output_jinja= */ "<|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n I am an assistant <|user|>\nAnother question<|assistant|>",
// /* .bos_token= */ "",
// /* .eos_token= */ "",
// },
{
/* .name= */ "MiniCPM-3B-OpenHermes-2.5-v2-GGUF",
/* .template_str= */ U8C("{% for message in messages %}{% if message['role'] == 'user' %}{{'<用户>' + message['content'].strip() + '<AI>'}}{% else %}{{message['content'].strip()}}{% endif %}{% endfor %}"),
Expand Down
Loading