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

mtmd : chat : Fix extra \n between text and media marker (#19595) 

* mtmd : chat : Fix extra \n between text and media marker

Thanks to @tugot17 for detecting and reporting the issue.

For vision models (e.g. LFM2.5-VL-1.6B and Qwen/Qwen3-VL-4B-Instruct) `llama-mtmd-cli` produces identical output to HF implementation.

However `llama-server` doesn't. I traced it down to extra newline
inserted after `<__media__>`.

This happens in `to_json_oaicompat`, that treats media markers as text
and joins all parts with `\n` separator.

PR introduces new type `media_marker` and uses it for media markers.
Extra logic is added to prevent insertion of newlines before and after
media markers.

With this change number of input tokens is identical to HF
implementation and as a result the output is also identical.

I explored other ways to address the issue
* remove completely `\n` between text parts in `to_json_oaicompat`
* merge text messages in server-common.cpp before sending them to `to_json_oaicompat`

Please propose alternative ways of fixing this issue.

* Refactor to use explicite per type ifs

* Update common/chat.cpp

Co-authored-by: Piotr Wilkin (ilintar) <piotr.wilkin@syndatis.com>

* Update common_chat_templates_apply_legacy

---------

Co-authored-by: Piotr Wilkin (ilintar) <piotr.wilkin@syndatis.com>
This commit is contained in:
Tarek Dakhran 2026-02-19 12:18:57 +01:00 committed by GitHub
parent 03fd9d3bb4
commit c5897995a7
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 17 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
common/chat.cpp
View File

@ -65,14 +65,25 @@ json common_chat_msg::to_json_oaicompat(bool concat_typed_text) const {
} else if (!content_parts.empty()) {
if (concat_typed_text) {
std::string text;
bool last_was_media_marker = false;
// join parts with newline, do not add newline before or after media markers
for (const auto & part : content_parts) {
if (part.type != "text") {
bool add_new_line = true;
if (part.type == "text") {
add_new_line = !last_was_media_marker && !text.empty();
last_was_media_marker = false;
} else if (part.type == "media_marker") {
add_new_line = false;
last_was_media_marker = true;
} else {
LOG_WRN("Ignoring content part type: %s\n", part.type.c_str());
continue;
}
if (!text.empty()) {
if (add_new_line) {
text += '\n';
}
text += part.text;
}
jmsg["content"] = text;
@ -319,7 +330,7 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
throw std::invalid_argument("Missing content part type: " + part.dump());
}
const auto & type = part.at("type");
if (type != "text") {
if (type != "text" && type != "media_marker") {
throw std::invalid_argument("Unsupported content part type: " + type.dump());
}
common_chat_msg_content_part msg_part;
@ -3307,7 +3318,7 @@ static common_chat_params common_chat_templates_apply_legacy(
for (const auto & msg : inputs.messages) {
auto content = msg.content;
for (const auto & part : msg.content_parts) {
if (part.type != "text") {
if (part.type != "text" && part.type != "media_marker") {
LOG_WRN("Ignoring non-text content part: %s\n", part.type.c_str());
continue;
}

6
tools/server/server-common.cpp
View File

@ -916,8 +916,7 @@ json oaicompat_chat_params_parse(
json image_url = json_value(p, "image_url", json::object());
handle_media(out_files, image_url, opt.media_path);
// replace this chunk with a marker
p["type"] = "text";
p["type"] = "media_marker";
p["text"] = mtmd_default_marker();
p.erase("image_url");
@ -938,8 +937,7 @@ json oaicompat_chat_params_parse(
// TODO: add audio_url support by reusing handle_media()
// replace this chunk with a marker
p["type"] = "text";
p["type"] = "media_marker";
p["text"] = mtmd_default_marker();
p.erase("input_audio");