CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

chat : support Magistral thinking (#16413)

Browse Source 
* feat: added a dedicated Magistral chat format that preserves [THINK] spans, parses reasoning before tool calls

* feat: new flow in the chat template test suite for Magistral
This commit is contained in:
Pascal
2025-10-03 20:51:48 +02:00
committed by GitHub
parent f6dcda3900
commit 128d522c04
3 changed files with 92 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
tests/test-chat.cpp
View File

@@ -411,6 +411,7 @@ const common_chat_msg message_assist_thoughts_unparsed_md = simple_assis
const common_chat_msg message_assist_thoughts_unparsed_md_partial = simple_assist_msg("<think>I'm\nthinking</think>Hello, world!\nWhat's up?\n```json\n{}");
const common_chat_msg message_assist_thoughts_unparsed_r7b = simple_assist_msg("<|START_THINKING|>I'm\nthinking<|END_THINKING|>Hello, world!\nWhat's up?");
const common_chat_msg message_assist_thoughts_unparsed_magistral = simple_assist_msg("[THINK]raisonnement[/THINK]Réponse");
const common_chat_msg message_assist_thoughts = simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking");
const common_chat_msg message_assist_thoughts_unopened_unparsed = simple_assist_msg("I'm\nthinking</think>Hello, world!\nWhat's up?");
const common_chat_msg message_assist_thoughts_no_content = simple_assist_msg("", "I'm\nthinking");
@@ -745,6 +746,17 @@ static void test_template_output_parsers() {
tmpls.get(), end_tokens, message_assist_call_id, tools,
"[TOOL_CALLS][{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}, \"id\": \"123456789\"}]");
}
{
assert_msg_equals(
simple_assist_msg("Réponse", "raisonnement"),
common_chat_parse(
message_assist_thoughts_unparsed_magistral.content,
/* is_partial= */ false,
{
/* .format = */ COMMON_CHAT_FORMAT_MAGISTRAL,
/* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
}));
}
{
auto tmpls = read_templates("models/templates/Qwen-QwQ-32B.jinja");
std::vector<std::string> end_tokens{ "<|im_end|>" };