Skip to content

LLM inference in C/C++

C++ ggml

chat : Fix Muse Glimmer tool-call first parser error (#29242)

Nicolas Mowen7ab4ee7baad2
diff --git a/common/parsers/muse-glimmer.cpp b/common/parsers/muse-glimmer.cpp
index add95697eb..d03bf2d58f 100644
--- a/common/parsers/muse-glimmer.cpp
+++ b/common/parsers/muse-glimmer.cpp
@@ -130,7 +130,7 @@ common_chat_params common_chat_params_init_muse_glimmer(const common_chat_templa
});
data.grammar_triggers = {
{ COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
- "<\\|start\\|>assistant( to=(?!self<\\|message\\|>)(?!user<\\|message\\|>)[^<]*?<\\|message\\|>)" },
+ "(?:^|<\\|start\\|>assistant)( to=(?!self<\\|message\\|>)(?!user<\\|message\\|>)[^<]*?<\\|message\\|>)" },
};
}
diff --git a/tests/test-chat.cpp b/tests/test-chat.cpp
index 13733c1d9c..f2728f7cca 100644
--- a/tests/test-chat.cpp
+++ b/tests/test-chat.cpp
@@ -6357,6 +6357,14 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
.expect(message_assist)
.run();
+ // A tool call as the first message of the turn: "<|start|>assistant" is the
+ // generation prompt, so the output starts at " to=".
+ tst.test(" to=special_function<|message|>" + call_markup)
+ .tools({ special_function_tool })
+ .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+ .expect(message_assist_call)
+ .run();
+
// "Inform then act": the model answers the user and calls a tool in ONE generation,
// closing the answer with <|eom|>. The answer must stop there rather than swallow it.
tst.test(" to=user<|message|>Hello, world!\nWhat's up?<|eom|>"