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 : nemotron thinking & toolcalling support (#15676)

Browse Source 
* feat: nemotron thinking & toolcalling support

* Trailing whitespaces

* Corrected template for Nemotron

* Template and parser fixes

* Final template and grammar changes

* Whitespace

* Always do lazy grammar processing since </think> tag will always be there.

* Allow extra content after toolcall

* Whitespace

* New tests: thinking + tools, tools + content, thinking + tools + content (new!)

* Whitespace

* Remove cURL test script
This commit is contained in:
Piotr Wilkin (ilintar)
2025-09-05 01:22:22 +02:00
committed by GitHub
parent 9e2b1e83c6
commit b2426e469e
4 changed files with 333 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

73
tests/test-chat.cpp
View File

@@ -420,6 +420,7 @@ const common_chat_msg message_assist_call_empty_args = simple_assist
const common_chat_msg message_assist_call_cutoff_args = simple_assist_msg("", "", "special_function", "{\"arg");
const common_chat_msg message_assist_call_thoughts = simple_assist_msg("", "I'm\nthinking", "special_function", "{\"arg1\":1}");
const common_chat_msg message_assist_call_thoughts_unparsed = simple_assist_msg("<think>I'm\nthinking</think>\n\n", "", "special_function", "{\"arg1\": 1}");
const common_chat_msg message_assist_call_thoughts_content = simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking", "special_function", "{\"arg1\": 1}");
const common_chat_msg message_assist_call_id = simple_assist_msg("", "", "special_function", "{\"arg1\":1}", /* .id = */ "123456789");
const common_chat_msg message_assist_call_idx = simple_assist_msg("", "", "special_function", "{\"arg1\":1}", /* .id = */ "0");
const common_chat_msg message_assist_thoughts_call_idx = simple_assist_msg("", "I'm\nthinking", "special_function", "{\"arg1\": 1}", /* id = */ "0");
@@ -436,6 +437,7 @@ static void test_msgs_oaicompat_json_conversion() {
message_assist_call,
message_assist_call_thoughts,
message_assist_call_thoughts_unparsed,
message_assist_call_thoughts_content,
message_assist_call_id,
message_assist_call_idx,
message_assist_call_python,
@@ -1755,6 +1757,77 @@ static void test_template_output_parsers() {
/* is_partial= */ false,
{COMMON_CHAT_FORMAT_SEED_OSS}));
}
{
auto tmpls = read_templates("models/templates/NVIDIA-Nemotron-Nano-v2.jinja");
std::vector<std::string> end_tokens{ "<SPECIAL_12>" };
assert_equals(COMMON_CHAT_FORMAT_NEMOTRON_V2, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
assert_equals(COMMON_CHAT_FORMAT_NEMOTRON_V2, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
// Test parsing regular content
assert_msg_equals(message_assist,
common_chat_parse(
"Hello, world!\nWhat's up?",
/* is_partial= */ false,
{COMMON_CHAT_FORMAT_NEMOTRON_V2}));
// Test parsing content with thinking
assert_msg_equals(message_assist_thoughts,
common_chat_parse(
"<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
/* is_partial= */ false,
{
/* .format = */ COMMON_CHAT_FORMAT_NEMOTRON_V2,
/* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
}));
// Test parsing tool calls
assert_msg_equals(message_assist_call,
common_chat_parse(
"<TOOLCALL>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]</TOOLCALL>",
/* is_partial= */ false,
{COMMON_CHAT_FORMAT_NEMOTRON_V2}));
// Test parsing tool calls with thinking
assert_msg_equals(message_assist_call_thoughts,
common_chat_parse(
"<think>I'm\nthinking</think><TOOLCALL>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]</TOOLCALL>",
/* is_partial= */ false,
{
/* .format = */ COMMON_CHAT_FORMAT_NEMOTRON_V2,
/* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK
}));
// Test tool calls with extra content
assert_msg_equals(message_assist_call_content,
common_chat_parse(
"<TOOLCALL>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]</TOOLCALL>Hello, world!\nWhat's up?",
/* is_partial= */ false,
{COMMON_CHAT_FORMAT_NEMOTRON_V2}
));
// Test tool calls with extra content AND thinking
assert_msg_equals(message_assist_call_thoughts_content,
common_chat_parse(
"<think>I'm\nthinking</think><TOOLCALL>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]</TOOLCALL>Hello, world!\nWhat's up?",
/* is_partial= */ false,
{
/* .format = */ COMMON_CHAT_FORMAT_NEMOTRON_V2,
/* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK
}));
// Test template generation for regular content
test_templates(tmpls.get(), end_tokens, message_assist, tools,
"Hello, world!\nWhat's up?\n",
/* expect_grammar_triggered= */ false);
// Test template generation for tool calls
test_templates(tmpls.get(), end_tokens, message_assist_call, tools,
"<TOOLCALL>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]</TOOLCALL>",
/* expect_grammar_triggered= */ true
);
}
}
static void test_msg_diffs_compute() {