Skip to content

chat: Fix streaming parser for granite models #15682

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 2 commits into from
Sep 19, 2025
+114 −13
Merged

chat: Fix streaming parser for granite models #15682

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
836e859
fix(chat): fix streaming parser for granite models
shun095 Aug 30, 2025
f16589b
tests: add test cases for Granite models chat parser
shun095 Aug 31, 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
36 changes: 23 additions & 13 deletions common/chat.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2025,15 +2025,28 @@ static common_chat_params common_chat_params_init_granite(const common_chat_temp

static void common_chat_parse_granite(common_chat_msg_parser & builder) {
// Parse thinking tags
static const common_regex start_think_regex(regex_escape("<think>"));
static const common_regex end_think_regex(regex_escape("</think>"));
// Granite models output partial tokens such as "<" and "<think".
// By leveraging try_consume_regex()/try_find_regex() throwing
// common_chat_msg_partial_exception for these partial tokens,
// processing is interrupted and the tokens are not passed to add_content().
if (auto res = builder.try_consume_regex(start_think_regex)) {
Copy link
Contributor Author

@shun095 shun095 Aug 30, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please let me know if there are some better way to implement this.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This isn't a corner of the code I'm intimately familiar with, but I see similar usage elsewhere to parse tool calls that may be partial, so I think this looks like the right approach.

// Restore position for try_parse_reasoning()
builder.move_to(res->groups[0].begin);
builder.try_find_regex(end_think_regex, std::string::npos, false);
// Restore position for try_parse_reasoning()
builder.move_to(res->groups[0].begin);
}
builder.try_parse_reasoning("<think>", "</think>");

// Parse response tags using regex
static const common_regex response_regex("<response>([\\s\\S]*?)</response>");
if (auto res = builder.try_find_regex(response_regex)) {
// Extract the content between the tags (capture group 1)
auto content = builder.str(res->groups[1]);
builder.add_content(content);
builder.move_to(res->groups[0].end);
// Parse response tags
static const common_regex start_response_regex(regex_escape("<response>"));
static const common_regex end_response_regex(regex_escape("</response>"));
// Granite models output partial tokens such as "<" and "<response".
// Same hack as reasoning parsing.
if (builder.try_consume_regex(start_response_regex)) {
builder.try_find_regex(end_response_regex);
}

if (!builder.syntax().parse_tool_calls) {
Expand All @@ -2047,13 +2060,10 @@ static void common_chat_parse_granite(common_chat_msg_parser & builder) {
builder.move_to(res->groups[0].end);

// Expect JSON array of tool calls
auto tool_calls_data = builder.consume_json();
if (tool_calls_data.json.is_array()) {
if (!builder.add_tool_calls(tool_calls_data.json)) {
builder.add_content("<|tool_call|>" + tool_calls_data.json.dump());
if (auto tool_call = builder.try_consume_json_with_dumped_args({{{"arguments"}}})) {
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Some of the earlier 3-series granite models have different tool calling behavior. I'm trying to verify whether any of them would return a single tool-call object (versus an array), but we may need to retain the clause that checks is_array.

if (!builder.add_tool_calls(tool_call->value) || tool_call->is_partial) {
throw common_chat_msg_partial_exception("incomplete tool call");
}
} else {
builder.add_content("<|tool_call|>" + tool_calls_data.json.dump());
}
} else {
builder.add_content(builder.consume_rest());
Expand Down
91 changes: 91 additions & 0 deletions tests/test-chat.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1400,6 +1400,12 @@ static void test_template_output_parsers() {
"Hello, world!\nWhat's up?",
/* is_partial= */ false,
{COMMON_CHAT_FORMAT_GRANITE}));
assert_msg_equals(
message_assist,
common_chat_parse(
"Hello, world!\nWhat's up?",
/* is_partial= */ true,
{COMMON_CHAT_FORMAT_GRANITE}));

// Test parsing content with thinking
assert_msg_equals(message_assist_thoughts,
Expand All @@ -1410,13 +1416,98 @@ static void test_template_output_parsers() {
/* .format = */ COMMON_CHAT_FORMAT_GRANITE,
/* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
}));
assert_msg_equals(message_assist_thoughts_unparsed_deepseek,
common_chat_parse(
"<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
/* is_partial= */ false,
{COMMON_CHAT_FORMAT_GRANITE}));
assert_msg_equals(message_assist_thoughts,
common_chat_parse(
"<think>I'm\nthinking</think><response>Hello, world!\nWhat's up?",
/* is_partial= */ true,
{
/* .format = */ COMMON_CHAT_FORMAT_GRANITE,
/* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
}));
assert_msg_equals(message_assist_thoughts,
common_chat_parse(
"<think>I'm\nthinking</think><response>Hello, world!\nWhat's up?</response>",
/* is_partial= */ false,
{
/* .format = */ COMMON_CHAT_FORMAT_GRANITE,
/* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
}));
assert_msg_equals(simple_assist_msg("<think>I'm\nthinking</think><response>Hello, world!\nWhat's up?</response>"),
common_chat_parse(
"<think>I'm\nthinking</think><response>Hello, world!\nWhat's up?</response>",
/* is_partial= */ false,
{COMMON_CHAT_FORMAT_GRANITE}));
assert_msg_equals(message_assist_empty,
common_chat_parse(
"<think",
/* is_partial= */ true,
{
/* .format = */ COMMON_CHAT_FORMAT_GRANITE,
/* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
}));
assert_msg_equals(message_assist_empty,
common_chat_parse(
"<think",
/* is_partial= */ true,
{COMMON_CHAT_FORMAT_GRANITE}));
assert_msg_equals(message_assist_thoughts_no_content,
common_chat_parse(
"<think>I'm\nthinking",
/* is_partial= */ true,
{
/* .format = */ COMMON_CHAT_FORMAT_GRANITE,
/* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
}));
assert_msg_equals(
message_assist_empty,
common_chat_parse(
"<think>I'm\nthinking</think><response",
/* is_partial= */ true,
{COMMON_CHAT_FORMAT_GRANITE}));

// Test parsing tool calls
assert_msg_equals(message_assist_call,
common_chat_parse(
"<|tool_call|>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]",
/* is_partial= */ false,
{COMMON_CHAT_FORMAT_GRANITE}));
assert_msg_equals(
message_assist_call_empty_args,
common_chat_parse(
"<|tool_call|>[{\"name\": \"special_function\"",
/* is_partial= */ true,
{COMMON_CHAT_FORMAT_GRANITE}));
assert_msg_equals(
message_assist_call_cutoff_args,
common_chat_parse(
"<|tool_call|>[{\"name\": \"special_function\", \"arguments\": {\"arg",
/* is_partial= */ true,
{COMMON_CHAT_FORMAT_GRANITE}));
assert_msg_equals(
message_assist_call_cutoff_args,
common_chat_parse(
"<|tool_call|>[{\"name\": \"special_function\", \"arguments\": {\"arg",
/* is_partial= */ true,
{
/* .format = */ COMMON_CHAT_FORMAT_GRANITE,
/* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
}));

// Test parsing tool calls with thinking
assert_msg_equals(
message_assist_call_thoughts,
common_chat_parse(
"<think>I'm\nthinking</think><|tool_call|>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}, {",
/* is_partial= */ true,
{
/* .format = */ COMMON_CHAT_FORMAT_GRANITE,
/* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
}));

// Test template generation for regular content
test_templates(tmpls.get(), end_tokens, message_assist, tools,
Expand Down
Loading