ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 17:20:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Deepseek R1 function calls (more formats) (#652)

Browse Source 
* Implement function calling / tools for ik_llama.cpp for Kimi K2

* Implement basic tool choice

* Backport llama.cpp tool calls support

* Enhance function calls with improved chat parser and string utilities

- Add new chat.h/chat.cpp and chat-parser.h/chat-parser.cpp for better chat handling
- Improve function calls parsing with fallback to llama.cpp builder pattern
- Add string utility functions (starts_with, ends_with, find_partial_stop)
- Update README with function calls testing instructions
- Enhance Kimi K2 parser and function calls documentation
- Add comprehensive test suite for function calls
- Update CMakeLists.txt and Makefile for new components

* Enhance function calling with unified streaming and parser improvements

- Fix streaming content cleanup to prevent function syntax in output
- Unify content extraction patterns with llama.cpp approach
- Improve Kimi K2 parser robustness and partial content handling
- Add comprehensive test coverage for function call scenarios
- Optimize chat message parsing and diff computation

* Replace hardcoded values in kimi_k2_parser.hpp with named constants

- Add compile-time constants for all token format markers
- Add compile-time constants for XML format markers
- Add compile-time constants for simple format patterns
- Replace all hardcoded string literals with named constants
- Use compile-time length calculation to avoid manual counting
- Improve maintainability and reduce magic numbers throughout parser

* Fix duplicate common_chat_parse definition

- Remove duplicate implementation from chat-parser.cpp
- Keep single implementation in chat.cpp following llama.cpp patterns
- Resolves linker error: multiple definition of common_chat_parse

* Fix JSON assertion failure in function call parsing

- Add proper validation that 'function' field is an object before accessing nested keys
- Handle missing 'arguments' field gracefully with default "{}"
- Prevents crash when parsing malformed tool call JSON structures

* Add comprehensive Qwen3 XML tool calling support with unit tests

- Implement Qwen3 XML parser with <tool_call>{"name": "func", "arguments": {...}}</tool_call> format
- Add model detection and routing for Qwen3 vs Kimi-K2 formats
- Create 8 comprehensive unit tests covering parsing, streaming, error handling
- Fix token format cleaning bug in kimi_k2_parser.hpp processing order
- Remove progressive parsing code and related utilities
- Add tool injection support for Qwen3 format in server utils

* Add DeepSeek R1 function calling support with comprehensive unit tests

- Implement complete DeepSeek R1 tool call parsing in common_chat_parser.cpp
- Add DeepSeek R1 model detection and tool injection in deepseek_r1_tools.hpp
- Update function_calls.hpp with DeepSeek R1 integration and content extraction
- Update documentation to reflect support for Kimi-K2, Qwen3, and DeepSeek R1 models
- Add comprehensive unit tests for DeepSeek R1 reasoning, tool calls, and integration
- Port exact implementation patterns from original llama.cpp for compatibility

Key features:
- Native DeepSeek R1 format: <｜tool▁calls▁begin｜>function<｜tool▁sep｜>name```json{}```<｜tool▁call▁end｜><｜tool▁calls▁end｜>
- Reasoning content extraction from <think>...</think> tags
- Multiple tool calls support with separate call blocks
- Model detection for deepseek-r1, deepseek_r1 naming patterns
- Integration with incremental parsing and streaming support

* Add partial parsing support for JSON and regex

- json-partial.h/cpp: JSON partial parsing functionality
- regex-partial.h/cpp: Regex partial parsing functionality

* Add format_chat integration tests for Qwen3 tool injection

- Add test_qwen3_format_chat_integration() to validate tool injection pipeline
- Test tool injection conditions and system message enhancement
- Verify JSON formatting and anti-preamble instructions
- Add comprehensive test documentation

Tests confirm tool injection works correctly - conversational preamble
issue is not in ik_llama.cpp but likely in UI configuration.

* Fix Qwen3 tool call parsing - pass model name to parser

Server was not passing model name to parse_chat_message_incremental(),
causing Qwen3 to fall back to Kimi-K2 parser and return tool calls
as content instead of proper tool_calls array.

* Fix non-streaming path to use model-specific parsing

Non-streaming responses were hardcoded to use Kimi-K2 format,
causing Qwen3 XML tool calls to be returned as content instead
of proper tool_calls array. Now uses same model detection as
streaming path for consistency.

* Update Qwen3 function call handling in server and tests

- Enhanced server function call detection and response formatting
- Improved test coverage for Qwen3 tool call scenarios
- Refined XML parsing for better tool execution support

* Add DeepSeek-R1 function call parsing support

Implements comprehensive parsing for all 4 DeepSeek-R1 function call formats:
- Format 1: Standard function call syntax (already supported)
- Format 2: Alternative function call patterns (already supported)
- Format 3: Tools array format - function\n```json\n{"tools": [...]}
- Format 4: XML wrapped format - <tool_call>function</think>Name\n```json\n{...}```</tool_call>

Key changes:
- Added parse_deepseek_r1_tools_array() following original parse_prefixed_json_tool_call_array pattern
- Added parse_deepseek_r1_xml_wrapped() following Hermes-2-Pro XML wrapper patterns
- Integrated both parsers into exception handling chain for robust fallback
- Added comprehensive TDD test coverage for all formats
- Anonymized all confidential information while preserving functionality

Resolves tool_calls_count=0 issue where DeepSeek-R1 models generated valid tool calls
but server failed to parse them correctly.

* Update function_calls.md documentation for DeepSeek-R1 Format 4

- Added Format 4 (XML wrapped) documentation with examples
- Updated implementation notes with correct parser order (3→4→1→2)
- Marked all DeepSeek-R1 formats as working (July 2025 update)
- Updated test status for Format 3 and 4 as passing
- Added parse_deepseek_r1_xml_wrapped() function reference
- Corrected implementation file line numbers

* Fix merge conflict in test-function-calls.cpp

- Removed incomplete merge conflict marker from line 3027
- Ensured all tests compile and pass successfully
- All DeepSeek-R1 formats (1-4) working correctly
- All streaming and content cleaning tests passing
This commit is contained in:
Anton Sokolchenko
2025-08-07 07:15:57 +02:00
committed by GitHub
parent d65d5fe29e
commit f4051d9c3e
7 changed files with 695 additions and 125 deletions
Download Patch File Download Diff File Expand all files Collapse all files

85
common/chat-parser.cpp
View File

@@ -208,90 +208,11 @@ void common_chat_msg_parser::parse_generic_format() {
}
void common_chat_msg_parser::parse_deepseek_r1_format() {
// DeepSeek R1 format supports <think> tags for reasoning content
try_parse_reasoning("<think>", "</think>");
if (!syntax_.enable_tool_calls) {
add_content(consume_rest());
return;
}
// DeepSeek R1 tool call patterns from original llama.cpp
static const common_regex tool_calls_begin("(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)");
static const common_regex tool_calls_end("<｜tool▁calls▁end｜>");
static const common_regex function_regex("(?:<｜tool▁call▁begin｜>)?function<｜tool▁sep｜>([^\n]+)\n```json\n");
static const common_regex close_regex("```[\\s\\r\\n]*<｜tool▁call▁end｜>");
parse_deepseek_r1_tool_calls(tool_calls_begin, function_regex, close_regex, tool_calls_end);
// Delegate to the main chat.cpp function which has the corrected implementation
// This follows the original llama.cpp pattern where chat-parser delegates to chat.cpp
common_chat_parse_deepseek_r1(*this);
}
void common_chat_msg_parser::parse_deepseek_r1_tool_calls(
const common_regex & tool_calls_begin,
const common_regex & function_regex,
const common_regex & close_regex,
const common_regex & tool_calls_end) {
// Helper function to wrap code as JSON arguments (ported from original llama.cpp)
auto wrap_code_as_arguments = [this](const std::string & code) -> std::string {
std::string arguments;
if (is_partial_) {
arguments = (json {{"code", code + healing_marker_}}).dump();
auto idx = arguments.find(healing_marker_);
if (idx != std::string::npos) {
arguments.resize(idx);
}
} else {
arguments = (json {{"code", code}}).dump();
}
return arguments;
};
auto parse_tool_calls = [&]() {
size_t from = std::string::npos;
while (true) {
auto res = try_find_regex(function_regex, from);
if (res) {
// Extract function name from regex group 1
std::string name = str(res->groups[1]);
from = std::string::npos;
if (name.empty()) {
from = res->groups[0].begin + 1;
continue;
}
auto maybe_raw_python = name == "python";
if (input_[pos_] == '{' || !maybe_raw_python) {
if (auto arguments = try_consume_json_with_dumped_args({{}})) {
if (!add_tool_call(name, "", arguments->value) || arguments->is_partial) {
throw common_chat_msg_partial_exception("incomplete tool call");
}
try_consume_regex(close_regex);
}
continue;
}
if (maybe_raw_python) {
auto arguments = wrap_code_as_arguments(consume_rest());
if (!add_tool_call(name, "", arguments)) {
throw common_chat_msg_partial_exception("incomplete tool call");
}
return;
}
throw common_chat_msg_partial_exception("incomplete tool call");
}
break;
}
try_consume_regex(tool_calls_end);
consume_spaces();
add_content(consume_rest());
};
if (auto res = try_find_regex(tool_calls_begin)) {
parse_tool_calls();
} else {
add_content(consume_rest());
}
}
void common_chat_msg_parser::finish() {
// Any final processing can go here

7
common/chat-parser.h
View File

@@ -113,13 +113,6 @@ private:
void parse_deepseek_r1_format();
void parse_generic_format();
// DeepSeek R1 specific tool call parsing
void parse_deepseek_r1_tool_calls(
const common_regex & tool_calls_begin,
const common_regex & function_regex,
const common_regex & close_regex,
const common_regex & tool_calls_end);
// JSON parsing utilities (enhanced streaming support)
struct json_parse_result {

260
common/chat.cpp
View File

@@ -104,7 +104,103 @@ static void common_chat_parse_generic(common_chat_msg_parser & builder) {
}
}
static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
// Helper function from original llama.cpp
static std::string wrap_code_as_arguments(common_chat_msg_parser & builder, const std::string & code) {
std::string arguments;
if (builder.is_partial()) {
arguments = (json {{"code", code + builder.healing_marker()}}).dump();
auto idx = arguments.find(builder.healing_marker());
if (idx != std::string::npos) {
arguments.resize(idx);
}
} else {
arguments = (json {{"code", code}}).dump();
}
return arguments;
}
// Forward declaration
static void parse_deepseek_r1_tools_array(common_chat_msg_parser & builder);
static void parse_deepseek_r1_xml_wrapped(common_chat_msg_parser & builder);
// Helper function from original llama.cpp for parsing JSON tool calls
static void parse_json_tool_calls(
common_chat_msg_parser & builder,
const std::optional<common_regex> & block_open,
const std::optional<common_regex> & function_regex_start_only,
const std::optional<common_regex> & function_regex,
const common_regex & close_regex,
const std::optional<common_regex> & block_close,
bool allow_raw_python = false,
const std::function<std::string(const common_chat_msg_parser::find_regex_result & fres)> & get_function_name = nullptr) {
auto parse_tool_calls = [&]() {
size_t from = std::string::npos;
auto first = true;
while (true) {
auto res = function_regex_start_only && first
? builder.try_consume_regex(*function_regex_start_only)
: function_regex
? builder.try_find_regex(*function_regex, from)
: std::nullopt;
if (res) {
std::string name;
if (get_function_name) {
name = get_function_name(*res);
} else {
if (res->groups.size() < 2) {
from = res->groups[0].begin + 1;
continue;
}
name = builder.str(res->groups[1]);
}
first = false;
if (name.empty()) {
// get_function_name signalled us that we should skip this match and treat it as content.
from = res->groups[0].begin + 1;
continue;
}
from = std::string::npos;
auto maybe_raw_python = name == "python" && allow_raw_python;
if (builder.input()[builder.pos()] == '{' || !maybe_raw_python) {
if (auto arguments = builder.try_consume_json_with_dumped_args({{}})) {
if (!builder.add_tool_call(name, "", arguments->value) || arguments->is_partial) {
throw common_chat_msg_partial_exception("incomplete tool call");
}
builder.try_consume_regex(close_regex);
}
continue;
}
if (maybe_raw_python) {
auto arguments = wrap_code_as_arguments(builder, builder.consume_rest());
if (!builder.add_tool_call(name, "", arguments)) {
throw common_chat_msg_partial_exception("incomplete tool call");
}
return;
}
throw common_chat_msg_partial_exception("incomplete tool call");
}
break;
}
if (block_close) {
builder.try_consume_regex(*block_close);
}
builder.consume_spaces();
builder.add_content(builder.consume_rest());
};
if (block_open) {
if (auto res = builder.try_find_regex(*block_open)) {
parse_tool_calls();
} else {
builder.add_content(builder.consume_rest());
}
} else {
parse_tool_calls();
}
}
void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
builder.try_parse_reasoning("<think>", "</think>");
if (!builder.syntax().enable_tool_calls) {
builder.add_content(builder.consume_rest());
@@ -113,25 +209,159 @@ static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
static const common_regex tool_calls_begin("(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)");
static const common_regex tool_calls_end("<｜tool▁calls▁end｜>");
// Primary regex for correct format with separator
static const common_regex function_regex("(?:<｜tool▁call▁begin｜>)?function<｜tool▁sep｜>([^\n]+)\n```json\n");
// Fallback regex for format without separator (some models generate this)
static const common_regex function_regex_no_sep("(?:<｜tool▁call▁begin｜>)?function<([^>]+)>\n```json\n");
// Third regex for new format: just "function" with no markers
static const common_regex function_regex_simple("function\n```json\n");
static const common_regex close_regex("```[\\s\\r\\n]*<｜tool▁call▁end｜>");
static const common_regex close_regex_simple("```"); // For simple format without end markers
// Simplified tool calls parsing for DEEPSEEK_R1
if (auto res = builder.try_find_regex(tool_calls_begin)) {
while (auto func_res = builder.try_find_regex(function_regex)) {
auto function_name = builder.str(func_res->groups[1]);
auto args_json = builder.try_consume_json();
if (args_json) {
builder.add_tool_call(function_name, "", args_json->json.dump());
builder.try_consume_regex(close_regex);
} else {
throw common_chat_msg_partial_exception("incomplete tool call JSON");
}
// Check for the new tools array format first (no DeepSeek markers)
auto original_pos = builder.pos();
// First, try the tools array format for content like "function\n```json\n{"tools": [...]}"
if (builder.try_find_regex(function_regex_simple)) {
builder.move_to(original_pos);
try {
parse_deepseek_r1_tools_array(builder);
return; // Success, we're done
} catch (const common_chat_msg_partial_exception&) {
// Fall through to try standard DeepSeek patterns
}
builder.try_consume_regex(tool_calls_end);
builder.add_content(builder.consume_rest());
}
// If tools array format didn't work, try XML-wrapped format
builder.move_to(original_pos);
try {
parse_deepseek_r1_xml_wrapped(builder);
return; // Success, we're done
} catch (const common_chat_msg_partial_exception&) {
// Fall through to try standard DeepSeek patterns
}
// If XML wrapper format didn't work, try standard DeepSeek patterns
builder.move_to(original_pos);
try {
parse_json_tool_calls(
builder,
/* block_open= */ tool_calls_begin,
/* function_regex_start_only= */ std::nullopt,
function_regex,
close_regex,
tool_calls_end);
} catch (const common_chat_msg_partial_exception&) {
// If primary regex fails and we're not in partial mode, try fallback regex
if (!builder.is_partial()) {
builder.move_to(original_pos);
try {
parse_json_tool_calls(
builder,
/* block_open= */ tool_calls_begin,
/* function_regex_start_only= */ std::nullopt,
function_regex_no_sep,
close_regex,
tool_calls_end);
} catch (const common_chat_msg_partial_exception&) {
// Try the simple format without markers as final fallback
builder.move_to(original_pos);
parse_json_tool_calls(
builder,
/* block_open= */ std::nullopt,
/* function_regex_start_only= */ std::nullopt,
function_regex_simple,
close_regex_simple,
std::nullopt);
}
} else {
throw; // Re-throw for partial mode
}
}
}
// Parse DeepSeek R1 tools array format following original llama.cpp parse_prefixed_json_tool_call_array pattern
static void parse_deepseek_r1_tools_array(common_chat_msg_parser & builder) {
static const common_regex prefix("function\n```json\n");
if (auto res = builder.try_find_regex(prefix)) {
// Parse JSON and manually process tools array to convert arguments to strings
auto json_result = builder.try_consume_json();
if (!json_result) {
throw common_chat_msg_partial_exception("invalid JSON");
}
// DeepSeek R1 format has "tools" array, manually process each tool
if (json_result->json.contains("tools") && json_result->json.at("tools").is_array()) {
// Manually create tool calls array with string arguments (following original pattern)
json tools_with_dumped_args = json::array();
for (const auto& tool : json_result->json.at("tools")) {
if (tool.contains("name") && tool.contains("arguments")) {
json formatted_tool;
formatted_tool["name"] = tool.at("name");
// Convert arguments object to string (this is what consume_json_with_dumped_args does)
formatted_tool["arguments"] = tool.at("arguments").dump();
tools_with_dumped_args.push_back(formatted_tool);
}
}
if (!builder.add_tool_calls(tools_with_dumped_args) || !json_result->healing_marker.marker.empty()) {
throw common_chat_msg_partial_exception("incomplete tool call array");
}
} else {
throw common_chat_msg_partial_exception("tools key not found or not array");
}
// Consume closing ```
builder.try_consume_regex(common_regex("```"));
} else {
builder.add_content(builder.consume_rest());
throw common_chat_msg_partial_exception("function prefix not found");
}
}
// Parse DeepSeek R1 XML-wrapped format following original Hermes-2-Pro pattern
static void parse_deepseek_r1_xml_wrapped(common_chat_msg_parser & builder) {
// Pattern for: <tool_call>\nfunction</think>FunctionName\n```json\n{...}\n```\n</tool_call>
static const common_regex xml_pattern(
"<tool_call>\\s*" // Opening XML tag
"function</think>([^\\n]+)" // Function name after "function</think>"
"\\s*```json\\s*" // JSON block start
);
if (auto res = builder.try_find_regex(xml_pattern)) {
// Extract function name from capture group
std::string function_name = builder.str(res->groups[1]);
// Parse JSON arguments
auto json_result = builder.try_consume_json();
if (!json_result) {
throw common_chat_msg_partial_exception("invalid JSON in XML wrapper");
}
// Create single tool call following original pattern
json tool_call;
tool_call["name"] = function_name;
tool_call["arguments"] = json_result->json.dump(); // Convert to string
json tool_calls_array = json::array();
tool_calls_array.push_back(tool_call);
if (!builder.add_tool_calls(tool_calls_array) || !json_result->healing_marker.marker.empty()) {
throw common_chat_msg_partial_exception("incomplete XML wrapped tool call");
}
// Consume closing ```\n</tool_call>
builder.try_consume_regex(common_regex("```\\s*</tool_call>"));
} else {
throw common_chat_msg_partial_exception("XML wrapper pattern not found");
}
}

3
common/chat.h
View File

@@ -162,3 +162,6 @@ common_chat_msg common_chat_parse(const std::string & input, bool is_partial, co
// Forward declare parser class
class common_chat_msg_parser;
// Format-specific parsing functions (accessible from chat-parser)
void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder);