Skip to content

Commit 1920345

Browse files
hksdpc255CISC
andauthored
common : Generalized XML-style tool-call parsing with streaming support (GLM 4.5/4.6 + MiniMax M2 + SeedOSS + Kimi-K2 + Qwen3-Coder + Apriel-1.5 + Xiaomi-MiMo) (ggml-org#16932)
* Add files via upload * fix unit test * fix crashes for --reasoning-format=none * Patch buggy official MiniMax-M2 chat template * add upstream minja fix: ochafik/minja#7 * Fix <think> token not generated * add test copied from ggml-org#16946 * cleanup * Hopes to fix the compilation error on CI * Delete chat template patching since it’s fixed by upstream Minja * Remove undeeded Minimax-M2 template patch ochafik/minja#7 (comment) * Add proper handling of optional parameters with test merged tests from: ggml-org@23d4bb7 * Fix making all tool parameters optional * Move xml tool parser to separate file * cleanup & add tests for GLM4.5 * add streaming tests & enhancement & cleanups Add streaming test for both GLM 4.5 and minimax-m2. Cleanup for preserved_tokens. Cleanup for grammar rule name. Enhance the parser's stability. * cleanup & add support for Kimi-K2 Qwen3-Coder Apriel-1.5 Xiaomi-MiMo * apply suggestions from reviewers * fix a misuse for data.grammar_lazy * fix grammar when tool have no argument * Fix `no triggers set for lazy grammar!` for GLM4.5/4.6. Insert additional stops for Kimi-K2 * update chat.cpp * fix grammar for GLM 4.5/4.6 * Try fix Jinja template for GLM * Try fix GLM-4.6.jinja * Update common/chat-parser-xml-toolcall.cpp Co-authored-by: Sigbjørn Skjæret <[email protected]> * Update tests/test-chat.cpp Co-authored-by: Sigbjørn Skjæret <[email protected]> * improve chat template for GLM, rename Kimi-K2 template to Kimi-K2-Thinking * Improve Kimi-K2 chat template * Fix unit test * Fix "Invalid tool call arguments passed" in a rare case. In a rare case, the model may emit a raw string that begins with a valid JSON string. This commit adds unit tests to cover that scenario and fixes the regression introduced during the Kimi-K2 adaptation. --------- Co-authored-by: Sigbjørn Skjæret <[email protected]>
1 parent 561a3e2 commit 1920345

17 files changed

+3171
-106
lines changed

common/CMakeLists.txt

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -50,6 +50,8 @@ add_library(${TARGET} STATIC
5050
base64.hpp
5151
chat-parser.cpp
5252
chat-parser.h
53+
chat-parser-xml-toolcall.h
54+
chat-parser-xml-toolcall.cpp
5355
chat.cpp
5456
chat.h
5557
common.cpp

common/chat-parser-xml-toolcall.cpp

Lines changed: 861 additions & 0 deletions
Large diffs are not rendered by default.

common/chat-parser-xml-toolcall.h

Lines changed: 45 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,45 @@
1+
#pragma once
2+
3+
#include "chat.h"
4+
5+
#include <nlohmann/json.hpp>
6+
7+
#include <optional>
8+
#include <string>
9+
#include <vector>
10+
11+
12+
// Sample config:
13+
// MiniMax-M2 (left): <minimax:tool_call>\n<invoke name="tool-name">\n<parameter name="key">value</parameter>\n...</invoke>\n...</minimax:tool_call>
14+
// GLM 4.5 (right): <tool_call>function_name\n<arg_key>key</arg_key>\n<arg_value>value</arg_value>\n</tool_call>
15+
struct xml_tool_call_format {
16+
std::string scope_start; // <minimax:tool_call>\n // \n // can be empty
17+
std::string tool_start; // <invoke name=\" // <tool_call>
18+
std::string tool_sep; // \">\n // \n // can be empty only for parse_xml_tool_calls
19+
std::string key_start; // <parameter name=\" // <arg_key>
20+
std::string key_val_sep; // \"> // </arg_key>\n<arg_value>
21+
std::string val_end; // </parameter>\n // </arg_value>\n
22+
std::string tool_end; // </invoke>\n // </tool_call>\n
23+
std::string scope_end; // </minimax:tool_call> // // can be empty
24+
// Set this if there can be dynamic spaces inside key_val_sep.
25+
// e.g. key_val_sep=</arg_key> key_val_sep2=<arg_value> for GLM4.5
26+
std::optional<std::string> key_val_sep2 = std::nullopt;
27+
// Set true if argval should only be raw string. e.g. Hello "world" hi
28+
// Set false if argval should only be json string. e.g. "Hello \"world\" hi"
29+
// Defaults to std::nullopt, both will be allowed.
30+
std::optional<bool> raw_argval = std::nullopt;
31+
std::optional<std::string> last_val_end = std::nullopt;
32+
std::optional<std::string> last_tool_end = std::nullopt;
33+
bool trim_raw_argval = false;
34+
bool allow_toolcall_in_think = false; // TODO: UNTESTED!!!
35+
};
36+
37+
// make a GBNF that accept any strings except those containing any of the forbidden strings.
38+
std::string make_gbnf_excluding(std::vector<std::string> forbids);
39+
40+
/**
41+
* Build grammar for xml-style tool call
42+
* form.scope_start and form.scope_end can be empty.
43+
* Requires data.format for model-specific hacks.
44+
*/
45+
void build_grammar_xml_tool_call(common_chat_params & data, const nlohmann::ordered_json & tools, const struct xml_tool_call_format & form);

common/chat-parser.h

Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,7 @@
11
#pragma once
22

33
#include "chat.h"
4+
#include "chat-parser-xml-toolcall.h"
45
#include "json-partial.h"
56
#include "regex-partial.h"
67

@@ -119,5 +120,14 @@ class common_chat_msg_parser {
119120
const std::vector<std::vector<std::string>> & content_paths = {}
120121
);
121122

123+
/**
124+
* Parse XML-Style tool call for given xml_tool_call_format. Return false for invalid syntax and get the position untouched.
125+
* form.scope_start, form.tool_sep and form.scope_end can be empty.
126+
*/
127+
bool try_consume_xml_tool_calls(const struct xml_tool_call_format & form);
128+
129+
// Parse content uses reasoning and XML-Style tool call
130+
void consume_reasoning_with_xml_tool_calls(const struct xml_tool_call_format & form, const std::string & start_think = "<think>", const std::string & end_think = "</think>");
131+
122132
void clear_tools();
123133
};

0 commit comments

Comments
 (0)