ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

server: streaming of tool calls and thoughts when --jinja is on (#12379)

Browse source 
* add common_json w/ support for truncated json healing

* add common_chat_msg_diff

* partial common_chat_parse

* refactor parser w/ optionals

* server: wire chat diffs in stream mode

* fix trigger of thinking models (must happen after thoughts are closed)

* fix functionary v3.2 raw python!

* rename: common_chat_syntax (now contains format)

* rm common_regex.at_start

* don't return empty <think></think>

* accommodate yet another deepseek r1 distill fantasy syntax (`<｜tool▁calls｜>`)

* fix QwQ 32B tool call parsing after thoughts (hermes2)

* better logs for grammar triggers

* consume spaces after parse_json_tool_calls

* fix required tool calls w/ thinking models that have pre-opened thinking tags

* fix thinking model's initial trigger + test qwq's template

* run most test_tool_call tests in stream + non-stream modes

* make functionary v3.2 parsing more strict (differentiate first match from others)

* send final diff from server, to close off raw python arguments

* support partial content streaming in Generic mode

* tool-call: allow content prelude before hermes2 tool calls (for Qwen2.5)

* Update function-calling.md

* Update tool_bench.py

* chat-parser: remove input from exception (llm output may contain PII)

---------

Co-authored-by: ochafik <ochafik@google.com>
Co-authored-by: Olivier Chafik <ochafik@users.noreply.github.com>
This commit is contained in:
Olivier Chafik 2025-05-25 01:48:08 +01:00 • committed by GitHub
parent a2d02d5793
commit f5cd27b71d
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
23 changed files with 3245 additions and 1091 deletions
Download patch file Download diff file Expand all files Collapse all files

37
common/json-partial.h Normal file
View file

@ -0,0 +1,37 @@
#pragma once
#include <json.hpp>
// Healing marker (empty if the JSON was fully parsed / wasn't healed).
struct common_healing_marker {
// Raw marker.
std::string marker;
// Cutting the `common_json.json.dump()` string at the (only) occurrence of this marker should yield the original partial JSON string (modulo spaces / if it had the same dump format).
std::string json_dump_marker;
};
// Represents a parsed JSON object, with its optional healing marker (a JSON dump fragment that can be used to find the position of healing in the JSON dump string)
struct common_json {
nlohmann::ordered_json json;
common_healing_marker healing_marker;
};
// Parse the JSON string, healing (closing) any partial JSON if `healing_marker` is not empty.
//
// Healing completes partial JSON strings by adding a (possibly modified) healing marker, then whatever is needed to close the JSON.
// This allows to parse the resulting healed JSON string, yet be able to cut it again if needed at the healing marker.
// (this is used when parsing JSON outputs from the models, then crafting partial JSONs for the partial tool calls in OAI format).
//
// For instance, parsing `{` with a healing marker `foo` will produce a healed JSON `{"foo":1}`, w/ json_dump_marker = `"foo"` (which can be used to break the JSON again).
bool common_json_parse(
const std::string & input,
const std::string & healing_marker,
common_json & out);
// Parse the JSON string (see overload above), but advancing an iterator to the end of the input when the (potentially partial) parsing succeeds.
bool common_json_parse(
std::string::const_iterator & it,
const std::string::const_iterator & end,
const std::string & healing_marker,
common_json & out);