Skip to content

vllm.parser.granite_toolcall

Granite 4.2 tool-call parser.

Granite 4.2 emits tool calls in an XML format::

<tool_call>
<function=func_name>
<parameter=key>value</parameter>
</function>
</tool_call>

Granite 4.2 shares the Nemotron V3 tool-call and reasoning format, so this parser subclasses :class:NemotronV3Parser to reuse its grammar, argument converter and streaming logic, exposed as a standalone --tool-call-parser.

It deliberately does NOT subclass :class:GraniteThinkingParser: that parser strips a leading newline from content (to clean up the chat template's \n</think>\n markup), which is a reasoning-side concern. A tool parser running on its own (no reasoning parser) must leave content untouched, so the leading-newline handling stays in the reasoning parser where it belongs.

Because the tool-call extraction is inherited unchanged, fixes to the shared grammar flow in automatically. The parser works on its own, with any reasoning parser, or with none, in both thinking modes.

Classes:

GraniteToolCallParser

Bases: NemotronV3Parser

Granite 4.2 tool-call parser.

Subclass of :class:NemotronV3Parser (whose tool-call and reasoning format Granite 4.2 shares) exposed as a standalone tool parser. The tool adapter generated from this class extracts tool calls through the shared grammar; reasoning/content splitting, when enabled, is handled by the separately configured reasoning parser.

Source code in vllm/parser/granite_toolcall.py 
33
34
35
36
37
38
39
40
41
class GraniteToolCallParser(NemotronV3Parser):
    """Granite 4.2 tool-call parser.

    Subclass of :class:`NemotronV3Parser` (whose tool-call and reasoning format
    Granite 4.2 shares) exposed as a standalone tool parser. The tool adapter
    generated from this class extracts tool calls through the shared grammar;
    reasoning/content splitting, when enabled, is handled by the separately
    configured reasoning parser.
    """