iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

[Bugfix][Parser] Prevent Inkling block-end leakage with tools (#51391)

Browse Source 
Signed-off-by: cherry77-cloud <1615405@qq.com>
Co-authored-by: Ben Browning <56071+bbrowning@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
cherry77-cloud
2026-08-08 12:34:32 -04:00
committed by GitHub
co-authored by Ben Browning Claude Opus 4.8
parent 58d3918e3e
commit 75231eff2f
3 changed files with 352 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files
vllm/parser/engine/streaming_parser_engine.py
+55 -30
View File
@@ -156,6 +156,13 @@ class StreamingParserEngine:
for (state, terminal), tr in config.transitions.items()
if tr.next_state in self._TOOL_STATES or state in self._TOOL_STATES
)
# TOOL_CALL_END may close an inner call rather than its lexical wrapper,
# as in MiniMax, so identify exits from state transitions instead.
self._tool_exit_terminals: frozenset[str] = frozenset(
terminal
for (state, terminal), tr in config.transitions.items()
if state in self._TOOL_STATES and tr.next_state not in self._TOOL_STATES
)
self.skip_tool_parsing = False
self.reset(initial_state=initial_state)
@@ -186,6 +193,7 @@ class StreamingParserEngine:
self._scanner.reset()
self._lexer.reset()
self._message_header_buffer = ""
self._in_skipped_tool_span = False
self._reset_args_state()
def feed(
@@ -316,39 +324,56 @@ class StreamingParserEngine:
if transition is None:
if self._has_drops and terminal == DROP_TERMINAL:
return []
# The projected skip state may not define the wrapper closer.
if self.skip_tool_parsing and terminal in self._tool_exit_terminals:
self._in_skipped_tool_span = False
return self._emit_for_state(value)
if self.skip_tool_parsing and terminal in self._tool_terminals:
if self.state == ParserState.MESSAGE_HEADER:
self.state = ParserState.CONTENT
self._message_header_buffer = ""
return [
SemanticEvent(
EventType.TEXT_CHUNK,
value=value,
tool_index=self.tool_index,
)
]
if EventType.REASONING_END in transition.events:
self.state = ParserState.CONTENT
return [
SemanticEvent(
EventType.REASONING_END,
value=value,
tool_index=self.tool_index,
),
SemanticEvent(
EventType.TEXT_CHUNK,
value=value,
tool_index=self.tool_index,
),
]
content_type = self.config.content_events.get(self.state)
if content_type is not None:
return [
SemanticEvent(content_type, value=value, tool_index=self.tool_index)
]
return []
# Inkling reuses one terminal for tool, text, and reasoning exits.
# Outside a forwarded tool span, apply its normal transition.
is_opener = transition.next_state in self._TOOL_STATES
is_exit = terminal in self._tool_exit_terminals
used_as_plain_closer = (
is_exit and not is_opener and not self._in_skipped_tool_span
)
if not used_as_plain_closer:
if is_opener:
self._in_skipped_tool_span = True
elif is_exit:
self._in_skipped_tool_span = False
if self.state == ParserState.MESSAGE_HEADER:
self.state = ParserState.CONTENT
self._message_header_buffer = ""
return [
SemanticEvent(
EventType.TEXT_CHUNK,
value=value,
tool_index=self.tool_index,
)
]
if EventType.REASONING_END in transition.events:
self.state = ParserState.CONTENT
return [
SemanticEvent(
EventType.REASONING_END,
value=value,
tool_index=self.tool_index,
),
SemanticEvent(
EventType.TEXT_CHUNK,
value=value,
tool_index=self.tool_index,
),
]
content_type = self.config.content_events.get(self.state)
if content_type is not None:
return [
SemanticEvent(
content_type, value=value, tool_index=self.tool_index
)
]
return []
if transition.skip_in_token_id_mode and self._ever_had_token_ids:
return self._emit_for_state(value)