[Bugfix] Drop incomplete tool-call markup in non-streaming to match streaming (#47562)
Signed-off-by: Suvrakamal Das <jyb2084.usa@gmail.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Ben Browning <56071+bbrowning@users.noreply.github.com>
This commit is contained in:
co-authored by
Claude
Ben Browning
parent
2f01039666
commit
e7cf4730d6
@@ -150,6 +150,15 @@ class ParserEngineToolAdapter(ToolParser):
|
||||
state so it can parse reasoning-stripped content (i.e. the output of
|
||||
:meth:`ReasoningParser.extract_reasoning`).
|
||||
|
||||
When no reasoning parser is configured at the serving layer, the
|
||||
owning :class:`~vllm.parser.abstract_parser.Parser` sets
|
||||
:attr:`skip_reasoning_parsing` so reasoning markup — plain content
|
||||
in that configuration — passes through verbatim instead of being
|
||||
consumed or reclassified by the engine. The engine honors the flag
|
||||
only for markers that are reasoning-exclusive in its config; a
|
||||
grammar like Inkling, whose block closer is shared with text and
|
||||
tool blocks, keeps parsing unchanged.
|
||||
|
||||
Subclasses set :attr:`_parser_engine_cls` to the concrete
|
||||
:class:`ParserEngine` class.
|
||||
"""
|
||||
@@ -166,6 +175,14 @@ class ParserEngineToolAdapter(ToolParser):
|
||||
super().__init__(tokenizer, tools)
|
||||
self._parser_engine = self._parser_engine_cls(tokenizer, tools, **kwargs) # type: ignore[call-arg]
|
||||
|
||||
@property
|
||||
def skip_reasoning_parsing(self) -> bool:
|
||||
return self._parser_engine.skip_reasoning_parsing
|
||||
|
||||
@skip_reasoning_parsing.setter
|
||||
def skip_reasoning_parsing(self, value: bool) -> None:
|
||||
self._parser_engine.skip_reasoning_parsing = value
|
||||
|
||||
def adjust_request(
|
||||
self,
|
||||
request: ChatCompletionRequest | ResponsesRequest,
|
||||
|
||||
@@ -175,6 +175,14 @@ class ParserEngine(Parser):
|
||||
def skip_tool_parsing(self, value: bool) -> None:
|
||||
self._engine.skip_tool_parsing = value
|
||||
|
||||
@property
|
||||
def skip_reasoning_parsing(self) -> bool:
|
||||
return self._engine.skip_reasoning_parsing
|
||||
|
||||
@skip_reasoning_parsing.setter
|
||||
def skip_reasoning_parsing(self, value: bool) -> None:
|
||||
self._engine.skip_reasoning_parsing = value
|
||||
|
||||
@property
|
||||
def reasoning_ended(self) -> bool:
|
||||
return self._reasoning_ended
|
||||
@@ -714,6 +722,7 @@ class ParserEngine(Parser):
|
||||
content_parts: list[str] = []
|
||||
reasoning_parts: list[str] = []
|
||||
|
||||
carried_deferred = self._deferred_content
|
||||
seen_tool_event = False
|
||||
suppress = self._suppress_tool_calls
|
||||
for event in events:
|
||||
@@ -750,7 +759,14 @@ class ParserEngine(Parser):
|
||||
tool_call_deltas = self._coalesce_tool_call_deltas(tool_call_deltas)
|
||||
|
||||
if self._deferred_content and (not seen_tool_event or not tool_call_deltas):
|
||||
content_parts.insert(0, self._deferred_content)
|
||||
# Deferred content carried in from a previous delta precedes this
|
||||
# delta's content; content deferred during this call (text after
|
||||
# an unpromoted tool event) follows it.
|
||||
deferred_now = self._deferred_content[len(carried_deferred) :]
|
||||
if carried_deferred:
|
||||
content_parts.insert(0, carried_deferred)
|
||||
if deferred_now:
|
||||
content_parts.append(deferred_now)
|
||||
self._deferred_content = ""
|
||||
|
||||
content_str = "".join(content_parts)
|
||||
|
||||
@@ -164,7 +164,12 @@ class StreamingParserEngine:
|
||||
if state in self._TOOL_STATES and tr.next_state not in self._TOOL_STATES
|
||||
)
|
||||
|
||||
self._reasoning_markup_terminals: frozenset[str] = (
|
||||
self._compute_reasoning_markup_terminals()
|
||||
)
|
||||
|
||||
self.skip_tool_parsing = False
|
||||
self.skip_reasoning_parsing = False
|
||||
self.reset(initial_state=initial_state)
|
||||
|
||||
@property
|
||||
@@ -358,6 +363,37 @@ class StreamingParserEngine:
|
||||
}
|
||||
)
|
||||
|
||||
_PLAIN_STATES = frozenset({ParserState.CONTENT, ParserState.REASONING})
|
||||
|
||||
_REASONING_EVENTS = frozenset({EventType.REASONING_START, EventType.REASONING_END})
|
||||
|
||||
def _compute_reasoning_markup_terminals(self) -> frozenset[str]:
|
||||
"""Terminals the ``skip_reasoning_parsing`` bypass may neutralize.
|
||||
|
||||
Only reasoning-exclusive markers qualify: every transition they
|
||||
participate in stays within CONTENT/REASONING and emits nothing
|
||||
but reasoning events. Inkling's ``<|end_message|>`` is labelled
|
||||
THINK_END yet also closes text, header, and tool blocks;
|
||||
bypassing a shared marker would eat that structure, so one impure
|
||||
marker disables the bypass for the whole config.
|
||||
"""
|
||||
markers = frozenset(
|
||||
terminal
|
||||
for (state, terminal), tr in self.config.transitions.items()
|
||||
if ParserState.REASONING in (state, tr.next_state)
|
||||
and tr.next_state not in self._TOOL_STATES
|
||||
)
|
||||
for (state, terminal), tr in self.config.transitions.items():
|
||||
if terminal not in markers:
|
||||
continue
|
||||
if (
|
||||
state not in self._PLAIN_STATES
|
||||
or tr.next_state not in self._PLAIN_STATES
|
||||
or not self._REASONING_EVENTS.issuperset(tr.events)
|
||||
):
|
||||
return frozenset()
|
||||
return markers
|
||||
|
||||
def _on_terminal(
|
||||
self, terminal: str, value: str, token_count: int = 0
|
||||
) -> list[SemanticEvent]:
|
||||
@@ -372,6 +408,9 @@ class StreamingParserEngine:
|
||||
self._in_skipped_tool_span = False
|
||||
return self._emit_for_state(value, token_count)
|
||||
|
||||
if self.skip_reasoning_parsing and terminal in self._reasoning_markup_terminals:
|
||||
return self._emit_for_state(value, token_count)
|
||||
|
||||
if self.skip_tool_parsing and terminal in self._tool_terminals:
|
||||
# Inkling reuses one terminal for tool, text, and reasoning exits.
|
||||
# Outside a forwarded tool span, apply its normal transition.
|
||||
|
||||
Reference in New Issue
Block a user