[Frontend] Support count_reasoning_tokens in the Streaming Parser Engine (#45802)

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Chauncey
2026-08-15 06:51:14 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 5cecfc0137
commit ac2ae8798c
16 changed files with 669 additions and 73 deletions
+80 -17
View File
@@ -167,6 +167,17 @@ class StreamingParserEngine:
self.skip_tool_parsing = False
self.reset(initial_state=initial_state)
@property
def reasoning_token_count(self) -> int:
return self._reasoning_token_count
def _record_reasoning_tokens(self, events: Sequence[SemanticEvent]) -> None:
self._reasoning_token_count += sum(
event.token_count
for event in events
if event.type == EventType.REASONING_CHUNK
)
def _reset_args_state(self) -> None:
self._args_buffer: str = ""
self._args_safe_end: int = 0
@@ -186,6 +197,7 @@ class StreamingParserEngine:
)
self.tool_index = -1
self._ever_had_token_ids = False
self._reasoning_token_count = 0
# DO NOT reset skip_tool_parsing here — callers set it before
# calling methods that trigger reset() (e.g. extract_reasoning),
# and clearing it silently breaks non-streaming tool-call-as-
@@ -193,6 +205,7 @@ class StreamingParserEngine:
self._scanner.reset()
self._lexer.reset()
self._message_header_buffer = ""
self._message_header_token_count = 0
self._in_skipped_tool_span = False
self._reset_args_state()
@@ -218,18 +231,30 @@ class StreamingParserEngine:
has_special = True
break
if not has_special:
return self._emit_for_state(delta_text)
events = self._emit_for_state(
delta_text, token_count=len(delta_token_ids)
)
self._record_reasoning_tokens(events)
return events
scanner_items = self._scanner.scan(delta_text, delta_token_ids)
if len(scanner_items) == 1 and isinstance(scanner_items[0], TextChunk):
lex_tokens = self._lexer.feed(scanner_items[0].text)
item = scanner_items[0]
lex_tokens = self._lexer.feed(item.text, item.token_texts, item.token_count)
if len(lex_tokens) == 1 and lex_tokens[0].terminal == CONTENT_TERMINAL:
text = lex_tokens[0].value
return self._emit_for_state(text)
return self._process_lex_tokens(lex_tokens)
events = self._emit_for_state(
lex_tokens[0].value,
token_count=lex_tokens[0].token_count,
)
else:
events = self._process_lex_tokens(lex_tokens)
self._record_reasoning_tokens(events)
return events
return self._process_scanner_items(scanner_items)
events = self._process_scanner_items(scanner_items)
self._record_reasoning_tokens(events)
return events
def _process_scanner_items(
self, items: Sequence[LexerInput]
@@ -240,7 +265,18 @@ class StreamingParserEngine:
events.extend(self._process_lex_tokens(self._lexer.flush()))
events.extend(self._on_terminal(item.terminal, item.text))
elif isinstance(item, TextChunk):
events.extend(self._process_lex_tokens(self._lexer.feed(item.text)))
if not item.text and item.token_count:
events.extend(
self._emit_for_state("", token_count=item.token_count)
)
else:
events.extend(
self._process_lex_tokens(
self._lexer.feed(
item.text, item.token_texts, item.token_count
)
)
)
return events
def finish(self) -> list[SemanticEvent]:
@@ -285,11 +321,14 @@ class StreamingParserEngine:
EventType.TEXT_CHUNK,
value=self._message_header_buffer,
tool_index=self.tool_index,
token_count=self._message_header_token_count,
)
)
self._message_header_buffer = ""
self._message_header_token_count = 0
self.state = ParserState.CONTENT
self._record_reasoning_tokens(events)
return events
def parse_complete(self, text: str) -> list[SemanticEvent]:
@@ -303,9 +342,11 @@ class StreamingParserEngine:
strict = self._token_id_terminal_names if self._ever_had_token_ids else None
for tok in tokens:
if tok.terminal == CONTENT_TERMINAL or (strict and tok.terminal in strict):
events.extend(self._on_content(tok.value))
events.extend(self._on_content(tok.value, tok.token_count))
else:
events.extend(self._on_terminal(tok.terminal, tok.value))
events.extend(
self._on_terminal(tok.terminal, tok.value, tok.token_count)
)
return events
_TOOL_STATES = frozenset(
@@ -317,7 +358,9 @@ class StreamingParserEngine:
}
)
def _on_terminal(self, terminal: str, value: str) -> list[SemanticEvent]:
def _on_terminal(
self, terminal: str, value: str, token_count: int = 0
) -> list[SemanticEvent]:
key = (self.state, terminal)
transition = self.config.transitions.get(key)
@@ -327,7 +370,7 @@ class StreamingParserEngine:
# 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)
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.
@@ -345,6 +388,7 @@ class StreamingParserEngine:
leaving_message_header = self.state == ParserState.MESSAGE_HEADER
if leaving_message_header:
self._message_header_buffer = ""
self._message_header_token_count = 0
# A tool terminal that implicitly ends reasoning must report
# that even from the header state, or the reasoning pass never
# hands the block to the tool pass.
@@ -381,13 +425,14 @@ class StreamingParserEngine:
return []
if transition.skip_in_token_id_mode and self._ever_had_token_ids:
return self._emit_for_state(value)
return self._emit_for_state(value, token_count)
return self._apply_transition(transition, value)
return self._apply_transition(transition, value, token_count)
def _emit_for_state(self, text: str) -> list[SemanticEvent]:
def _emit_for_state(self, text: str, token_count: int = 0) -> list[SemanticEvent]:
if self.state == ParserState.MESSAGE_HEADER:
self._message_header_buffer += text
self._message_header_token_count += token_count
return []
if self.state == ParserState.TOOL_ARGS:
if self.config.tool_args_json:
@@ -397,26 +442,36 @@ class StreamingParserEngine:
EventType.ARG_VALUE_CHUNK,
value=text,
tool_index=self.tool_index,
token_count=token_count,
)
]
content_type = self.config.content_events.get(self.state)
if content_type is not None:
return [SemanticEvent(content_type, value=text, tool_index=self.tool_index)]
return [
SemanticEvent(
content_type,
value=text,
tool_index=self.tool_index,
token_count=token_count,
)
]
return []
def _on_content(self, text: str) -> list[SemanticEvent]:
def _on_content(self, text: str, token_count: int = 0) -> list[SemanticEvent]:
if not text:
return []
return self._emit_for_state(text)
return self._emit_for_state(text, token_count)
def _apply_transition(
self,
transition: Transition,
value: str,
token_count: int = 0,
) -> list[SemanticEvent]:
events: list[SemanticEvent] = []
previous_state = self.state
message_header = ""
message_header_token_count = 0
if (
self.state == ParserState.TOOL_ARGS
@@ -434,7 +489,9 @@ class StreamingParserEngine:
if previous_state == ParserState.MESSAGE_HEADER:
message_header = self._message_header_buffer
message_header_token_count = self._message_header_token_count
self._message_header_buffer = ""
self._message_header_token_count = 0
self.state = transition.next_state
@@ -454,6 +511,12 @@ class StreamingParserEngine:
event_type,
value=event_value,
tool_index=self.tool_index,
token_count=(
message_header_token_count
if previous_state == ParserState.MESSAGE_HEADER
and event_type == EventType.TEXT_CHUNK
else token_count
),
)
)