[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
+19 -1
View File
@@ -47,6 +47,11 @@ class ParserEngineReasoningAdapter(ReasoningParser):
def __init__(self, tokenizer: TokenizerLike, *args, **kwargs) -> None:
super().__init__(tokenizer, *args, **kwargs)
self._parser_engine = self._parser_engine_cls(tokenizer, **kwargs) # type: ignore[call-arg]
self._parser_engine_kwargs = kwargs
self._counting_parser_engine: ParserEngine | None = None
# TODO: Remove once Responses finalization reuses accumulated streaming
# parser results instead of reparsing the complete output.
self._streaming_count_valid = False
@contextmanager
def _skip_tool_parsing(self) -> Iterator[None]:
@@ -71,6 +76,7 @@ class ParserEngineReasoningAdapter(ReasoningParser):
model_output: str,
request: ChatCompletionRequest | ResponsesRequest,
) -> tuple[str | None, str | None]:
self._streaming_count_valid = False
with self._skip_tool_parsing():
return self._parser_engine.extract_reasoning(model_output, request)
@@ -83,6 +89,7 @@ class ParserEngineReasoningAdapter(ReasoningParser):
current_token_ids: Sequence[int],
delta_token_ids: Sequence[int],
) -> DeltaMessage | None:
self._streaming_count_valid = True
with self._skip_tool_parsing():
return self._parser_engine.extract_reasoning_streaming(
previous_text,
@@ -122,7 +129,18 @@ class ParserEngineReasoningAdapter(ReasoningParser):
return self._parser_engine.get_streaming_fallback_content(text, request)
def count_reasoning_tokens(self, token_ids: Sequence[int]) -> int:
return self._parser_engine.count_reasoning_tokens(token_ids)
if self._streaming_count_valid:
return self._parser_engine.count_reasoning_tokens(token_ids)
if not token_ids:
return 0
if self._counting_parser_engine is None:
self._counting_parser_engine = self._parser_engine_cls(
self.model_tokenizer, **self._parser_engine_kwargs
) # type: ignore[call-arg]
self._counting_parser_engine._single_pass_parse(
self.model_tokenizer.decode(token_ids), token_ids
)
return self._counting_parser_engine.count_reasoning_tokens(token_ids)
class ParserEngineToolAdapter(ToolParser):
+1
View File
@@ -24,3 +24,4 @@ class SemanticEvent:
type: EventType
value: str = ""
tool_index: int = -1
token_count: int = 0
+81 -8
View File
@@ -24,6 +24,7 @@ class TerminalDef:
class LexToken:
terminal: str
value: str
token_count: int = 0
class LexerShape:
@@ -104,6 +105,7 @@ class IncrementalLexer:
self.terminals = shape.terminals
self.content_terminal = content_terminal
self.buffer = ""
self._token_counts: list[int] = []
self._literal_strings = shape.literal_strings
self._max_literal_len = shape.max_literal_len
@@ -114,15 +116,23 @@ class IncrementalLexer:
def reset(self) -> None:
self.buffer = ""
self._token_counts.clear()
def feed(self, text: str) -> list[LexToken]:
def feed(
self,
text: str,
token_texts: tuple[str, ...] = (),
token_count: int = 0,
) -> list[LexToken]:
char_token_counts = self._char_token_counts(text, token_texts, token_count)
if not self.buffer and self._has_only_literals and self._literal_first_chars:
for ch in text:
if ch in self._literal_first_chars:
break
else:
return [LexToken(self.content_terminal, text)]
return [LexToken(self.content_terminal, text, sum(char_token_counts))]
self.buffer += text
self._token_counts.extend(char_token_counts)
return self._drain()
def flush(self) -> list[LexToken]:
@@ -130,10 +140,46 @@ class IncrementalLexer:
if self.buffer:
tokens.extend(self._drain(final=True))
if self.buffer:
tokens.append(LexToken(self.content_terminal, self.buffer))
tokens.append(
LexToken(self.content_terminal, self.buffer, sum(self._token_counts))
)
self.buffer = ""
self._token_counts.clear()
return tokens
@staticmethod
def _char_token_counts(
text: str,
token_texts: tuple[str, ...],
token_count: int,
) -> list[int]:
counts = [0] * len(text)
if not text:
return counts
if token_texts:
pos = 0
assigned = 0
for token_text in token_texts:
if not token_text:
continue
found = text.find(token_text, pos)
if found < 0:
continue
counts[found] += 1
assigned += 1
pos = found + len(token_text)
missing = token_count - assigned
if missing > 0:
counts[0] += missing
elif token_count:
counts[0] = token_count
return counts
def _pop_token_count(self, length: int) -> int:
token_count = sum(self._token_counts[:length])
del self._token_counts[:length]
return token_count
def _drain(self, *, final: bool = False) -> list[LexToken]:
tokens: list[LexToken] = []
first_chars = self._literal_first_chars
@@ -150,8 +196,11 @@ class IncrementalLexer:
has_potential = True
break
if not has_potential:
tokens.append(LexToken(content_terminal, self.buffer))
tokens.append(
LexToken(content_terminal, self.buffer, sum(self._token_counts))
)
self.buffer = ""
self._token_counts.clear()
break
best_match: tuple[str, str, int] | None = None
@@ -175,7 +224,13 @@ class IncrementalLexer:
longer_match = True
break
if not longer_match:
tokens.append(LexToken(best_match[0], best_match[1]))
tokens.append(
LexToken(
best_match[0],
best_match[1],
self._pop_token_count(best_match[2]),
)
)
self.buffer = self.buffer[best_match[2] :]
continue
break
@@ -183,15 +238,33 @@ class IncrementalLexer:
break
if best_match is not None:
tokens.append(LexToken(best_match[0], best_match[1]))
tokens.append(
LexToken(
best_match[0],
best_match[1],
self._pop_token_count(best_match[2]),
)
)
self.buffer = self.buffer[best_match[2] :]
else:
content_end = self._find_content_boundary()
if content_end > 0:
tokens.append(LexToken(content_terminal, self.buffer[:content_end]))
tokens.append(
LexToken(
content_terminal,
self.buffer[:content_end],
self._pop_token_count(content_end),
)
)
self.buffer = self.buffer[content_end:]
else:
tokens.append(LexToken(content_terminal, self.buffer[0]))
tokens.append(
LexToken(
content_terminal,
self.buffer[0],
self._pop_token_count(1),
)
)
self.buffer = self.buffer[1:]
return tokens
+2 -17
View File
@@ -625,23 +625,8 @@ class ParserEngine(Parser):
return None
def count_reasoning_tokens(self, token_ids: Sequence[int]) -> int:
start_id = self._reasoning_start_token_id
end_id = self._reasoning_end_token_id
if start_id is None or end_id is None:
return 0
count = 0
depth = 0
for token_id in token_ids:
if token_id == start_id:
depth += 1
continue
if token_id == end_id:
if depth > 0:
depth -= 1
continue
if depth > 0:
count += 1
return count
"""Return reasoning tokens observed by the parser engine so far."""
return self._engine.reasoning_token_count
# ── Single-pass parse helper ────────────────────────────────────────
+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
),
)
)
+121 -17
View File
@@ -14,6 +14,8 @@ DROP_TERMINAL = "__DROP__"
@dataclass(slots=True)
class TextChunk:
text: str
token_texts: tuple[str, ...] = ()
token_count: int = 0
@dataclass(slots=True)
@@ -49,18 +51,30 @@ class TokenIDScanner:
self.tokenizer = tokenizer
self._token_text_cache: dict[int, str] = {}
self._deferred_terminals: list[PreLexedTerminal] = []
self._deferred_prefix_token_counts: list[int] = []
self._deferred_trailing_token_count = 0
self._deferred_post_text: str = ""
def reset(self) -> None:
"""Clear mutable state for reuse. Preserves the token text cache."""
self._deferred_terminals.clear()
self._deferred_prefix_token_counts.clear()
self._deferred_trailing_token_count = 0
self._deferred_post_text = ""
def _decode_token(self, token_id: int) -> str:
if token_id not in self._token_text_cache:
self._token_text_cache[token_id] = self.tokenizer.decode([token_id])
token_text = self.tokenizer.decode([token_id])
self._token_text_cache[token_id] = (
token_text if isinstance(token_text, str) else ""
)
return self._token_text_cache[token_id]
def _decode_tokens(self, token_ids: Sequence[int]) -> tuple[str, ...]:
if self.tokenizer is None:
return ()
return tuple(self._decode_token(tid) for tid in token_ids)
_EMPTY: tuple[LexerInput, ...] = ()
def scan(
@@ -70,13 +84,23 @@ class TokenIDScanner:
) -> Sequence[LexerInput]:
prefix_items: list[LexerInput] = []
effective_text = delta_text
deferred_trailing_count = 0
if self._deferred_terminals:
prefix_items, effective_text = self._resolve_deferred(delta_text)
prefix_items, effective_text, deferred_trailing_count = (
self._resolve_deferred(delta_text, len(delta_token_ids))
)
if not self.token_id_to_terminal:
if effective_text:
prefix_items.append(TextChunk(effective_text))
token_texts = self._decode_tokens(delta_token_ids)
prefix_items.append(
TextChunk(
effective_text,
token_texts,
deferred_trailing_count + len(token_texts),
)
)
return prefix_items
has_special = False
@@ -89,14 +113,23 @@ class TokenIDScanner:
if not has_special:
if effective_text:
if not prefix_items:
return [TextChunk(effective_text)]
prefix_items.append(TextChunk(effective_text))
token_texts = self._decode_tokens(delta_token_ids)
return [TextChunk(effective_text, token_texts, len(token_texts))]
token_texts = self._decode_tokens(delta_token_ids)
prefix_items.append(
TextChunk(
effective_text,
token_texts,
deferred_trailing_count + len(token_texts),
)
)
return prefix_items or self._EMPTY
token_texts = [self._decode_token(tid) for tid in delta_token_ids]
decoded_token_texts = [self._decode_token(tid) for tid in delta_token_ids]
results: list[LexerInput] = []
text_accum: list[str] = []
token_text_accum: list[str] = []
for idx, tid in enumerate(delta_token_ids):
terminal = self.token_id_to_terminal.get(tid)
@@ -104,16 +137,31 @@ class TokenIDScanner:
if text_accum:
joined = "".join(text_accum)
if joined:
results.append(TextChunk(joined))
results.append(
TextChunk(
joined,
tuple(token_text_accum),
len(token_text_accum),
)
)
text_accum.clear()
results.append(PreLexedTerminal(terminal, tid, token_texts[idx]))
token_text_accum.clear()
results.append(
PreLexedTerminal(terminal, tid, decoded_token_texts[idx])
)
else:
text_accum.append(token_texts[idx])
text_accum.append(decoded_token_texts[idx])
token_text_accum.append(decoded_token_texts[idx])
if text_accum:
joined = "".join(text_accum)
if joined:
results.append(TextChunk(joined))
results.append(
TextChunk(joined, tuple(token_text_accum), len(token_text_accum))
)
if deferred_trailing_count:
results.insert(0, TextChunk("", token_count=deferred_trailing_count))
if effective_text:
results = self._recover_holdback_text(effective_text, results)
@@ -124,9 +172,15 @@ class TokenIDScanner:
# transition before the preceding text has arrived. The
# deferred terminals will be resolved against the actual
# delta_text in a subsequent scan() or flushed by finish().
prefix_token_count = 0
for r in results:
if isinstance(r, TextChunk):
prefix_token_count += r.token_count
if isinstance(r, PreLexedTerminal):
self._deferred_terminals.append(r)
self._deferred_prefix_token_counts.append(prefix_token_count)
prefix_token_count = 0
self._deferred_trailing_token_count += prefix_token_count
results = []
return prefix_items + results
@@ -136,16 +190,30 @@ class TokenIDScanner:
return []
results: list[LexerInput] = []
if self._deferred_post_text:
results.append(TextChunk(self._deferred_post_text))
prefix_count = (
self._deferred_prefix_token_counts[0]
if self._deferred_prefix_token_counts
else 0
)
results.append(
TextChunk(self._deferred_post_text, token_count=prefix_count)
)
self._deferred_post_text = ""
results.extend(self._deferred_terminals)
if self._deferred_trailing_token_count:
results.append(
TextChunk("", token_count=self._deferred_trailing_token_count)
)
self._deferred_terminals.clear()
self._deferred_prefix_token_counts.clear()
self._deferred_trailing_token_count = 0
return results
def _resolve_deferred(
self,
delta_text: str,
) -> tuple[list[LexerInput], str]:
current_token_count: int = 0,
) -> tuple[list[LexerInput], str, int]:
"""Resolve deferred terminals against new delta_text.
When a previous ``scan()`` deferred a terminal (its text hadn't
@@ -160,7 +228,11 @@ class TokenIDScanner:
should be scanned with the current delta's token IDs.
"""
deferred = self._deferred_terminals
prefix_token_counts = self._deferred_prefix_token_counts
trailing_token_count = self._deferred_trailing_token_count
self._deferred_terminals = []
self._deferred_prefix_token_counts = []
self._deferred_trailing_token_count = 0
results: list[LexerInput] = []
remaining = delta_text
@@ -171,10 +243,13 @@ class TokenIDScanner:
# Duplicate-text deferred terminals resolve left-to-right via
# find(); correct when each terminal text appears once in sequence.
for terminal in deferred:
for idx, terminal in enumerate(deferred):
prefix_token_count = prefix_token_counts[idx]
pos = remaining.find(terminal.text)
if pos > 0:
results.append(TextChunk(remaining[:pos]))
results.append(
TextChunk(remaining[:pos], token_count=prefix_token_count)
)
results.append(terminal)
remaining = remaining[pos + len(terminal.text) :]
elif pos == 0:
@@ -185,10 +260,16 @@ class TokenIDScanner:
# only the terminal provides a reliable split point.
if remaining:
self._deferred_post_text += remaining
prefix_token_count += current_token_count
remaining = ""
self._deferred_terminals.append(terminal)
self._deferred_prefix_token_counts.append(prefix_token_count)
return results, remaining
if self._deferred_terminals:
self._deferred_trailing_token_count = trailing_token_count
trailing_token_count = 0
return results, remaining, trailing_token_count
def _recover_holdback_text(
self,
@@ -257,6 +338,16 @@ class TokenIDScanner:
if not anchors:
return [TextChunk(delta_text)]
token_groups: list[list[str]] = [[] for _ in range(len(anchors) + 1)]
count_groups = [0] * (len(anchors) + 1)
group_idx = 0
for item in results:
if isinstance(item, PreLexedTerminal):
group_idx += 1
else:
token_groups[group_idx].extend(item.token_texts)
count_groups[group_idx] += item.token_count
# Resolve positions right-to-left: each anchor gets the
# rightmost occurrence that is still before the next anchor.
positions: list[int] = [-1] * len(anchors)
@@ -274,7 +365,13 @@ class TokenIDScanner:
pos = positions[i]
if pos >= consumed:
if pos > consumed:
new_results.append(TextChunk(delta_text[consumed:pos]))
new_results.append(
TextChunk(
delta_text[consumed:pos],
tuple(token_groups[i]),
count_groups[i],
)
)
new_results.append(anchor)
consumed = pos + len(anchor.text)
else:
@@ -291,6 +388,13 @@ class TokenIDScanner:
self._deferred_post_text += delta_text[consumed:]
consumed = len(delta_text)
self._deferred_terminals.append(anchor)
self._deferred_prefix_token_counts.append(count_groups[i])
if consumed < len(delta_text):
new_results.append(TextChunk(delta_text[consumed:]))
new_results.append(
TextChunk(
delta_text[consumed:],
tuple(token_groups[-1]),
count_groups[-1],
)
)
return new_results