[Frontend] [Parser] Port DeepSeek V4 to streaming parser engine framework (#45877)

Signed-off-by: Ben Browning <bbrownin@redhat.com>
This commit is contained in:
Ben Browning
2026-07-03 20:55:23 -04:00
committed by GitHub
parent d6d39c111e
commit fb5291b35b
22 changed files with 1900 additions and 671 deletions
+2 -1
View File
@@ -111,7 +111,8 @@ class ParserEngineReasoningAdapter(ReasoningParser):
return self._parser_engine.reasoning_ended
def finish_streaming(self) -> DeltaMessage | None:
return self._parser_engine.finish_streaming()
with self._skip_tool_parsing():
return self._parser_engine.finish_streaming()
def get_streaming_fallback_content(
self,
+7 -3
View File
@@ -108,7 +108,11 @@ class ParserEngine(Parser):
parser_engine_config, tokenizer, vocab=self.vocab
)
self._reasoning_ended: bool = False
self._has_reasoning = (
"THINK_END" in parser_engine_config.token_id_terminals
or parser_engine_config.initial_state == ParserState.REASONING
)
self._reasoning_ended: bool = not self._has_reasoning
self._streaming_initialized: bool = False
self._prompt_streaming_prepared: bool = False
@@ -188,7 +192,7 @@ class ParserEngine(Parser):
def _reset(self, initial_state: ParserState | None = None) -> None:
self._engine.reset(initial_state=initial_state)
self._reasoning_ended = False
self._reasoning_ended = not self._has_reasoning
self._tool_slots.clear()
self._deferred_content = ""
self._deferred_reasoning = ""
@@ -255,7 +259,7 @@ class ParserEngine(Parser):
types = extract_types_from_schema(schema)
as_str = json.dumps(value, ensure_ascii=False)
coerced = coerce_to_schema_type(as_str, types)
if coerced != value:
if type(coerced) is not type(value) or coerced != value:
return coerced, True
return value, False
+12
View File
@@ -7,6 +7,8 @@ names so that :class:`ReasoningParserManager` and
:class:`ToolParserManager` can load them lazily.
"""
from vllm.parser.deepseek_v4 import DeepSeekV4Parser
from vllm.parser.deepseek_v32 import DeepSeekV32Parser
from vllm.parser.engine.adapters import make_adapters
from vllm.parser.gemma4 import Gemma4Parser
from vllm.parser.glm47_moe import Glm47MoeParser
@@ -16,6 +18,16 @@ from vllm.parser.nemotron_v3 import NemotronV3Parser
from vllm.parser.qwen3 import Qwen3Parser
from vllm.parser.seed_oss import SeedOssParser
(
DeepSeekV32ParserReasoningAdapter,
DeepSeekV32ParserToolAdapter,
) = make_adapters(DeepSeekV32Parser)
(
DeepSeekV4ParserReasoningAdapter,
DeepSeekV4ParserToolAdapter,
) = make_adapters(DeepSeekV4Parser)
(
MinimaxM2ParserReasoningAdapter,
MinimaxM2ParserToolAdapter,
@@ -22,14 +22,13 @@ from vllm.parser.engine.parser_engine_config import (
Transition,
)
from vllm.parser.engine.token_id_scanner import (
DROP_TERMINAL,
LexerInput,
PreLexedTerminal,
TextChunk,
TokenIDScanner,
)
DROP_TERMINAL = "__DROP__"
@dataclass(slots=True)
class _DropInfo:
+11 -1
View File
@@ -8,6 +8,8 @@ from __future__ import annotations
from collections.abc import Sequence
from dataclasses import dataclass
DROP_TERMINAL = "__DROP__"
@dataclass(slots=True)
class TextChunk:
@@ -277,7 +279,15 @@ class TokenIDScanner:
consumed = pos + len(anchor.text)
else:
has_later_valid = any(p >= 0 for p in positions[i + 1 :])
if not has_later_valid and consumed < len(delta_text):
# DROP anchors (EOS, etc.) may have text that never
# arrives in delta_text (stripped by detokenizer).
# Don't defer remaining content waiting for text
# that will never come.
if (
not has_later_valid
and consumed < len(delta_text)
and anchor.terminal != DROP_TERMINAL
):
self._deferred_post_text += delta_text[consumed:]
consumed = len(delta_text)
self._deferred_terminals.append(anchor)