[Frontend] Port seed_oss to the streaming parser engine as a Qwen3 subclass (#46314)

Signed-off-by: EazyReal <8047065+EazyReal@users.noreply.github.com>
This commit is contained in:
Maxwill Lin
2026-06-24 20:08:42 -04:00
committed by GitHub
parent b69816043a
commit cd347298e8
13 changed files with 338 additions and 1435 deletions
@@ -13,6 +13,7 @@ from vllm.parser.glm47_moe import Glm47MoeParser
from vllm.parser.minimax_m2 import MinimaxM2Parser
from vllm.parser.nemotron_v3 import NemotronV3Parser
from vllm.parser.qwen3 import Qwen3Parser
from vllm.parser.seed_oss import SeedOssParser
(
MinimaxM2ParserReasoningAdapter,
@@ -34,6 +35,11 @@ from vllm.parser.qwen3 import Qwen3Parser
Qwen3ParserToolAdapter,
) = make_adapters(Qwen3Parser)
(
SeedOssParserReasoningAdapter,
SeedOssParserToolAdapter,
) = make_adapters(SeedOssParser)
(
Glm47MoeParserReasoningAdapter,
Glm47MoeParserToolAdapter,
+40 -13
View File
@@ -38,6 +38,8 @@ if TYPE_CHECKING:
from vllm.tokenizers import TokenizerLike
from vllm.tool_parsers.abstract_tool_parser import Tool
THINK_START = "<think>"
THINK_END = "</think>"
TOOL_CALL_START = "<tool_call>"
TOOL_CALL_END = "</tool_call>"
FUNC_PREFIX = "<function="
@@ -75,17 +77,25 @@ def _qwen3_arg_converter(raw_args: str, partial: bool) -> str:
@functools.cache
def qwen3_config(thinking: bool = True) -> ParserEngineConfig:
def qwen3_config(
thinking: bool = True,
*,
name: str = "qwen3",
think_start: str = THINK_START,
think_end: str = THINK_END,
tool_start: str = TOOL_CALL_START,
tool_end: str = TOOL_CALL_END,
) -> ParserEngineConfig:
return ParserEngineConfig(
name="qwen3",
name=name,
initial_state=ParserState.REASONING if thinking else ParserState.CONTENT,
terminals={
# Reasoning terminals
"THINK_START": "<think>",
"THINK_END": "</think>",
"THINK_START": think_start,
"THINK_END": think_end,
# Tool call terminals
"TOOL_START": TOOL_CALL_START,
"TOOL_END": TOOL_CALL_END,
"TOOL_START": tool_start,
"TOOL_END": tool_end,
"FUNC_PREFIX": FUNC_PREFIX,
"FUNC_END": FUNC_END,
"PARAM_START": PARAM_START,
@@ -93,10 +103,10 @@ def qwen3_config(thinking: bool = True) -> ParserEngineConfig:
"CLOSE_ANGLE": ">",
},
token_id_terminals={
"THINK_START": "<think>",
"THINK_END": "</think>",
"TOOL_START": TOOL_CALL_START,
"TOOL_END": TOOL_CALL_END,
"THINK_START": think_start,
"THINK_END": think_end,
"TOOL_START": tool_start,
"TOOL_END": tool_end,
},
transitions={
# -- Reasoning transitions --
@@ -185,8 +195,18 @@ class Qwen3Parser(ParserEngine):
- ``<tool_call>`` as implicit reasoning end
- Unpaired ``<tool_call>`` token ID detection for ``is_reasoning_end``
Subclasses that share the grammar but differ only in the four wrapper
token strings (reasoning + tool-call) override the class attributes
below; everything else is inherited unchanged.
"""
CONFIG_NAME = "qwen3"
THINK_START = THINK_START
THINK_END = THINK_END
TOOL_START = TOOL_CALL_START
TOOL_END = TOOL_CALL_END
def __init__(
self,
tokenizer: TokenizerLike,
@@ -197,7 +217,14 @@ class Qwen3Parser(ParserEngine):
self.thinking_enabled = chat_kwargs.get("enable_thinking", True)
kwargs.setdefault(
"parser_engine_config",
qwen3_config(thinking=self.thinking_enabled),
qwen3_config(
thinking=self.thinking_enabled,
name=self.CONFIG_NAME,
think_start=self.THINK_START,
think_end=self.THINK_END,
tool_start=self.TOOL_START,
tool_end=self.TOOL_END,
),
)
super().__init__(
tokenizer,
@@ -205,8 +232,8 @@ class Qwen3Parser(ParserEngine):
**kwargs,
)
vocab = self.vocab
self._tool_call_token_id: int | None = vocab.get("<tool_call>")
self._tool_call_end_token_id: int | None = vocab.get("</tool_call>")
self._tool_call_token_id: int | None = vocab.get(self.TOOL_START)
self._tool_call_end_token_id: int | None = vocab.get(self.TOOL_END)
def extract_reasoning(
self,
+28
View File
@@ -0,0 +1,28 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""seed_oss parser for tool calls and reasoning.
seed_oss shares the Qwen3 XML grammar exactly; only the four wrapper
token strings differ::
<think> -> <seed:think>
</think> -> </seed:think>
<tool_call> -> <seed:tool_call>
</tool_call> -> </seed:tool_call>
``<function=...>`` and ``<parameter=...>`` are byte-identical, so the
entire transition table and ``_qwen3_arg_converter`` are inherited from
:class:`Qwen3Parser` unchanged.
"""
from __future__ import annotations
from vllm.parser.qwen3 import Qwen3Parser
class SeedOssParser(Qwen3Parser):
CONFIG_NAME = "seed_oss"
THINK_START = "<seed:think>"
THINK_END = "</seed:think>"
TOOL_START = "<seed:tool_call>"
TOOL_END = "</seed:tool_call>"