[Tool] adjust_request to reasoning parser, and Gemma4 fixes (#39027)

Signed-off-by: Ben Browning <bbrownin@redhat.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Ben Browning
2026-04-08 19:04:04 +00:00
committed by GitHub
co-authored by Claude Opus 4.6 Cursor
parent e24e0a43a4
commit 8477fe427d
12 changed files with 878 additions and 16 deletions
+2 -1
View File
@@ -170,7 +170,8 @@ class AnthropicServingMessages(OpenAIServingChat):
else:
cls._convert_message_content(msg, openai_msg, openai_messages)
openai_messages.append(openai_msg)
if not (msg.role == "user" and "content" not in openai_msg):
openai_messages.append(openai_msg)
@classmethod
def _convert_message_content(
+2
View File
@@ -379,6 +379,7 @@ async def init_app_state(
enable_auto_tools=args.enable_auto_tool_choice,
exclude_tools_when_tool_choice_none=args.exclude_tools_when_tool_choice_none,
tool_parser=args.tool_call_parser,
reasoning_parser=args.structured_outputs_config.reasoning_parser,
default_chat_template_kwargs=args.default_chat_template_kwargs,
log_error_stack=args.log_error_stack,
)
@@ -481,6 +482,7 @@ async def init_render_app_state(
enable_auto_tools=args.enable_auto_tool_choice,
exclude_tools_when_tool_choice_none=args.exclude_tools_when_tool_choice_none,
tool_parser=args.tool_call_parser,
reasoning_parser=args.structured_outputs_config.reasoning_parser,
default_chat_template_kwargs=args.default_chat_template_kwargs,
log_error_stack=args.log_error_stack,
)
@@ -594,6 +594,7 @@ class OpenAIServingResponses(OpenAIServing):
default_template_kwargs=None,
tool_dicts=tool_dicts,
tool_parser=self.parser.tool_parser_cls if self.parser else None,
reasoning_parser=self.parser.reasoning_parser_cls if self.parser else None,
)
return messages, engine_inputs
@@ -618,6 +619,7 @@ class OpenAIServingResponses(OpenAIServing):
default_template_kwargs=None,
tool_dicts=tool_dicts,
tool_parser=tool_parser,
reasoning_parser=self.parser.reasoning_parser_cls if self.parser else None,
)
return engine_inputs
+13
View File
@@ -44,6 +44,7 @@ from vllm.inputs import (
)
from vllm.logger import init_logger
from vllm.parser import ParserManager
from vllm.reasoning.abs_reasoning_parsers import ReasoningParser
from vllm.renderers import BaseRenderer, merge_kwargs
from vllm.renderers.inputs.preprocess import (
extract_prompt_components,
@@ -74,6 +75,7 @@ class OpenAIServingRender:
enable_auto_tools: bool = False,
exclude_tools_when_tool_choice_none: bool = False,
tool_parser: str | None = None,
reasoning_parser: str | None = None,
default_chat_template_kwargs: dict[str, Any] | None = None,
log_error_stack: bool = False,
) -> None:
@@ -94,6 +96,11 @@ class OpenAIServingRender:
enable_auto_tools=enable_auto_tools,
model_name=model_config.model,
)
self.reasoning_parser: type[ReasoningParser] | None = (
ParserManager.get_reasoning_parser(
reasoning_parser_name=reasoning_parser,
)
)
self.default_chat_template_kwargs: dict[str, Any] = (
default_chat_template_kwargs or {}
)
@@ -245,6 +252,7 @@ class OpenAIServingRender:
default_template_kwargs=self.default_chat_template_kwargs,
tool_dicts=tool_dicts,
tool_parser=tool_parser,
reasoning_parser=self.reasoning_parser,
)
else:
# For GPT-OSS.
@@ -503,6 +511,7 @@ class OpenAIServingRender:
default_template_kwargs: dict[str, Any] | None,
tool_dicts: list[dict[str, Any]] | None = None,
tool_parser: type[ToolParser] | None = None,
reasoning_parser: type[ReasoningParser] | None = None,
*,
skip_mm_cache: bool = False,
) -> tuple[list[ConversationMessage], list[EngineInput]]:
@@ -539,6 +548,10 @@ class OpenAIServingRender:
skip_mm_cache=skip_mm_cache,
)
if reasoning_parser is not None:
tokenizer = renderer.get_tokenizer()
request = reasoning_parser(tokenizer).adjust_request(request=request)
# tool parsing is done only if a tool_parser has been set and if
# tool_choice is not "none" (if tool_choice is "none" but a tool_parser
# is set, we want to prevent parsing a tool_call hallucinated by the LLM
+9
View File
@@ -496,6 +496,15 @@ class DelegatingParser(Parser):
# No tool calls
return [], content
def adjust_request(
self, request: ChatCompletionRequest | ResponsesRequest
) -> ChatCompletionRequest | ResponsesRequest:
if self._reasoning_parser is not None:
request = self._reasoning_parser.adjust_request(request)
if self._tool_parser is not None:
request = self._tool_parser.adjust_request(request)
return request
def extract_reasoning_streaming(
self,
previous_text: str,
+8 -2
View File
@@ -6,7 +6,7 @@ import os
from abc import abstractmethod
from collections.abc import Callable, Iterable, Sequence
from functools import cached_property
from typing import TYPE_CHECKING
from typing import TYPE_CHECKING, cast
from vllm.entrypoints.mcp.tool_server import ToolServer
from vllm.logger import init_logger
@@ -150,6 +150,12 @@ class ReasoningParser:
previously been parsed and extracted (see constructor)
"""
def adjust_request(
self, request: "ChatCompletionRequest | ResponsesRequest"
) -> "ChatCompletionRequest | ResponsesRequest":
"""Adjust request parameters; override in subclasses as needed."""
return request
def prepare_structured_tag(
self,
original_tag: str | None,
@@ -298,7 +304,7 @@ class ReasoningParserManager:
if isinstance(name, str):
names = [name]
elif is_list_of(name, str):
names = name
names = cast(list[str], name)
else:
names = [class_name]
+35 -3
View File
@@ -52,6 +52,16 @@ class Gemma4ReasoningParser(BaseThinkingReasoningParser):
# skip_special_tokens=True).
self._reasoning_text: str = ""
self._prefix_stripped: bool = False
self.new_turn_token_id = self.vocab["<|turn>"]
self.tool_call_token_id = self.vocab["<|tool_call>"]
self.tool_response_token_id = self.vocab["<|tool_response>"]
def adjust_request(
self, request: "ChatCompletionRequest | ResponsesRequest"
) -> "ChatCompletionRequest | ResponsesRequest":
"""Disable special-token stripping to preserve boundary tokens."""
request.skip_special_tokens = False
return request
@property
def start_token(self) -> str:
@@ -63,6 +73,29 @@ class Gemma4ReasoningParser(BaseThinkingReasoningParser):
"""The token that ends reasoning content."""
return "<channel|>"
def is_reasoning_end(self, input_ids: Sequence[int]) -> bool:
start_token_id = self.start_token_id
end_token_id = self.end_token_id
new_turn_token_id = self.new_turn_token_id
tool_call_token_id = self.tool_call_token_id
tool_response_token_id = self.tool_response_token_id
# Search from the end of input_ids to find the last match.
for i in range(len(input_ids) - 1, -1, -1):
if input_ids[i] == start_token_id:
return False
if input_ids[i] == tool_call_token_id:
# We're generating a tool call, so reasoning must be ended.
return True
if input_ids[i] in (new_turn_token_id, tool_response_token_id):
# We found a new turn or tool response token so don't consider
# reasoning ended yet, since the model starts new reasoning
# after these tokens.
return False
if input_ids[i] == end_token_id:
return True
return False
# ------------------------------------------------------------------
# Non-streaming path
# ------------------------------------------------------------------
@@ -159,11 +192,10 @@ class Gemma4ReasoningParser(BaseThinkingReasoningParser):
result.reasoning = stripped
return result
else:
# This entire delta was prefix — suppress it.
# Don't set _prefix_stripped yet; there may be more
# prefix chars to consume in the next delta.
if len(self._reasoning_text) >= prefix_len:
self._prefix_stripped = True
result.reasoning = ""
return result
return None
# Case 2: Accumulated text is a strict prefix of
+4 -2
View File
@@ -122,14 +122,16 @@ def _parse_gemma4_args(args_str: str, *, partial: bool = False) -> dict:
# Parse value
if i >= n:
result[key] = ""
if not partial:
result[key] = ""
break
# Skip whitespace after ':'
while i < n and args_str[i] in (" ", "\n", "\t"):
i += 1
if i >= n:
result[key] = ""
if not partial:
result[key] = ""
break
# String value: <|"|>...<|"|>