[Bugfix][Frontend] Force reasoning mode for GLM-5.3 chat templates in the GLM MoE parser (#56994)
Signed-off-by: Shijin Zhang <75300765+Dovis01@users.noreply.github.com> Co-authored-by: Chauncey <chaunceyjiang@gmail.com>
This commit is contained in:
@@ -1445,7 +1445,7 @@ class TestDetectMergeInlineSystem:
|
|||||||
AnthropicServingMessages._should_merge_inline_system(online_renderer)
|
AnthropicServingMessages._should_merge_inline_system(online_renderer)
|
||||||
is expected
|
is expected
|
||||||
)
|
)
|
||||||
assert mock_logger.warning.called is expected
|
assert mock_logger.warning_once.called is expected
|
||||||
|
|
||||||
|
|
||||||
# ======================================================================
|
# ======================================================================
|
||||||
|
|||||||
@@ -1,6 +1,8 @@
|
|||||||
# SPDX-License-Identifier: Apache-2.0
|
# SPDX-License-Identifier: Apache-2.0
|
||||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||||
|
|
||||||
|
import copy
|
||||||
|
|
||||||
import pytest
|
import pytest
|
||||||
from transformers import AutoTokenizer
|
from transformers import AutoTokenizer
|
||||||
|
|
||||||
@@ -227,3 +229,81 @@ def test_is_reasoning_end_full_prompt(
|
|||||||
token_ids = glm45_tokenizer.convert_tokens_to_ids(tokens)
|
token_ids = glm45_tokenizer.convert_tokens_to_ids(tokens)
|
||||||
check_is_reasoning_end = parser.is_reasoning_end(token_ids)
|
check_is_reasoning_end = parser.is_reasoning_end(token_ids)
|
||||||
assert check_is_reasoning_end == is_reasoning_end
|
assert check_is_reasoning_end == is_reasoning_end
|
||||||
|
|
||||||
|
|
||||||
|
GLM53_TEMPLATE = (
|
||||||
|
"[gMASK]<sop>\n"
|
||||||
|
"{%- set effective_reasoning_effort = reasoning_effort if reasoning_effort is"
|
||||||
|
" defined and reasoning_effort in ['low', 'high'] else 'max' -%}\n"
|
||||||
|
"<|system|>Reasoning Effort: {{ effective_reasoning_effort | capitalize }}\n"
|
||||||
|
"{% for tc in m.tool_calls %}\n"
|
||||||
|
"{{- '<tool_call>' + tc.name -}}\n"
|
||||||
|
"{% set _args = tc.arguments %}"
|
||||||
|
"{% for k, v in _args.items() %}"
|
||||||
|
"<arg_key>{{ k }}</arg_key><arg_value>{{ v }}</arg_value>"
|
||||||
|
"{% endfor %}</tool_call>\n"
|
||||||
|
"{% endfor %}\n"
|
||||||
|
"<|assistant|>{{- '<think>' -}}"
|
||||||
|
)
|
||||||
|
|
||||||
|
GLM53_LEAK = {
|
||||||
|
"output": "Simple question.</think>2 + 2 = **4**",
|
||||||
|
"reasoning": "Simple question.",
|
||||||
|
"content": "2 + 2 = **4**",
|
||||||
|
"is_reasoning_end": True,
|
||||||
|
}
|
||||||
|
|
||||||
|
|
||||||
|
@pytest.fixture()
|
||||||
|
def glm53_style_tokenizer(glm45_tokenizer):
|
||||||
|
tokenizer = copy.copy(glm45_tokenizer)
|
||||||
|
tokenizer.chat_template = GLM53_TEMPLATE
|
||||||
|
return tokenizer
|
||||||
|
|
||||||
|
|
||||||
|
@pytest.mark.parametrize(
|
||||||
|
"disable_kwargs", [{"enable_thinking": False}, {"thinking": False}]
|
||||||
|
)
|
||||||
|
def test_glm53_template_forces_reasoning(disable_kwargs: dict, glm53_style_tokenizer):
|
||||||
|
parser_cls = ReasoningParserManager.get_reasoning_parser(parser_name)
|
||||||
|
parser = parser_cls(glm53_style_tokenizer, chat_template_kwargs=disable_kwargs)
|
||||||
|
assert parser._parser_engine.thinking_enabled
|
||||||
|
|
||||||
|
output = glm53_style_tokenizer.tokenize(GLM53_LEAK["output"])
|
||||||
|
output_tokens: list[str] = [
|
||||||
|
glm53_style_tokenizer.convert_tokens_to_string([token]) for token in output
|
||||||
|
]
|
||||||
|
reasoning, content = run_reasoning_extraction(parser, output_tokens)
|
||||||
|
assert reasoning == GLM53_LEAK["reasoning"]
|
||||||
|
assert content == GLM53_LEAK["content"]
|
||||||
|
|
||||||
|
output_ids = glm53_style_tokenizer.convert_tokens_to_ids(output)
|
||||||
|
assert parser.is_reasoning_end(output_ids) == GLM53_LEAK["is_reasoning_end"]
|
||||||
|
|
||||||
|
|
||||||
|
def test_glm53_template_forces_reasoning_streaming(glm53_style_tokenizer):
|
||||||
|
parser = ReasoningParserManager.get_reasoning_parser(parser_name)(
|
||||||
|
glm53_style_tokenizer, chat_template_kwargs={"enable_thinking": False}
|
||||||
|
)
|
||||||
|
output = glm53_style_tokenizer.tokenize(GLM53_LEAK["output"])
|
||||||
|
output_tokens: list[str] = [
|
||||||
|
glm53_style_tokenizer.convert_tokens_to_string([token]) for token in output
|
||||||
|
]
|
||||||
|
reasoning, content = run_reasoning_extraction(parser, output_tokens, streaming=True)
|
||||||
|
assert reasoning == GLM53_LEAK["reasoning"]
|
||||||
|
assert content == GLM53_LEAK["content"]
|
||||||
|
|
||||||
|
|
||||||
|
def test_glm47_template_honors_thinking_disable(glm45_tokenizer):
|
||||||
|
parser = ReasoningParserManager.get_reasoning_parser(parser_name)(
|
||||||
|
glm45_tokenizer, chat_template_kwargs={"enable_thinking": False}
|
||||||
|
)
|
||||||
|
assert not parser._parser_engine.thinking_enabled
|
||||||
|
|
||||||
|
output = glm45_tokenizer.tokenize(GLM53_LEAK["output"])
|
||||||
|
output_tokens: list[str] = [
|
||||||
|
glm45_tokenizer.convert_tokens_to_string([token]) for token in output
|
||||||
|
]
|
||||||
|
reasoning, content = run_reasoning_extraction(parser, output_tokens)
|
||||||
|
assert reasoning is None
|
||||||
|
assert content == GLM53_LEAK["output"]
|
||||||
|
|||||||
@@ -20,6 +20,7 @@ import regex as re
|
|||||||
|
|
||||||
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
|
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
|
||||||
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
|
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
|
||||||
|
from vllm.logger import init_logger
|
||||||
from vllm.parser.engine.events import EventType
|
from vllm.parser.engine.events import EventType
|
||||||
from vllm.parser.engine.parser_engine import ParserEngine
|
from vllm.parser.engine.parser_engine import ParserEngine
|
||||||
from vllm.parser.engine.parser_engine_config import (
|
from vllm.parser.engine.parser_engine_config import (
|
||||||
@@ -28,6 +29,8 @@ from vllm.parser.engine.parser_engine_config import (
|
|||||||
Transition,
|
Transition,
|
||||||
)
|
)
|
||||||
|
|
||||||
|
logger = init_logger(__name__)
|
||||||
|
|
||||||
if TYPE_CHECKING:
|
if TYPE_CHECKING:
|
||||||
from vllm.tokenizers import TokenizerLike
|
from vllm.tokenizers import TokenizerLike
|
||||||
from vllm.tool_parsers.abstract_tool_parser import Tool
|
from vllm.tool_parsers.abstract_tool_parser import Tool
|
||||||
@@ -77,6 +80,23 @@ def _glm47_arg_converter(raw_args: str, partial: bool) -> str:
|
|||||||
return json.dumps(params, ensure_ascii=False)
|
return json.dumps(params, ensure_ascii=False)
|
||||||
|
|
||||||
|
|
||||||
|
def _glm53_always_thinks(tokenizer: TokenizerLike) -> bool:
|
||||||
|
# GLM-5.3 keeps the GLM-4.x prompt and tool-call format but replaces the
|
||||||
|
# enable_thinking toggle with an always-on "Reasoning Effort:" header, so
|
||||||
|
# its generation prompt always opens <think> and the model output carries
|
||||||
|
# no opening tag.
|
||||||
|
template = getattr(tokenizer, "chat_template", None)
|
||||||
|
return (
|
||||||
|
isinstance(template, str)
|
||||||
|
and "[gMASK]<sop>" in template
|
||||||
|
and "Reasoning Effort:" in template
|
||||||
|
and "enable_thinking" not in template
|
||||||
|
and "<tool_call>" in template
|
||||||
|
and "<arg_key>" in template
|
||||||
|
and "<arg_value>" in template
|
||||||
|
)
|
||||||
|
|
||||||
|
|
||||||
@functools.cache
|
@functools.cache
|
||||||
def glm47_moe_config(thinking: bool = True) -> ParserEngineConfig:
|
def glm47_moe_config(thinking: bool = True) -> ParserEngineConfig:
|
||||||
arg_tag_transitions = {
|
arg_tag_transitions = {
|
||||||
@@ -192,6 +212,16 @@ class Glm47MoeParser(ParserEngine):
|
|||||||
chat_kwargs = kwargs.get("chat_template_kwargs", {}) or {}
|
chat_kwargs = kwargs.get("chat_template_kwargs", {}) or {}
|
||||||
thinking = chat_kwargs.get("thinking", None)
|
thinking = chat_kwargs.get("thinking", None)
|
||||||
enable_thinking = chat_kwargs.get("enable_thinking", None)
|
enable_thinking = chat_kwargs.get("enable_thinking", None)
|
||||||
|
if (thinking is False or enable_thinking is False) and _glm53_always_thinks(
|
||||||
|
tokenizer
|
||||||
|
):
|
||||||
|
logger.warning_once(
|
||||||
|
"Ignoring enable_thinking/thinking: the GLM-5.3 chat template "
|
||||||
|
"has no thinking switch, so reasoning is always on and "
|
||||||
|
"disabling extraction would leak it into the content."
|
||||||
|
)
|
||||||
|
thinking = None
|
||||||
|
enable_thinking = None
|
||||||
self.thinking_enabled = (
|
self.thinking_enabled = (
|
||||||
True
|
True
|
||||||
if thinking is None and enable_thinking is None
|
if thinking is None and enable_thinking is None
|
||||||
|
|||||||
Reference in New Issue
Block a user