[Frontend] Let an unset --tool-strict-level fall back to the tool parser's default (#61328)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Bugen Zhao
2026-10-15 15:48:38 +08:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 91ed2859c7
commit 8fa5ae844a
26 changed files with 181 additions and 90 deletions
+1 -1
View File
@@ -160,7 +160,7 @@ class AnthropicServingMessages(OpenAIServingChat):
reasoning_parser: str = "",
enable_auto_tools: bool = False,
tool_parser: str | None = None,
tool_strict_level: str = "auto",
tool_strict_level: str | None = None,
enable_prompt_tokens_details: bool = False,
enable_force_include_usage: bool = False,
enable_log_outputs: bool = False,
+1 -1
View File
@@ -251,7 +251,7 @@ class CohereServingChatV2(OpenAIServingChat):
reasoning_parser: str = "",
enable_auto_tools: bool = False,
tool_parser: str | None = None,
tool_strict_level: str = "auto",
tool_strict_level: str | None = None,
enable_prompt_tokens_details: bool = False,
enable_force_include_usage: bool = False,
enable_log_outputs: bool = False,
+3 -2
View File
@@ -128,13 +128,14 @@ class BaseFrontendArgs:
"""Special the tool parser plugin write to parse the model-generated tool
into OpenAI API format, the name register in this plugin can be used in
`--tool-call-parser`."""
tool_strict_level: ToolStrictLevelName = "auto"
tool_strict_level: ToolStrictLevelName | None = None
"""Server-side floor for structural-tag based tool calling, applied on top
of the per-tool `strict` field. `auto` follows the request's tool choice
and per-tool strictness; `function` constrains the
tool-call envelope (markup and function name) for every request with
tools; `parameter` additionally pins argument schemas, as if every tool
were `strict: true`."""
were `strict: true`. When unset, the tool parser's default level applies,
which is `auto` for most parsers."""
tool_server: str | None = None
"""Comma-separated list of host:port pairs (IPv4, IPv6, or hostname).
Examples: 127.0.0.1:8000, [::1]:8000, localhost:1234. Or `demo` for
@@ -132,7 +132,7 @@ class OpenAIServingChat(GenerateBaseServing):
enable_auto_tools: bool = False,
exclude_tools_when_tool_choice_none: bool = False,
tool_parser: str | None = None,
tool_strict_level: str = "auto",
tool_strict_level: str | None = None,
enable_prompt_tokens_details: bool = False,
enable_force_include_usage: bool = False,
enable_log_outputs: bool = False,
+1 -1
View File
@@ -113,7 +113,7 @@ class OpenAIServingResponses(GenerateBaseServing):
reasoning_parser: str = "",
enable_auto_tools: bool = False,
tool_parser: str | None = None,
tool_strict_level: str = "auto",
tool_strict_level: str | None = None,
tool_server: ToolServer | None = None,
enable_prompt_tokens_details: bool = False,
enable_force_include_usage: bool = False,
+4 -6
View File
@@ -118,8 +118,9 @@ class Parser:
# Subclasses should override these if they use specific parser classes
reasoning_parser_cls: type[ReasoningParser] | None = None
tool_parser_cls: type[ToolParser] | None = None
# Server-side floor for tool-call structural tags (--tool-strict-level).
tool_strict_level: ToolStrictLevel = ToolStrictLevel.AUTO
# Server-side floor for tool-call structural tags (--tool-strict-level), or
# None for the tool parser's default level.
tool_strict_level: ToolStrictLevel | None = None
always_adjust_request: bool = False
def __init__(
@@ -519,10 +520,7 @@ class DelegatingParser(Parser):
is_auto = request.tool_choice == "auto"
single_call = request.parallel_tool_calls is False
strict_level = self.tool_strict_level
if (
strict_level == ToolStrictLevel.AUTO
and tool_parser.default_tool_strict_level is not None
):
if strict_level is None:
strict_level = tool_parser.default_tool_strict_level
if single_call and not (is_auto and structured_outputs is not None):
# Limiting the call count needs the call envelope in the grammar.
+9 -4
View File
@@ -104,7 +104,7 @@ class ParserManager:
enable_auto_tools: bool = False,
model_name: str | None = None,
is_harmony: bool = False,
tool_strict_level: str = "auto",
tool_strict_level: str | None = None,
tokenizer: TokenizerLike | None = None,
) -> type[Parser] | None:
"""Get a Parser that handles both reasoning and tool parsing.
@@ -119,7 +119,8 @@ class ParserManager:
is_harmony: Whether the selected model uses the Harmony format.
If True, HarmonyParser is always returned.
tool_strict_level: Server-side floor for tool-call structural
tags (``--tool-strict-level``).
tags (``--tool-strict-level``), or ``None`` for the tool
parser's default level.
tokenizer: Tokenizer the composed parser is validated against
at startup (see :meth:`validate_parser_tokenizer`). ``None``
skips the check (e.g. ``skip_tokenizer_init``).
@@ -155,7 +156,7 @@ class ParserManager:
enable_auto_tools: bool,
model_name: str | None,
is_harmony: bool,
tool_strict_level: str,
tool_strict_level: str | None,
tokenizer: TokenizerLike | None,
) -> type[Parser] | None:
if not tool_parser_name and not reasoning_parser_name:
@@ -169,7 +170,11 @@ class ParserManager:
if reasoning_parser_cls is None and tool_parser_cls is None:
return None
strict_level = ToolStrictLevel.from_name(tool_strict_level)
strict_level = (
None
if tool_strict_level is None
else ToolStrictLevel.from_name(tool_strict_level)
)
if is_harmony:
from vllm.parser.harmony import HarmonyParser
+1 -1
View File
@@ -69,7 +69,7 @@ class OnlineDerenderer:
exclude_tools_when_tool_choice_none: bool = False,
tool_parser: str | None = None,
reasoning_parser: str | None = None,
tool_strict_level: str = "auto",
tool_strict_level: str | None = None,
default_chat_template_kwargs: dict[str, Any] | None = None,
log_error_stack: bool = False,
) -> None:
+1 -1
View File
@@ -165,7 +165,7 @@ class OnlineRenderer:
exclude_tools_when_tool_choice_none: bool = False,
tool_parser: str | None = None,
reasoning_parser: str | None = None,
tool_strict_level: str = "auto",
tool_strict_level: str | None = None,
default_chat_template_kwargs: dict[str, Any] | None = None,
log_error_stack: bool = False,
) -> None:
+2 -2
View File
@@ -60,9 +60,9 @@ class ToolParser:
# their parsed tool-call syntax matches a builtin xgrammar format.
structural_tag_model: str | None = None
# Strict level applied when the operator leaves ``--tool-strict-level``
# at auto. Operator flags override it, and
# unset. Operator flags override it, and
# ``VLLM_ENFORCE_STRICT_TOOL_CALLING=0`` still disables structural tags.
default_tool_strict_level: ToolStrictLevel | None = None
default_tool_strict_level: ToolStrictLevel = ToolStrictLevel.AUTO
engine_based_streaming: bool = False
def __init_subclass__(cls, **kwargs: Any) -> None: