chore: import upstream snapshot with attribution
This commit is contained in:
wehub-resource-sync
@@ -0,0 +1,377 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
# Adapted from
|
||||
# https://github.com/lm-sys/FastChat/blob/168ccc29d3f7edc50823016105c024fe2282732a/fastchat/protocol/openai_api_protocol.py
|
||||
import time
|
||||
from http import HTTPStatus
|
||||
from typing import Any, ClassVar, Literal, TypeAlias
|
||||
|
||||
import regex as re
|
||||
from pydantic import (
|
||||
BaseModel,
|
||||
ConfigDict,
|
||||
Field,
|
||||
model_serializer,
|
||||
model_validator,
|
||||
)
|
||||
|
||||
from vllm.entrypoints.chat_utils import make_tool_call_id
|
||||
from vllm.exceptions import VLLMValidationError
|
||||
from vllm.logger import init_logger
|
||||
from vllm.utils import random_uuid
|
||||
from vllm.utils.import_utils import resolve_obj_by_qualname
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class OpenAIBaseModel(BaseModel):
|
||||
# OpenAI API does allow extra fields
|
||||
model_config = ConfigDict(extra="allow")
|
||||
|
||||
# Cache class field names
|
||||
field_names: ClassVar[set[str] | None] = None
|
||||
|
||||
@model_validator(mode="wrap")
|
||||
@classmethod
|
||||
def __log_extra_fields__(cls, data, handler):
|
||||
result = handler(data)
|
||||
if not isinstance(data, dict):
|
||||
return result
|
||||
field_names = cls.field_names
|
||||
if field_names is None:
|
||||
# Get all class field names and their potential aliases
|
||||
field_names = set()
|
||||
for field_name, field in cls.model_fields.items():
|
||||
field_names.add(field_name)
|
||||
if alias := getattr(field, "alias", None):
|
||||
field_names.add(alias)
|
||||
cls.field_names = field_names
|
||||
|
||||
# Compare against both field names and aliases
|
||||
if any(k not in field_names for k in data):
|
||||
logger.debug(
|
||||
"The following fields were present in the request but ignored: %s",
|
||||
data.keys() - field_names,
|
||||
)
|
||||
return result
|
||||
|
||||
|
||||
class ErrorInfo(OpenAIBaseModel):
|
||||
message: str
|
||||
type: str
|
||||
param: str | None = None
|
||||
code: int
|
||||
|
||||
|
||||
class ErrorResponse(OpenAIBaseModel):
|
||||
error: ErrorInfo
|
||||
|
||||
|
||||
class ModelPermission(OpenAIBaseModel):
|
||||
id: str = Field(default_factory=lambda: f"modelperm-{random_uuid()}")
|
||||
object: str = "model_permission"
|
||||
created: int = Field(default_factory=lambda: int(time.time()))
|
||||
allow_create_engine: bool = False
|
||||
allow_sampling: bool = True
|
||||
allow_logprobs: bool = True
|
||||
allow_search_indices: bool = False
|
||||
allow_view: bool = True
|
||||
allow_fine_tuning: bool = False
|
||||
organization: str = "*"
|
||||
group: str | None = None
|
||||
is_blocking: bool = False
|
||||
|
||||
|
||||
class ModelCard(OpenAIBaseModel):
|
||||
id: str
|
||||
object: str = "model"
|
||||
created: int = Field(default_factory=lambda: int(time.time()))
|
||||
owned_by: str = "vllm"
|
||||
root: str | None = None
|
||||
parent: str | None = None
|
||||
max_model_len: int | None = None
|
||||
permission: list[ModelPermission] = Field(default_factory=list)
|
||||
|
||||
|
||||
class ModelList(OpenAIBaseModel):
|
||||
object: str = "list"
|
||||
data: list[ModelCard] = Field(default_factory=list)
|
||||
|
||||
|
||||
class PromptTokenUsageInfo(OpenAIBaseModel):
|
||||
cached_tokens: int | None = None
|
||||
multimodal_tokens: dict[str, int] | None = None
|
||||
"""Prompt tokens contributed by each input modality, keyed by modality name
|
||||
(e.g. `image`, `audio`, `video`). A breakdown of the multimodal
|
||||
placeholder tokens already counted in `prompt_tokens`; `None` when the
|
||||
request has no multimodal input."""
|
||||
|
||||
|
||||
class UsageInfo(OpenAIBaseModel):
|
||||
prompt_tokens: int = 0
|
||||
total_tokens: int = 0
|
||||
completion_tokens: int | None = 0
|
||||
prompt_tokens_details: PromptTokenUsageInfo | None = None
|
||||
|
||||
|
||||
class PerRequestTimingMetrics(OpenAIBaseModel):
|
||||
time_to_first_token_ms: float | None = None
|
||||
generation_time_ms: float | None = None
|
||||
queue_time_ms: float | None = None
|
||||
mean_itl_ms: float | None = None
|
||||
tokens_per_second: float | None = None
|
||||
|
||||
|
||||
class RequestResponseMetadata(BaseModel):
|
||||
request_id: str
|
||||
final_usage_info: UsageInfo | None = None
|
||||
|
||||
|
||||
class JsonSchemaResponseFormat(OpenAIBaseModel):
|
||||
name: str
|
||||
description: str | None = None
|
||||
# schema is the field in openai but that causes conflicts with pydantic so
|
||||
# instead use json_schema with an alias
|
||||
json_schema: dict[str, Any] | None = Field(default=None, alias="schema")
|
||||
strict: bool | None = None
|
||||
|
||||
|
||||
class LegacyStructuralTag(OpenAIBaseModel):
|
||||
begin: str
|
||||
# schema is the field, but that causes conflicts with pydantic so
|
||||
# instead use structural_tag_schema with an alias
|
||||
structural_tag_schema: dict[str, Any] | None = Field(default=None, alias="schema")
|
||||
end: str
|
||||
|
||||
|
||||
class LegacyStructuralTagResponseFormat(OpenAIBaseModel):
|
||||
type: Literal["structural_tag"]
|
||||
structures: list[LegacyStructuralTag]
|
||||
triggers: list[str]
|
||||
|
||||
|
||||
class StructuralTagResponseFormat(OpenAIBaseModel):
|
||||
type: Literal["structural_tag"]
|
||||
format: Any
|
||||
|
||||
|
||||
AnyStructuralTagResponseFormat: TypeAlias = (
|
||||
LegacyStructuralTagResponseFormat | StructuralTagResponseFormat
|
||||
)
|
||||
|
||||
|
||||
class ResponseFormat(OpenAIBaseModel):
|
||||
# type must be "json_schema", "json_object", or "text"
|
||||
type: Literal["text", "json_object", "json_schema"]
|
||||
json_schema: JsonSchemaResponseFormat | None = None
|
||||
|
||||
|
||||
AnyResponseFormat: TypeAlias = (
|
||||
ResponseFormat | StructuralTagResponseFormat | LegacyStructuralTagResponseFormat
|
||||
)
|
||||
|
||||
|
||||
def validate_structural_tag_response_format(
|
||||
response_format: AnyStructuralTagResponseFormat | dict[str, Any],
|
||||
) -> None:
|
||||
"""Validate structural tags before they are sent to the engine.
|
||||
|
||||
Engine-side validation reports malformed structural tags as generation
|
||||
failures. OpenAI request parsing should classify them as bad requests.
|
||||
"""
|
||||
import json
|
||||
|
||||
from pydantic import TypeAdapter, ValidationError
|
||||
|
||||
if isinstance(response_format, dict):
|
||||
try:
|
||||
response_format = TypeAdapter(
|
||||
AnyStructuralTagResponseFormat
|
||||
).validate_python(response_format)
|
||||
except ValidationError as exc:
|
||||
raise VLLMValidationError(
|
||||
"Invalid response_format structural_tag specification.",
|
||||
parameter="response_format",
|
||||
) from exc
|
||||
|
||||
try:
|
||||
payload = json.dumps(response_format.model_dump(by_alias=True))
|
||||
validate_structural_tag_payload(payload, parameter="response_format")
|
||||
except (TypeError, ValueError) as exc:
|
||||
raise VLLMValidationError(
|
||||
"Invalid response_format structural_tag specification.",
|
||||
parameter="response_format",
|
||||
) from exc
|
||||
|
||||
|
||||
def validate_structural_tag_payload(payload: Any, *, parameter: str) -> None:
|
||||
from vllm.sampling_params import SamplingParams, StructuredOutputsParams
|
||||
from vllm.v1.structured_output.backend_xgrammar import validate_xgrammar_grammar
|
||||
|
||||
if isinstance(payload, str) and not payload:
|
||||
raise VLLMValidationError(
|
||||
f"Invalid {parameter} structural_tag specification.",
|
||||
parameter=parameter,
|
||||
)
|
||||
|
||||
try:
|
||||
validate_xgrammar_grammar(
|
||||
SamplingParams(
|
||||
structured_outputs=StructuredOutputsParams(structural_tag=payload)
|
||||
)
|
||||
)
|
||||
except (TypeError, ValueError) as exc:
|
||||
raise VLLMValidationError(
|
||||
f"Invalid {parameter} structural_tag specification.",
|
||||
parameter=parameter,
|
||||
) from exc
|
||||
|
||||
|
||||
def validate_structured_outputs_structural_tag(
|
||||
structured_outputs: Any,
|
||||
) -> None:
|
||||
from vllm.sampling_params import StructuredOutputsParams
|
||||
|
||||
if isinstance(structured_outputs, StructuredOutputsParams):
|
||||
structural_tag = structured_outputs.structural_tag
|
||||
elif isinstance(structured_outputs, dict):
|
||||
structural_tag = structured_outputs.get("structural_tag")
|
||||
else:
|
||||
return
|
||||
if structural_tag is not None:
|
||||
validate_structural_tag_payload(
|
||||
structural_tag,
|
||||
parameter="structured_outputs",
|
||||
)
|
||||
|
||||
|
||||
class StreamOptions(OpenAIBaseModel):
|
||||
include_usage: bool | None = False
|
||||
continuous_usage_stats: bool | None = False
|
||||
|
||||
|
||||
class FunctionDefinition(OpenAIBaseModel):
|
||||
name: str
|
||||
description: str | None = None
|
||||
parameters: dict[str, Any] | None = None
|
||||
strict: bool | None = None
|
||||
defer_loading: bool | None = None
|
||||
|
||||
@model_serializer(mode="wrap")
|
||||
def _serialize(self, handler):
|
||||
data = handler(self)
|
||||
if self.strict is None:
|
||||
data.pop("strict", None)
|
||||
if self.defer_loading is None:
|
||||
data.pop("defer_loading", None)
|
||||
return data
|
||||
|
||||
|
||||
# extra="forbid" is a workaround to have kwargs as a field,
|
||||
# see https://github.com/pydantic/pydantic/issues/3125
|
||||
class LogitsProcessorConstructor(BaseModel):
|
||||
qualname: str
|
||||
args: list[Any] | None = None
|
||||
kwargs: dict[str, Any] | None = None
|
||||
|
||||
model_config = ConfigDict(extra="forbid")
|
||||
|
||||
|
||||
LogitsProcessors = list[str | LogitsProcessorConstructor]
|
||||
|
||||
|
||||
def get_logits_processors(
|
||||
processors: LogitsProcessors | None, pattern: str | None
|
||||
) -> list[Any] | None:
|
||||
if processors and pattern:
|
||||
logits_processors = []
|
||||
for processor in processors:
|
||||
qualname = processor if isinstance(processor, str) else processor.qualname
|
||||
if not re.match(pattern, qualname):
|
||||
raise ValueError(
|
||||
f"Logits processor '{qualname}' is not allowed by this "
|
||||
"server. See --logits-processor-pattern engine argument "
|
||||
"for more information."
|
||||
)
|
||||
try:
|
||||
logits_processor = resolve_obj_by_qualname(qualname)
|
||||
except Exception as e:
|
||||
raise ValueError(
|
||||
f"Logits processor '{qualname}' could not be resolved: {e}"
|
||||
) from e
|
||||
if isinstance(processor, LogitsProcessorConstructor):
|
||||
logits_processor = logits_processor(
|
||||
*processor.args or [], **processor.kwargs or {}
|
||||
)
|
||||
logits_processors.append(logits_processor)
|
||||
return logits_processors
|
||||
elif processors:
|
||||
raise ValueError(
|
||||
"The `logits_processors` argument is not supported by this "
|
||||
"server. See --logits-processor-pattern engine argument "
|
||||
"for more information."
|
||||
)
|
||||
return None
|
||||
|
||||
|
||||
class FunctionCall(OpenAIBaseModel):
|
||||
# Internal field to preserve native tool call ID from tool parser.
|
||||
# Excluded from serialization to maintain OpenAI API compatibility
|
||||
# (function object should only contain 'name' and 'arguments').
|
||||
id: str | None = Field(default=None, exclude=True)
|
||||
name: str
|
||||
arguments: str
|
||||
|
||||
|
||||
class ToolCall(OpenAIBaseModel):
|
||||
id: str = Field(default_factory=make_tool_call_id)
|
||||
type: Literal["function"] = "function"
|
||||
function: FunctionCall
|
||||
|
||||
|
||||
class DeltaFunctionCall(BaseModel):
|
||||
name: str | None = None
|
||||
arguments: str | None = None
|
||||
|
||||
|
||||
# a tool call delta where everything is optional
|
||||
class DeltaToolCall(OpenAIBaseModel):
|
||||
id: str | None = None
|
||||
type: Literal["function"] | None = None
|
||||
index: int
|
||||
function: DeltaFunctionCall | None = None
|
||||
|
||||
|
||||
class ExtractedToolCallInformation(BaseModel):
|
||||
# indicate if tools were called
|
||||
tools_called: bool
|
||||
|
||||
# extracted tool calls
|
||||
tool_calls: list[ToolCall]
|
||||
|
||||
# content - per OpenAI spec, content AND tool calls can be returned rarely
|
||||
# But some models will do this intentionally
|
||||
content: str | None = None
|
||||
|
||||
|
||||
class DeltaMessage(OpenAIBaseModel):
|
||||
role: str | None = None
|
||||
content: str | None = None
|
||||
reasoning: str | None = None
|
||||
tool_calls: list[DeltaToolCall] = Field(default_factory=list)
|
||||
|
||||
@model_serializer(mode="wrap")
|
||||
def _serialize(self, handler):
|
||||
data = handler(self)
|
||||
if len(data.get("tool_calls", [])) == 0:
|
||||
data.pop("tool_calls", None)
|
||||
return data
|
||||
|
||||
|
||||
class GenerationError(Exception):
|
||||
"""raised when finish_reason indicates internal server error (500)"""
|
||||
|
||||
def __init__(self, message: str = "Internal server error"):
|
||||
super().__init__(message)
|
||||
self.status_code = HTTPStatus.INTERNAL_SERVER_ERROR
|
||||
Reference in New Issue
Block a user