Added real /v1/embeddings support for messages + chat_template_kw (#45173)

Signed-off-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>
This commit is contained in:
Taneem Ibrahim
2026-06-15 09:08:10 +08:00
committed by GitHub
parent c621af1690
commit 2c764c089a
5 changed files with 400 additions and 10 deletions
+107 -3
View File
@@ -10,17 +10,19 @@ import builtins
import struct
import time
from collections.abc import Sequence
from typing import Literal, TypeAlias
from typing import Annotated, Any, Literal, TypeAlias
import pybase64 as base64
from pydantic import BaseModel, Field
from pydantic import BaseModel, Field, model_validator
from vllm import PoolingParams
from vllm.entrypoints.chat_utils import ChatCompletionMessageParam
from vllm.entrypoints.openai.engine.protocol import OpenAIBaseModel, UsageInfo
from vllm.utils import random_uuid
from ..base.protocol import (
ChatRequestMixin,
ChatRequestOptionsMixin,
CompletionRequestMixin,
EmbeddingTokenizeParamsMixin,
EmbedRequestMixin,
@@ -42,12 +44,34 @@ class EmbeddingCompletionRequest(
)
def _is_chat_message(value: Any) -> bool:
return isinstance(value, dict) and isinstance(value.get("role"), str)
def _is_chat_messages(value: Any) -> bool:
return (
isinstance(value, list)
and bool(value)
and all(_is_chat_message(item) for item in value)
)
def _is_batched_chat_messages(value: Any) -> bool:
return (
isinstance(value, list)
and bool(value)
and all(_is_chat_messages(item) for item in value)
)
class EmbeddingChatRequest(
PoolingBasicRequestMixin,
ChatRequestMixin,
EmbedRequestMixin,
EmbeddingTokenizeParamsMixin,
):
"""OpenAI embeddings request with one top-level chat conversation."""
def to_pooling_params(self):
return PoolingParams(
task="embed",
@@ -56,7 +80,87 @@ class EmbeddingChatRequest(
)
EmbeddingRequest: TypeAlias = EmbeddingCompletionRequest | EmbeddingChatRequest
class EmbeddingBatchChatRequest(
PoolingBasicRequestMixin,
ChatRequestOptionsMixin,
EmbedRequestMixin,
EmbeddingTokenizeParamsMixin,
):
"""OpenAI embeddings request with batched top-level chat conversations.
Mirrors ``BatchChatCompletionRequest`` by keeping batched conversations in
``messages`` instead of introducing a separate batch-specific field.
"""
messages: list[Annotated[list[ChatCompletionMessageParam], Field(min_length=1)]] = (
Field(..., min_length=1)
)
def to_pooling_params(self):
return PoolingParams(
task="embed",
dimensions=self.dimensions,
use_activation=self.use_activation,
)
class EmbeddingChatInputRequest(
EmbeddingChatRequest,
):
"""OpenAI embeddings request with one chat conversation in ``input``."""
input: list[ChatCompletionMessageParam]
@model_validator(mode="before")
@classmethod
def normalize_input_messages(cls, data):
if not isinstance(data, dict):
return data
if "messages" in data or "input" not in data:
return data
input_data = data["input"]
if not _is_chat_messages(input_data):
return data
normalized = dict(data)
normalized["messages"] = input_data
return normalized
class EmbeddingBatchChatInputRequest(EmbeddingBatchChatRequest):
"""OpenAI embeddings request with batched chat conversations in ``input``."""
input: list[Annotated[list[ChatCompletionMessageParam], Field(min_length=1)]] = (
Field(..., min_length=1)
)
@model_validator(mode="before")
@classmethod
def normalize_input_messages(cls, data):
if not isinstance(data, dict):
return data
if "messages" in data or "input" not in data:
return data
input_data = data["input"]
if not _is_batched_chat_messages(input_data):
return data
normalized = dict(data)
normalized["messages"] = input_data
return normalized
EmbeddingRequest: TypeAlias = (
EmbeddingCompletionRequest
| EmbeddingChatRequest
| EmbeddingBatchChatRequest
| EmbeddingChatInputRequest
| EmbeddingBatchChatInputRequest
)
# ---------------------------------------------------------------------------