[Frontend] Binary embedding response does not return metadata by setting encoding_format to bytes_only. (#30249)
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io> Signed-off-by: wang.yuqi <noooop@126.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
This commit is contained in:
co-authored by
gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
parent
408cf42f67
commit
2e660c2434
@@ -59,8 +59,8 @@ async def create_embedding(
|
||||
return JSONResponse(content=generator.model_dump())
|
||||
elif isinstance(generator, EmbeddingBytesResponse):
|
||||
return StreamingResponse(
|
||||
content=generator.body,
|
||||
headers={"metadata": generator.metadata},
|
||||
content=generator.content,
|
||||
headers=generator.headers,
|
||||
media_type=generator.media_type,
|
||||
)
|
||||
|
||||
|
||||
@@ -203,6 +203,6 @@ class EmbeddingResponse(OpenAIBaseModel):
|
||||
|
||||
|
||||
class EmbeddingBytesResponse(OpenAIBaseModel):
|
||||
body: list[bytes]
|
||||
metadata: str
|
||||
content: list[bytes]
|
||||
headers: dict[str, str] | None = None
|
||||
media_type: str = "application/octet-stream"
|
||||
|
||||
@@ -163,29 +163,35 @@ class EmbeddingMixin(OpenAIServing):
|
||||
usage=usage,
|
||||
)
|
||||
|
||||
def encode_bytes():
|
||||
body, items, usage = encode_pooling_bytes(
|
||||
def encode_bytes(bytes_only: bool) -> EmbeddingBytesResponse:
|
||||
content, items, usage = encode_pooling_bytes(
|
||||
pooling_outputs=final_res_batch_checked,
|
||||
embed_dtype=embed_dtype,
|
||||
endianness=endianness,
|
||||
)
|
||||
|
||||
metadata = {
|
||||
"id": ctx.request_id,
|
||||
"created": ctx.created_time,
|
||||
"model": ctx.model_name,
|
||||
"data": items,
|
||||
"usage": usage,
|
||||
}
|
||||
return EmbeddingBytesResponse(
|
||||
body=body,
|
||||
metadata=json.dumps(metadata),
|
||||
headers = (
|
||||
None
|
||||
if bytes_only
|
||||
else {
|
||||
"metadata": json.dumps(
|
||||
{
|
||||
"id": ctx.request_id,
|
||||
"created": ctx.created_time,
|
||||
"model": ctx.model_name,
|
||||
"data": items,
|
||||
"usage": usage,
|
||||
}
|
||||
)
|
||||
}
|
||||
)
|
||||
|
||||
return EmbeddingBytesResponse(content=content, headers=headers)
|
||||
|
||||
if encoding_format == "float" or encoding_format == "base64":
|
||||
return encode_float_base64()
|
||||
elif encoding_format == "bytes":
|
||||
return encode_bytes()
|
||||
elif encoding_format == "bytes" or encoding_format == "bytes_only":
|
||||
return encode_bytes(bytes_only=encoding_format == "bytes_only")
|
||||
else:
|
||||
assert_never(encoding_format)
|
||||
|
||||
|
||||
@@ -55,8 +55,8 @@ async def create_pooling(request: PoolingRequest, raw_request: Request):
|
||||
return JSONResponse(content=generator.model_dump())
|
||||
elif isinstance(generator, PoolingBytesResponse):
|
||||
return StreamingResponse(
|
||||
content=generator.body,
|
||||
headers={"metadata": generator.metadata},
|
||||
content=generator.content,
|
||||
headers=generator.headers,
|
||||
media_type=generator.media_type,
|
||||
)
|
||||
|
||||
|
||||
@@ -143,6 +143,6 @@ class PoolingResponse(OpenAIBaseModel):
|
||||
|
||||
|
||||
class PoolingBytesResponse(OpenAIBaseModel):
|
||||
body: list[bytes]
|
||||
metadata: str
|
||||
content: list[bytes]
|
||||
headers: dict[str, str] | None = None
|
||||
media_type: str = "application/octet-stream"
|
||||
|
||||
@@ -314,29 +314,38 @@ class OpenAIServingPooling(OpenAIServing):
|
||||
usage=usage,
|
||||
)
|
||||
|
||||
def encode_bytes():
|
||||
body, items, usage = encode_pooling_bytes(
|
||||
def encode_bytes(bytes_only: bool) -> PoolingBytesResponse:
|
||||
content, items, usage = encode_pooling_bytes(
|
||||
pooling_outputs=final_res_batch,
|
||||
embed_dtype=embed_dtype,
|
||||
endianness=endianness,
|
||||
)
|
||||
|
||||
metadata = {
|
||||
"id": request_id,
|
||||
"created": created_time,
|
||||
"model": model_name,
|
||||
"data": items,
|
||||
"usage": usage,
|
||||
}
|
||||
headers = (
|
||||
None
|
||||
if bytes_only
|
||||
else {
|
||||
"metadata": json.dumps(
|
||||
{
|
||||
"id": request_id,
|
||||
"created": created_time,
|
||||
"model": model_name,
|
||||
"data": items,
|
||||
"usage": usage,
|
||||
}
|
||||
)
|
||||
}
|
||||
)
|
||||
|
||||
return PoolingBytesResponse(
|
||||
body=body,
|
||||
metadata=json.dumps(metadata),
|
||||
content=content,
|
||||
headers=headers,
|
||||
)
|
||||
|
||||
if encoding_format == "float" or encoding_format == "base64":
|
||||
return encode_float_base64()
|
||||
elif encoding_format == "bytes":
|
||||
return encode_bytes()
|
||||
elif encoding_format == "bytes" or encoding_format == "bytes_only":
|
||||
return encode_bytes(bytes_only=encoding_format == "bytes_only")
|
||||
else:
|
||||
assert_never(encoding_format)
|
||||
|
||||
|
||||
@@ -2,15 +2,19 @@
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
import base64
|
||||
import io
|
||||
import math
|
||||
import sys
|
||||
from dataclasses import dataclass
|
||||
from typing import Literal
|
||||
from typing import TYPE_CHECKING, Any, Literal
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from typing_extensions import assert_never
|
||||
|
||||
from vllm import PoolingRequestOutput
|
||||
if TYPE_CHECKING:
|
||||
from vllm import PoolingRequestOutput
|
||||
else:
|
||||
PoolingRequestOutput = Any
|
||||
|
||||
sys_byteorder = sys.byteorder
|
||||
|
||||
@@ -27,6 +31,14 @@ EMBED_DTYPE_TO_TORCH_DTYPE = {
|
||||
"fp8_e5m2": torch.float8_e5m2,
|
||||
}
|
||||
|
||||
EMBED_DTYPE_TO_N_BYTES = {
|
||||
"float32": 4,
|
||||
"float16": 2,
|
||||
"bfloat16": 2,
|
||||
"fp8_e4m3": 1,
|
||||
"fp8_e5m2": 1,
|
||||
}
|
||||
|
||||
|
||||
EMBED_DTYPE_TO_TORCH_DTYPE_VIEW = {
|
||||
"float32": torch.float32,
|
||||
@@ -50,7 +62,7 @@ ENDIANNESS = ["native", "big", "little"]
|
||||
|
||||
EmbedDType = Literal["float32", "float16", "bfloat16", "fp8_e4m3", "fp8_e5m2"]
|
||||
Endianness = Literal["native", "big", "little"]
|
||||
EncodingFormat = Literal["float", "base64", "bytes"]
|
||||
EncodingFormat = Literal["float", "base64", "bytes", "bytes_only"]
|
||||
|
||||
|
||||
def tensor2base64(x: torch.Tensor) -> str:
|
||||
@@ -114,7 +126,7 @@ def encode_pooling_output(
|
||||
elif encoding_format == "base64":
|
||||
embedding_bytes = tensor2binary(output.outputs.data, embed_dtype, endianness)
|
||||
return base64.b64encode(embedding_bytes).decode("utf-8")
|
||||
elif encoding_format == "bytes":
|
||||
elif encoding_format == "bytes" or encoding_format == "bytes_only":
|
||||
return tensor2binary(output.outputs.data, embed_dtype, endianness)
|
||||
assert_never(encoding_format)
|
||||
|
||||
@@ -129,6 +141,29 @@ class MetadataItem:
|
||||
shape: tuple[int, ...]
|
||||
|
||||
|
||||
def build_metadata_items(
|
||||
embed_dtype: EmbedDType,
|
||||
endianness: Endianness,
|
||||
shape: tuple[int, ...],
|
||||
n_request: int,
|
||||
):
|
||||
n_bytes = EMBED_DTYPE_TO_N_BYTES[embed_dtype]
|
||||
size = math.prod(shape)
|
||||
items = [
|
||||
MetadataItem(
|
||||
index=i,
|
||||
embed_dtype=embed_dtype,
|
||||
endianness=endianness,
|
||||
start=i * size * n_bytes,
|
||||
end=(i + 1) * size * n_bytes,
|
||||
shape=shape,
|
||||
)
|
||||
for i in range(n_request)
|
||||
]
|
||||
|
||||
return items
|
||||
|
||||
|
||||
def encode_pooling_bytes(
|
||||
pooling_outputs: list[PoolingRequestOutput],
|
||||
embed_dtype: EmbedDType,
|
||||
|
||||
Reference in New Issue
Block a user