[Frontend] Binary embedding response does not return metadata by setting...

[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>

examples/pooling/embed/embedding_requests_bytes_client.py

@@ -16,6 +16,7 @@ from vllm.utils.serial_utils import (
     EMBED_DTYPE_TO_TORCH_DTYPE,
     ENDIANNESS,
     MetadataItem,
+    build_metadata_items,
     decode_pooling_output,
 )
 
@@ -38,6 +39,11 @@ def parse_args():
 def main(args):
     api_url = f"http://{args.host}:{args.port}/v1/embeddings"
     model_name = args.model
+    embedding_size = 0
+
+    input_texts = [
+        "The best thing about vLLM is that it supports many different models",
+    ] * 2
 
     # The OpenAI client does not support the bytes encoding_format.
     # The OpenAI client does not support the embed_dtype and endianness parameters.
@@ -45,7 +51,7 @@ def main(args):
         for endianness in ENDIANNESS:
             prompt = {
                 "model": model_name,
-                "input": "vLLM is great!",
+                "input": input_texts,
                 "encoding_format": "bytes",
                 "embed_dtype": embed_dtype,
                 "endianness": endianness,
@@ -57,7 +63,34 @@ def main(args):
 
             embedding = decode_pooling_output(items=items, body=body)
             embedding = [x.to(torch.float32) for x in embedding]
-            embedding = torch.cat(embedding)
+            embedding = torch.stack(embedding)
+            embedding_size = embedding.shape[-1]
+            print(embed_dtype, endianness, embedding.shape)
+
+    # The vllm server always sorts the returned embeddings in the order of input. So
+    # returning metadata is not necessary. You can set encoding_format to bytes_only
+    # to let the server not return metadata.
+    for embed_dtype in EMBED_DTYPE_TO_TORCH_DTYPE:
+        for endianness in ENDIANNESS:
+            prompt = {
+                "model": model_name,
+                "input": input_texts,
+                "encoding_format": "bytes_only",
+                "embed_dtype": embed_dtype,
+                "endianness": endianness,
+            }
+            response = post_http_request(prompt=prompt, api_url=api_url)
+            body = response.content
+
+            items = build_metadata_items(
+                embed_dtype=embed_dtype,
+                endianness=endianness,
+                shape=(embedding_size,),
+                n_request=len(input_texts),
+            )
+            embedding = decode_pooling_output(items=items, body=body)
+            embedding = [x.to(torch.float32) for x in embedding]
+            embedding = torch.stack(embedding)
             print(embed_dtype, endianness, embedding.shape)
 
 

tests/entrypoints/pooling/embed/test_online.py

@@ -24,6 +24,7 @@ from vllm.utils.serial_utils import (
     ENDIANNESS,
     MetadataItem,
     binary2tensor,
+    build_metadata_items,
     decode_pooling_output,
 )
 
@@ -344,6 +345,55 @@ async def test_bytes_embed_dtype_and_endianness(
             )
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+async def test_bytes_only_embed_dtype_and_endianness(
+    server: RemoteOpenAIServer, client: openai.AsyncOpenAI, model_name: str
+):
+    input_texts = [
+        "The best thing about vLLM is that it supports many different models",
+    ] * 2
+
+    responses_float = await client.embeddings.create(
+        input=input_texts, model=model_name, encoding_format="float"
+    )
+    float_data = [d.embedding for d in responses_float.data]
+    embedding_size = len(float_data[0])
+
+    for embed_dtype in list(EMBED_DTYPE_TO_TORCH_DTYPE.keys()):
+        for endianness in ENDIANNESS:
+            responses_bytes = requests.post(
+                server.url_for("/v1/embeddings"),
+                json={
+                    "model": model_name,
+                    "input": input_texts,
+                    "encoding_format": "bytes_only",
+                    "embed_dtype": embed_dtype,
+                    "endianness": endianness,
+                },
+            )
+
+            assert "metadata" not in responses_bytes.headers
+            body = responses_bytes.content
+            items = build_metadata_items(
+                embed_dtype=embed_dtype,
+                endianness=endianness,
+                shape=(embedding_size,),
+                n_request=len(input_texts),
+            )
+
+            bytes_data = decode_pooling_output(items=items, body=body)
+            bytes_data = [x.to(torch.float32).tolist() for x in bytes_data]
+
+            check_embeddings_close(
+                embeddings_0_lst=float_data,
+                embeddings_1_lst=bytes_data,
+                name_0="float_data",
+                name_1="bytes_data",
+                tol=1e-2,
+            )
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
 @pytest.mark.parametrize("param_name", ["encoding_format", "embed_dtype", "endianness"])

tests/entrypoints/pooling/pooling/test_online.py

@@ -18,6 +18,7 @@ from vllm.utils.serial_utils import (
     ENDIANNESS,
     MetadataItem,
     binary2tensor,
+    build_metadata_items,
     decode_pooling_output,
 )
 
@@ -352,6 +353,61 @@ async def test_bytes_embed_dtype_and_endianness(
             )
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("model_name", [MODEL_NAME])
+async def test_bytes_only_embed_dtype_and_endianness(
+    server: RemoteOpenAIServer, model_name: str
+):
+    input_texts = [
+        "The best thing about vLLM is that it supports many different models",
+    ] * 2
+
+    url = server.url_for("pooling")
+    float_response = requests.post(
+        url,
+        json={
+            "model": model_name,
+            "input": input_texts,
+            "encoding_format": "float",
+        },
+    )
+    responses_float = PoolingResponse.model_validate(float_response.json())
+    float_data = [np.array(d.data).squeeze(-1).tolist() for d in responses_float.data]
+    n_tokens = responses_float.usage.prompt_tokens // len(input_texts)
+
+    for embed_dtype in list(EMBED_DTYPE_TO_TORCH_DTYPE.keys()):
+        for endianness in ENDIANNESS:
+            responses_bytes = requests.post(
+                url,
+                json={
+                    "model": model_name,
+                    "input": input_texts,
+                    "encoding_format": "bytes_only",
+                    "embed_dtype": embed_dtype,
+                    "endianness": endianness,
+                },
+            )
+
+            assert "metadata" not in responses_bytes.headers
+            body = responses_bytes.content
+            items = build_metadata_items(
+                embed_dtype=embed_dtype,
+                endianness=endianness,
+                shape=(n_tokens, 1),
+                n_request=len(input_texts),
+            )
+            bytes_data = decode_pooling_output(items=items, body=body)
+            bytes_data = [x.to(torch.float32).view(-1).tolist() for x in bytes_data]
+
+            check_embeddings_close(
+                embeddings_0_lst=float_data,
+                embeddings_1_lst=bytes_data,
+                name_0="float_data",
+                name_1="bytes_data",
+                tol=1e-2,
+            )
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize("model_name", [MODEL_NAME])
 @pytest.mark.parametrize("param_name", ["encoding_format", "embed_dtype", "endianness"])

vllm/entrypoints/pooling/embed/api_router.py

@@ -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,
         )
 

vllm/entrypoints/pooling/embed/protocol.py

@@ -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"

vllm/entrypoints/pooling/embed/serving.py

@@ -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 = {
+            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(
-                body=body,
-                metadata=json.dumps(metadata),
                     )
+                }
+            )
+
+            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)
 

vllm/entrypoints/pooling/pooling/api_router.py

@@ -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,
         )
 

vllm/entrypoints/pooling/pooling/protocol.py

@@ -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"

vllm/entrypoints/pooling/pooling/serving.py

@@ -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 = {
+            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)
 

vllm/utils/serial_utils.py

@@ -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,