[Render] Add /derender endpoints for disaggregated postprocessing (#43606)

Signed-off-by: Martin Hickey <martin.hickey@ie.ibm.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Martin Hickey
2026-06-13 13:55:33 +08:00
committed by GitHub
co-authored by Isotr0py
parent 1033ffac2e
commit 43f0e024bc
7 changed files with 898 additions and 9 deletions
+244 -2
View File
@@ -1,5 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import time
from collections.abc import Sequence
from http import HTTPStatus
from typing import Any, cast
@@ -11,11 +12,24 @@ from vllm.entrypoints.chat_utils import (
ChatTemplateContentFormatOption,
ConversationMessage,
)
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
from vllm.entrypoints.openai.completion.protocol import CompletionRequest
from vllm.entrypoints.openai.chat_completion.protocol import (
ChatCompletionLogProbs,
ChatCompletionRequest,
ChatCompletionResponse,
ChatCompletionResponseChoice,
ChatMessage,
)
from vllm.entrypoints.openai.completion.protocol import (
CompletionLogProbs,
CompletionRequest,
CompletionResponse,
CompletionResponseChoice,
)
from vllm.entrypoints.openai.engine.protocol import (
ErrorResponse,
UsageInfo,
)
from vllm.entrypoints.openai.engine.serving import resolve_token_id_placeholder
from vllm.entrypoints.openai.models.serving import OpenAIModelRegistry
from vllm.entrypoints.openai.parser.harmony_utils import (
build_harmony_preamble,
@@ -26,7 +40,10 @@ from vllm.entrypoints.openai.parser.harmony_utils import (
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
from vllm.entrypoints.serve.disagg.mm_serde import encode_mm_kwargs_item
from vllm.entrypoints.serve.disagg.protocol import (
DerenderChatRequest,
DerenderCompletionRequest,
GenerateRequest,
GenerateResponseChoice,
MultiModalFeatures,
PlaceholderRangeInfo,
)
@@ -51,6 +68,7 @@ from vllm.renderers.inputs.preprocess import (
parse_model_prompt,
prompt_to_seq,
)
from vllm.tokenizers import TokenizerLike
from vllm.utils import random_uuid
from vllm.utils.mistral import is_mistral_tokenizer, is_mistral_tool_parser
from vllm.utils.mistral import mt as _mt
@@ -58,6 +76,90 @@ from vllm.utils.mistral import mt as _mt
logger = init_logger(__name__)
def _resolve_logprobs(
logprobs: ChatCompletionLogProbs, tokenizer: TokenizerLike
) -> ChatCompletionLogProbs:
"""Resolve all token_id:N placeholders in a ChatCompletionLogProbs object."""
if logprobs.content is None:
return logprobs
resolved_content = []
for entry in logprobs.content:
token_str, token_bytes = resolve_token_id_placeholder(entry.token, tokenizer)
resolved_top = []
for top in entry.top_logprobs:
top_str, top_bytes = resolve_token_id_placeholder(top.token, tokenizer)
resolved_top.append(
top.model_copy(update={"token": top_str, "bytes": top_bytes})
)
resolved_content.append(
entry.model_copy(
update={
"token": token_str,
"bytes": token_bytes,
"top_logprobs": resolved_top,
}
)
)
return ChatCompletionLogProbs(content=resolved_content)
def _convert_chat_logprobs_to_completion_logprobs(
logprobs: ChatCompletionLogProbs,
) -> CompletionLogProbs:
"""Convert ChatCompletionLogProbs (per-token objects) to CompletionLogProbs
(parallel flat lists) as required by the /v1/completions response schema."""
if logprobs.content is None:
return CompletionLogProbs()
tokens: list[str] = []
token_logprobs: list[float | None] = []
top_logprobs_list: list[dict[str, float] | None] = []
text_offset: list[int] = []
offset = 0
for entry in logprobs.content:
text_offset.append(offset)
tokens.append(entry.token)
token_logprobs.append(entry.logprob)
top_logprobs_list.append(
{t.token: t.logprob for t in entry.top_logprobs}
if entry.top_logprobs
else None
)
offset += len(entry.token)
return CompletionLogProbs(
text_offset=text_offset,
token_logprobs=token_logprobs,
tokens=tokens,
top_logprobs=top_logprobs_list,
)
def _build_chat_choice(
choice: GenerateResponseChoice, tokenizer: TokenizerLike
) -> ChatCompletionResponseChoice:
"""Detokenize and resolve logprobs for a single GenerateResponseChoice.
Raises:
ValueError: if choice.token_ids is empty or None.
"""
if not choice.token_ids:
raise ValueError(f"choice {choice.index} has empty or null token_ids")
decoded_text = tokenizer.decode(choice.token_ids, skip_special_tokens=True)
resolved_logprobs = (
_resolve_logprobs(choice.logprobs, tokenizer)
if choice.logprobs is not None
else None
)
return ChatCompletionResponseChoice(
index=choice.index,
message=ChatMessage(role="assistant", content=decoded_text),
logprobs=resolved_logprobs,
finish_reason=choice.finish_reason,
)
class OpenAIServingRender:
def __init__(
self,
@@ -427,6 +529,146 @@ class OpenAIServingRender:
return messages, [engine_input]
async def derender_chat_response(
self,
request: DerenderChatRequest,
) -> ChatCompletionResponse | ErrorResponse:
"""Postprocess a GenerateResponse into a ChatCompletionResponse.
This is the symmetric inverse of render_chat_request: it detokenizes
output token IDs, resolves token_id:N logprob placeholders, and
formats the result as an OpenAI-compatible chat completion response.
"""
error_check_ret = await self._check_model(request)
if error_check_ret is not None:
return error_check_ret
tokenizer = self.renderer.get_tokenizer()
gen = request.generate_response
choices: list[ChatCompletionResponseChoice] = []
try:
for choice in gen.choices:
choices.append(_build_chat_choice(choice, tokenizer))
except ValueError as exc:
return self.create_error_response(str(exc))
prompt_tokens = (
request.prompt_tokens if request.prompt_tokens is not None else 0
)
completion_tokens = sum(len(ch.token_ids) for ch in gen.choices if ch.token_ids)
usage = UsageInfo(
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
total_tokens=prompt_tokens + completion_tokens,
)
logger.debug(
"derender_chat request_id=%s model=%s choices=%d completion_tokens=%d",
gen.request_id,
request.model,
len(choices),
completion_tokens,
)
return ChatCompletionResponse(
id=gen.request_id,
model=request.model,
created=int(time.time()),
choices=choices,
usage=usage,
prompt_logprobs=gen.prompt_logprobs,
kv_transfer_params=gen.kv_transfer_params,
)
async def derender_completion_response(
self,
request: DerenderCompletionRequest,
) -> CompletionResponse | ErrorResponse:
"""Postprocess a list of GenerateResponses into a CompletionResponse.
Mirrors the multi-prompt completions case: one GenerateResponse per
prompt, parallel to the list[GenerateRequest] from /v1/completions/render.
"""
error_check_ret = await self._check_model(request)
if error_check_ret is not None:
return error_check_ret
n = len(request.generate_responses)
prompt_tokens_list: list[int] = (
request.prompt_tokens if request.prompt_tokens is not None else [0] * n
)
tokenizer = self.renderer.get_tokenizer()
choices: list[CompletionResponseChoice] = []
total_prompt_tokens = 0
total_completion_tokens = 0
index = 0
for gen, pt in zip(request.generate_responses, prompt_tokens_list):
for choice in gen.choices:
if not choice.token_ids:
return self.create_error_response(
f"choice {choice.index} in response {gen.request_id} "
"has empty or null token_ids"
)
decoded_text = tokenizer.decode(
choice.token_ids, skip_special_tokens=True
)
completion_logprobs = None
if choice.logprobs is not None:
resolved = _resolve_logprobs(choice.logprobs, tokenizer)
completion_logprobs = _convert_chat_logprobs_to_completion_logprobs(
resolved
)
choices.append(
CompletionResponseChoice(
index=index,
text=decoded_text,
finish_reason=choice.finish_reason,
logprobs=completion_logprobs,
)
)
total_completion_tokens += len(choice.token_ids)
index += 1
total_prompt_tokens += pt
if not request.generate_responses:
return self.create_error_response("generate_responses must not be empty")
first = request.generate_responses[0]
kv_params = first.kv_transfer_params
if any(
r.kv_transfer_params != kv_params for r in request.generate_responses[1:]
):
logger.warning(
"derender_completion: kv_transfer_params differ across responses; "
"setting to None on the aggregated response"
)
kv_params = None
usage = UsageInfo(
prompt_tokens=total_prompt_tokens,
completion_tokens=total_completion_tokens,
total_tokens=total_prompt_tokens + total_completion_tokens,
)
logger.debug(
"derender_completion request_id=%s model=%s choices=%d"
" completion_tokens=%d",
first.request_id,
request.model,
len(choices),
total_completion_tokens,
)
return CompletionResponse(
id=first.request_id,
model=request.model,
created=int(time.time()),
choices=choices,
usage=usage,
kv_transfer_params=kv_params,
)
def create_error_response(
self,
message: str | Exception,