[RL][Frontend] Return weight-version spans on OpenAI completions and chat (#61347)

Signed-off-by: Shuolei Wang <shuoleiwang123@gmail.com>
Co-authored-by: aoshen02 <aoshen@inferact.ai>
This commit is contained in:
Shuolei Wang
2026-10-17 02:08:07 +00:00
committed by GitHub
co-authored by aoshen02
parent 6b6a6c5c98
commit b3a7e2ff6e
7 changed files with 139 additions and 4 deletions
@@ -40,6 +40,7 @@ from vllm.entrypoints.serve.engine.protocol import OpenAIBaseModel, UsageInfo
from vllm.exceptions import VLLMValidationError
from vllm.logger import init_logger
from vllm.logprobs import Logprob
from vllm.outputs import WeightVersionSpan
from vllm.renderers import ChatParams, TokenizeParams, merge_kwargs
from vllm.renderers.chat_utils import (
ChatCompletionMessageParam,
@@ -129,6 +130,8 @@ class ChatCompletionResponseChoice(OpenAIBaseModel):
# ``None`` if (a) the request was aborted before any forward pass,
# or (b) ``enable_return_routed_experts`` is off server-side.
routed_experts: str | None = None
# Contiguous weight version spans over output-token indexes.
weight_versions: list[WeightVersionSpan] | None = None
class ChatCompletionResponse(OpenAIBaseModel):
@@ -168,6 +171,8 @@ class ChatCompletionResponseStreamChoice(OpenAIBaseModel):
stop_reason: int | str | None = None
# not part of the OpenAI spec but for tracing the tokens
token_ids: list[int] | None = None
# Set on the final chunk only; covers the whole output.
weight_versions: list[WeightVersionSpan] | None = None
class ChatCompletionStreamResponse(OpenAIBaseModel):
@@ -430,6 +435,15 @@ class ChatCompletionRequest(OpenAIBaseModel):
"need to map generated text back to input tokens."
),
)
return_weight_versions: bool = Field(
default=False,
description=(
"If true, each choice includes weight_versions: the weight-version "
"label that sampled each half-open range of its output tokens. "
"When streaming, only the last chunk of each choice includes it. "
"Beam search does not return it."
),
)
routed_experts_prompt_start: int = Field(
default=0,
ge=0,
@@ -651,17 +651,21 @@ class OpenAIServingChat(GenerateBaseServing):
delta_text = output.text
if (
nothing_generated = (
not delta_text
and not output.token_ids
and not previous_num_tokens[i]
):
)
if nothing_generated and output.finish_reason is None:
# Chunked prefill case, don't return empty chunks
continue
delta_message: DeltaMessage | None
if parser is not None:
if nothing_generated:
# Finished without output: send the final chunk unparsed.
delta_message = DeltaMessage()
elif parser is not None:
delta_message = parser.parse_delta(
delta_text=delta_text,
delta_token_ids=as_list(output.token_ids),
@@ -773,6 +777,11 @@ class OpenAIServingChat(GenerateBaseServing):
token_ids=(
as_list(output.token_ids) if include_token_ids else None
),
weight_versions=(
output.weight_versions
if request.return_weight_versions
else None
),
)
finish_reason_sent[i] = True
@@ -1097,6 +1106,9 @@ class OpenAIServingChat(GenerateBaseServing):
else None
),
routed_experts=routed_experts_b64,
weight_versions=(
output.weight_versions if request.return_weight_versions else None
),
)
choice_data = maybe_filter_parallel_tool_calls(choice_data, request)
@@ -26,6 +26,7 @@ from vllm.entrypoints.serve.engine.protocol import OpenAIBaseModel, UsageInfo
from vllm.exceptions import VLLMValidationError
from vllm.logger import init_logger
from vllm.logprobs import Logprob
from vllm.outputs import WeightVersionSpan
from vllm.renderers import TokenizeParams
from vllm.sampling_params import (
BeamSearchParams,
@@ -178,6 +179,15 @@ class CompletionRequest(OpenAIBaseModel):
"need to map generated text back to input tokens."
),
)
return_weight_versions: bool = Field(
default=False,
description=(
"If true, each choice includes weight_versions: the weight-version "
"label that sampled each half-open range of its output tokens. "
"When streaming, only the last chunk of each choice includes it. "
"Beam search does not return it."
),
)
routed_experts_prompt_start: int = Field(
default=0,
ge=0,
@@ -677,6 +687,8 @@ class CompletionResponseChoice(OpenAIBaseModel):
# ``None`` if (a) the request was aborted before any forward pass,
# or (b) ``enable_return_routed_experts`` is off server-side.
routed_experts: str | None = None
# Contiguous weight version spans over output-token indexes.
weight_versions: list[WeightVersionSpan] | None = None
class CompletionResponse(OpenAIBaseModel):
@@ -716,6 +728,8 @@ class CompletionResponseStreamChoice(OpenAIBaseModel):
# prompt tokens is put into choice to align with CompletionResponseChoice
prompt_token_ids: list[int] | None = None
token_ids: list[int] | None = None
# Set on the final chunk only; covers the whole output.
weight_versions: list[WeightVersionSpan] | None = None
class CompletionStreamResponse(OpenAIBaseModel):
@@ -376,6 +376,7 @@ class OpenAIServingCompletion(GenerateBaseServing):
not delta_text
and not delta_token_ids
and not previous_num_tokens[i]
and finish_reason is None
):
# Chunked prefill case, don't return empty chunks
continue
@@ -421,6 +422,11 @@ class OpenAIServingCompletion(GenerateBaseServing):
if request.return_token_ids
else None
),
weight_versions=(
output.weight_versions
if request.return_weight_versions
else None
),
)
],
)
@@ -594,6 +600,11 @@ class OpenAIServingCompletion(GenerateBaseServing):
as_list(output.token_ids) if request.return_token_ids else None
),
routed_experts=routed_experts_b64,
weight_versions=(
output.weight_versions
if request.return_weight_versions
else None
),
)
choices.append(choice_data)