[Chore] Remove redundant RequestPrompt (#30612)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-12-14 09:22:37 +00:00
committed by GitHub
parent f569c654e1
commit dcb31196da
12 changed files with 188 additions and 253 deletions
+36 -19
View File
@@ -61,7 +61,7 @@ from vllm.entrypoints.openai.tool_parsers import ToolParser
from vllm.entrypoints.openai.tool_parsers.mistral_tool_parser import MistralToolCall
from vllm.entrypoints.openai.utils import maybe_filter_parallel_tool_calls
from vllm.entrypoints.utils import get_max_tokens, should_include_usage
from vllm.inputs.data import TokensPrompt as EngineTokensPrompt
from vllm.inputs.data import TokensPrompt
from vllm.logger import init_logger
from vllm.logprobs import Logprob
from vllm.outputs import CompletionOutput, RequestOutput
@@ -234,11 +234,7 @@ class OpenAIServingChat(OpenAIServing):
)
if error_check_ret is not None:
return error_check_ret
(
conversation,
request_prompts,
engine_prompts,
) = await self._preprocess_chat(
conversation, engine_prompts = await self._preprocess_chat(
request,
tokenizer,
request.messages,
@@ -254,11 +250,7 @@ class OpenAIServingChat(OpenAIServing):
)
else:
# For GPT-OSS.
(
conversation,
request_prompts,
engine_prompts,
) = self._make_request_with_harmony(request)
conversation, engine_prompts = self._make_request_with_harmony(request)
except (ValueError, TypeError, RuntimeError, jinja2.TemplateError) as e:
logger.exception("Error in preprocessing prompt inputs")
return self.create_error_response(f"{e} {e.__cause__}")
@@ -278,7 +270,7 @@ class OpenAIServingChat(OpenAIServing):
generators: list[AsyncGenerator[RequestOutput, None]] = []
try:
for i, engine_prompt in enumerate(engine_prompts):
prompt_text, _, _ = self._get_prompt_components(request_prompts[i])
prompt_text, _, _ = self._get_prompt_components(engine_prompt)
# If we are creating sub requests for multiple prompts, ensure that they
# have unique request ids.
sub_request_id = (
@@ -313,7 +305,7 @@ class OpenAIServingChat(OpenAIServing):
self._log_inputs(
sub_request_id,
request_prompts[i],
engine_prompt,
params=sampling_params,
lora_request=lora_request,
)
@@ -537,7 +529,7 @@ class OpenAIServingChat(OpenAIServing):
request_id: str,
model_name: str,
conversation: list[ConversationMessage],
tokenizer: TokenizerLike,
tokenizer: TokenizerLike | None,
request_metadata: RequestResponseMetadata,
) -> AsyncGenerator[str, None]:
created_time = int(time.time())
@@ -591,6 +583,11 @@ class OpenAIServingChat(OpenAIServing):
try:
if self.reasoning_parser:
if tokenizer is None:
raise ValueError(
"Tokenizer not available when `skip_tokenizer_init=True`"
)
reasoning_parser = self.reasoning_parser(
tokenizer,
chat_template_kwargs=request.chat_template_kwargs, # type: ignore
@@ -604,6 +601,11 @@ class OpenAIServingChat(OpenAIServing):
# Prepare the tool parser if it's needed
try:
if tool_choice_auto and self.tool_parser:
if tokenizer is None:
raise ValueError(
"Tokenizer not available when `skip_tokenizer_init=True`"
)
tool_parsers: list[ToolParser | None] = [
self.tool_parser(tokenizer)
] * num_choices
@@ -1317,7 +1319,7 @@ class OpenAIServingChat(OpenAIServing):
request_id: str,
model_name: str,
conversation: list[ConversationMessage],
tokenizer: TokenizerLike,
tokenizer: TokenizerLike | None,
request_metadata: RequestResponseMetadata,
) -> ErrorResponse | ChatCompletionResponse:
created_time = int(time.time())
@@ -1367,6 +1369,11 @@ class OpenAIServingChat(OpenAIServing):
reasoning = None
if self.tool_parser is not None:
if tokenizer is None:
raise ValueError(
"Tokenizer not available when `skip_tokenizer_init=True`"
)
tool_parser = self.tool_parser(tokenizer)
# NOTE: We use token_ids for openai tool parser
tool_call_info = tool_parser.extract_tool_calls(
@@ -1409,6 +1416,11 @@ class OpenAIServingChat(OpenAIServing):
if self.reasoning_parser:
try:
if tokenizer is None:
raise ValueError(
"Tokenizer not available when `skip_tokenizer_init=True`"
)
reasoning_parser = self.reasoning_parser(
tokenizer,
chat_template_kwargs=request.chat_template_kwargs, # type: ignore
@@ -1648,7 +1660,7 @@ class OpenAIServingChat(OpenAIServing):
self,
logprobs: dict[int, Logprob],
top_logprobs: int | None,
tokenizer: TokenizerLike,
tokenizer: TokenizerLike | None,
should_return_as_token_id: bool,
) -> list[ChatCompletionLogProb]:
return [
@@ -1672,7 +1684,7 @@ class OpenAIServingChat(OpenAIServing):
self,
token_ids: GenericSequence[int],
top_logprobs: GenericSequence[dict[int, Logprob] | None],
tokenizer: TokenizerLike,
tokenizer: TokenizerLike | None,
num_output_top_logprobs: int | None = None,
return_as_token_id: bool | None = None,
) -> ChatCompletionLogProbs:
@@ -1690,6 +1702,11 @@ class OpenAIServingChat(OpenAIServing):
if should_return_as_token_id:
token = f"token_id:{token_id}"
else:
if tokenizer is None:
raise ValueError(
"Tokenizer not available when `skip_tokenizer_init=True`"
)
token = tokenizer.decode(token_id)
logprobs_content.append(
@@ -1800,10 +1817,10 @@ class OpenAIServingChat(OpenAIServing):
# Render prompt token ids.
prompt_token_ids = render_for_completion(messages)
engine_prompt = EngineTokensPrompt(prompt_token_ids=prompt_token_ids)
engine_prompt = TokensPrompt(prompt_token_ids=prompt_token_ids)
# Add cache_salt if provided in the request
if request.cache_salt is not None:
engine_prompt["cache_salt"] = request.cache_salt
return messages, [prompt_token_ids], [engine_prompt]
return messages, [engine_prompt]