[Refactor] Consolidate sequence normalization and enc-dec parsing (#33928)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-02-06 15:43:47 +00:00
committed by GitHub
parent 4707f7ebb4
commit cd8b405bd0
38 changed files with 1271 additions and 863 deletions
@@ -67,12 +67,13 @@ from vllm.entrypoints.openai.parser.harmony_utils import (
)
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 EmbedsPrompt, TokensPrompt
from vllm.inputs.data import TokensPrompt
from vllm.logger import init_logger
from vllm.logprobs import Logprob
from vllm.outputs import CompletionOutput, RequestOutput
from vllm.parser import ParserManager
from vllm.reasoning import ReasoningParser
from vllm.renderers.inputs import TokPrompt
from vllm.sampling_params import BeamSearchParams, SamplingParams
from vllm.tokenizers import TokenizerLike
from vllm.tokenizers.mistral import (
@@ -218,10 +219,7 @@ class OpenAIServingChat(OpenAIServing):
async def render_chat_request(
self,
request: ChatCompletionRequest,
) -> (
tuple[list[ConversationMessage], list[TokensPrompt | EmbedsPrompt]]
| ErrorResponse
):
) -> tuple[list[ConversationMessage], list[TokPrompt]] | ErrorResponse:
"""
render chat request by validating and preprocessing inputs.
@@ -380,7 +378,7 @@ class OpenAIServingChat(OpenAIServing):
generators: list[AsyncGenerator[RequestOutput, None]] = []
try:
for i, engine_prompt in enumerate(engine_prompts):
prompt_text = engine_prompt.get("prompt")
prompt_text = self._extract_prompt_text(engine_prompt)
# If we are creating sub requests for multiple prompts, ensure that they
# have unique request ids.
@@ -389,10 +387,10 @@ class OpenAIServingChat(OpenAIServing):
)
max_tokens = get_max_tokens(
max_model_len=self.max_model_len,
request=request,
prompt=engine_prompt,
default_sampling_params=self.default_sampling_params,
self.max_model_len,
request,
self._extract_prompt_len(engine_prompt),
self.default_sampling_params,
)
sampling_params: SamplingParams | BeamSearchParams
+10 -10
View File
@@ -34,10 +34,10 @@ from vllm.entrypoints.openai.engine.serving import (
from vllm.entrypoints.openai.models.serving import OpenAIServingModels
from vllm.entrypoints.utils import get_max_tokens, should_include_usage
from vllm.exceptions import VLLMValidationError
from vllm.inputs.data import EmbedsPrompt, TokensPrompt
from vllm.logger import init_logger
from vllm.logprobs import Logprob
from vllm.outputs import RequestOutput
from vllm.renderers.inputs import TokPrompt
from vllm.sampling_params import BeamSearchParams, SamplingParams
from vllm.tokenizers import TokenizerLike
from vllm.utils.async_utils import merge_async_iterators
@@ -78,7 +78,7 @@ class OpenAIServingCompletion(OpenAIServing):
async def render_completion_request(
self,
request: CompletionRequest,
) -> list[TokensPrompt | EmbedsPrompt] | ErrorResponse:
) -> list[TokPrompt] | ErrorResponse:
"""
render completion request by validating and preprocessing inputs.
@@ -160,13 +160,13 @@ class OpenAIServingCompletion(OpenAIServing):
generators: list[AsyncGenerator[RequestOutput, None]] = []
try:
for i, engine_prompt in enumerate(engine_prompts):
prompt_text = engine_prompt.get("prompt")
prompt_text = self._extract_prompt_text(engine_prompt)
max_tokens = get_max_tokens(
max_model_len=self.max_model_len,
request=request,
prompt=engine_prompt,
default_sampling_params=self.default_sampling_params,
self.max_model_len,
request,
self._extract_prompt_len(engine_prompt),
self.default_sampling_params,
)
sampling_params: SamplingParams | BeamSearchParams
@@ -277,7 +277,7 @@ class OpenAIServingCompletion(OpenAIServing):
# with the inputs token IDs
if final_res.prompt is None:
engine_prompt = engine_prompts[i]
final_res.prompt = engine_prompt.get("prompt")
final_res.prompt = self._extract_prompt_text(engine_prompt)
final_res_batch_checked = cast(list[RequestOutput], final_res_batch)
@@ -313,7 +313,7 @@ class OpenAIServingCompletion(OpenAIServing):
async def completion_stream_generator(
self,
request: CompletionRequest,
engine_prompts: list[TokensPrompt | EmbedsPrompt],
engine_prompts: list[TokPrompt],
result_generator: AsyncIterator[tuple[int, RequestOutput]],
request_id: str,
created_time: int,
@@ -347,7 +347,7 @@ class OpenAIServingCompletion(OpenAIServing):
prompt_text = res.prompt
if prompt_text is None:
engine_prompt = engine_prompts[prompt_idx]
prompt_text = engine_prompt.get("prompt")
prompt_text = self._extract_prompt_text(engine_prompt)
# Prompt details are excluded from later streamed outputs
if prompt_token_ids is not None:
+68 -41
View File
@@ -96,11 +96,7 @@ from vllm.entrypoints.serve.tokenize.protocol import (
)
from vllm.entrypoints.utils import get_max_tokens, sanitize_message
from vllm.exceptions import VLLMValidationError
from vllm.inputs.data import EmbedsPrompt, PromptType, TokensPrompt
from vllm.inputs.parse import (
get_prompt_components,
is_explicit_encoder_decoder_prompt,
)
from vllm.inputs.data import PromptType, SingletonPrompt, TokensPrompt
from vllm.logger import init_logger
from vllm.logprobs import Logprob, PromptLogprobs
from vllm.lora.request import LoRARequest
@@ -108,6 +104,14 @@ from vllm.multimodal import MultiModalDataDict
from vllm.outputs import CompletionOutput, PoolingRequestOutput, RequestOutput
from vllm.pooling_params import PoolingParams
from vllm.renderers import ChatParams, TokenizeParams, merge_kwargs
from vllm.renderers.inputs import TokPrompt
from vllm.renderers.inputs.preprocess import (
SingletonDictPrompt,
extract_prompt_components,
extract_prompt_len,
parse_model_prompt,
prompt_to_seq,
)
from vllm.sampling_params import BeamSearchParams, SamplingParams
from vllm.tokenizers import TokenizerLike
from vllm.tool_parsers import ToolParser
@@ -203,7 +207,7 @@ class ServeContext(Generic[RequestT]):
request_id: str
created_time: int = field(default_factory=lambda: int(time.time()))
lora_request: LoRARequest | None = None
engine_prompts: list[TokensPrompt | EmbedsPrompt] | None = None
engine_prompts: list[TokPrompt] | None = None
result_generator: AsyncGenerator[tuple[int, PoolingRequestOutput], None] | None = (
None
@@ -247,7 +251,7 @@ class OpenAIServing:
async def beam_search(
self,
prompt: PromptType,
prompt: TokPrompt,
request_id: str,
params: BeamSearchParams,
lora_request: LoRARequest | None = None,
@@ -271,20 +275,12 @@ class OpenAIServing:
eos_token_id: int = tokenizer.eos_token_id # type: ignore
if is_explicit_encoder_decoder_prompt(prompt):
raise NotImplementedError
if isinstance(prompt, dict) and "encoder_prompt" in prompt:
raise NotImplementedError("Encoder-decoder prompt not supported")
prompt_text: str | None
prompt_token_ids: list[int]
multi_modal_data: MultiModalDataDict | None
if isinstance(prompt, str):
prompt_text = prompt
prompt_token_ids = []
multi_modal_data = None
else:
prompt_text = prompt.get("prompt") # type: ignore
prompt_token_ids = prompt.get("prompt_token_ids", []) # type: ignore
multi_modal_data = prompt.get("multi_modal_data") # type: ignore
prompt_text: str | None = prompt.get("prompt") # type: ignore
prompt_token_ids: list[int] = prompt.get("prompt_token_ids", []) # type: ignore
multi_modal_data: MultiModalDataDict | None = prompt.get("multi_modal_data") # type: ignore
mm_processor_kwargs: dict[str, Any] | None = None
@@ -963,22 +959,40 @@ class OpenAIServing:
request: RendererRequest,
prompt_input: str | list[str] | list[int] | list[list[int]] | None,
prompt_embeds: bytes | list[bytes] | None,
) -> list[TokensPrompt | EmbedsPrompt]:
) -> list[TokPrompt]:
renderer = self.renderer
tok_params = request.build_tok_params(self.model_config)
model_config = self.model_config
in_prompts = await renderer.render_completions_async(
prompt_input, prompt_embeds
)
engine_prompts = await renderer.tokenize_prompts_async(in_prompts, tok_params)
tok_params = request.build_tok_params(model_config)
prompts = list[SingletonPrompt | bytes]()
if prompt_embeds is not None: # embeds take higher priority
prompts.extend(prompt_to_seq(prompt_embeds))
if prompt_input is not None:
prompts.extend(prompt_to_seq(prompt_input))
parsed_prompts = [
(
prompt
if isinstance(prompt, bytes)
else parse_model_prompt(model_config, prompt)
)
for prompt in prompts
]
in_prompts = await renderer.render_prompts_async(parsed_prompts)
extra_items = {
k: v
for k in ("mm_processor_kwargs", "cache_salt")
if (v := getattr(request, k, None)) is not None
}
for prompt in engine_prompts:
prompt.update(extra_items) # type: ignore
for in_prompt in in_prompts:
target_prompt: SingletonDictPrompt = in_prompt.get( # type: ignore
"encoder_prompt", in_prompt
)
target_prompt.update(extra_items) # type: ignore
engine_prompts = await renderer.tokenize_prompts_async(in_prompts, tok_params)
return engine_prompts
@@ -991,7 +1005,7 @@ class OpenAIServing:
default_template_kwargs: dict[str, Any] | None,
tool_dicts: list[dict[str, Any]] | None = None,
tool_parser: Callable[[TokenizerLike], ToolParser] | None = None,
) -> tuple[list[ConversationMessage], list[TokensPrompt | EmbedsPrompt]]:
) -> tuple[list[ConversationMessage], list[TokPrompt]]:
from vllm.tokenizers.mistral import MistralTokenizer
renderer = self.renderer
@@ -1009,17 +1023,21 @@ class OpenAIServing:
default_template, default_template_content_format
).with_defaults(default_template_kwargs)
conversation, prompt = await renderer.render_messages_async(
conversation, in_prompt = await renderer.render_messages_async(
messages, chat_params
)
engine_prompt = await renderer.tokenize_prompt_async(prompt, tok_params)
target_prompt: SingletonDictPrompt = in_prompt.get( # type: ignore
"encoder_prompt", in_prompt
)
extra_items = {
k: v
for k in ("mm_processor_kwargs", "cache_salt")
if (v := getattr(request, k, None)) is not None
}
engine_prompt.update(extra_items) # type: ignore
target_prompt.update(extra_items) # type: ignore
engine_prompt = await renderer.tokenize_prompt_async(target_prompt, tok_params)
# tool parsing is done only if a tool_parser has been set and if
# tool_choice is not "none" (if tool_choice is "none" but a tool_parser
@@ -1040,6 +1058,15 @@ class OpenAIServing:
return conversation, [engine_prompt]
def _extract_prompt_components(self, prompt: object):
return extract_prompt_components(self.model_config, prompt)
def _extract_prompt_text(self, prompt: object):
return self._extract_prompt_components(prompt).text
def _extract_prompt_len(self, prompt: object):
return extract_prompt_len(self.model_config, prompt)
async def _render_next_turn(
self,
request: ResponsesRequest,
@@ -1067,7 +1094,7 @@ class OpenAIServing:
async def _generate_with_builtin_tools(
self,
request_id: str,
engine_prompt: TokensPrompt | EmbedsPrompt,
engine_prompt: TokPrompt,
sampling_params: SamplingParams,
tok_params: TokenizeParams,
context: ConversationContext,
@@ -1075,7 +1102,7 @@ class OpenAIServing:
priority: int = 0,
trace_headers: Mapping[str, str] | None = None,
):
prompt_text = engine_prompt.get("prompt")
prompt_text = self._extract_prompt_text(engine_prompt)
orig_priority = priority
sub_request = 0
@@ -1145,12 +1172,12 @@ class OpenAIServing:
context.chat_template_content_format,
)
engine_prompt = engine_prompts[0]
prompt_text = engine_prompt.get("prompt")
prompt_text = self._extract_prompt_text(engine_prompt)
sampling_params.max_tokens = get_max_tokens(
self.max_model_len,
context.request,
engine_prompt,
self._extract_prompt_len(engine_prompt),
self.default_sampling_params, # type: ignore
)
@@ -1161,20 +1188,20 @@ class OpenAIServing:
def _log_inputs(
self,
request_id: str,
inputs: PromptType,
inputs: PromptType | TokPrompt,
params: SamplingParams | PoolingParams | BeamSearchParams | None,
lora_request: LoRARequest | None,
) -> None:
if self.request_logger is None:
return
prompt, prompt_token_ids, prompt_embeds = get_prompt_components(inputs)
components = self._extract_prompt_components(inputs)
self.request_logger.log_inputs(
request_id,
prompt,
prompt_token_ids,
prompt_embeds,
components.text,
components.token_ids,
components.embeds,
params=params,
lora_request=lora_request,
)
+5 -5
View File
@@ -116,13 +116,13 @@ from vllm.entrypoints.openai.responses.utils import (
)
from vllm.entrypoints.utils import get_max_tokens
from vllm.exceptions import VLLMValidationError
from vllm.inputs.data import EmbedsPrompt, TokensPrompt
from vllm.inputs.parse import get_prompt_len
from vllm.inputs.data import TokensPrompt
from vllm.logger import init_logger
from vllm.logprobs import Logprob as SampleLogprob
from vllm.logprobs import SampleLogprobs
from vllm.outputs import CompletionOutput
from vllm.parser import ParserManager
from vllm.renderers.inputs import TokPrompt
from vllm.sampling_params import SamplingParams, StructuredOutputsParams
from vllm.tokenizers import TokenizerLike
from vllm.utils import random_uuid
@@ -292,10 +292,10 @@ class OpenAIServingResponses(OpenAIServing):
def _validate_generator_input(
self,
engine_prompt: TokensPrompt | EmbedsPrompt,
engine_prompt: TokPrompt,
) -> ErrorResponse | None:
"""Add validations to the input to the generator here."""
prompt_len = get_prompt_len(engine_prompt)
prompt_len = self._extract_prompt_len(engine_prompt)
if self.max_model_len <= prompt_len:
error_message = (
f"The engine prompt length {prompt_len} "
@@ -442,7 +442,7 @@ class OpenAIServingResponses(OpenAIServing):
default_max_tokens = get_max_tokens(
self.max_model_len,
request,
engine_prompt,
self._extract_prompt_len(engine_prompt),
self.default_sampling_params,
)
@@ -7,7 +7,7 @@ import time
import zlib
from collections.abc import AsyncGenerator, Callable
from functools import cached_property
from typing import Literal, TypeAlias, TypeVar, cast
from typing import Final, Literal, TypeAlias, TypeVar, cast
import numpy as np
from fastapi import Request
@@ -37,12 +37,13 @@ from vllm.entrypoints.openai.speech_to_text.protocol import (
TranslationStreamResponse,
)
from vllm.exceptions import VLLMValidationError
from vllm.inputs.data import ExplicitEncoderDecoderPrompt, PromptType
from vllm.inputs.parse import is_explicit_encoder_decoder_prompt
from vllm.inputs.data import PromptType
from vllm.logger import init_logger
from vllm.logprobs import FlatLogprobs, Logprob
from vllm.model_executor.models import SupportsTranscription, supports_transcription
from vllm.outputs import RequestOutput
from vllm.renderers.inputs import EncoderDecoderDictPrompt
from vllm.renderers.inputs.preprocess import parse_enc_dec_prompt
from vllm.tokenizers import get_tokenizer
from vllm.utils.import_utils import PlaceholderModule
@@ -94,7 +95,7 @@ class OpenAISpeechToText(OpenAIServing):
)
self.default_sampling_params = self.model_config.get_diff_sampling_param()
self.task_type = task_type
self.task_type: Final = task_type
self.asr_config = self.model_cls.get_speech_to_text_config(
self.model_config, task_type
@@ -298,35 +299,26 @@ class OpenAISpeechToText(OpenAIServing):
to_language=to_language,
)
if request.response_format == "verbose_json":
if not is_explicit_encoder_decoder_prompt(prompt):
raise VLLMValidationError(
"Expected prompt to be an encoder-decoder prompt",
parameter="prompt",
value=type(prompt).__name__,
)
prompt = self._preprocess_verbose_prompt(prompt)
prompt = self._preprocess_verbose_prompt(parse_enc_dec_prompt(prompt))
prompts.append(prompt)
return prompts, duration
def _repl_verbose_text(self, text: str):
return text.replace("<|notimestamps|>", "<|0.00|>")
def _preprocess_verbose_prompt(self, prompt: ExplicitEncoderDecoderPrompt):
def _preprocess_verbose_prompt(self, prompt: EncoderDecoderDictPrompt):
dec_prompt = prompt["decoder_prompt"]
if isinstance(dec_prompt, str):
prompt["decoder_prompt"] = self._repl_verbose_text(dec_prompt)
elif isinstance(dec_prompt, dict) and "prompt" in dec_prompt:
dec_prompt["prompt"] = self._repl_verbose_text(dec_prompt["prompt"])
else:
if not (isinstance(dec_prompt, dict) and "prompt" in dec_prompt):
raise VLLMValidationError(
"Expected decoder_prompt to contain text",
parameter="decoder_prompt",
value=type(dec_prompt).__name__,
)
dec_prompt["prompt"] = dec_prompt["prompt"].replace(
"<|notimestamps|>", "<|0.00|>"
)
return prompt
def _get_verbose_segments(