[Renderer] Move InputPreprocessor into Renderer (2/2) (#34560)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-02-17 05:29:01 -08:00
committed by GitHub
parent c61a98f529
commit 574fe75245
32 changed files with 984 additions and 1054 deletions
+3 -12
View File
@@ -28,11 +28,10 @@ from vllm.entrypoints.pooling.utils import (
encode_pooling_output_base64,
encode_pooling_output_float,
)
from vllm.inputs.data import TokensPrompt
from vllm.inputs.data import ProcessorInputs, TokensPrompt, token_inputs
from vllm.logger import init_logger
from vllm.outputs import PoolingOutput, PoolingRequestOutput
from vllm.pooling_params import PoolingParams
from vllm.renderers.inputs import TokPrompt
from vllm.utils.async_utils import merge_async_iterators
from vllm.utils.collection_utils import chunk_list
from vllm.utils.serial_utils import EmbedDType, Endianness
@@ -256,7 +255,7 @@ class OpenAIServingEmbedding(OpenAIServing):
chunk_request_id = f"{ctx.request_id}-prompt-{prompt_idx}-chunk-{chunk_idx}"
# Create engine prompt for this chunk
chunk_engine_prompt = TokensPrompt(prompt_token_ids=chunk_tokens)
chunk_engine_prompt = token_inputs(chunk_tokens)
# Log the chunk
self._log_inputs(
@@ -266,16 +265,12 @@ class OpenAIServingEmbedding(OpenAIServing):
lora_request=ctx.lora_request,
)
tok_params = ctx.request.build_tok_params(self.model_config)
tokenization_kwargs = tok_params.get_encode_kwargs()
# Create generator for this chunk and wrap it to return indices
original_generator = self.engine_client.encode(
chunk_engine_prompt,
pooling_params,
chunk_request_id,
lora_request=ctx.lora_request,
tokenization_kwargs=tokenization_kwargs,
trace_headers=trace_headers,
priority=ctx.request.priority,
)
@@ -362,7 +357,7 @@ class OpenAIServingEmbedding(OpenAIServing):
async def _create_single_prompt_generator(
self,
ctx: EmbeddingServeContext,
engine_prompt: TokPrompt,
engine_prompt: ProcessorInputs,
pooling_params: PoolingParams,
trace_headers: Mapping[str, str] | None,
prompt_index: int,
@@ -377,16 +372,12 @@ class OpenAIServingEmbedding(OpenAIServing):
lora_request=ctx.lora_request,
)
tok_params = ctx.request.build_tok_params(self.model_config)
tokenization_kwargs = tok_params.get_encode_kwargs()
# Return the original generator without wrapping
return self.engine_client.encode(
engine_prompt,
pooling_params,
request_id_item,
lora_request=ctx.lora_request,
tokenization_kwargs=tokenization_kwargs,
trace_headers=trace_headers,
priority=ctx.request.priority,
)
+2 -7
View File
@@ -33,10 +33,9 @@ from vllm.entrypoints.pooling.utils import (
encode_pooling_output_base64,
encode_pooling_output_float,
)
from vllm.inputs import PromptType
from vllm.inputs import ProcessorInputs
from vllm.logger import init_logger
from vllm.outputs import PoolingRequestOutput
from vllm.renderers.inputs import TokPrompt
from vllm.renderers.inputs.preprocess import prompt_to_seq
from vllm.utils.async_utils import merge_async_iterators
from vllm.utils.serial_utils import EmbedDType, EncodingFormat, Endianness
@@ -93,7 +92,7 @@ class OpenAIServingPooling(OpenAIServing):
"dimensions is currently not supported"
)
engine_prompts: Sequence[PromptType | TokPrompt]
engine_prompts: Sequence[ProcessorInputs]
if use_io_processor := isinstance(request, IOProcessorRequest):
if self.io_processor is None:
raise ValueError(
@@ -152,9 +151,6 @@ class OpenAIServingPooling(OpenAIServing):
else:
pooling_params = request.to_pooling_params() # type: ignore
tok_params = request.build_tok_params(self.model_config)
tokenization_kwargs = tok_params.get_encode_kwargs()
for i, engine_prompt in enumerate(engine_prompts):
request_id_item = f"{request_id}-{i}"
@@ -176,7 +172,6 @@ class OpenAIServingPooling(OpenAIServing):
pooling_params,
request_id_item,
lora_request=lora_request,
tokenization_kwargs=tokenization_kwargs,
trace_headers=trace_headers,
priority=request.priority,
)
+13 -7
View File
@@ -35,7 +35,7 @@ from vllm.entrypoints.pooling.score.utils import (
get_score_prompt,
validate_score_input,
)
from vllm.inputs.data import TokensPrompt
from vllm.inputs.data import ProcessorInputs, TokensPrompt, token_inputs
from vllm.logger import init_logger
from vllm.lora.request import LoRARequest
from vllm.outputs import PoolingRequestOutput, ScoringRequestOutput
@@ -108,12 +108,15 @@ class ServingScores(OpenAIServing):
*(encode_async(t, **tokenization_kwargs) for t in input_texts)
)
engine_prompts: list[TokensPrompt] = []
engine_prompts: list[ProcessorInputs] = []
for tok_result, input_text in zip(tokenized_prompts, input_texts):
text_token_prompt = self._validate_input(request, tok_result, input_text)
engine_prompts.append(
TokensPrompt(prompt_token_ids=text_token_prompt["prompt_token_ids"])
token_inputs(
text_token_prompt["prompt_token_ids"],
prompt=input_text,
)
)
# Schedule the request and get the result generator.
@@ -125,7 +128,7 @@ class ServingScores(OpenAIServing):
self._log_inputs(
request_id_item,
input_texts[i],
engine_prompt,
params=pooling_params,
lora_request=lora_request,
)
@@ -207,12 +210,15 @@ class ServingScores(OpenAIServing):
*(encode_async(t, **tokenization_kwargs) for t in input_texts)
)
engine_prompts: list[TokensPrompt] = []
engine_prompts: list[ProcessorInputs] = []
for tok_result, input_text in zip(tokenized_prompts, input_texts):
text_token_prompt = self._validate_input(request, tok_result, input_text)
engine_prompts.append(
TokensPrompt(prompt_token_ids=text_token_prompt["prompt_token_ids"])
token_inputs(
text_token_prompt["prompt_token_ids"],
prompt=input_text,
)
)
# Schedule the request and get the result generator.
@@ -225,7 +231,7 @@ class ServingScores(OpenAIServing):
self._log_inputs(
request_id_item,
input_texts[i],
engine_prompt,
params=pooling_params,
lora_request=lora_request,
)