[Renderer] Move InputPreprocessor into Renderer (1/2) (#34510)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Signed-off-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
This commit is contained in:
Cyrus Leung
2026-02-14 10:14:21 -08:00
committed by GitHub
parent b3c14229b0
commit 73391a1baa
39 changed files with 456 additions and 458 deletions
+16 -29
View File
@@ -72,7 +72,7 @@ from vllm.outputs import (
)
from vllm.platforms import current_platform
from vllm.pooling_params import PoolingParams
from vllm.renderers import ChatParams, TokenizeParams, merge_kwargs
from vllm.renderers import ChatParams, merge_kwargs
from vllm.renderers.inputs import DictPrompt, TokPrompt
from vllm.renderers.inputs.preprocess import (
conversation_to_seq,
@@ -384,7 +384,7 @@ class LLM:
return parallel_config.world_size
def reset_mm_cache(self) -> None:
self.input_processor.clear_mm_cache()
self.renderer.clear_mm_cache()
self.llm_engine.reset_mm_cache()
def get_default_sampling_params(self) -> SamplingParams:
@@ -876,19 +876,6 @@ class LLM:
return outputs
def _get_cmpl_tok_params(self, tokenization_kwargs: dict[str, Any] | None):
model_config = self.model_config
encoder_config = model_config.encoder_config or {}
return TokenizeParams(
max_total_tokens=model_config.max_model_len,
do_lower_case=encoder_config.get("do_lower_case", False),
# For Whisper, special tokens should be provided by the user based
# on the task and language of their request. Also needed to avoid
# appending an EOS token to the prompt which disrupts generation.
add_special_tokens=not model_config.is_encoder_decoder,
).with_kwargs(tokenization_kwargs)
def _preprocess_cmpl(
self,
prompts: Sequence[PromptType],
@@ -910,20 +897,12 @@ class LLM:
parsed_prompts = [
parse_model_prompt(model_config, prompt) for prompt in prompts
]
tok_params = self._get_cmpl_tok_params(tokenization_kwargs)
tok_params = renderer.default_cmpl_tok_params.with_kwargs(
**(tokenization_kwargs or {})
)
return renderer.render_cmpl(parsed_prompts, tok_params)
def _get_chat_tok_params(self, tokenization_kwargs: dict[str, Any] | None):
model_config = self.model_config
encoder_config = model_config.encoder_config or {}
return TokenizeParams(
max_total_tokens=model_config.max_model_len,
do_lower_case=encoder_config.get("do_lower_case", False),
add_special_tokens=False,
).with_kwargs(tokenization_kwargs)
def _preprocess_chat(
self,
conversations: Sequence[list[ChatCompletionMessageParam]],
@@ -961,7 +940,9 @@ class LLM:
),
),
)
tok_params = self._get_chat_tok_params(tokenization_kwargs)
tok_params = renderer.default_chat_tok_params.with_kwargs(
**(tokenization_kwargs or {})
)
_, engine_prompts = renderer.render_chat(
conversations,
@@ -1653,7 +1634,10 @@ class LLM:
architecture=architecture,
)
tok_params = self._get_cmpl_tok_params(tokenization_kwargs)
renderer = self.renderer
tok_params = renderer.default_cmpl_tok_params.with_kwargs(
**(tokenization_kwargs or {})
)
encode_kwargs = tok_params.get_encode_kwargs()
if model_config.is_cross_encoder:
@@ -1970,7 +1954,10 @@ class LLM:
dict(truncate_prompt_tokens=params.truncate_prompt_tokens),
)
tok_params = self._get_cmpl_tok_params(tokenization_kwargs)
renderer = self.renderer
tok_params = renderer.default_cmpl_tok_params.with_kwargs(
**(tokenization_kwargs or {})
)
tokenization_kwargs = tok_params.get_encode_kwargs()
engine_request = self.input_processor.process_inputs(
+2 -2
View File
@@ -8,11 +8,11 @@ from typing import Literal, cast
import numpy as np
from vllm.engine.protocol import EngineClient
from vllm.engine.protocol import EngineClient, StreamingInput
from vllm.entrypoints.logger import RequestLogger
from vllm.entrypoints.openai.engine.serving import OpenAIServing
from vllm.entrypoints.openai.models.serving import OpenAIServingModels
from vllm.inputs.data import PromptType, StreamingInput
from vllm.inputs.data import PromptType
from vllm.logger import init_logger
from vllm.model_executor.models.interfaces import SupportsRealtime