[Bugfix][Pooling] Align CrossEncoder token type ids after truncation (#47772)

Signed-off-by: Ting Sun <suntcrick@gmail.com>
This commit is contained in:
Ting SUN
2026-07-08 03:59:22 +00:00
committed by GitHub
parent dd127d82ed
commit 0303f37a54
2 changed files with 85 additions and 3 deletions
@@ -2,6 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import weakref
from types import SimpleNamespace
import pytest
import torch
@@ -9,7 +10,10 @@ import torch
from tests.models.utils import softmax
from vllm import LLM, PoolingParams
from vllm.distributed import cleanup_dist_env_and_memory
from vllm.entrypoints.pooling.scoring.io_processor import CrossEncoderIOProcessor
from vllm.entrypoints.pooling.scoring.typing import ScoringData
from vllm.platforms import current_platform
from vllm.renderers import TokenizeParams
MODEL_NAME = "tomaarsen/Qwen3-Reranker-0.6B-seq-cls"
PROMPT = "The chef prepared a delicious meal."
@@ -141,6 +145,45 @@ def test_max_tokens_per_doc(llm: LLM):
assert with_limit_tokens < no_limit_tokens
def test_token_type_ids_follow_post_tokenization():
processor = object.__new__(CrossEncoderIOProcessor)
processor.tokenizer = SimpleNamespace(truncation_side="right", pad_token_id=-1)
processor.renderer = SimpleNamespace(process_for_engine=lambda prompt, _: prompt)
processor.model_config = None
processor.get_score_prompt = lambda **_: (
"",
{
"prompt_token_ids": list(range(32)),
"token_type_ids": [0] * 16 + [1] * 16,
},
)
engine_inputs, pooling_params = processor._pre_process(
ScoringData(data_1=["query"], data_2=["document"]),
TokenizeParams(
max_total_tokens=None,
truncate_prompt_tokens=16,
truncation_side="left",
),
PoolingParams(task="classify", extra_kwargs={"cache_salt": "salt"}),
)
assert engine_inputs[0]["prompt_token_ids"] == list(range(16, 32))
assert pooling_params[0].extra_kwargs == {
"cache_salt": "salt",
"compressed_token_type_ids": 0,
}
engine_inputs, pooling_params = processor._pre_process(
ScoringData(data_1=["query"], data_2=["document"]),
TokenizeParams(max_total_tokens=None, pad_prompt_tokens=40),
PoolingParams(task="classify"),
)
assert engine_inputs[0]["prompt_token_ids"] == list(range(32)) + [-1] * 8
assert pooling_params[0].extra_kwargs == {"compressed_token_type_ids": 16}
def test_pooling_params(llm: LLM):
def get_outputs(use_activation):
outputs = llm.score(
@@ -36,6 +36,39 @@ from .utils import (
ScoringServeContext: TypeAlias = PoolingServeContext[ScoringRequest]
def _apply_post_tokenization_to_token_type_ids(
tokenizer: Any,
tok_params: TokenizeParams,
token_type_ids: list[int],
) -> list[int]:
pad_length = tok_params.pad_prompt_tokens
if pad_length is not None and pad_length < 0:
pad_length = tok_params.max_input_tokens
if pad_length is not None and pad_length > len(token_type_ids):
pad_token_type_id = token_type_ids[-1] if token_type_ids else 0
token_type_ids = token_type_ids + [pad_token_type_id] * (
pad_length - len(token_type_ids)
)
max_length = tok_params.truncate_prompt_tokens
if max_length is not None and max_length < 0:
max_length = tok_params.max_input_tokens
if max_length is None or max_length >= len(token_type_ids):
return token_type_ids
if max_length == 0:
return token_type_ids[:0]
side = tok_params.truncation_side or (
tokenizer.truncation_side if tokenizer is not None else None
)
if side == "left":
return token_type_ids[-max_length:]
return token_type_ids[:max_length]
class ScoringIOProcessor(PoolingIOProcessor):
name: str
pooling_task: PoolingTask
@@ -469,9 +502,16 @@ class CrossEncoderIOProcessor(ScoringIOProcessor):
else None,
)
if token_type_ids := engine_prompt.pop("token_type_ids", None):
token_type_ids = engine_prompt.pop("token_type_ids", None)
tok_params.apply_post_tokenization(self.tokenizer, engine_prompt)
if token_type_ids is not None:
params = pooling_params.clone()
compressed = compress_token_type_ids(token_type_ids)
compressed = compress_token_type_ids(
_apply_post_tokenization_to_token_type_ids(
self.tokenizer, tok_params, token_type_ids
)
)
params.extra_kwargs = {
**(params.extra_kwargs or {}),
"compressed_token_type_ids": compressed,
@@ -480,7 +520,6 @@ class CrossEncoderIOProcessor(ScoringIOProcessor):
else:
pooling_params_list.append(pooling_params)
tok_params.apply_post_tokenization(self.tokenizer, engine_prompt)
if engine_prompt_extras:
target_prompt = extract_target_prompt(self.model_config, engine_prompt)
target_prompt.update(engine_prompt_extras)