[Bugfix] Honor STEP token pooling in DispatchPooler.for_seq_cls (#55307)
Signed-off-by: weihan <130119942+krsish@users.noreply.github.com>
This commit is contained in:
@@ -3,12 +3,15 @@
|
||||
"""Unit tests for sequence and token pooling methods and their factories."""
|
||||
|
||||
from dataclasses import dataclass
|
||||
from types import SimpleNamespace
|
||||
from unittest.mock import patch
|
||||
|
||||
import numpy as np
|
||||
import pytest
|
||||
import torch
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
from vllm.config import PoolerConfig, set_current_vllm_config
|
||||
from vllm.model_executor.layers.pooler import PoolingParamsUpdate
|
||||
from vllm.model_executor.layers.pooler.seqwise.methods import (
|
||||
CLSPool,
|
||||
@@ -16,6 +19,7 @@ from vllm.model_executor.layers.pooler.seqwise.methods import (
|
||||
MeanPool,
|
||||
get_seq_pooling_method,
|
||||
)
|
||||
from vllm.model_executor.layers.pooler.special import DispatchPooler
|
||||
from vllm.model_executor.layers.pooler.tokwise.methods import (
|
||||
AllPool,
|
||||
StepPool,
|
||||
@@ -300,6 +304,50 @@ class _FakeVllmConfig:
|
||||
scheduler_config: _FakeSchedulerConfig
|
||||
|
||||
|
||||
@pytest.mark.parametrize("tok_pooling_type", ["ALL", "STEP"])
|
||||
def test_dispatch_seq_cls_honors_token_pooling_type(tok_pooling_type):
|
||||
"""Filter STEP rows before classification; preserve all rows for resolved ALL."""
|
||||
pooler_config = PoolerConfig(
|
||||
seq_pooling_type="LAST", tok_pooling_type=tok_pooling_type
|
||||
)
|
||||
config = SimpleNamespace(
|
||||
scheduler_config=_FakeSchedulerConfig(),
|
||||
model_config=SimpleNamespace(
|
||||
pooler_config=pooler_config,
|
||||
head_dtype=None,
|
||||
hf_config=PretrainedConfig(num_labels=3),
|
||||
),
|
||||
)
|
||||
classifier = torch.nn.Linear(4, 3)
|
||||
with set_current_vllm_config(config):
|
||||
pooler = DispatchPooler.for_seq_cls(pooler_config, classifier=classifier)
|
||||
|
||||
token_ids = [[2, 99, 3, 99], [4, 99], [5, 6]]
|
||||
params = [
|
||||
PoolingParams(
|
||||
task="token_classify",
|
||||
use_activation=False,
|
||||
step_tag_id=99 if tok_pooling_type == "STEP" else None,
|
||||
)
|
||||
for _ in token_ids
|
||||
]
|
||||
metadata = _make_metadata([4, 2, 2], token_ids=token_ids, pooling_params=params)
|
||||
hidden = torch.arange(32, dtype=torch.float32).reshape(8, 4)
|
||||
outputs = pooler(hidden, metadata)
|
||||
|
||||
expected_rows = (
|
||||
[[1, 3], [5], []]
|
||||
if tok_pooling_type == "STEP"
|
||||
else [[0, 1, 2, 3], [4, 5], [6, 7]]
|
||||
)
|
||||
assert len(outputs) == len(expected_rows)
|
||||
for output, rows in zip(outputs, expected_rows):
|
||||
torch.testing.assert_close(output, classifier(hidden[rows]))
|
||||
assert pooler.get_pooling_updates("token_classify").requires_token_ids == (
|
||||
tok_pooling_type == "STEP"
|
||||
)
|
||||
|
||||
|
||||
class TestAllPool:
|
||||
@staticmethod
|
||||
def _make_all_pool(
|
||||
|
||||
Reference in New Issue
Block a user