A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 21:00 UTC
[CI] Fix SPLADE pooler test broken by #38139 (#38495)
Signed-off-by: haosdent <haosdent@gmail.com>
haosdent committed
Mar 30, 2026 at 15:48 UTC
a08b7733fd3c07ad7b28e15a3fd3e75a503989ec
1 file changed
+8
-4
tests/models/language/pooling/test_splade_sparse_pooler.py
+8
-4
@@ -1,8 +1,6 @@
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
-import types
5
-
4
import pytest
5
import torch
6
import torch.nn as nn
@@ -11,6 +9,8 @@ from vllm.model_executor.models.bert import (
9
BertMLMHead,
10
SPLADESparsePooler,
11
)
12
+from vllm.pooling_params import PoolingParams
13
+from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates
14
15
# ---------------------------------------------------------------------
16
# Functional test: SPLADE formula correctness (no HF download needed)
@@ -38,8 +38,12 @@ def test_splade_pooler_matches_reference_formula(B, T, H, V):
38
],
39
dtype=torch.long,
40
)
41
- meta = types.SimpleNamespace(
42
- prompt_lens=prompt_lens_tenser, prompt_token_ids=token_ids
41
+ meta = PoolingMetadata(
42
+ prompt_lens=prompt_lens_tenser,
43
+ prompt_token_ids=token_ids,
44
+ prompt_token_ids_cpu=token_ids,
45
+ pooling_params=[PoolingParams(task="embed")] * B,
46
+ pooling_states=[PoolingStates() for _ in range(B)],
47
)
48
49
# MLM head (prefer BertMLMHead, fallback to Linear if unavailable)