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 10, 02:45 UTC
[Refactor] Remove unused file (#32422)
Wentao Ye committed
Jan 15, 2026 at 17:59 UTC
aca5c514876303913c686866cd7571638e512064
1 file changed
-28
vllm/entrypoints/pooling/embed/conftest.py
deleted
-28
@@ -1,28 +0,0 @@
1
-# SPDX-License-Identifier: Apache-2.0
2
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
-"""Pytest configuration for vLLM pooling embed tests."""
4
-
5
-import warnings
6
-
7
-import torch
8
-
9
-from vllm.platforms import current_platform
10
-
11
-
12
-def pytest_collection_modifyitems(config, items):
13
- """Configure ROCm-specific settings based on collected tests."""
14
- if not current_platform.is_rocm():
15
- return
16
-
17
- # Disable Flash/MemEfficient SDP on ROCm to avoid HF Transformers
18
- # accuracy issues: https://github.com/vllm-project/vllm/issues/30167
19
- # TODO: Remove once ROCm SDP accuracy issues are resolved on HuggingFace
20
- torch.backends.cuda.enable_flash_sdp(False)
21
- torch.backends.cuda.enable_mem_efficient_sdp(False)
22
- torch.backends.cuda.enable_math_sdp(True)
23
- warnings.warn(
24
- "ROCm: Disabled flash_sdp and mem_efficient_sdp, enabled math_sdp "
25
- "to avoid HuggingFace Transformers accuracy issues",
26
- UserWarning,
27
- stacklevel=1,
28
- )