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, 16:30 UTC
[ROCm][CI] Fix ModernBERT token classification test numerical accuracy on ROCm (#31820)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Jan 6, 2026 at 17:21 UTC
2a42ae790d3f42a381b782a504947255aa0db090
1 file changed
+29
tests/models/language/pooling/conftest.py
new
+29
@@ -0,0 +1,29 @@
1
+# SPDX-License-Identifier: Apache-2.0
2
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
+"""Pytest configuration for vLLM language generation tests."""
4
+
5
+import warnings
6
+
7
+import torch
8
+
9
+from vllm.platforms import current_platform
10
+
11
+
12
+def pytest_sessionstart(session):
13
+ """Configure ROCm-specific settings before test session starts."""
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
+ torch.set_float32_matmul_precision("high")
24
+ warnings.warn(
25
+ "ROCm: Disabled flash_sdp and mem_efficient_sdp, enabled math_sdp "
26
+ "to avoid HuggingFace Transformers accuracy issues",
27
+ UserWarning,
28
+ stacklevel=1,
29
+ )