A high-throughput and memory-efficient inference and serving engine for LLMs
[CI/Build] Skip test_use_trtllm_attention on non-CUDA platforms (#44687)
Signed-off-by: Dan Blanaru <48605845+DanBlanaru@users.noreply.github.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Dan Blanaru committed
Jun 11, 2026 at 00:18 UTC
82d6b59f0411b70ad2b5a74a24f808270e0af588
1 file changed
+7
tests/kernels/attention/test_use_trtllm_attention.py
+7
@@ -6,12 +6,19 @@ from unittest.mock import patch
6
import pytest
7
import torch
8
9
+from vllm.platforms import current_platform
10
from vllm.utils.flashinfer import (
11
can_use_trtllm_attention,
12
supports_trtllm_attention,
13
use_trtllm_attention,
14
)
15
16
+if not current_platform.is_cuda():
17
+ pytest.skip(
18
+ "TRTLLM attention is only supported on CUDA platforms.",
19
+ allow_module_level=True,
20
+ )
21
+
22
MODEL_CONFIGS = {
23
"Llama-3-70B": dict(num_qo_heads=64, num_kv_heads=8),
24
"Llama-3-8B": dict(num_qo_heads=32, num_kv_heads=8),