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 12, 21:00 UTC
[CI/Build][AMD] Skip if flash_attn_varlen_func not available in test_aiter_flash_attn.py (#29043)
Signed-off-by: Randall Smith <ransmith@amd.com> Co-authored-by: Randall Smith <ransmith@amd.com>
rasmith committed
Nov 20, 2025 at 14:39 UTC
3d84ef9054af190ce68333be3e4d16fe928be754
1 file changed
+3
tests/kernels/attention/test_aiter_flash_attn.py
+3
@@ -6,6 +6,7 @@ import pytest
6
import torch
7
8
import vllm.v1.attention.backends.rocm_aiter_fa # noqa: F401
9
+from vllm.attention.utils.fa_utils import is_flash_attn_varlen_func_available
10
from vllm.platforms import current_platform
11
12
NUM_HEADS = [(4, 4), (8, 2)]
@@ -100,6 +101,8 @@ def test_varlen_with_paged_kv(
101
num_blocks: int,
102
q_dtype: torch.dtype | None,
103
) -> None:
104
+ if not is_flash_attn_varlen_func_available():
105
+ pytest.skip("flash_attn_varlen_func required to run this test.")
106
torch.set_default_device("cuda")
107
current_platform.seed_everything(0)
108
num_seqs = len(seq_lens)