[ROCm][CI] fix flashinfer import check (#48647)
Signed-off-by: Divakar Verma <divakar.verma@amd.com>
Divakar Verma committed
Jul 14, 2026 at 16:54 UTC
05d4f8bba3aac85814c3fc42cfc60bef21bb2bb4
1 file changed
+6
-4
tests/v1/attention/test_flashinfer_dcp_spec_reorder.py
+6
-4
@@ -3,11 +3,16 @@
3
"""FlashInfer GQA builder: reorder threshold under DCP with spec decode."""
4
5
import pytest
6
+
7
+from vllm.platforms import current_platform
8
+
9
+if not current_platform.is_cuda():
10
+ pytest.skip("FlashInfer backend requires a CUDA platform.", allow_module_level=True)
11
+
12
import torch
13
14
from tests.v1.attention.utils import create_vllm_config
15
from vllm.config import SpeculativeConfig, set_current_vllm_config
10
-from vllm.platforms import current_platform
16
from vllm.v1.attention.backends import flashinfer as flashinfer_backend
17
from vllm.v1.attention.backends.flashinfer import (
18
FlashInferDecodeKernel,
@@ -16,9 +21,6 @@ from vllm.v1.attention.backends.flashinfer import (
21
from vllm.v1.attention.backends.utils import PerLayerParameters
22
from vllm.v1.kv_cache_interface import FullAttentionSpec
23
19
-if not current_platform.is_cuda():
20
- pytest.skip("FlashInfer backend requires a CUDA platform.", allow_module_level=True)
21
-
24
25
def test_flashinfer_gqa_dcp_spec_decode_clamps_reorder_threshold(monkeypatch):
26
"""trtllm-gen decode receives no cp_rank/global-seq-len information, so its