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 8, 06:15 UTC
[ROCm][CI] Guard sparse MLA backend imports for ROCm compatibility in tests (#34538)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Feb 14, 2026 at 09:32 UTC
b3c14229b032a8bbf93d450a52c9a404ddaea429
1 file changed
+11
tests/v1/attention/test_sparse_mla_backends.py
+11
@@ -23,6 +23,17 @@ from vllm import _custom_ops as ops
23
from vllm.config import set_current_vllm_config
24
from vllm.model_executor.layers.linear import ColumnParallelLinear
25
from vllm.platforms import current_platform
26
+
27
+# TODO: Integrate ROCMAiterMLASparseBackend for ROCm.
28
+# The ROCm sparse MLA backend (rocm_aiter_mla_sparse.py) has a compatible
29
+# forward_mqa interface but needs validation on ROCm hardware.
30
+if not current_platform.is_cuda():
31
+ pytest.skip(
32
+ "Sparse MLA backend tests currently only support CUDA. "
33
+ "ROCm support requires integrating ROCMAiterMLASparseBackend.",
34
+ allow_module_level=True,
35
+ )
36
+
37
from vllm.utils.math_utils import cdiv
38
from vllm.v1.attention.backends.mla.flashinfer_mla_sparse import (
39
FlashInferMLASparseBackend,