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 9, 20:45 UTC
[ROCm][CI] fix get_valid_backends (#32787)
Signed-off-by: Divakar Verma <divakar.verma@amd.com>
Divakar Verma committed
Jan 21, 2026 at 22:27 UTC
49d9653852c49893efabb5c2bff5309268517a41
1 file changed
+8
-2
tests/v1/spec_decode/test_acceptance_length.py
+8
-2
@@ -83,8 +83,14 @@ EXCLUDED_BACKENDS = {AttentionBackendEnum.FLEX_ATTENTION}
83
84
85
def get_available_attention_backends() -> list[str]:
86
- if not hasattr(current_platform, "get_valid_backends"):
87
- return ["FLASH_ATTN"]
86
+ # Check if get_valid_backends is actually defined in the platform class
87
+ # (not just returning None from __getattr__)
88
+ get_valid_backends = getattr(current_platform.__class__, "get_valid_backends", None)
89
+ if get_valid_backends is None:
90
+ if current_platform.is_rocm():
91
+ return ["TRITON_ATTN"]
92
+ else:
93
+ return ["FLASH_ATTN"]
94
95
device_capability = current_platform.get_device_capability()
96
if device_capability is None: