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][Docs] Add comment explaining TRITON_ATTN fallback for ROCm (#32835)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Jan 22, 2026 at 00:11 UTC
a810299838739b32fea9ecdb6eb32bc845aca6c5
1 file changed
+2
tests/v1/spec_decode/test_acceptance_length.py
+2
@@ -88,6 +88,8 @@ def get_available_attention_backends() -> list[str]:
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
+ # ROCm uses Triton as its default attention backend since
92
+ # Flash Attention is not supported.
93
return ["TRITON_ATTN"]
94
else:
95
return ["FLASH_ATTN"]