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 13, 12:00 UTC
Prefer FlashAttention MLA as default over FlashMLA (#27363)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Nov 11, 2025 at 11:13 UTC
684f2545851ee0ee49be9a80545ed497324f1a96
1 file changed
+2
-2
vllm/platforms/cuda.py
+2
-2
@@ -55,15 +55,15 @@ def _get_backend_priorities(
55
return [
56
AttentionBackendEnum.CUTLASS_MLA,
57
AttentionBackendEnum.FLASHINFER_MLA,
58
- AttentionBackendEnum.FLASHMLA,
58
AttentionBackendEnum.FLASH_ATTN_MLA,
59
+ AttentionBackendEnum.FLASHMLA,
60
AttentionBackendEnum.TRITON_MLA,
61
AttentionBackendEnum.FLASHMLA_SPARSE,
62
]
63
else:
64
return [
65
- AttentionBackendEnum.FLASHMLA,
65
AttentionBackendEnum.FLASH_ATTN_MLA,
66
+ AttentionBackendEnum.FLASHMLA,
67
AttentionBackendEnum.FLASHINFER_MLA,
68
AttentionBackendEnum.TRITON_MLA,
69
AttentionBackendEnum.FLASHMLA_SPARSE,