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 16, 02:00 UTC
[Bugfix][Kernel] Give unique name to BlockSparseFlashAttention (#12040)
Signed-off-by: Chen Zhang <zhangch99@outlook.com>
Chen Zhang committed
Jan 14, 2025 at 23:45 UTC
a2d2acb4c8d240b1e5946afe2736e497ce5b71a2
2 files changed
+2
-2
vllm/attention/backends/blocksparse_attn.py
+1
-2
@@ -89,8 +89,7 @@ class BlocksparseFlashAttentionBackend(AttentionBackend):
89
90
@staticmethod
91
def get_name() -> str:
92
- # For attention layer compatibility
93
- return "FLASH_ATTN"
92
+ return "BLOCK_SPARSE_FLASH_ATTN"
93
94
@staticmethod
95
def get_impl_cls() -> Type["BlocksparseFlashAttentionImpl"]:
vllm/platforms/interface.py
+1
@@ -33,6 +33,7 @@ class _Backend(enum.Enum):
33
HPU_ATTN = enum.auto()
34
PALLAS = enum.auto()
35
IPEX = enum.auto()
36
+ BLOCK_SPARSE_FLASH_ATTN = enum.auto()
37
NO_ATTENTION = enum.auto()
38
39