Implements dual-chunk-flash-attn backend for dual chunk attention with sparse attention support (#11844)
This commit is contained in:
@@ -222,6 +222,10 @@ class CudaPlatformBase(Platform):
|
||||
elif selected_backend == _Backend.XFORMERS:
|
||||
logger.info("Using XFormers backend.")
|
||||
return "vllm.attention.backends.xformers.XFormersBackend"
|
||||
elif selected_backend == _Backend.DUAL_CHUNK_FLASH_ATTN:
|
||||
logger.info("Using DualChunkFlashAttention backend.")
|
||||
return ("vllm.attention.backends.dual_chunk_flash_attn."
|
||||
"DualChunkFlashAttentionBackend")
|
||||
elif selected_backend == _Backend.FLASH_ATTN:
|
||||
pass
|
||||
elif selected_backend:
|
||||
|
||||
@@ -51,6 +51,7 @@ class _Backend(enum.Enum):
|
||||
PALLAS_VLLM_V1 = enum.auto()
|
||||
IPEX = enum.auto()
|
||||
BLOCK_SPARSE_FLASH_ATTN = enum.auto()
|
||||
DUAL_CHUNK_FLASH_ATTN = enum.auto()
|
||||
NO_ATTENTION = enum.auto()
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user