Add tree attention backend for v1 (part 1) (#20401)

Signed-off-by: Giancarlo Delfin <gdelfin@meta.com>
This commit is contained in:
Giancarlo Delfin
2025-08-03 22:13:26 -07:00
committed by GitHub
parent c2e75b3c11
commit aa7012eb6d
12 changed files with 1098 additions and 25 deletions
+4
View File
@@ -270,6 +270,7 @@ class CudaPlatformBase(Platform):
FLEX_ATTENTION_V1 = "vllm.v1.attention.backends.flex_attention.FlexAttentionBackend" # noqa: E501
TRITON_ATTN_VLLM_V1 = "vllm.v1.attention.backends.triton_attn.TritonAttentionBackend" # noqa: E501
FLASH_ATTN_V1 = "vllm.v1.attention.backends.flash_attn.FlashAttentionBackend" # noqa: E501
TREE_ATTN_V1 = "vllm.v1.attention.backends.tree_attn.TreeAttentionBackend" # noqa: E501
if selected_backend == _Backend.FLASHINFER:
logger.info_once("Using FlashInfer backend on V1 engine.")
@@ -287,6 +288,9 @@ class CudaPlatformBase(Platform):
elif selected_backend == _Backend.FLASH_ATTN:
logger.info_once("Using Flash Attention backend on V1 engine.")
return FLASH_ATTN_V1
elif selected_backend == _Backend.TREE_ATTN:
logger.info_once("Using Tree Attention backend on V1 engine.")
return TREE_ATTN_V1
from vllm.attention.selector import is_attn_backend_supported
+1
View File
@@ -62,6 +62,7 @@ class _Backend(enum.Enum):
DIFFERENTIAL_FLASH_ATTN = enum.auto()
NO_ATTENTION = enum.auto()
FLEX_ATTENTION = enum.auto()
TREE_ATTN = enum.auto()
class PlatformEnum(enum.Enum):