chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[diffusion] feat: support SageSparseLinearAttention attention backend (#17399)

Browse Source
This commit is contained in:
HuangJi
2026-01-21 18:13:51 +08:00
committed by GitHub
parent 1b97fa769b
commit e776239afd
6 changed files with 479 additions and 131 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/multimodal_gen/runtime/platforms/cuda.py
+3
View File
@@ -243,6 +243,9 @@ class CudaPlatformBase(Platform):
elif selected_backend == AttentionBackendEnum.SLA_ATTN:
logger.info("Using Sparse Linear Attention backend")
return "sglang.multimodal_gen.runtime.layers.attention.backends.sparse_linear_attn.SparseLinearAttentionBackend"
elif selected_backend == AttentionBackendEnum.SAGE_SLA_ATTN:
logger.info("Using Sage Sparse Linear Attention backend")
return "sglang.multimodal_gen.runtime.layers.attention.backends.sparse_linear_attn.SageSparseLinearAttentionBackend"
elif selected_backend in [
AttentionBackendEnum.FA,
]:
python/sglang/multimodal_gen/runtime/platforms/interface.py
+1
View File
@@ -34,6 +34,7 @@ class AttentionBackendEnum(enum.Enum):
VMOBA_ATTN = enum.auto()
AITER = enum.auto()
SLA_ATTN = enum.auto()
SAGE_SLA_ATTN = enum.auto()
NO_ATTENTION = enum.auto()
def __str__(self):