minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Diffusion] Support select fa2 backend in hopper (#17514)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2026-01-22 08:23:53 +08:00
committed by GitHub
parent e6ccb2949b
commit 590969ee9c
1 changed files with 7 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/multimodal_gen/runtime/platforms/cuda.py
+7
View File
@@ -246,6 +246,13 @@ class CudaPlatformBase(Platform):
elif selected_backend == AttentionBackendEnum.SAGE_SLA_ATTN: elif selected_backend == AttentionBackendEnum.SAGE_SLA_ATTN:
logger.info("Using Sage Sparse Linear Attention backend") logger.info("Using Sage Sparse Linear Attention backend")
return "sglang.multimodal_gen.runtime.layers.attention.backends.sparse_linear_attn.SageSparseLinearAttentionBackend" return "sglang.multimodal_gen.runtime.layers.attention.backends.sparse_linear_attn.SageSparseLinearAttentionBackend"
elif selected_backend == AttentionBackendEnum.FA2:
from sglang.multimodal_gen.runtime.layers.attention.backends.flash_attn_2 import ( # noqa: F401
FlashAttention2Backend,
)
logger.info("Using FlashAttention2 backend")
return "sglang.multimodal_gen.runtime.layers.attention.backends.flash_attn_2.FlashAttention2Backend"
elif selected_backend in [ elif selected_backend in [
AttentionBackendEnum.FA, AttentionBackendEnum.FA,
]: ]: