[diffusion] feat: add AITER Sage attention backend (#20178)

This commit is contained in:
Aleksi Vesanto
2026-03-11 12:17:45 +08:00
committed by GitHub
parent 21a0015aa3
commit c8bbe5010a
7 changed files with 98 additions and 1 deletions
@@ -34,6 +34,7 @@ class AttentionBackendEnum(enum.Enum):
SPARSE_VIDEO_GEN_2_ATTN = enum.auto()
VMOBA_ATTN = enum.auto()
AITER = enum.auto()
AITER_SAGE = enum.auto()
SLA_ATTN = enum.auto()
SAGE_SLA_ATTN = enum.auto()
NO_ATTENTION = enum.auto()
@@ -115,6 +115,16 @@ class RocmPlatform(Platform):
logger.info("Using AITer backend on ROCm.")
return "sglang.multimodal_gen.runtime.layers.attention.backends.aiter.AITerBackend"
elif selected_backend == AttentionBackendEnum.AITER_SAGE:
if dtype in (torch.float16, torch.bfloat16):
logger.info("Using AITER Sage backend on ROCm.")
return "sglang.multimodal_gen.runtime.layers.attention.backends.aiter_sage.AITERSageBackend"
else:
logger.warning(
"AITER Sage backend only supports bf16/fp16 inputs but got dtype=%s.",
dtype,
)
elif selected_backend in (
AttentionBackendEnum.SLIDING_TILE_ATTN,
AttentionBackendEnum.SAGE_ATTN,