vllm_omni.diffusion.attention.backends.ring.ring_globals ¶

FA3_SUPPORTED_CUDA_MAJORS: frozenset[int] | None = (
    frozenset({9})
)

HAS_AITER = True

HAS_FA3 = True

HAS_FA4 = True

HAS_FLASHINFER = True

HAS_FLASH_ATTN = True

HAS_FLASH_ATTN_HOPPER = HAS_FA3

HAS_NPU = True

HAS_SAGE_ATTENTION = True

HAS_SPARSE_SAGE_ATTENTION = True

fa3_fwd_version = version('fa3_fwd')

flash3_attn_func = fa3_attn_func

flash_attn_forward_hopper = fa3_fwd_func

logger = logging.getLogger(__name__)

published_majors = getattr(
    fa3_fwd_interface, "SUPPORTED_CUDA_MAJORS", None
)