vllm_omni.diffusion.attention.backends.cudnn_attn ¶

CuDNNAttentionBackend ¶

accept_output_buffer: bool = True

supports_prefix_kv_slicing: bool = True

get_impl_cls() -> type[CuDNNAttentionImpl]

get_name() -> str

get_supported_head_sizes() -> list[int]

supports_attention_mask(
    attention_spec: object | None = None,
) -> bool

supports_head_size(head_size: int) -> bool

backend_explicit = bool(
    extra_impl_args.get("backend_explicit", False)
)

causal = causal

softmax_scale = softmax_scale

forward_cuda(
    query: Tensor,
    key: Tensor,
    value: Tensor,
    attn_metadata: AttentionMetadata | None = None,
) -> Tensor