Class XFormersAttnProcessor

models_diffusers/attention_processor.py:1081–1175

r""" Processor for implementing memory efficient attention using xFormers. Args: attention_op (`Callable`, *optional*, defaults to `None`): The base [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to

class XFormersAttnProcessor:
    r"""
    Processor for implementing memory efficient attention using xFormers.

    Args:
        attention_op (`Callable`, *optional*, defaults to `None`):
            The base
            [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to
            use as the attention operator. It is recommended to set to `None`, and allow xFormers to choose the best
            operator.
    """

    def __init__(self, attention_op: Optional[Callable] = None):
        self.attention_op = attention_op

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        temb: Optional[torch.FloatTensor] = None,
        scale: float = 1.0,
    ) -> torch.FloatTensor:
        residual = hidden_states

        args = () if USE_PEFT_BACKEND else (scale,)

        if attn.spatial_norm is not None:
            hidden_states = attn.spatial_norm(hidden_states, temb)

        input_ndim = hidden_states.ndim

        if input_ndim == 4:
            batch_size, channel, height, width = hidden_states.shape
            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

        batch_size, key_tokens, _ = (
            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
        )

        attention_mask = attn.prepare_attention_mask(attention_mask, key_tokens, batch_size)
        if attention_mask is not None:
            # expand our mask's singleton query_tokens dimension:
            #   [batch*heads,            1, key_tokens] ->
            #   [batch*heads, query_tokens, key_tokens]
            # so that it can be added as a bias onto the attention scores that xformers computes:
            #   [batch*heads, query_tokens, key_tokens]
            # we do this explicitly because xformers doesn't broadcast the singleton dimension for us.
            _, query_tokens, _ = hidden_states.shape
            attention_mask = attention_mask.expand(-1, query_tokens, -1)

        if attn.group_norm is not None:
            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        query = attn.to_q(hidden_states, *args)

        if encoder_hidden_states is None: