src/models/attention_processor.py

from typing import Callable, List, Optional, Tuple, Union

import torch
import torch.nn.functional as F
from diffusers.models.attention_processor import Attention
from diffusers.utils import logging
from diffusers.utils.import_utils import is_torch_npu_available, is_xformers_available
from diffusers.utils.torch_utils import is_torch_version, maybe_allow_in_graph
from einops import rearrange
from torch import nn

logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

class FlashTripo2AttnProcessor2_0:
    r"""
    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). This is
    used in the Tripo2DiT model. It applies a s normalization layer and rotary embedding on query and key vector.
    """

    def __init__(self, topk=True):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError(
                "AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
            )
        self.topk = topk

    def qkv(self, attn, q, k, v, attn_mask, dropout_p, is_causal):
        if k.shape[-2] == 3072:
            topk = 1024
        elif k.shape[-2] == 512:
            topk = 256
        else:
            topk = k.shape[-2] // 3

        if self.topk is True:
            q1 = q[:, :, ::100, :]
            sim = q1 @ k.transpose(-1, -2)
            sim = torch.mean(sim, -2)
            topk_ind = torch.topk(sim, dim=-1, k=topk).indices.squeeze(-2).unsqueeze(-1)
            topk_ind = topk_ind.expand(-1, -1, -1, v.shape[-1])
            v0 = torch.gather(v, dim=-2, index=topk_ind)
            k0 = torch.gather(k, dim=-2, index=topk_ind)
            out = F.scaled_dot_product_attention(q, k0, v0)
        elif self.topk is False:
            out = F.scaled_dot_product_attention(q, k, v)
        else:
            idx, counts = self.topk
            start = 0
            outs = []
            for grid_coord, count in zip(idx, counts):
                end = start + count
                q_chunk = q[:, :, start:end, :]
                q1 = q_chunk[:, :, ::50, :]
                sim = q1 @ k.transpose(-1, -2)
                sim = torch.mean(sim, -2)
                topk_ind = torch.topk(sim, dim=-1, k=topk).indices.squeeze(-2).unsqueeze(-1)
                topk_ind = topk_ind.expand(-1, -1, -1, v.shape[-1])
                v0 = torch.gather(v, dim=-2, index=topk_ind)
                k0 = torch.gather(k, dim=-2, index=topk_ind)
                out = F.scaled_dot_product_attention(q_chunk, k0, v0)
                outs.append(out)
                start += count
            out = torch.cat(outs, dim=-2)
        self.topk = False
        return out

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        temb: Optional[torch.Tensor] = None,
        image_rotary_emb: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        from diffusers.models.embeddings import apply_rotary_emb

        residual = hidden_states
        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, sequence_length, _ = (
            hidden_states.shape
            if encoder_hidden_states is None
            else encoder_hidden_states.shape
        )

        if attention_mask is not None:
            attention_mask = attn.prepare_attention_mask(
                attention_mask, sequence_length, batch_size
            )
            # scaled_dot_product_attention expects attention_mask shape to be
            # (batch, heads, source_length, target_length)
            attention_mask = attention_mask.view(
                batch_size, attn.heads, -1, attention_mask.shape[-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)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        elif attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(
                encoder_hidden_states
            )

        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        # NOTE that tripo2 split heads first then split qkv or kv, like .view(..., attn.heads, 3, dim)
        # instead of .view(..., 3, attn.heads, dim). So we need to re-split here.
        if not attn.is_cross_attention:
            qkv = torch.cat((query, key, value), dim=-1)
            split_size = qkv.shape[-1] // attn.heads // 3
            qkv = qkv.view(batch_size, -1, attn.heads, split_size * 3)
            query, key, value = torch.split(qkv, split_size, dim=-1)
        else:
            kv = torch.cat((key, value), dim=-1)
            split_size = kv.shape[-1] // attn.heads // 2
            kv = kv.view(batch_size, -1, attn.heads, split_size * 2)
            key, value = torch.split(kv, split_size, dim=-1)

        head_dim = key.shape[-1]

        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        if attn.norm_q is not None:
            query = attn.norm_q(query)
        if attn.norm_k is not None:
            key = attn.norm_k(key)

        # Apply RoPE if needed
        if image_rotary_emb is not None:
            query = apply_rotary_emb(query, image_rotary_emb)
            if not attn.is_cross_attention:
                key = apply_rotary_emb(key, image_rotary_emb)

        # flashvdm topk
        hidden_states = self.qkv(attn, query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False)   

        hidden_states = hidden_states.transpose(1, 2).reshape(
            batch_size, -1, attn.heads * head_dim
        )
        hidden_states = hidden_states.to(query.dtype)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(
                batch_size, channel, height, width
            )

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states

class TripoSGAttnProcessor2_0:
    r"""
    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). This is
    used in the TripoSG model. It applies a s normalization layer and rotary embedding on query and key vector.
    """

    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError(
                "AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
            )

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        temb: Optional[torch.Tensor] = None,
        image_rotary_emb: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        from diffusers.models.embeddings import apply_rotary_emb

        residual = hidden_states
        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, sequence_length, _ = (
            hidden_states.shape
            if encoder_hidden_states is None
            else encoder_hidden_states.shape
        )

        if attention_mask is not None:
            attention_mask = attn.prepare_attention_mask(
                attention_mask, sequence_length, batch_size
            )
            # scaled_dot_product_attention expects attention_mask shape to be
            # (batch, heads, source_length, target_length)
            attention_mask = attention_mask.view(
                batch_size, attn.heads, -1, attention_mask.shape[-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)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        elif attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(
                encoder_hidden_states
            )

        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        # NOTE that pre-trained models split heads first then split qkv or kv, like .view(..., attn.heads, 3, dim)
        # instead of .view(..., 3, attn.heads, dim). So we need to re-split here.
        if not attn.is_cross_attention:
            qkv = torch.cat((query, key, value), dim=-1)
            split_size = qkv.shape[-1] // attn.heads // 3
            qkv = qkv.view(batch_size, -1, attn.heads, split_size * 3)
            query, key, value = torch.split(qkv, split_size, dim=-1)
        else:
            kv = torch.cat((key, value), dim=-1)
            split_size = kv.shape[-1] // attn.heads // 2
            kv = kv.view(batch_size, -1, attn.heads, split_size * 2)
            key, value = torch.split(kv, split_size, dim=-1)

        head_dim = key.shape[-1]

        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        if attn.norm_q is not None:
            query = attn.norm_q(query)
        if attn.norm_k is not None:
            key = attn.norm_k(key)

        # Apply RoPE if needed
        if image_rotary_emb is not None:
            query = apply_rotary_emb(query, image_rotary_emb)
            if not attn.is_cross_attention:
                key = apply_rotary_emb(key, image_rotary_emb)

        # the output of sdp = (batch, num_heads, seq_len, head_dim)
        # TODO: add support for attn.scale when we move to Torch 2.1
        hidden_states = F.scaled_dot_product_attention(
            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
        )

        hidden_states = hidden_states.transpose(1, 2).reshape(
            batch_size, -1, attn.heads * head_dim
        )
        hidden_states = hidden_states.to(query.dtype)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(
                batch_size, channel, height, width
            )

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states


class FusedTripoSGAttnProcessor2_0:
    r"""
    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0) with fused
    projection layers. This is used in the HunyuanDiT model. It applies a s normalization layer and rotary embedding on
    query and key vector.
    """

    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError(
                "FusedTripoSGAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
            )

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        temb: Optional[torch.Tensor] = None,
        image_rotary_emb: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        from diffusers.models.embeddings import apply_rotary_emb

        residual = hidden_states
        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, sequence_length, _ = (
            hidden_states.shape
            if encoder_hidden_states is None
            else encoder_hidden_states.shape
        )

        if attention_mask is not None:
            attention_mask = attn.prepare_attention_mask(
                attention_mask, sequence_length, batch_size
            )
            # scaled_dot_product_attention expects attention_mask shape to be
            # (batch, heads, source_length, target_length)
            attention_mask = attention_mask.view(
                batch_size, attn.heads, -1, attention_mask.shape[-1]
            )

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

        # NOTE that pre-trained split heads first, then split qkv
        if encoder_hidden_states is None:
            qkv = attn.to_qkv(hidden_states)
            split_size = qkv.shape[-1] // attn.heads // 3
            qkv = qkv.view(batch_size, -1, attn.heads, split_size * 3)
            query, key, value = torch.split(qkv, split_size, dim=-1)
        else:
            if attn.norm_cross:
                encoder_hidden_states = attn.norm_encoder_hidden_states(
                    encoder_hidden_states
                )
            query = attn.to_q(hidden_states)

            kv = attn.to_kv(encoder_hidden_states)
            split_size = kv.shape[-1] // attn.heads // 2
            kv = kv.view(batch_size, -1, attn.heads, split_size * 2)
            key, value = torch.split(kv, split_size, dim=-1)

        head_dim = key.shape[-1]

        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        if attn.norm_q is not None:
            query = attn.norm_q(query)
        if attn.norm_k is not None:
            key = attn.norm_k(key)

        # Apply RoPE if needed
        if image_rotary_emb is not None:
            query = apply_rotary_emb(query, image_rotary_emb)
            if not attn.is_cross_attention:
                key = apply_rotary_emb(key, image_rotary_emb)

        # the output of sdp = (batch, num_heads, seq_len, head_dim)
        # TODO: add support for attn.scale when we move to Torch 2.1
        hidden_states = F.scaled_dot_product_attention(
            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
        )

        hidden_states = hidden_states.transpose(1, 2).reshape(
            batch_size, -1, attn.heads * head_dim
        )
        hidden_states = hidden_states.to(query.dtype)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(
                batch_size, channel, height, width
            )

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states

# Modified from https://github.com/VAST-AI-Research/MIDI-3D/blob/main/midi/models/attention_processor.py#L264
class PartCrafterAttnProcessor:
    r"""
    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). This is
    used in the PartCrafter model. It applies a normalization layer and rotary embedding on query and key vector.
    """

    def __init__(self):
        if not hasattr(F, "scaled_dot_product_attention"):
            raise ImportError(
                "AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
            )


    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        temb: Optional[torch.Tensor] = None,
        image_rotary_emb: Optional[torch.Tensor] = None,
        num_parts: Optional[Union[int, torch.Tensor]] = None,
    ) -> torch.Tensor:
        from diffusers.models.embeddings import apply_rotary_emb

        residual = hidden_states
        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, sequence_length, _ = (
            hidden_states.shape
            if encoder_hidden_states is None
            else encoder_hidden_states.shape
        )

        if attention_mask is not None:
            attention_mask = attn.prepare_attention_mask(
                attention_mask, sequence_length, batch_size
            )
            # scaled_dot_product_attention expects attention_mask shape to be
            # (batch, heads, source_length, target_length)
            attention_mask = attention_mask.view(
                batch_size, attn.heads, -1, attention_mask.shape[-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)

        if encoder_hidden_states is None:
            encoder_hidden_states = hidden_states
        elif attn.norm_cross:
            encoder_hidden_states = attn.norm_encoder_hidden_states(
                encoder_hidden_states
            )

        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        # NOTE that pre-trained models split heads first then split qkv or kv, like .view(..., attn.heads, 3, dim)
        # instead of .view(..., 3, attn.heads, dim). So we need to re-split here.
        if not attn.is_cross_attention:
            qkv = torch.cat((query, key, value), dim=-1)
            split_size = qkv.shape[-1] // attn.heads // 3
            qkv = qkv.view(batch_size, -1, attn.heads, split_size * 3)
            query, key, value = torch.split(qkv, split_size, dim=-1)
        else:
            kv = torch.cat((key, value), dim=-1)
            split_size = kv.shape[-1] // attn.heads // 2
            kv = kv.view(batch_size, -1, attn.heads, split_size * 2)
            key, value = torch.split(kv, split_size, dim=-1)

        head_dim = key.shape[-1]

        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)

        if attn.norm_q is not None:
            query = attn.norm_q(query)
        if attn.norm_k is not None:
            key = attn.norm_k(key)

        # Apply RoPE if needed
        if image_rotary_emb is not None:
            query = apply_rotary_emb(query, image_rotary_emb)
            if not attn.is_cross_attention:
                key = apply_rotary_emb(key, image_rotary_emb)

        if isinstance(num_parts, torch.Tensor):
            # Assume list in training, do not consider classifier-free guidance
            idx = 0
            hidden_states_list = []
            for n_p in num_parts:
                k = key[idx : idx + n_p]
                v = value[idx : idx + n_p]
                q = query[idx : idx + n_p]
                idx += n_p
                if k.shape[2] == q.shape[2]:
                    # Assuming self-attention
                    # Here 'b' is always 1
                    k = rearrange(
                        k, "(b ni) h nt c -> b h (ni nt) c", ni=n_p
                    ) # [b, h, ni*nt, c]
                    v = rearrange(
                        v, "(b ni) h nt c -> b h (ni nt) c", ni=n_p
                    ) # [b, h, ni*nt, c]
                else:
                    # Assuming cross-attention
                    # Here 'b' is always 1
                    k = k[::n_p]     # [b, h, nt, c]
                    v = v[::n_p]     # [b, h, nt, c]
                # Here 'b' is always 1
                q = rearrange(
                    q, "(b ni) h nt c -> b h (ni nt) c", ni=n_p
                ) # [b, h, ni*nt, c]
                # the output of sdp = (batch, num_heads, seq_len, head_dim)
                h_s = F.scaled_dot_product_attention(
                    q, k, v,
                    dropout_p=0.0,
                    is_causal=False,
                )
                h_s = h_s.transpose(1, 2).reshape(
                    n_p, -1, attn.heads * head_dim
                )
                h_s = h_s.to(query.dtype)
                hidden_states_list.append(h_s)
            hidden_states = torch.cat(hidden_states_list, dim=0)

        elif isinstance(num_parts, int):
            # Assume single instance
            if key.shape[2] == query.shape[2]:
                # Assuming self-attention
                # Here we need 'b' when using classifier-free guidance
                key = rearrange(
                    key, "(b ni) h nt c -> b h (ni nt) c", ni=num_parts
                ) # [b, h, ni*nt, c]
                value = rearrange(
                    value, "(b ni) h nt c -> b h (ni nt) c", ni=num_parts
                ) # [b, h, ni*nt, c]
            else:
                # Assuming cross-attention
                # Here we need 'b' when using classifier-free guidance
                # Control signal is repeated ni times within each (b, ni)
                # We select only the first instance per group
                key = key[::num_parts]     # [b, h, nt, c]
                value = value[::num_parts] # [b, h, nt, c]
            query = rearrange(
                query, "(b ni) h nt c -> b h (ni nt) c", ni=num_parts
            ) # [b, h, ni*nt, c]

            # the output of sdp = (batch, num_heads, seq_len, head_dim)
            hidden_states = F.scaled_dot_product_attention(
                query,
                key,
                value,
                dropout_p=0.0,
                is_causal=False,
            )
            hidden_states = hidden_states.transpose(1, 2).reshape(
                batch_size, -1, attn.heads * head_dim
            )
            hidden_states = hidden_states.to(query.dtype)

        else:
            raise ValueError(
                "num_parts must be a torch.Tensor or int, but got {}".format(type(num_parts))
            )
        
        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(
                batch_size, channel, height, width
            )

        if attn.residual_connection:
            hidden_states = hidden_states + residual

        hidden_states = hidden_states / attn.rescale_output_factor

        return hidden_states