models/customer_attention_processor.py

# Copyright 2024 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Optional, Union, Tuple

import torch
import torch.nn.functional as F
from torch import nn

from diffusers.utils import logging
from diffusers.models.attention_processor import Attention

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


class CustomLiteLAProcessor2_0:
    """Attention processor used typically in processing the SD3-like self-attention projections. add rms norm for query and key and apply RoPE"""

    def __init__(self):
        self.kernel_func = nn.ReLU(inplace=False)
        self.eps = 1e-15
        self.pad_val = 1.0

    def apply_rotary_emb(
        self,
        x: torch.Tensor,
        freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]],
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings
        to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are
        reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting
        tensors contain rotary embeddings and are returned as real tensors.

        Args:
            x (`torch.Tensor`):
                Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply
            freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],)

        Returns:
            Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings.
        """
        cos, sin = freqs_cis  # [S, D]
        cos = cos[None, None]
        sin = sin[None, None]
        cos, sin = cos.to(x.device), sin.to(x.device)

        x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
        x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
        out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)

        return out

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: torch.FloatTensor = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
        rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
        *args,
        **kwargs,
    ) -> torch.FloatTensor:
        hidden_states_len = hidden_states.shape[1]

        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)
        if encoder_hidden_states is not None:
            context_input_ndim = encoder_hidden_states.ndim
            if context_input_ndim == 4:
                batch_size, channel, height, width = encoder_hidden_states.shape
                encoder_hidden_states = encoder_hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

        batch_size = hidden_states.shape[0]

        # `sample` projections.
        dtype = hidden_states.dtype
        query = attn.to_q(hidden_states)
        key = attn.to_k(hidden_states)
        value = attn.to_v(hidden_states)

        # `context` projections.
        has_encoder_hidden_state_proj = hasattr(attn, "add_q_proj") and hasattr(attn, "add_k_proj") and hasattr(attn, "add_v_proj")
        if encoder_hidden_states is not None and has_encoder_hidden_state_proj:
            encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
            encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
            encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)

            # attention
            if not attn.is_cross_attention:
                query = torch.cat([query, encoder_hidden_states_query_proj], dim=1)
                key = torch.cat([key, encoder_hidden_states_key_proj], dim=1)
                value = torch.cat([value, encoder_hidden_states_value_proj], dim=1)
            else:
                query = hidden_states
                key = encoder_hidden_states
                value = encoder_hidden_states

        inner_dim = key.shape[-1]
        head_dim = inner_dim // attn.heads

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

        # RoPE需要 [B, H, S, D] 输入
        # 此时 query是 [B, H, D, S], 需要转成 [B, H, S, D] 才能应用RoPE
        query = query.permute(0, 1, 3, 2)  # [B, H, S, D]  (从 [B, H, D, S])

        # Apply query and key normalization if needed
        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 rotary_freqs_cis is not None:
            query = self.apply_rotary_emb(query, rotary_freqs_cis)
            if not attn.is_cross_attention:
                key = self.apply_rotary_emb(key, rotary_freqs_cis)
            elif rotary_freqs_cis_cross is not None and has_encoder_hidden_state_proj:
                key = self.apply_rotary_emb(key, rotary_freqs_cis_cross)

        # 此时 query是 [B, H, S, D]，需要还原成 [B, H, D, S]
        query = query.permute(0, 1, 3, 2)  # [B, H, D, S]

        if attention_mask is not None:
            # attention_mask: [B, S] -> [B, 1, S, 1]
            attention_mask = attention_mask[:, None, :, None].to(key.dtype)  # [B, 1, S, 1]
            query = query * attention_mask.permute(0, 1, 3, 2)  # [B, H, S, D] * [B, 1, S, 1]
            if not attn.is_cross_attention:
                key = key * attention_mask  # key: [B, h, S, D] 与 mask [B, 1, S, 1] 相乘
                value = value * attention_mask.permute(0, 1, 3, 2)  # 如果 value 是 [B, h, D, S]，那么需调整mask以匹配S维度

        if attn.is_cross_attention and encoder_attention_mask is not None and has_encoder_hidden_state_proj:
            encoder_attention_mask = encoder_attention_mask[:, None, :, None].to(key.dtype)  # [B, 1, S_enc, 1]
            # 此时 key: [B, h, S_enc, D], value: [B, h, D, S_enc]
            key = key * encoder_attention_mask  # [B, h, S_enc, D] * [B, 1, S_enc, 1]
            value = value * encoder_attention_mask.permute(0, 1, 3, 2)  # [B, h, D, S_enc] * [B, 1, 1, S_enc]

        query = self.kernel_func(query)
        key = self.kernel_func(key)

        query, key, value = query.float(), key.float(), value.float()

        value = F.pad(value, (0, 0, 0, 1), mode="constant", value=self.pad_val)

        vk = torch.matmul(value, key)

        hidden_states = torch.matmul(vk, query)

        if hidden_states.dtype in [torch.float16, torch.bfloat16]:
            hidden_states = hidden_states.float()

        hidden_states = hidden_states[:, :, :-1] / (hidden_states[:, :, -1:] + self.eps)

        hidden_states = hidden_states.view(batch_size, attn.heads * head_dim, -1).permute(0, 2, 1)

        hidden_states = hidden_states.to(dtype)
        if encoder_hidden_states is not None:
            encoder_hidden_states = encoder_hidden_states.to(dtype)

        # Split the attention outputs.
        if encoder_hidden_states is not None and not attn.is_cross_attention and has_encoder_hidden_state_proj:
            hidden_states, encoder_hidden_states = (
                hidden_states[:, : hidden_states_len],
                hidden_states[:, hidden_states_len:],
            )

        # linear proj
        hidden_states = attn.to_out[0](hidden_states)
        # dropout
        hidden_states = attn.to_out[1](hidden_states)
        if encoder_hidden_states is not None and not attn.context_pre_only and not attn.is_cross_attention and hasattr(attn, "to_add_out"):
            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)

        if input_ndim == 4:
            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
        if encoder_hidden_states is not None and context_input_ndim == 4:
            encoder_hidden_states = encoder_hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

        if torch.get_autocast_gpu_dtype() == torch.float16:
            hidden_states = hidden_states.clip(-65504, 65504)
            if encoder_hidden_states is not None:
                encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)

        return hidden_states, encoder_hidden_states


class CustomerAttnProcessor2_0:
    r"""
    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
    """

    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 apply_rotary_emb(
        self,
        x: torch.Tensor,
        freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]],
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings
        to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are
        reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting
        tensors contain rotary embeddings and are returned as real tensors.

        Args:
            x (`torch.Tensor`):
                Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply
            freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],)

        Returns:
            Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings.
        """
        cos, sin = freqs_cis  # [S, D]
        cos = cos[None, None]
        sin = sin[None, None]
        cos, sin = cos.to(x.device), sin.to(x.device)

        x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
        x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
        out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)

        return out

    def __call__(
        self,
        attn: Attention,
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: torch.FloatTensor = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
        rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
        *args,
        **kwargs,
    ) -> torch.Tensor:

        residual = hidden_states
        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
        )
        
        has_encoder_hidden_state_proj = hasattr(attn, "add_q_proj") and hasattr(attn, "add_k_proj") and hasattr(attn, "add_v_proj")

        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)

        inner_dim = key.shape[-1]
        head_dim = inner_dim // attn.heads

        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 rotary_freqs_cis is not None:
            query = self.apply_rotary_emb(query, rotary_freqs_cis)
            if not attn.is_cross_attention:
                key = self.apply_rotary_emb(key, rotary_freqs_cis)
            elif rotary_freqs_cis_cross is not None and has_encoder_hidden_state_proj:
                key = self.apply_rotary_emb(key, rotary_freqs_cis_cross)

        if attn.is_cross_attention and encoder_attention_mask is not None and has_encoder_hidden_state_proj:
            # attention_mask: N x S1
            # encoder_attention_mask: N x S2
            # cross attention 整合attention_mask和encoder_attention_mask
            combined_mask = attention_mask[:, :, None] * encoder_attention_mask[:, None, :]
            attention_mask = torch.where(combined_mask == 1, 0.0, -torch.inf)
            attention_mask = attention_mask[:, None, :, :].expand(-1, attn.heads, -1, -1).to(query.dtype)

        elif not attn.is_cross_attention and 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])

        # 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