model.py

from dataclasses import dataclass
import numpy as np
import torch

from torch import Tensor, nn
from einops import rearrange

from layers import (DoubleStreamBlock, EmbedND, LastLayer,
                                 MLPEmbedder, SingleStreamBlock,
                                 timestep_embedding)

import torch.distributed as dist
from diffusers.models.embeddings import get_1d_sincos_pos_embed_from_grid

from accelerate.logging import get_logger
logger = get_logger(__name__, log_level="INFO")




@dataclass
class FluxParams:
    in_channels: int
    vec_in_dim: int
    context_in_dim: int
    hidden_size: int
    mlp_ratio: float
    num_heads: int
    depth: int
    depth_single_blocks: int
    axes_dim: list[int]
    theta: int
    qkv_bias: bool
    guidance_embed: bool
     

class Flux(nn.Module):
    """
    Transformer model for flow matching on sequences.
    """
    _supports_gradient_checkpointing = True

    def __init__(self, params: FluxParams):
        super().__init__()

        self.params = params
        self.in_channels = params.in_channels
        self.out_channels = self.in_channels
        if params.hidden_size % params.num_heads != 0:
            raise ValueError(
                f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
            )
        pe_dim = params.hidden_size // params.num_heads
        if sum(params.axes_dim) != pe_dim:
            raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
        self.hidden_size = params.hidden_size
        self.num_heads = params.num_heads
        self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)

        self.img_in = nn.Linear(self.in_channels, self.hidden_size, bias=True)
        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size)
        self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size)
        self.guidance_in = (
            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size) if params.guidance_embed else nn.Identity()
        )
        self.txt_in = nn.Linear(params.context_in_dim, self.hidden_size)


        self.double_blocks = nn.ModuleList(
            [
                DoubleStreamBlock(
                    self.hidden_size,
                    self.num_heads,
                    mlp_ratio=params.mlp_ratio,
                    qkv_bias=params.qkv_bias
                )
                for i in range(1, params.depth+1)
            ]
        )

        self.single_blocks = nn.ModuleList(
            [
                SingleStreamBlock(
                    self.hidden_size, 
                    self.num_heads, 
                    mlp_ratio=params.mlp_ratio
                )
                for i in range(1, params.depth_single_blocks+1)
            ]
        )

        self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels)
        self.gradient_checkpointing = True 

    def _set_gradient_checkpointing(self, module, value=False):
        if hasattr(module, "gradient_checkpointing"):
            module.gradient_checkpointing = value
    
    @property
    def attn_processors(self):
        # set recursively
        processors = {}

        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors):
            if hasattr(module, "set_processor"):
                processors[f"{name}.processor"] = module.processor

            for sub_name, child in module.named_children():
                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)

            return processors

        for name, module in self.named_children():
            fn_recursive_add_processors(name, module, processors)

        return processors

    def set_attn_processor(self, processor):
        r"""
        Sets the attention processor to use to compute attention.

        Parameters:
            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
                The instantiated processor class or a dictionary of processor classes that will be set as the processor
                for **all** `Attention` layers.

                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
                processor. This is strongly recommended when setting trainable attention processors.

        """
        count = len(self.attn_processors.keys())

        if isinstance(processor, dict) and len(processor) != count:
            raise ValueError(
                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
            )

        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
            if hasattr(module, "set_processor"):
                if not isinstance(processor, dict):
                    module.set_processor(processor)
                else:
                    module.set_processor(processor.pop(f"{name}.processor"))

            for sub_name, child in module.named_children():
                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)

        for name, module in self.named_children():
            fn_recursive_attn_processor(name, module, processor)

    def forward(
        self,
        img: Tensor,
        img_ids: Tensor,
        txt: Tensor,
        txt_ids: Tensor,
        timesteps: Tensor,
        y: Tensor,
        block_controlnet_hidden_states=None,
        guidance: Tensor = None,
        image_proj: Tensor = None, 
        ip_scale: Tensor = 1.0, 
        return_intermediate: bool = False,
    ):

        if return_intermediate:
            intermediate_double = []
            intermediate_single = []
            
        # running on sequences img
        img = self.img_in(img)
        vec = self.time_in(timestep_embedding(timesteps, 256))
        if self.params.guidance_embed:
            if guidance is None:
                raise ValueError("Didn't get guidance strength for guidance distilled model.")
            vec = vec + self.guidance_in(timestep_embedding(guidance, 256))
        vec = vec + self.vector_in(y)
        txt = self.txt_in(txt)

        ids = torch.cat((txt_ids, img_ids), dim=1)
        pe = self.pe_embedder(ids)
        
        if block_controlnet_hidden_states is not None:
            controlnet_depth = len(block_controlnet_hidden_states)


        for index_block, block in enumerate(self.double_blocks):

            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward
                
                img, txt = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(block),
                    img,
                    txt,
                    vec,
                    pe,
                    image_proj,
                    ip_scale,
                    use_reentrant=False
                )

            else:
                img, txt = block(
                    img=img, 
                    txt=txt, 
                    vec=vec, 
                    pe=pe, 
                    image_proj=image_proj,
                    ip_scale=ip_scale
                )
                    
    
            if return_intermediate:
                intermediate_double.append(
                    [img, txt]
                )
          
            if block_controlnet_hidden_states is not None:
                img = img + block_controlnet_hidden_states[index_block % 2]

        img = torch.cat((txt, img), dim=1)
        txt_dim = txt.shape[1]
        for index_block, block in enumerate(self.single_blocks):
    
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

                # ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                img = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(block),
                    img,
                    vec,
                    pe,
                    use_reentrant=False
                )

            else:
                img = block(img, vec=vec, pe=pe)


            # if return_intermediate:
            img_ = img[:, txt.shape[1]:, ...]
            txt_ = img[:, :txt.shape[1], ...]

            if return_intermediate:
                intermediate_single.append(
                    [img_, txt_]
                )
                
            img = torch.cat([txt_, img_], dim=1)

        img = img[:, txt.shape[1] :, ...]
        img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)    
        if return_intermediate:
            return img, intermediate_double, intermediate_single
        else:
            return img