model/flux.py

import math
from collections import OrderedDict
from functools import partial

import torch
from einops import rearrange, repeat
from scepter.modules.model.base_model import BaseModel
from scepter.modules.model.registry import BACKBONES
from scepter.modules.utils.config import dict_to_yaml
from scepter.modules.utils.distribute import we
from scepter.modules.utils.file_system import FS
from torch import Tensor, nn
from torch.utils.checkpoint import checkpoint_sequential

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

@BACKBONES.register_class()
class Flux(BaseModel):
    """
    Transformer backbone Diffusion model with RoPE.
    """
    para_dict = {
        "IN_CHANNELS": {
            "value": 64,
            "description": "model's input channels."
        },
        "OUT_CHANNELS": {
            "value": 64,
            "description": "model's output channels."
        },
        "HIDDEN_SIZE": {
            "value": 1024,
            "description": "model's hidden size."
        },
        "NUM_HEADS": {
            "value": 16,
            "description": "number of heads in the transformer."
        },
        "AXES_DIM": {
            "value": [16, 56, 56],
            "description": "dimensions of the axes of the positional encoding."
        },
        "THETA": {
            "value": 10_000,
            "description": "theta for positional encoding."
        },
        "VEC_IN_DIM": {
            "value": 768,
            "description": "dimension of the vector input."
        },
        "GUIDANCE_EMBED": {
            "value": False,
            "description": "whether to use guidance embedding."
        },
        "CONTEXT_IN_DIM": {
            "value": 4096,
            "description": "dimension of the context input."
        },
        "MLP_RATIO": {
            "value": 4.0,
            "description": "ratio of mlp hidden size to hidden size."
        },
        "QKV_BIAS": {
            "value": True,
            "description": "whether to use bias in qkv projection."
        },
        "DEPTH": {
            "value": 19,
            "description": "number of transformer blocks."
        },
        "DEPTH_SINGLE_BLOCKS": {
            "value": 38,
            "description": "number of transformer blocks in the single stream block."
        },
        "USE_GRAD_CHECKPOINT": {
            "value": False,
            "description": "whether to use gradient checkpointing."
        },
        "ATTN_BACKEND": {
            "value": "pytorch",
            "description": "backend for the transformer blocks, 'pytorch' or 'flash_attn'."
        }
    }
    def __init__(
            self,
            cfg,
            logger = None
    ):
        super().__init__(cfg, logger=logger)
        self.in_channels = cfg.IN_CHANNELS
        self.out_channels = cfg.get("OUT_CHANNELS", self.in_channels)
        hidden_size = cfg.get("HIDDEN_SIZE", 1024)
        num_heads = cfg.get("NUM_HEADS", 16)
        axes_dim = cfg.AXES_DIM
        theta = cfg.THETA
        vec_in_dim = cfg.VEC_IN_DIM
        self.guidance_embed = cfg.GUIDANCE_EMBED
        context_in_dim = cfg.CONTEXT_IN_DIM
        mlp_ratio = cfg.MLP_RATIO
        qkv_bias = cfg.QKV_BIAS
        depth = cfg.DEPTH
        depth_single_blocks = cfg.DEPTH_SINGLE_BLOCKS
        self.use_grad_checkpoint = cfg.get("USE_GRAD_CHECKPOINT", False)
        self.attn_backend = cfg.get("ATTN_BACKEND", "pytorch")
        self.lora_model = cfg.get("DIFFUSERS_LORA_MODEL", None)
        self.swift_lora_model = cfg.get("SWIFT_LORA_MODEL", None)
        self.pretrain_adapter = cfg.get("PRETRAIN_ADAPTER", None)

        if hidden_size % num_heads != 0:
            raise ValueError(
                f"Hidden size {hidden_size} must be divisible by num_heads {num_heads}"
            )
        pe_dim = hidden_size // num_heads
        if sum(axes_dim) != pe_dim:
            raise ValueError(f"Got {axes_dim} but expected positional dim {pe_dim}")
        self.hidden_size = hidden_size
        self.num_heads = num_heads
        self.pe_embedder = EmbedND(dim=pe_dim, theta=theta, axes_dim= 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(vec_in_dim, self.hidden_size)
        self.guidance_in = (
            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size) if self.guidance_embed else nn.Identity()
        )
        self.txt_in = nn.Linear(context_in_dim, self.hidden_size)

        self.double_blocks = nn.ModuleList(
            [
                DoubleStreamBlock(
                    self.hidden_size,
                    self.num_heads,
                    mlp_ratio=mlp_ratio,
                    qkv_bias=qkv_bias,
                    backend=self.attn_backend
                )
                for _ in range(depth)
            ]
        )

        self.single_blocks = nn.ModuleList(
            [
                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=mlp_ratio, backend=self.attn_backend)
                for _ in range(depth_single_blocks)
            ]
        )

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

    def prepare_input(self, x, context, y, x_shape=None):
        # x.shape [6, 16, 16, 16] target is [6, 16, 768, 1360]
        bs, c, h, w = x.shape
        x = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
        x_id = torch.zeros(h // 2, w // 2, 3)
        x_id[..., 1] = x_id[..., 1] + torch.arange(h // 2)[:, None]
        x_id[..., 2] = x_id[..., 2] + torch.arange(w // 2)[None, :]
        x_ids = repeat(x_id, "h w c -> b (h w) c", b=bs)
        txt_ids = torch.zeros(bs, context.shape[1], 3)
        return x, x_ids.to(x), context.to(x), txt_ids.to(x), y.to(x), h, w

    def unpack(self, x: Tensor, height: int, width: int) -> Tensor:
        return rearrange(
            x,
            "b (h w) (c ph pw) -> b c (h ph) (w pw)",
            h=math.ceil(height/2),
            w=math.ceil(width/2),
            ph=2,
            pw=2,
        )

    def merge_diffuser_lora(self, ori_sd, lora_sd, scale = 1.0):
        key_map = {
            "single_blocks.{}.linear1.weight": {"key_list": [
                ["transformer.single_transformer_blocks.{}.attn.to_q.lora_A.weight",
                 "transformer.single_transformer_blocks.{}.attn.to_q.lora_B.weight"],
                ["transformer.single_transformer_blocks.{}.attn.to_k.lora_A.weight",
                 "transformer.single_transformer_blocks.{}.attn.to_k.lora_B.weight"],
                ["transformer.single_transformer_blocks.{}.attn.to_v.lora_A.weight",
                 "transformer.single_transformer_blocks.{}.attn.to_v.lora_B.weight"],
                ["transformer.single_transformer_blocks.{}.proj_mlp.lora_A.weight",
                 "transformer.single_transformer_blocks.{}.proj_mlp.lora_B.weight"]
            ], "num": 38},
            "single_blocks.{}.modulation.lin.weight": {"key_list": [
                ["transformer.single_transformer_blocks.{}.norm.linear.lora_A.weight",
                 "transformer.single_transformer_blocks.{}.norm.linear.lora_B.weight"],
            ], "num": 38},
            "single_blocks.{}.linear2.weight": {"key_list": [
                ["transformer.single_transformer_blocks.{}.proj_out.lora_A.weight",
                 "transformer.single_transformer_blocks.{}.proj_out.lora_B.weight"],
            ], "num": 38},
            "double_blocks.{}.txt_attn.qkv.weight": {"key_list": [
                ["transformer.transformer_blocks.{}.attn.add_q_proj.lora_A.weight",
                 "transformer.transformer_blocks.{}.attn.add_q_proj.lora_B.weight"],
                ["transformer.transformer_blocks.{}.attn.add_k_proj.lora_A.weight",
                 "transformer.transformer_blocks.{}.attn.add_k_proj.lora_B.weight"],
                ["transformer.transformer_blocks.{}.attn.add_v_proj.lora_A.weight",
                 "transformer.transformer_blocks.{}.attn.add_v_proj.lora_B.weight"],
            ], "num": 19},
            "double_blocks.{}.img_attn.qkv.weight": {"key_list": [
                ["transformer.transformer_blocks.{}.attn.to_q.lora_A.weight",
                 "transformer.transformer_blocks.{}.attn.to_q.lora_B.weight"],
                ["transformer.transformer_blocks.{}.attn.to_k.lora_A.weight",
                 "transformer.transformer_blocks.{}.attn.to_k.lora_B.weight"],
                ["transformer.transformer_blocks.{}.attn.to_v.lora_A.weight",
                 "transformer.transformer_blocks.{}.attn.to_v.lora_B.weight"],
            ], "num": 19},
            "double_blocks.{}.img_attn.proj.weight": {"key_list": [
                ["transformer.transformer_blocks.{}.attn.to_out.0.lora_A.weight",
                 "transformer.transformer_blocks.{}.attn.to_out.0.lora_B.weight"]
            ], "num": 19},
            "double_blocks.{}.txt_attn.proj.weight": {"key_list": [
                ["transformer.transformer_blocks.{}.attn.to_add_out.lora_A.weight",
                 "transformer.transformer_blocks.{}.attn.to_add_out.lora_B.weight"]
            ], "num": 19},
            "double_blocks.{}.img_mlp.0.weight": {"key_list": [
                ["transformer.transformer_blocks.{}.ff.net.0.proj.lora_A.weight",
                 "transformer.transformer_blocks.{}.ff.net.0.proj.lora_B.weight"]
            ], "num": 19},
            "double_blocks.{}.img_mlp.2.weight": {"key_list": [
                ["transformer.transformer_blocks.{}.ff.net.2.lora_A.weight",
                 "transformer.transformer_blocks.{}.ff.net.2.lora_B.weight"]
            ], "num": 19},
            "double_blocks.{}.txt_mlp.0.weight": {"key_list": [
                ["transformer.transformer_blocks.{}.ff_context.net.0.proj.lora_A.weight",
                 "transformer.transformer_blocks.{}.ff_context.net.0.proj.lora_B.weight"]
            ], "num": 19},
            "double_blocks.{}.txt_mlp.2.weight": {"key_list": [
                ["transformer.transformer_blocks.{}.ff_context.net.2.lora_A.weight",
                 "transformer.transformer_blocks.{}.ff_context.net.2.lora_B.weight"]
            ], "num": 19},
            "double_blocks.{}.img_mod.lin.weight": {"key_list": [
                ["transformer.transformer_blocks.{}.norm1.linear.lora_A.weight",
                 "transformer.transformer_blocks.{}.norm1.linear.lora_B.weight"]
            ], "num": 19},
            "double_blocks.{}.txt_mod.lin.weight": {"key_list": [
                ["transformer.transformer_blocks.{}.norm1_context.linear.lora_A.weight",
                 "transformer.transformer_blocks.{}.norm1_context.linear.lora_B.weight"]
            ], "num": 19}
        }
        for k, v in key_map.items():
            key_list = v["key_list"]
            block_num = v["num"]
            for block_id in range(block_num):
                current_weight_list = []
                for k_list in key_list:
                    current_weight = torch.matmul(lora_sd[k_list[0].format(block_id)].permute(1, 0),
                                                  lora_sd[k_list[1].format(block_id)].permute(1, 0)).permute(1, 0)
                    current_weight_list.append(current_weight)
                current_weight = torch.cat(current_weight_list, dim=0)
                ori_sd[k.format(block_id)] += scale*current_weight
        return ori_sd

    def merge_swift_lora(self, ori_sd, lora_sd, scale = 1.0):
        have_lora_keys = {}
        for k, v in lora_sd.items():
            k = k[len("model."):] if k.startswith("model.") else k
            ori_key = k.split("lora")[0] + "weight"
            if ori_key not in ori_sd:
                raise f"{ori_key} should in the original statedict"
            if ori_key not in have_lora_keys:
                have_lora_keys[ori_key] = {}
            if "lora_A" in k:
                have_lora_keys[ori_key]["lora_A"] = v
            elif "lora_B" in k:
                have_lora_keys[ori_key]["lora_B"] = v
            else:
                raise NotImplementedError
        for key, v in have_lora_keys.items():
            current_weight = torch.matmul(v["lora_A"].permute(1, 0), v["lora_B"].permute(1, 0)).permute(1, 0)
            ori_sd[key] += scale * current_weight
        return ori_sd


    def load_pretrained_model(self, pretrained_model):
        if next(self.parameters()).device.type == 'meta':
            map_location = we.device_id
        else:
            map_location = "cpu"
        if self.lora_model is not None:
            map_location = we.device_id
        if pretrained_model is not None:
            with FS.get_from(pretrained_model, wait_finish=True) as local_model:
                if local_model.endswith('safetensors'):
                    from safetensors.torch import load_file as load_safetensors
                    sd = load_safetensors(local_model, device=map_location)
                else:
                    sd = torch.load(local_model, map_location=map_location)
            if "state_dict" in sd:
                sd = sd["state_dict"]
            if "model" in sd:
                sd = sd["model"]["model"]

            if self.lora_model is not None:
                with FS.get_from(self.lora_model, wait_finish=True) as local_model:
                    if local_model.endswith('safetensors'):
                        from safetensors.torch import load_file as load_safetensors
                        lora_sd = load_safetensors(local_model, device=map_location)
                    else:
                        lora_sd = torch.load(local_model, map_location=map_location)
                sd = self.merge_diffuser_lora(sd, lora_sd)
            if self.swift_lora_model is not None:
                with FS.get_from(self.swift_lora_model, wait_finish=True) as local_model:
                    if local_model.endswith('safetensors'):
                        from safetensors.torch import load_file as load_safetensors
                        lora_sd = load_safetensors(local_model, device=map_location)
                    else:
                        lora_sd = torch.load(local_model, map_location=map_location)
                sd = self.merge_swift_lora(sd, lora_sd)

            adapter_ckpt = {}
            if self.pretrain_adapter is not None:
                with FS.get_from(self.pretrain_adapter, wait_finish=True) as local_adapter:
                    if local_model.endswith('safetensors'):
                        from safetensors.torch import load_file as load_safetensors
                        adapter_ckpt = load_safetensors(local_adapter, device=map_location)
                    else:
                        adapter_ckpt = torch.load(local_adapter, map_location=map_location)
            sd.update(adapter_ckpt)


            new_ckpt = OrderedDict()
            for k, v in sd.items():
                if k in ("img_in.weight"):
                    model_p = self.state_dict()[k]
                    if v.shape != model_p.shape:
                        model_p.zero_()
                        model_p[:, :64].copy_(v[:, :64])
                        new_ckpt[k] = torch.nn.parameter.Parameter(model_p)
                    else:
                        new_ckpt[k] = v
                else:
                    new_ckpt[k] = v


            missing, unexpected = self.load_state_dict(new_ckpt, strict=False, assign=True)
            self.logger.info(
                f'Restored from {pretrained_model} with {len(missing)} missing and {len(unexpected)} unexpected keys'
            )
            if len(missing) > 0:
                self.logger.info(f'Missing Keys:\n {missing}')
            if len(unexpected) > 0:
                self.logger.info(f'\nUnexpected Keys:\n {unexpected}')

    def forward(
        self,
        x: Tensor,
        t: Tensor,
        cond: dict = {},
        guidance: Tensor | None = None,
        gc_seg: int = 0
    ) -> Tensor:
        x, x_ids, txt, txt_ids, y, h, w = self.prepare_input(x, cond["context"], cond["y"])
        # running on sequences img
        x = self.img_in(x)
        vec = self.time_in(timestep_embedding(t, 256))
        if self.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, x_ids), dim=1)
        pe = self.pe_embedder(ids)
        kwargs = dict(
            vec=vec,
            pe=pe,
            txt_length=txt.shape[1],
        )
        x = torch.cat((txt, x), 1)
        if self.use_grad_checkpoint and gc_seg >= 0:
            x = checkpoint_sequential(
                functions=[partial(block, **kwargs) for block in self.double_blocks],
                segments=gc_seg if gc_seg > 0 else len(self.double_blocks),
                input=x,
                use_reentrant=False
            )
        else:
            for block in self.double_blocks:
                x = block(x, **kwargs)

        kwargs = dict(
            vec=vec,
            pe=pe,
        )

        if self.use_grad_checkpoint and gc_seg >= 0:
            x = checkpoint_sequential(
                functions=[partial(block, **kwargs) for block in self.single_blocks],
                segments=gc_seg if gc_seg > 0 else len(self.single_blocks),
                input=x,
                use_reentrant=False
            )
        else:
            for block in self.single_blocks:
                x = block(x, **kwargs)
        x = x[:, txt.shape[1] :, ...]
        x = self.final_layer(x, vec)  # (N, T, patch_size ** 2 * out_channels) 6 64 64
        x = self.unpack(x, h, w)
        return x

    @staticmethod
    def get_config_template():
        return dict_to_yaml('MODEL',
                            __class__.__name__,
                            Flux.para_dict,
                            set_name=True)