lyrasd_model/lyrasdxl_pipeline_base.py

import inspect
import os
import time
from typing import Any, Callable, Dict, List, Optional, Union, Tuple

import gc
import torch
import numpy as np
from glob import glob

from diffusers import StableDiffusionXLPipeline, UNet2DConditionModel
from diffusers.loaders import TextualInversionLoaderMixin
from diffusers.image_processor import VaeImageProcessor
from diffusers.models import AutoencoderKL
from diffusers.schedulers import (DPMSolverMultistepScheduler,
                                  EulerAncestralDiscreteScheduler,
                                  EulerDiscreteScheduler,
                                  KarrasDiffusionSchedulers)
from diffusers.models.embeddings import TimestepEmbedding, Timesteps
from diffusers.utils.torch_utils import randn_tensor
from diffusers.utils import logging
from PIL import Image
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPTextModelWithProjection
from .lyrasd_vae_model import LyraSdVaeModel
from .module.lyrasd_ip_adapter import LyraIPAdapter
from .lora_util import add_text_lora_layer, add_xltext_lora_layer, add_lora_to_opt_model, load_state_dict
from safetensors.torch import load_file


class LyraSDXLPipelineBase(TextualInversionLoaderMixin):
    def __init__(self, device=torch.device("cuda"), dtype=torch.float16, num_channels_unet=4, num_channels_latents=4, vae_scale_factor=8, vae_scaling_factor=0.13025) -> None:
        self.device = device
        self.dtype = dtype

        self.num_channels_unet = num_channels_unet
        self.num_channels_latents = num_channels_latents
        self.vae_scale_factor = vae_scale_factor
        self.vae_scaling_factor = vae_scaling_factor

        self.unet_cache = {}
        self.unet_in_channels = 4

        self.controlnet_cache = {}
        self.controlnet_add_embedding = {}

        self.loaded_lora = {}
        self.loaded_lora_strength = {}

        self.scheduler = None

        self.init_pipe()

    def init_pipe(self):
        self.vae = LyraSdVaeModel(
            scale_factor=self.vae_scale_factor, scaling_factor=self.vae_scaling_factor, is_upcast=True)

        self.unet = torch.classes.lyrasd.XLUnet2dConditionalModelOp(
            "fp16",
            self.num_channels_unet,
            self.num_channels_latents)

        self.default_sample_size = 128
        self.addition_time_embed_dim = 256
        flip_sin_to_cos, freq_shift = True, 0
        self.projection_class_embeddings_input_dim, self.time_embed_dim = 2816, 1280

        self.add_time_proj = Timesteps(
            self.addition_time_embed_dim, flip_sin_to_cos, freq_shift).to(self.dtype).to(self.device)

        self.add_embedding = TimestepEmbedding(
            self.projection_class_embeddings_input_dim, self.time_embed_dim).to(self.dtype).to(self.device)

        self.image_processor = VaeImageProcessor(
            vae_scale_factor=self.vae_scale_factor)

        self.mask_processor = VaeImageProcessor(
            vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True
        )

        self.control_image_processor = VaeImageProcessor(
            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
        )

        self.feature_extractor = CLIPImageProcessor()

    def reload_pipe(self, model_path):
        self.tokenizer = CLIPTokenizer.from_pretrained(
            model_path, subfolder="tokenizer")
        self.text_encoder = CLIPTextModel.from_pretrained(
            model_path, subfolder="text_encoder").to(self.dtype).to(self.device)

        self.tokenizer_2 = CLIPTokenizer.from_pretrained(
            model_path, subfolder="tokenizer_2")
        self.text_encoder_2 = CLIPTextModelWithProjection.from_pretrained(
            model_path, subfolder="text_encoder_2").to(self.dtype).to(self.device)
        
        self.reload_unet_model_v2(model_path)
        self.reload_vae_model_v2(model_path)

        if not self.scheduler:
            self.scheduler = EulerAncestralDiscreteScheduler.from_pretrained(
                model_path, subfolder="scheduler")

    def load_embedding_weight(self, model, weight_path, unet_file_format="fp16"):
        bin_list = glob(weight_path)
        sate_dicts = model.state_dict()
        dtype = np.float32 if unet_file_format == "fp32" else np.float16
        for bin_file in bin_list:
            weight = torch.from_numpy(np.fromfile(bin_file, dtype=dtype)).to(
                self.dtype).to(self.device)
            key = '.'.join(os.path.basename(bin_file).split('.')[1:-1])
            weight = weight.reshape(sate_dicts[key].shape)
            sate_dicts.update({key: weight})
        model.load_state_dict(sate_dicts)

    @property
    def _execution_device(self):
        if not hasattr(self.unet, "_hf_hook"):
            return self.device
        for module in self.unet.modules():
            if (
                hasattr(module, "_hf_hook")
                and hasattr(module._hf_hook, "execution_device")
                and module._hf_hook.execution_device is not None
            ):
                return torch.device(module._hf_hook.execution_device)
        return self.device

    def reload_unet_model(self, unet_path, unet_file_format='fp32'):
        if len(unet_path) > 0 and unet_path[-1] != "/":
            unet_path = unet_path + "/"
        self.unet.reload_unet_model(unet_path, unet_file_format)
        self.load_embedding_weight(
            self.add_embedding, f"{unet_path}add_embedding*", unet_file_format=unet_file_format)

    def reload_vae_model(self, vae_path, vae_file_format='fp32'):
        if len(vae_path) > 0 and vae_path[-1] != "/":
            vae_path = vae_path + "/"
        return self.vae.reload_vae_model(vae_path, vae_file_format)

    def load_lora(self, lora_model_path, lora_name, lora_strength, lora_file_format='fp32'):
        if len(lora_model_path) > 0 and lora_model_path[-1] != "/":
            lora_model_path = lora_model_path + "/"
        lora = add_xltext_lora_layer(
            self.text_encoder, self.text_encoder_2, lora_model_path, lora_strength, lora_file_format)

        self.loaded_lora[lora_name] = lora
        self.unet.load_lora(lora_model_path, lora_name,
                            lora_strength, lora_file_format)

    def unload_lora(self, lora_name, clean_cache=False):
        for layer_data in self.loaded_lora[lora_name]:
            layer = layer_data['layer']
            added_weight = layer_data['added_weight']
            layer.weight.data -= added_weight
        self.unet.unload_lora(lora_name, clean_cache)
        del self.loaded_lora[lora_name]
        gc.collect()
        torch.cuda.empty_cache()

    def load_lora_v2(self, lora_model_path, lora_name, lora_strength):
        if lora_name in self.loaded_lora:
            state_dict = self.loaded_lora[lora_name]
        else:
            state_dict = load_state_dict(lora_model_path)
            self.loaded_lora[lora_name] = state_dict
        self.loaded_lora_strength[lora_name] = lora_strength
        add_lora_to_opt_model(state_dict, self.unet, self.text_encoder,
                              self.text_encoder_2, lora_strength)

    def unload_lora_v2(self, lora_name, clean_cache=False):
        state_dict = self.loaded_lora[lora_name]
        lora_strength = self.loaded_lora_strength[lora_name]
        add_lora_to_opt_model(state_dict, self.unet, self.text_encoder,
                              self.text_encoder_2,  -1.0 * lora_strength)
        del self.loaded_lora_strength[lora_name]

        if clean_cache:
            del self.loaded_lora[lora_name]
            gc.collect()
            torch.cuda.empty_cache()

    def clean_lora_cache(self):
        self.unet.clean_lora_cache()

    def get_loaded_lora(self):
        return self.unet.get_loaded_lora()

    def _get_aug_emb(self, time_ids, text_embeds, dtype):
        time_embeds = self.add_time_proj(time_ids.flatten())
        time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
        add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
        add_embeds = add_embeds.to(dtype)
        aug_emb = self.add_embedding(add_embeds)
        return aug_emb

    def load_ip_adapter(self, dir_ip_adapter, ip_plus, image_encoder_path, num_ip_tokens, ip_projection_dim,  dir_face_in=None, num_fp_tokens=1, fp_projection_dim=None, sdxl=True):
        self.ip_adapter_helper = LyraIPAdapter(self, sdxl, "cuda", dir_ip_adapter, ip_plus, image_encoder_path,
                                               num_ip_tokens, ip_projection_dim, dir_face_in, num_fp_tokens, fp_projection_dim)

    def reload_unet_model_v2(self, model_path):
        checkpoint_file = os.path.join(
            model_path, "unet/diffusion_pytorch_model.bin")
        if not os.path.exists(checkpoint_file):
            checkpoint_file = os.path.join(
                model_path, "unet/diffusion_pytorch_model.safetensors")
        if checkpoint_file in self.unet_cache:
            state_dict = self.unet_cache[checkpoint_file]
        else:
            if "safetensors" in checkpoint_file:
                state_dict = load_file(checkpoint_file)
            else:
                state_dict = torch.load(checkpoint_file, map_location="cpu")

            for key in state_dict:
                if len(state_dict[key].shape) == 4:
                    # converted_unet_checkpoint[key] = converted_unet_checkpoint[key].to(torch.float16).to("cuda").permute(0,2,3,1).contiguous().cpu()
                    state_dict[key] = state_dict[key].to(
                        torch.float16).permute(0, 2, 3, 1).contiguous()
                state_dict[key] = state_dict[key].to(torch.float16)
            self.unet_cache[checkpoint_file] = state_dict

        self.unet.reload_unet_model_from_cache(state_dict, "cpu")
        self.load_embedding_weight_v2(self.add_embedding, state_dict)

    def load_embedding_weight_v2(self, model, state_dict):
        sub_state_dict = {}
        for k in state_dict:
            if k.startswith("add_embedding"):
                v = state_dict[k]
                sub_k = ".".join(k.split(".")[1:])
                sub_state_dict[sub_k] = v

        model.load_state_dict(sub_state_dict)

    def reload_vae_model_v2(self, model_path):
        self.vae.reload_vae_model_v2(model_path)

    def load_controlnet_model_v2(self, model_name, controlnet_path):
        checkpoint_file = os.path.join(
            controlnet_path, "diffusion_pytorch_model.bin")
        if not os.path.exists(checkpoint_file):
            checkpoint_file = os.path.join(
                controlnet_path, "diffusion_pytorch_model.safetensors")
        if checkpoint_file in self.controlnet_cache:
            state_dict = self.controlnet_cache[checkpoint_file]
        else:
            if "safetensors" in checkpoint_file:
                state_dict = load_file(checkpoint_file)
            else:
                state_dict = torch.load(checkpoint_file, map_location="cpu")

            for key in state_dict:
                if len(state_dict[key].shape) == 4:
                    # converted_unet_checkpoint[key] = converted_unet_checkpoint[key].to(torch.float16).to("cuda").permute(0,2,3,1).contiguous().cpu()
                    state_dict[key] = state_dict[key].to(
                        torch.float16).permute(0, 2, 3, 1).contiguous()
                state_dict[key] = state_dict[key].to(torch.float16)
            self.controlnet_cache[checkpoint_file] = state_dict

        self.unet.load_controlnet_model_from_state_dict(
            model_name, state_dict, "cpu")

        add_embedding = TimestepEmbedding(
            self.projection_class_embeddings_input_dim, self.time_embed_dim).to(self.dtype).to(self.device)

        self.load_embedding_weight_v2(add_embedding, state_dict)
        self.controlnet_add_embedding[model_name] = add_embedding

    def unload_controlnet_model(self, model_name):
        self.unet.unload_controlnet_model(model_name, True)
        del self.controlnet_add_embedding[model_name]

    def get_loaded_controlnet(self):
        return self.unet.get_loaded_controlnet()