backend/diffusion_engine/sdxl.py

import safetensors.torch as sf
import torch
from huggingface_guess import model_list

from backend import memory_management, utils
from backend.args import dynamic_args
from backend.diffusion_engine.base import ForgeDiffusionEngine, ForgeObjects
from backend.nn.unet import Timestep
from backend.patcher.clip import CLIP
from backend.patcher.unet import UnetPatcher
from backend.patcher.vae import VAE
from backend.text_processing.classic_engine import ClassicTextProcessingEngine
from modules.shared import opts


class StableDiffusionXL(ForgeDiffusionEngine):
    matched_guesses = [model_list.SDXL]

    def __init__(self, estimated_config, huggingface_components):
        super().__init__(estimated_config, huggingface_components)

        clip = CLIP(model_dict={"clip_l": huggingface_components["text_encoder"], "clip_g": huggingface_components["text_encoder_2"]}, tokenizer_dict={"clip_l": huggingface_components["tokenizer"], "clip_g": huggingface_components["tokenizer_2"]})

        vae = VAE(model=huggingface_components["vae"])

        unet = UnetPatcher.from_model(model=huggingface_components["unet"], diffusers_scheduler=huggingface_components["scheduler"], config=estimated_config)

        self.text_processing_engine_l = ClassicTextProcessingEngine(
            text_encoder=clip.cond_stage_model.clip_l,
            tokenizer=clip.tokenizer.clip_l,
            embedding_dir=dynamic_args["embedding_dir"],
            embedding_key="clip_l",
            embedding_expected_shape=2048,
            text_projection=False,
            minimal_clip_skip=2,
            clip_skip=2,
            return_pooled=False,
            final_layer_norm=False,
        )

        self.text_processing_engine_g = ClassicTextProcessingEngine(
            text_encoder=clip.cond_stage_model.clip_g,
            tokenizer=clip.tokenizer.clip_g,
            embedding_dir=dynamic_args["embedding_dir"],
            embedding_key="clip_g",
            embedding_expected_shape=2048,
            text_projection=True,
            minimal_clip_skip=2,
            clip_skip=2,
            return_pooled=True,
            final_layer_norm=False,
        )

        self.embedder = Timestep(256)

        self.forge_objects = ForgeObjects(unet=unet, clip=clip, vae=vae, clipvision=None)
        self.forge_objects_original = self.forge_objects.shallow_copy()
        self.forge_objects_after_applying_lora = self.forge_objects.shallow_copy()

        # WebUI Legacy
        self.is_sdxl = True

    def set_clip_skip(self, clip_skip):
        self.text_processing_engine_l.clip_skip = clip_skip
        self.text_processing_engine_g.clip_skip = clip_skip

    @torch.inference_mode()
    def get_learned_conditioning(self, prompt: list[str]):
        memory_management.load_model_gpu(self.forge_objects.clip.patcher)

        cond_l = self.text_processing_engine_l(prompt)
        cond_g, clip_pooled = self.text_processing_engine_g(prompt)

        width = getattr(prompt, "width", 1024) or 1024
        height = getattr(prompt, "height", 1024) or 1024
        is_negative_prompt = getattr(prompt, "is_negative_prompt", False)

        crop_w = opts.sdxl_crop_left
        crop_h = opts.sdxl_crop_top
        target_width = width
        target_height = height

        out = [self.embedder(torch.Tensor([height])), self.embedder(torch.Tensor([width])), self.embedder(torch.Tensor([crop_h])), self.embedder(torch.Tensor([crop_w])), self.embedder(torch.Tensor([target_height])), self.embedder(torch.Tensor([target_width]))]

        flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0).repeat(clip_pooled.shape[0], 1).to(clip_pooled)

        force_zero_negative_prompt = is_negative_prompt and all(x == "" for x in prompt)

        if force_zero_negative_prompt:
            clip_pooled = torch.zeros_like(clip_pooled)
            cond_l = torch.zeros_like(cond_l)
            cond_g = torch.zeros_like(cond_g)

        # Ensure cond_l and cond_g have the same size
        max_len = max(cond_l.shape[1], cond_g.shape[1])
        cond_l = torch.cat([cond_l, cond_l.new_zeros(cond_l.size(0), max_len - cond_l.shape[1], cond_l.size(2))], dim=1)
        cond_g = torch.cat([cond_g, cond_g.new_zeros(cond_g.size(0), max_len - cond_g.shape[1], cond_g.size(2))], dim=1)

        cond = dict(
            crossattn=torch.cat([cond_l, cond_g], dim=2),
            vector=torch.cat([clip_pooled, flat], dim=1),
        )

        return cond

    @torch.inference_mode()
    def get_prompt_lengths_on_ui(self, prompt):
        _, token_count = self.text_processing_engine_l.process_texts([prompt])
        return token_count, self.text_processing_engine_l.get_target_prompt_token_count(token_count)

    @torch.inference_mode()
    def encode_first_stage(self, x):
        sample = self.forge_objects.vae.encode(x.movedim(1, -1) * 0.5 + 0.5)
        sample = self.forge_objects.vae.first_stage_model.process_in(sample)
        return sample.to(x)

    @torch.inference_mode()
    def decode_first_stage(self, x):
        sample = self.forge_objects.vae.first_stage_model.process_out(x)
        sample = self.forge_objects.vae.decode(sample).movedim(-1, 1) * 2.0 - 1.0
        return sample.to(x)

    def save_checkpoint(self, filename):
        sd = {}
        sd.update(utils.get_state_dict_after_quant(self.forge_objects.unet.model.diffusion_model, prefix="model.diffusion_model."))
        sd.update(model_list.SDXL.process_clip_state_dict_for_saving(self, utils.get_state_dict_after_quant(self.forge_objects.clip.cond_stage_model, prefix="")))
        sd.update(utils.get_state_dict_after_quant(self.forge_objects.vae.first_stage_model, prefix="first_stage_model."))
        sf.save_file(sd, filename)
        return filename


class StableDiffusionXLRefiner(ForgeDiffusionEngine):
    matched_guesses = [model_list.SDXLRefiner]

    def __init__(self, estimated_config, huggingface_components):
        super().__init__(estimated_config, huggingface_components)

        clip = CLIP(
            model_dict={"clip_g": huggingface_components["text_encoder"]},
            tokenizer_dict={
                "clip_g": huggingface_components["tokenizer"],
            },
        )

        vae = VAE(model=huggingface_components["vae"])

        unet = UnetPatcher.from_model(model=huggingface_components["unet"], diffusers_scheduler=huggingface_components["scheduler"], config=estimated_config)

        self.text_processing_engine_g = ClassicTextProcessingEngine(
            text_encoder=clip.cond_stage_model.clip_g,
            tokenizer=clip.tokenizer.clip_g,
            embedding_dir=dynamic_args["embedding_dir"],
            embedding_key="clip_g",
            embedding_expected_shape=2048,
            text_projection=True,
            minimal_clip_skip=2,
            clip_skip=2,
            return_pooled=True,
            final_layer_norm=False,
        )

        self.embedder = Timestep(256)

        self.forge_objects = ForgeObjects(unet=unet, clip=clip, vae=vae, clipvision=None)
        self.forge_objects_original = self.forge_objects.shallow_copy()
        self.forge_objects_after_applying_lora = self.forge_objects.shallow_copy()

        # WebUI Legacy
        self.is_sdxl = True

    def set_clip_skip(self, clip_skip):
        self.text_processing_engine_g.clip_skip = clip_skip

    @torch.inference_mode()
    def get_learned_conditioning(self, prompt: list[str]):
        memory_management.load_model_gpu(self.forge_objects.clip.patcher)

        cond_g, clip_pooled = self.text_processing_engine_g(prompt)

        width = getattr(prompt, "width", 1024) or 1024
        height = getattr(prompt, "height", 1024) or 1024
        is_negative_prompt = getattr(prompt, "is_negative_prompt", False)

        crop_w = opts.sdxl_crop_left
        crop_h = opts.sdxl_crop_top
        aesthetic = opts.sdxl_refiner_low_aesthetic_score if is_negative_prompt else opts.sdxl_refiner_high_aesthetic_score

        out = [self.embedder(torch.Tensor([height])), self.embedder(torch.Tensor([width])), self.embedder(torch.Tensor([crop_h])), self.embedder(torch.Tensor([crop_w])), self.embedder(torch.Tensor([aesthetic]))]

        flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0).repeat(clip_pooled.shape[0], 1).to(clip_pooled)

        force_zero_negative_prompt = is_negative_prompt and all(x == "" for x in prompt)

        if force_zero_negative_prompt:
            clip_pooled = torch.zeros_like(clip_pooled)
            cond_g = torch.zeros_like(cond_g)

        cond = dict(
            crossattn=cond_g,
            vector=torch.cat([clip_pooled, flat], dim=1),
        )

        return cond

    @torch.inference_mode()
    def get_prompt_lengths_on_ui(self, prompt):
        _, token_count = self.text_processing_engine_g.process_texts([prompt])
        return token_count, self.text_processing_engine_g.get_target_prompt_token_count(token_count)

    @torch.inference_mode()
    def encode_first_stage(self, x):
        sample = self.forge_objects.vae.encode(x.movedim(1, -1) * 0.5 + 0.5)
        sample = self.forge_objects.vae.first_stage_model.process_in(sample)
        return sample.to(x)

    @torch.inference_mode()
    def decode_first_stage(self, x):
        sample = self.forge_objects.vae.first_stage_model.process_out(x)
        sample = self.forge_objects.vae.decode(sample).movedim(-1, 1) * 2.0 - 1.0
        return sample.to(x)

    def save_checkpoint(self, filename):
        sd = {}
        sd.update(utils.get_state_dict_after_quant(self.forge_objects.unet.model.diffusion_model, prefix="model.diffusion_model."))
        sd.update(model_list.SDXLRefiner.process_clip_state_dict_for_saving(self, utils.get_state_dict_after_quant(self.forge_objects.clip.cond_stage_model, prefix="")))
        sd.update(utils.get_state_dict_after_quant(self.forge_objects.vae.first_stage_model, prefix="first_stage_model."))
        sf.save_file(sd, filename)
        return filename