backend/loader.py

import os
import torch
import logging
import importlib

import backend.args
import huggingface_guess

from diffusers import DiffusionPipeline
from transformers import modeling_utils

from backend import memory_management
from backend.utils import read_arbitrary_config, load_torch_file
from backend.state_dict import try_filter_state_dict, load_state_dict
from backend.operations import using_forge_operations
from backend.nn.vae import IntegratedAutoencoderKL
from backend.nn.clip import IntegratedCLIP
from backend.nn.unet import IntegratedUNet2DConditionModel

from backend.diffusion_engine.sd15 import StableDiffusion
from backend.diffusion_engine.sd20 import StableDiffusion2
from backend.diffusion_engine.sdxl import StableDiffusionXL
from backend.diffusion_engine.flux import Flux


possible_models = [StableDiffusion, StableDiffusion2, StableDiffusionXL, Flux]


logging.getLogger("diffusers").setLevel(logging.ERROR)
dir_path = os.path.dirname(__file__)


def load_huggingface_component(guess, component_name, lib_name, cls_name, repo_path, state_dict):
    config_path = os.path.join(repo_path, component_name)

    if component_name in ['feature_extractor', 'safety_checker']:
        return None

    if lib_name in ['transformers', 'diffusers']:
        if component_name in ['scheduler']:
            cls = getattr(importlib.import_module(lib_name), cls_name)
            return cls.from_pretrained(os.path.join(repo_path, component_name))
        if component_name.startswith('tokenizer'):
            cls = getattr(importlib.import_module(lib_name), cls_name)
            comp = cls.from_pretrained(os.path.join(repo_path, component_name))
            comp._eventual_warn_about_too_long_sequence = lambda *args, **kwargs: None
            return comp
        if cls_name in ['AutoencoderKL']:
            assert isinstance(state_dict, dict) and len(state_dict) > 16, 'You do not have VAE state dict!'

            config = IntegratedAutoencoderKL.load_config(config_path)

            with using_forge_operations(device=memory_management.cpu, dtype=memory_management.vae_dtype()):
                model = IntegratedAutoencoderKL.from_config(config)

            load_state_dict(model, state_dict, ignore_start='loss.')
            return model
        if component_name.startswith('text_encoder') and cls_name in ['CLIPTextModel', 'CLIPTextModelWithProjection']:
            assert isinstance(state_dict, dict) and len(state_dict) > 16, 'You do not have CLIP state dict!'

            from transformers import CLIPTextConfig, CLIPTextModel
            config = CLIPTextConfig.from_pretrained(config_path)

            to_args = dict(device=memory_management.cpu, dtype=memory_management.text_encoder_dtype())

            with modeling_utils.no_init_weights():
                with using_forge_operations(**to_args, manual_cast_enabled=True):
                    model = IntegratedCLIP(CLIPTextModel, config, add_text_projection=True).to(**to_args)

            load_state_dict(model, state_dict, ignore_errors=[
                'transformer.text_projection.weight',
                'transformer.text_model.embeddings.position_ids',
                'logit_scale'
            ], log_name=cls_name)

            return model
        if cls_name == 'T5EncoderModel':
            assert isinstance(state_dict, dict) and len(state_dict) > 16, 'You do not have T5 state dict!'

            from backend.nn.t5 import IntegratedT5
            config = read_arbitrary_config(config_path)

            dtype = memory_management.text_encoder_dtype()
            sd_dtype = memory_management.state_dict_dtype(state_dict)

            if sd_dtype in [torch.float8_e4m3fn, torch.float8_e5m2]:
                dtype = sd_dtype
                print(f'Using Detected T5 Data Type: {dtype}')
            else:
                print(f'Using Default T5 Data Type: {dtype}')

            with modeling_utils.no_init_weights():
                with using_forge_operations(device=memory_management.cpu, dtype=dtype, manual_cast_enabled=True):
                    model = IntegratedT5(config)

            load_state_dict(model, state_dict, log_name=cls_name, ignore_errors=['transformer.encoder.embed_tokens.weight', 'logit_scale'])

            return model
        if cls_name in ['UNet2DConditionModel', 'FluxTransformer2DModel']:
            assert isinstance(state_dict, dict) and len(state_dict) > 16, 'You do not have model state dict!'

            model_loader = None
            if cls_name == 'UNet2DConditionModel':
                model_loader = lambda c: IntegratedUNet2DConditionModel.from_config(c)
            if cls_name == 'FluxTransformer2DModel':
                from backend.nn.flux import IntegratedFluxTransformer2DModel
                model_loader = lambda c: IntegratedFluxTransformer2DModel(**c)

            unet_config = guess.unet_config.copy()
            state_dict_size = memory_management.state_dict_size(state_dict)
            state_dict_dtype = memory_management.state_dict_dtype(state_dict)

            storage_dtype = memory_management.unet_dtype(model_params=state_dict_size, supported_dtypes=guess.supported_inference_dtypes)

            unet_storage_dtype_overwrite = backend.args.dynamic_args.get('forge_unet_storage_dtype')

            if unet_storage_dtype_overwrite is not None:
                storage_dtype = unet_storage_dtype_overwrite
            else:
                if state_dict_dtype in [torch.float8_e4m3fn, torch.float8_e5m2, 'nf4', 'fp4', 'gguf']:
                    print(f'Using Detected UNet Type: {state_dict_dtype}')
                    storage_dtype = state_dict_dtype
                    if state_dict_dtype in ['nf4', 'fp4', 'gguf']:
                        print(f'Using pre-quant state dict!')

                        if state_dict_dtype in ['gguf']:
                            from gguf.constants import GGMLQuantizationType

                            type_counts = {}

                            for k, v in state_dict.items():
                                gguf_type = getattr(v, 'gguf_type', None)
                                if gguf_type is not None:
                                    type_name = GGMLQuantizationType(gguf_type).name
                                    if type_name in type_counts:
                                        type_counts[type_name] += 1
                                    else:
                                        type_counts[type_name] = 1

                            print(f'Using GGUF state dict: {type_counts}')

            load_device = memory_management.get_torch_device()
            computation_dtype = memory_management.get_computation_dtype(load_device, supported_dtypes=guess.supported_inference_dtypes)
            offload_device = memory_management.unet_offload_device()

            if storage_dtype in ['nf4', 'fp4', 'gguf']:
                initial_device = memory_management.unet_inital_load_device(parameters=state_dict_size, dtype=computation_dtype)
                with using_forge_operations(device=initial_device, dtype=computation_dtype, manual_cast_enabled=False, bnb_dtype=storage_dtype):
                    model = model_loader(unet_config)
            else:
                initial_device = memory_management.unet_inital_load_device(parameters=state_dict_size, dtype=storage_dtype)
                need_manual_cast = storage_dtype != computation_dtype
                to_args = dict(device=initial_device, dtype=storage_dtype)

                with using_forge_operations(**to_args, manual_cast_enabled=need_manual_cast):
                    model = model_loader(unet_config).to(**to_args)

            load_state_dict(model, state_dict)

            if hasattr(model, '_internal_dict'):
                model._internal_dict = unet_config
            else:
                model.config = unet_config

            model.storage_dtype = storage_dtype
            model.computation_dtype = computation_dtype
            model.load_device = load_device
            model.initial_device = initial_device
            model.offload_device = offload_device

            return model

    print(f'Skipped: {component_name} = {lib_name}.{cls_name}')
    return None


def replace_state_dict(sd, asd, guess):
    vae_key_prefix = guess.vae_key_prefix[0]
    text_encoder_key_prefix = guess.text_encoder_key_prefix[0]

    if "decoder.conv_in.weight" in asd:
        keys_to_delete = [k for k in sd if k.startswith(vae_key_prefix)]
        for k in keys_to_delete:
            del sd[k]
        for k, v in asd.items():
            sd[vae_key_prefix + k] = v

    if 'text_model.encoder.layers.0.layer_norm1.weight' in asd:
        keys_to_delete = [k for k in sd if k.startswith(f"{text_encoder_key_prefix}clip_l.")]
        for k in keys_to_delete:
            del sd[k]
        for k, v in asd.items():
            sd[f"{text_encoder_key_prefix}clip_l.transformer.{k}"] = v

    if 'encoder.block.0.layer.0.SelfAttention.k.weight' in asd:
        keys_to_delete = [k for k in sd if k.startswith(f"{text_encoder_key_prefix}t5xxl.")]
        for k in keys_to_delete:
            del sd[k]
        for k, v in asd.items():
            sd[f"{text_encoder_key_prefix}t5xxl.transformer.{k}"] = v

    return sd


def preprocess_state_dict(sd):
    if any("double_block" in k for k in sd.keys()):
        if not any(k.startswith("model.diffusion_model") for k in sd.keys()):
            sd = {f"model.diffusion_model.{k}": v for k, v in sd.items()}

    return sd


def split_state_dict(sd, additional_state_dicts: list = None):
    sd = load_torch_file(sd)
    sd = preprocess_state_dict(sd)
    guess = huggingface_guess.guess(sd)

    if isinstance(additional_state_dicts, list):
        for asd in additional_state_dicts:
            asd = load_torch_file(asd)
            sd = replace_state_dict(sd, asd, guess)

    guess.clip_target = guess.clip_target(sd)

    state_dict = {
        guess.unet_target: try_filter_state_dict(sd, guess.unet_key_prefix),
        guess.vae_target: try_filter_state_dict(sd, guess.vae_key_prefix)
    }

    sd = guess.process_clip_state_dict(sd)

    for k, v in guess.clip_target.items():
        state_dict[v] = try_filter_state_dict(sd, [k + '.'])

    state_dict['ignore'] = sd

    print_dict = {k: len(v) for k, v in state_dict.items()}
    print(f'StateDict Keys: {print_dict}')

    del state_dict['ignore']

    return state_dict, guess


@torch.no_grad()
def forge_loader(sd, additional_state_dicts=None):
    try:
        state_dicts, estimated_config = split_state_dict(sd, additional_state_dicts=additional_state_dicts)
    except:
        raise ValueError('Failed to recognize model type!')
    
    repo_name = estimated_config.huggingface_repo

    local_path = os.path.join(dir_path, 'huggingface', repo_name)
    config: dict = DiffusionPipeline.load_config(local_path)
    huggingface_components = {}
    for component_name, v in config.items():
        if isinstance(v, list) and len(v) == 2:
            lib_name, cls_name = v
            component_sd = state_dicts.get(component_name, None)
            component = load_huggingface_component(estimated_config, component_name, lib_name, cls_name, local_path, component_sd)
            if component_sd is not None:
                del state_dicts[component_name]
            if component is not None:
                huggingface_components[component_name] = component

    for M in possible_models:
        if any(isinstance(estimated_config, x) for x in M.matched_guesses):
            return M(estimated_config=estimated_config, huggingface_components=huggingface_components)

    print('Failed to recognize model type!')
    return None