finetrainers/models/cogview4/base_specification.py

import os
from typing import Any, Dict, List, Optional, Tuple

import torch
from accelerate import init_empty_weights
from diffusers import (
    AutoencoderKL,
    CogView4Pipeline,
    CogView4Transformer2DModel,
    FlowMatchEulerDiscreteScheduler,
)
from diffusers.models.autoencoders.vae import DiagonalGaussianDistribution
from transformers import AutoTokenizer, GlmModel

from ... import data
from ... import functional as FF
from ...logging import get_logger
from ...processors import CogView4GLMProcessor, ProcessorMixin
from ...typing import ArtifactType, SchedulerType
from ...utils import get_non_null_items
from ..modeling_utils import ModelSpecification


logger = get_logger()


class CogView4LatentEncodeProcessor(ProcessorMixin):
    r"""
    Processor to encode image/video into latents using the LTX VAE.

    Args:
        output_names (`List[str]`):
            The names of the outputs that the processor returns. The outputs are in the following order:
            - latents: The latents of the input image/video.
            - original_size: The original size of the input image/video.
            - target_size: The target size of the input image/video.
            - crop_coords: The top-left crop coordinates of the input image/video.
    """

    def __init__(self, output_names: List[str]):
        super().__init__()

        self.output_names = output_names
        assert len(self.output_names) == 4

    def forward(
        self,
        vae: AutoencoderKL,
        image: Optional[torch.Tensor] = None,
        video: Optional[torch.Tensor] = None,
        generator: Optional[torch.Generator] = None,
        compute_posterior: bool = True,
        _original_height: Optional[int] = None,
        _original_width: Optional[int] = None,
    ) -> Dict[str, torch.Tensor]:
        device = vae.device
        dtype = vae.dtype

        if video is not None:
            # TODO(aryan): perhaps better would be to flatten(0, 1), but need to account for reshaping sigmas accordingly
            image = video[:, 0]  # [B, F, C, H, W] -> [B, 1, C, H, W]

        assert image.ndim == 4, f"Expected 4D tensor, got {image.ndim}D tensor"
        image = image.to(device=device, dtype=vae.dtype)

        if compute_posterior:
            latents = vae.encode(image).latent_dist.sample(generator=generator)
            latents = latents.to(dtype=dtype)
        else:
            if vae.use_slicing and image.shape[0] > 1:
                encoded_slices = [vae._encode(x_slice) for x_slice in image.split(1)]
                moments = torch.cat(encoded_slices)
            else:
                moments = vae._encode(image)
            latents = moments.to(dtype=dtype)

        batch_size = latents.size(0)
        target_height = image.size(2)
        target_width = image.size(3)
        original_size = torch.tensor([(_original_height, _original_width)], device=device, dtype=dtype).repeat(
            batch_size, 1
        )
        target_size = torch.tensor([(target_height, target_width)], device=device, dtype=dtype).repeat(batch_size, 1)
        crop_coords = torch.tensor([(0, 0)], device=device, dtype=dtype).repeat(batch_size, 1)

        return {
            self.output_names[0]: latents,
            self.output_names[1]: original_size,
            self.output_names[2]: target_size,
            self.output_names[3]: crop_coords,
        }


class CogView4ModelSpecification(ModelSpecification):
    def __init__(
        self,
        pretrained_model_name_or_path: str = "THUDM/CogView4-6B",
        tokenizer_id: Optional[str] = None,
        text_encoder_id: Optional[str] = None,
        transformer_id: Optional[str] = None,
        vae_id: Optional[str] = None,
        text_encoder_dtype: torch.dtype = torch.bfloat16,
        transformer_dtype: torch.dtype = torch.bfloat16,
        vae_dtype: torch.dtype = torch.bfloat16,
        revision: Optional[str] = None,
        cache_dir: Optional[str] = None,
        condition_model_processors: List[ProcessorMixin] = None,
        latent_model_processors: List[ProcessorMixin] = None,
        **kwargs,
    ) -> None:
        super().__init__(
            pretrained_model_name_or_path=pretrained_model_name_or_path,
            tokenizer_id=tokenizer_id,
            text_encoder_id=text_encoder_id,
            transformer_id=transformer_id,
            vae_id=vae_id,
            text_encoder_dtype=text_encoder_dtype,
            transformer_dtype=transformer_dtype,
            vae_dtype=vae_dtype,
            revision=revision,
            cache_dir=cache_dir,
        )

        if condition_model_processors is None:
            condition_model_processors = [CogView4GLMProcessor(["encoder_hidden_states"])]
        if latent_model_processors is None:
            latent_model_processors = [
                CogView4LatentEncodeProcessor(["latents", "original_size", "target_size", "crop_coords"])
            ]

        self.condition_model_processors = condition_model_processors
        self.latent_model_processors = latent_model_processors

    @property
    def _resolution_dim_keys(self):
        return {"latents": (2, 3)}

    def load_condition_models(self) -> Dict[str, torch.nn.Module]:
        if self.tokenizer_id is not None:
            tokenizer = AutoTokenizer.from_pretrained(
                self.tokenizer_id, revision=self.revision, cache_dir=self.cache_dir
            )
        else:
            tokenizer = AutoTokenizer.from_pretrained(
                self.pretrained_model_name_or_path,
                subfolder="tokenizer",
                revision=self.revision,
                cache_dir=self.cache_dir,
            )

        if self.text_encoder_id is not None:
            text_encoder = GlmModel.from_pretrained(
                self.text_encoder_id,
                torch_dtype=self.text_encoder_dtype,
                revision=self.revision,
                cache_dir=self.cache_dir,
            )
        else:
            text_encoder = GlmModel.from_pretrained(
                self.pretrained_model_name_or_path,
                subfolder="text_encoder",
                torch_dtype=self.text_encoder_dtype,
                revision=self.revision,
                cache_dir=self.cache_dir,
            )

        return {"tokenizer": tokenizer, "text_encoder": text_encoder}

    def load_latent_models(self) -> Dict[str, torch.nn.Module]:
        if self.vae_id is not None:
            vae = AutoencoderKL.from_pretrained(
                self.vae_id,
                torch_dtype=self.vae_dtype,
                revision=self.revision,
                cache_dir=self.cache_dir,
            )
        else:
            vae = AutoencoderKL.from_pretrained(
                self.pretrained_model_name_or_path,
                subfolder="vae",
                torch_dtype=self.vae_dtype,
                revision=self.revision,
                cache_dir=self.cache_dir,
            )

        return {"vae": vae}

    def load_diffusion_models(self) -> Dict[str, torch.nn.Module]:
        if self.transformer_id is not None:
            transformer = CogView4Transformer2DModel.from_pretrained(
                self.transformer_id,
                torch_dtype=self.transformer_dtype,
                revision=self.revision,
                cache_dir=self.cache_dir,
            )
        else:
            transformer = CogView4Transformer2DModel.from_pretrained(
                self.pretrained_model_name_or_path,
                subfolder="transformer",
                torch_dtype=self.transformer_dtype,
                revision=self.revision,
                cache_dir=self.cache_dir,
            )

        scheduler = FlowMatchEulerDiscreteScheduler()

        return {"transformer": transformer, "scheduler": scheduler}

    def load_pipeline(
        self,
        tokenizer: Optional[AutoTokenizer] = None,
        text_encoder: Optional[GlmModel] = None,
        transformer: Optional[CogView4Transformer2DModel] = None,
        vae: Optional[AutoencoderKL] = None,
        scheduler: Optional[FlowMatchEulerDiscreteScheduler] = None,
        enable_slicing: bool = False,
        enable_tiling: bool = False,
        enable_model_cpu_offload: bool = False,
        training: bool = False,
        **kwargs,
    ) -> CogView4Pipeline:
        components = {
            "tokenizer": tokenizer,
            "text_encoder": text_encoder,
            "transformer": transformer,
            "vae": vae,
            # Load the scheduler based on CogView4's config instead of using the default initialization being used for training
            # "scheduler": scheduler,
        }
        components = get_non_null_items(components)

        pipe = CogView4Pipeline.from_pretrained(
            self.pretrained_model_name_or_path, **components, revision=self.revision, cache_dir=self.cache_dir
        )
        pipe.text_encoder.to(self.text_encoder_dtype)
        pipe.vae.to(self.vae_dtype)

        if not training:
            pipe.transformer.to(self.transformer_dtype)

        if enable_slicing:
            pipe.vae.enable_slicing()
        if enable_tiling:
            pipe.vae.enable_tiling()
        if enable_model_cpu_offload:
            pipe.enable_model_cpu_offload()

        return pipe

    @torch.no_grad()
    def prepare_conditions(
        self,
        tokenizer: AutoTokenizer,
        text_encoder: GlmModel,
        caption: str,
        max_sequence_length: int = 1024,
        **kwargs,
    ) -> Dict[str, Any]:
        conditions = {
            "tokenizer": tokenizer,
            "text_encoder": text_encoder,
            "caption": caption,
            "max_sequence_length": max_sequence_length,
            **kwargs,
        }
        input_keys = set(conditions.keys())
        conditions = super().prepare_conditions(**conditions)
        conditions = {k: v for k, v in conditions.items() if k not in input_keys}
        return conditions

    @torch.no_grad()
    def prepare_latents(
        self,
        vae: AutoencoderKL,
        image: Optional[torch.Tensor] = None,
        video: Optional[torch.Tensor] = None,
        generator: Optional[torch.Generator] = None,
        compute_posterior: bool = True,
        _original_height: Optional[int] = None,
        _original_width: Optional[int] = None,
        **kwargs,
    ) -> Dict[str, torch.Tensor]:
        conditions = {
            "vae": vae,
            "image": image,
            "video": video,
            "generator": generator,
            "compute_posterior": compute_posterior,
            "_original_height": _original_height,
            "_original_width": _original_width,
            **kwargs,
        }
        input_keys = set(conditions.keys())
        conditions = super().prepare_latents(**conditions)
        conditions = {k: v for k, v in conditions.items() if k not in input_keys}
        return conditions

    def forward(
        self,
        transformer: CogView4Transformer2DModel,
        condition_model_conditions: Dict[str, torch.Tensor],
        latent_model_conditions: Dict[str, torch.Tensor],
        sigmas: torch.Tensor,
        generator: Optional[torch.Generator] = None,
        compute_posterior: bool = True,
        **kwargs,
    ) -> Tuple[torch.Tensor, ...]:
        if compute_posterior:
            latents = latent_model_conditions.pop("latents")
        else:
            posterior = DiagonalGaussianDistribution(latent_model_conditions.pop("latents"))
            latents = posterior.sample(generator=generator)
            del posterior

        latents = (latents - self.vae_config.shift_factor) * self.vae_config.scaling_factor
        noise = torch.zeros_like(latents).normal_(generator=generator)
        timesteps = (sigmas.flatten() * 1000.0).long()

        base_image_sequence_length = 256
        base_shift = 0.25
        max_shift = 0.75

        image_sequence_length = latents.size(2) * latents.size(3) // self.transformer_config.patch_size**2
        mu = (image_sequence_length / base_image_sequence_length) ** 0.5
        mu = mu * max_shift + base_shift
        shifted_sigmas = mu / (mu + (1 / sigmas - 1) ** 1.0)
        noisy_latents = FF.flow_match_xt(latents, noise, shifted_sigmas)

        latent_model_conditions["hidden_states"] = noisy_latents.to(latents)

        pred = transformer(
            **latent_model_conditions,
            **condition_model_conditions,
            timestep=timesteps,
            return_dict=False,
        )[0]
        target = FF.flow_match_target(noise, latents)

        # NOTE: shifted_sigmas loss weighting seems to work better than sigmas. Needs more investigation
        # but let's keep it this way for now. Longer training runs should reveal more insights.
        # return pred, target, sigmas
        return pred, target, shifted_sigmas

    def validation(
        self,
        pipeline: CogView4Pipeline,
        prompt: str,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 50,
        generator: Optional[torch.Generator] = None,
        **kwargs,
    ) -> List[ArtifactType]:
        generation_kwargs = {
            "prompt": prompt,
            "height": height,
            "width": width,
            "num_inference_steps": num_inference_steps,
            "generator": generator,
            "return_dict": True,
            "output_type": "pil",
        }
        generation_kwargs = get_non_null_items(generation_kwargs)
        image = pipeline(**generation_kwargs).images[0]
        return [data.ImageArtifact(value=image)]

    def _save_lora_weights(
        self,
        directory: str,
        transformer_state_dict: Optional[Dict[str, torch.Tensor]] = None,
        scheduler: Optional[SchedulerType] = None,
        *args,
        **kwargs,
    ) -> None:
        # TODO(aryan): this needs refactoring
        if transformer_state_dict is not None:
            CogView4Pipeline.save_lora_weights(directory, transformer_state_dict, safe_serialization=True)
        if scheduler is not None:
            scheduler.save_pretrained(os.path.join(directory, "scheduler"))

    def _save_model(
        self,
        directory: str,
        transformer: CogView4Transformer2DModel,
        transformer_state_dict: Optional[Dict[str, torch.Tensor]] = None,
        scheduler: Optional[SchedulerType] = None,
    ) -> None:
        # TODO(aryan): this needs refactoring
        if transformer_state_dict is not None:
            with init_empty_weights():
                transformer_copy = CogView4Transformer2DModel.from_config(transformer.config)
            transformer_copy.load_state_dict(transformer_state_dict, strict=True, assign=True)
            transformer_copy.save_pretrained(os.path.join(directory, "transformer"))
        if scheduler is not None:
            scheduler.save_pretrained(os.path.join(directory, "scheduler"))