Upload 4 files

CogVideoX/CogVideoX_cli.py

@@ -0,0 +1,90 @@
+import os
+import torch
+from diffusers import CogVideoXDPMScheduler
+from pipeline_rgba import CogVideoXPipeline
+from diffusers.utils import export_to_video
+import argparse
+import numpy as np
+from rgba_utils import *
+
+def main(args):
+    # 1. load pipeline  
+    pipe = CogVideoXPipeline.from_pretrained(args.model_path, torch_dtype=torch.bfloat16)
+    pipe.scheduler = CogVideoXDPMScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
+    pipe.enable_sequential_cpu_offload()
+    pipe.vae.enable_slicing()
+    pipe.vae.enable_tiling()
+
+
+    # 2. define prompt and arguments
+    pipeline_args = {
+        "prompt": args.prompt,
+        "guidance_scale": args.guidance_scale,
+        "num_inference_steps": args.num_inference_steps,
+        "height": args.height,
+        "width": args.width,
+        "num_frames": args.num_frames,
+        "output_type": "latent",
+        "use_dynamic_cfg":True,
+    }
+
+    # 3. prepare rgbx utils
+    # breakpoint()
+    seq_length = 2 * (
+        (args.height // pipe.vae_scale_factor_spatial // 2)
+        * (args.width // pipe.vae_scale_factor_spatial // 2)
+        * ((args.num_frames - 1) // pipe.vae_scale_factor_temporal + 1)
+    )
+    # seq_length = 35100
+
+    prepare_for_rgba_inference(
+        pipe.transformer,
+        rgba_weights_path=args.lora_path,
+        device="cuda",
+        dtype=torch.bfloat16,
+        text_length=226,
+        seq_length=seq_length, # this is for the creation of attention mask.
+    )
+
+    # 4. run inference
+    generator = torch.manual_seed(args.seed) if args.seed else None
+    frames_latents = pipe(**pipeline_args, generator=generator).frames
+
+    frames_latents_rgb, frames_latents_alpha = frames_latents.chunk(2, dim=1)
+
+    frames_rgb = decode_latents(pipe, frames_latents_rgb)
+    frames_alpha = decode_latents(pipe, frames_latents_alpha)
+
+
+    pooled_alpha = np.max(frames_alpha, axis=-1, keepdims=True)
+    frames_alpha_pooled = np.repeat(pooled_alpha, 3, axis=-1)
+    premultiplied_rgb = frames_rgb * frames_alpha_pooled
+
+    if os.path.exists(args.output_path) == False:
+        os.makedirs(args.output_path)
+
+    export_to_video(premultiplied_rgb[0], os.path.join(args.output_path, "rgb.mp4"), fps=args.fps)
+    export_to_video(frames_alpha[0], os.path.join(args.output_path, "alpha.mp4"), fps=args.fps)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Generate a video from a text prompt")
+    parser.add_argument("--prompt", type=str, required=True, help="The description of the video to be generated")
+    parser.add_argument("--lora_path", type=str, default="/hpc2hdd/home/lwang592/projects/CogVideo/sat/outputs/training/ckpts-5b-attn_rebias-partial_lora-8gpu-wo_t2a/lora-rgba-12-21-19-11/5000/rgba_lora.safetensors", help="The path of the LoRA weights to be used")
+    
+    parser.add_argument(
+        "--model_path", type=str, default="THUDM/CogVideoX-5B", help="Path of the pre-trained model use"
+    )
+
+
+    parser.add_argument("--output_path", type=str, default="./output", help="The path save generated video")
+    parser.add_argument("--guidance_scale", type=float, default=6.0, help="The scale for classifier-free guidance")
+    parser.add_argument("--num_inference_steps", type=int, default=50, help="Inference steps")
+    parser.add_argument("--num_frames", type=int, default=49, help="Number of steps for the inference process")
+    parser.add_argument("--width", type=int, default=720, help="Number of steps for the inference process")
+    parser.add_argument("--height", type=int, default=480, help="Number of steps for the inference process")
+    parser.add_argument("--fps", type=int, default=8, help="Number of steps for the inference process")
+    parser.add_argument("--seed", type=int, default=None, help="The seed for reproducibility")
+    args = parser.parse_args()
+
+    main(args)

CogVideoX/CogVideoX_pipeline_rgba.py

@@ -0,0 +1,744 @@
+# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import math
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.loaders import CogVideoXLoraLoaderMixin
+from diffusers.models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
+from diffusers.models.embeddings import get_3d_rotary_pos_embed
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
+from diffusers.utils import logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+from diffusers.pipelines.cogvideo.pipeline_output import CogVideoXPipelineOutput
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers import CogVideoXPipeline
+        >>> from diffusers.utils import export_to_video
+
+        >>> # Models: "THUDM/CogVideoX-2b" or "THUDM/CogVideoX-5b"
+        >>> pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=torch.float16).to("cuda")
+        >>> prompt = (
+        ...     "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. "
+        ...     "The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other "
+        ...     "pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, "
+        ...     "casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. "
+        ...     "The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical "
+        ...     "atmosphere of this unique musical performance."
+        ... )
+        >>> video = pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
+        >>> export_to_video(video, "output.mp4", fps=8)
+        ```
+"""
+
+
+# Similar to diffusers.pipelines.hunyuandit.pipeline_hunyuandit.get_resize_crop_region_for_grid
+def get_resize_crop_region_for_grid(src, tgt_width, tgt_height):
+    tw = tgt_width
+    th = tgt_height
+    h, w = src
+    r = h / w
+    if r > (th / tw):
+        resize_height = th
+        resize_width = int(round(th / h * w))
+    else:
+        resize_width = tw
+        resize_height = int(round(tw / w * h))
+
+    crop_top = int(round((th - resize_height) / 2.0))
+    crop_left = int(round((tw - resize_width) / 2.0))
+
+    return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width)
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
+    r"""
+    Pipeline for text-to-video generation using CogVideoX.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            Frozen text-encoder. CogVideoX uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel); specifically the
+            [t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant.
+        tokenizer (`T5Tokenizer`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+        transformer ([`CogVideoXTransformer3DModel`]):
+            A text conditioned `CogVideoXTransformer3DModel` to denoise the encoded video latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded video latents.
+    """
+
+    _optional_components = []
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+
+    _callback_tensor_inputs = [
+        "latents",
+        "prompt_embeds",
+        "negative_prompt_embeds",
+    ]
+
+    def __init__(
+        self,
+        tokenizer: T5Tokenizer,
+        text_encoder: T5EncoderModel,
+        vae: AutoencoderKLCogVideoX,
+        transformer: CogVideoXTransformer3DModel,
+        scheduler: Union[CogVideoXDDIMScheduler, CogVideoXDPMScheduler],
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+        )
+        self.vae_scale_factor_spatial = (
+            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+        )
+        self.vae_scale_factor_temporal = (
+            self.vae.config.temporal_compression_ratio if hasattr(self, "vae") and self.vae is not None else 4
+        )
+        self.vae_scaling_factor_image = (
+            self.vae.config.scaling_factor if hasattr(self, "vae") and self.vae is not None else 0.7
+        )
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def prepare_latents(
+        self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
+    ):
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        shape = (
+            batch_size,
+            (num_frames - 1) // self.vae_scale_factor_temporal + 1,
+            num_channels_latents,
+            height // self.vae_scale_factor_spatial,
+            width // self.vae_scale_factor_spatial,
+        )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def decode_latents(self, latents: torch.Tensor) -> torch.Tensor:
+        latents = latents.permute(0, 2, 1, 3, 4)  # [batch_size, num_channels, num_frames, height, width]
+        latents = 1 / self.vae_scaling_factor_image * latents
+
+        frames = self.vae.decode(latents).sample
+        return frames
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    # Copied from diffusers.pipelines.latte.pipeline_latte.LattePipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        negative_prompt,
+        callback_on_step_end_tensor_inputs,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+    def fuse_qkv_projections(self) -> None:
+        r"""Enables fused QKV projections."""
+        self.fusing_transformer = True
+        self.transformer.fuse_qkv_projections()
+
+    def unfuse_qkv_projections(self) -> None:
+        r"""Disable QKV projection fusion if enabled."""
+        if not self.fusing_transformer:
+            logger.warning("The Transformer was not initially fused for QKV projections. Doing nothing.")
+        else:
+            self.transformer.unfuse_qkv_projections()
+            self.fusing_transformer = False
+
+    def _prepare_rotary_positional_embeddings(
+        self,
+        height: int,
+        width: int,
+        num_frames: int,
+        device: torch.device,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        grid_height = height // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+        grid_width = width // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+        base_size_width = 720 // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+        base_size_height = 480 // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+
+        grid_crops_coords = get_resize_crop_region_for_grid(
+            (grid_height, grid_width), base_size_width, base_size_height
+        )
+        freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
+            embed_dim=self.transformer.config.attention_head_dim,
+            crops_coords=grid_crops_coords,
+            grid_size=(grid_height, grid_width),
+            temporal_size=num_frames,
+        )
+
+        freqs_cos = freqs_cos.to(device=device)
+        freqs_sin = freqs_sin.to(device=device)
+        return freqs_cos, freqs_sin
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 480,
+        width: int = 720,
+        num_frames: int = 49,
+        num_inference_steps: int = 50,
+        timesteps: Optional[List[int]] = None,
+        guidance_scale: float = 6,
+        use_dynamic_cfg: bool = False,
+        num_videos_per_prompt: int = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: str = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 226,
+    ) -> Union[CogVideoXPipelineOutput, Tuple]:
+        """
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            height (`int`, *optional*, defaults to self.transformer.config.sample_height * self.vae_scale_factor_spatial):
+                The height in pixels of the generated image. This is set to 480 by default for the best results.
+            width (`int`, *optional*, defaults to self.transformer.config.sample_height * self.vae_scale_factor_spatial):
+                The width in pixels of the generated image. This is set to 720 by default for the best results.
+            num_frames (`int`, defaults to `48`):
+                Number of frames to generate. Must be divisible by self.vae_scale_factor_temporal. Generated video will
+                contain 1 extra frame because CogVideoX is conditioned with (num_seconds * fps + 1) frames where
+                num_seconds is 6 and fps is 8. However, since videos can be saved at any fps, the only condition that
+                needs to be satisfied is that of divisibility mentioned above.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, *optional*, defaults to 7.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+                of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int`, defaults to `226`):
+                Maximum sequence length in encoded prompt. Must be consistent with
+                `self.transformer.config.max_text_seq_length` otherwise may lead to poor results.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipelineOutput`] or `tuple`:
+            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        if num_frames > 49:
+            raise ValueError(
+                "The number of frames must be less than 49 for now due to static positional embeddings. This will be updated in the future to remove this limitation."
+            )
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        num_videos_per_prompt = 1
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            negative_prompt,
+            callback_on_step_end_tensor_inputs,
+            prompt_embeds,
+            negative_prompt_embeds,
+        )
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._interrupt = False
+
+        # 2. Default call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            negative_prompt,
+            do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+        self._num_timesteps = len(timesteps)
+
+        # 5. Prepare latents.
+        latent_channels = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            latent_channels,
+            num_frames,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        ).repeat(1,2,1,1,1) # Luozhou
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Create rotary embeds if required
+        image_rotary_emb = (
+            self._prepare_rotary_positional_embeddings(height, width, latents.size(1) // 2, device) # Luozhou
+            if self.transformer.config.use_rotary_positional_embeddings
+            else None
+        )
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            # for DPM-solver++
+            old_pred_original_sample = None
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    image_rotary_emb=image_rotary_emb,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                # perform guidance
+                if use_dynamic_cfg:
+                    self._guidance_scale = 1 + guidance_scale * (
+                        (1 - math.cos(math.pi * ((num_inference_steps - t.item()) / num_inference_steps) ** 5.0)) / 2
+                    )
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                if not isinstance(self.scheduler, CogVideoXDPMScheduler):
+                    latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+                else:
+                    latents, old_pred_original_sample = self.scheduler.step(
+                        noise_pred,
+                        old_pred_original_sample,
+                        t,
+                        timesteps[i - 1] if i > 0 else None,
+                        latents,
+                        **extra_step_kwargs,
+                        return_dict=False,
+                    )
+                latents = latents.to(prompt_embeds.dtype)
+
+                # call the callback, if provided
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+        if not output_type == "latent":
+            video = self.decode_latents(latents)
+            video = self.video_processor.postprocess_video(video=video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return CogVideoXPipelineOutput(frames=video)

CogVideoX/CogVideoX_rgba_utils.py

@@ -0,0 +1,313 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Any, Dict, Optional, Tuple, Union
+from diffusers.utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from safetensors.torch import load_file
+
+logger = logging.get_logger(__name__)
+
+@torch.no_grad()
+def decode_latents(pipe, latents):
+    video = pipe.decode_latents(latents)
+    video = pipe.video_processor.postprocess_video(video=video, output_type="np")
+    return video
+
+def create_attention_mask(text_length: int, seq_length: int, device: torch.device, dtype: torch.dtype) -> torch.Tensor:
+    """
+    Create an attention mask to block text from attending to alpha.
+
+    Args:
+        text_length: Length of the text sequence.
+        seq_length: Length of the other sequence.
+        device: The device where the mask will be stored.
+        dtype: The data type of the mask tensor.
+
+    Returns:
+        An attention mask tensor.
+    """
+    total_length = text_length + seq_length
+    dense_mask = torch.ones((total_length, total_length), dtype=torch.bool)
+    dense_mask[:text_length, text_length + seq_length // 2:] = False
+    return dense_mask.to(device=device, dtype=dtype)
+
+class RGBALoRACogVideoXAttnProcessor:
+    r"""
+    Processor for implementing scaled dot-product attention for the CogVideoX model. 
+    It applies a rotary embedding on query and key vectors, but does not include spatial normalization.
+    """
+
+    def __init__(self, device, dtype, attention_mask, lora_rank=128, lora_alpha=1.0, latent_dim=3072):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("CogVideoXAttnProcessor requires PyTorch 2.0 or later.")
+
+        # Initialize LoRA layers
+        self.lora_alpha = lora_alpha
+        self.lora_rank = lora_rank
+
+        # Helper function to create LoRA layers
+        def create_lora_layer(in_dim, mid_dim, out_dim):
+            return nn.Sequential(
+                nn.Linear(in_dim, mid_dim, bias=False, device=device, dtype=dtype),
+                nn.Linear(mid_dim, out_dim, bias=False, device=device, dtype=dtype)
+            )
+
+        self.to_q_lora = create_lora_layer(latent_dim, lora_rank, latent_dim)
+        self.to_k_lora = create_lora_layer(latent_dim, lora_rank, latent_dim)
+        self.to_v_lora = create_lora_layer(latent_dim, lora_rank, latent_dim)
+        self.to_out_lora = create_lora_layer(latent_dim, lora_rank, latent_dim)
+
+        # Store attention mask
+        self.attention_mask = attention_mask
+
+    def _apply_lora(self, hidden_states, seq_len, query, key, value, scaling):
+        """Applies LoRA updates to query, key, and value tensors."""
+        query_delta = self.to_q_lora(hidden_states).to(query.device)
+        query[:, -seq_len // 2:, :] += query_delta[:, -seq_len // 2:, :] * scaling
+
+        key_delta = self.to_k_lora(hidden_states).to(key.device)
+        key[:, -seq_len // 2:, :] += key_delta[:, -seq_len // 2:, :] * scaling
+
+        value_delta = self.to_v_lora(hidden_states).to(value.device)
+        value[:, -seq_len // 2:, :] += value_delta[:, -seq_len // 2:, :] * scaling
+
+        return query, key, value
+
+    def _apply_rotary_embedding(self, query, key, image_rotary_emb, seq_len, text_seq_length, attn):
+        """Applies rotary embeddings to query and key tensors."""
+        from diffusers.models.embeddings import apply_rotary_emb
+
+        # Apply rotary embedding to RGB and alpha sections
+        query[:, :, text_seq_length:text_seq_length + seq_len // 2] = apply_rotary_emb(
+            query[:, :, text_seq_length:text_seq_length + seq_len // 2], image_rotary_emb)
+        query[:, :, text_seq_length + seq_len // 2:] = apply_rotary_emb(
+            query[:, :, text_seq_length + seq_len // 2:], image_rotary_emb)
+        
+        if not attn.is_cross_attention:
+            key[:, :, text_seq_length:text_seq_length + seq_len // 2] = apply_rotary_emb(
+                key[:, :, text_seq_length:text_seq_length + seq_len // 2], image_rotary_emb)
+            key[:, :, text_seq_length + seq_len // 2:] = apply_rotary_emb(
+                key[:, :, text_seq_length + seq_len // 2:], image_rotary_emb)
+
+        return query, key
+
+    def __call__(
+        self,
+        attn,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        # Concatenate encoder and decoder hidden states
+        text_seq_length = encoder_hidden_states.size(1)
+        hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+
+        batch_size, sequence_length, _ = hidden_states.shape
+        seq_len = hidden_states.shape[1] - text_seq_length
+        scaling = self.lora_alpha / self.lora_rank
+
+        # Apply LoRA to query, key, value
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        query, key, value = self._apply_lora(hidden_states, seq_len, query, key, value, scaling)
+
+        # Reshape query, key, value for multi-head attention
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # Normalize query and key if required
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply rotary embeddings if provided
+        if image_rotary_emb is not None:
+            query, key = self._apply_rotary_embedding(query, key, image_rotary_emb, seq_len, text_seq_length, attn)
+
+        # Compute scaled dot-product attention
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=self.attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        # Reshape the output tensor back to the original shape
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+
+        # Apply linear projection and LoRA to the output
+        original_hidden_states = attn.to_out[0](hidden_states)
+        hidden_states_delta = self.to_out_lora(hidden_states).to(hidden_states.device)
+        original_hidden_states[:, -seq_len // 2:, :] += hidden_states_delta[:, -seq_len // 2:, :] * scaling
+
+        # Apply dropout
+        hidden_states = attn.to_out[1](original_hidden_states)
+
+        # Split back into encoder and decoder hidden states
+        encoder_hidden_states, hidden_states = hidden_states.split(
+            [text_seq_length, hidden_states.size(1) - text_seq_length], dim=1
+        )
+
+        return hidden_states, encoder_hidden_states
+
+def prepare_for_rgba_inference(
+    model, rgba_weights_path: str, device: torch.device, dtype: torch.dtype,
+    lora_rank: int = 128, lora_alpha: float = 1.0, text_length: int = 226, seq_length: int = 35100
+):
+    def load_lora_sequential_weights(lora_layer, lora_layers, prefix):
+        lora_layer[0].load_state_dict({'weight': lora_layers[f"{prefix}.lora_A.weight"]})
+        lora_layer[1].load_state_dict({'weight': lora_layers[f"{prefix}.lora_B.weight"]})
+
+
+    rgba_weights = load_file(rgba_weights_path)
+    aux_emb = rgba_weights['domain_emb']
+
+    attention_mask = create_attention_mask(text_length, seq_length, device, dtype)
+    attn_procs = {}
+
+    for name in model.attn_processors.keys():
+        attn_processor = RGBALoRACogVideoXAttnProcessor(
+            device=device, dtype=dtype, attention_mask=attention_mask,
+            lora_rank=lora_rank, lora_alpha=lora_alpha
+        )
+        
+        index = name.split('.')[1]
+        base_prefix = f'transformer.transformer_blocks.{index}.attn1'
+
+        for lora_layer, prefix in [
+            (attn_processor.to_q_lora, f'{base_prefix}.to_q'),
+            (attn_processor.to_k_lora, f'{base_prefix}.to_k'),
+            (attn_processor.to_v_lora, f'{base_prefix}.to_v'),
+            (attn_processor.to_out_lora, f'{base_prefix}.to_out.0'),
+        ]:
+            load_lora_sequential_weights(lora_layer, rgba_weights, prefix)
+
+        attn_procs[name] = attn_processor
+
+    model.set_attn_processor(attn_procs)
+
+    def custom_forward(self):
+        def forward(
+            hidden_states: torch.Tensor,
+            encoder_hidden_states: torch.Tensor,
+            timestep: Union[int, float, torch.LongTensor],
+            timestep_cond: Optional[torch.Tensor] = None,
+            ofs: Optional[Union[int, float, torch.LongTensor]] = None,
+            image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+            attention_kwargs: Optional[Dict[str, Any]] = None,
+            return_dict: bool = True,
+        ):
+            if attention_kwargs is not None:
+                attention_kwargs = attention_kwargs.copy()
+                lora_scale = attention_kwargs.pop("scale", 1.0)
+            else:
+                lora_scale = 1.0
+
+            if USE_PEFT_BACKEND:
+                # weight the lora layers by setting `lora_scale` for each PEFT layer
+                scale_lora_layers(self, lora_scale)
+            else:
+                if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                    logger.warning(
+                        "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                    )
+
+            batch_size, num_frames, channels, height, width = hidden_states.shape
+
+            # 1. Time embedding
+            timesteps = timestep
+            t_emb = self.time_proj(timesteps)
+
+            # timesteps does not contain any weights and will always return f32 tensors
+            # but time_embedding might actually be running in fp16. so we need to cast here.
+            # there might be better ways to encapsulate this.
+            t_emb = t_emb.to(dtype=hidden_states.dtype)
+            emb = self.time_embedding(t_emb, timestep_cond)
+
+            if self.ofs_embedding is not None:
+                ofs_emb = self.ofs_proj(ofs)
+                ofs_emb = ofs_emb.to(dtype=hidden_states.dtype)
+                ofs_emb = self.ofs_embedding(ofs_emb)
+                emb = emb + ofs_emb
+
+            # 2. Patch embedding
+            hidden_states = self.patch_embed(encoder_hidden_states, hidden_states)
+            hidden_states = self.embedding_dropout(hidden_states)
+
+            text_seq_length = encoder_hidden_states.shape[1]
+            encoder_hidden_states = hidden_states[:, :text_seq_length]
+            hidden_states = hidden_states[:, text_seq_length:]
+
+            hidden_states[:, hidden_states.size(1) // 2:, :] += aux_emb.expand(batch_size, -1, -1).to(hidden_states.device, dtype=hidden_states.dtype)
+
+            # 3. Transformer blocks
+            for i, block in enumerate(self.transformer_blocks):
+                if torch.is_grad_enabled() and self.gradient_checkpointing:
+
+                    def create_custom_forward(module):
+                        def custom_forward(*inputs):
+                            return module(*inputs)
+
+                        return custom_forward
+
+                    ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                    hidden_states, encoder_hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(block),
+                        hidden_states,
+                        encoder_hidden_states,
+                        emb,
+                        image_rotary_emb,
+                        **ckpt_kwargs,
+                    )
+                else:
+                    hidden_states, encoder_hidden_states = block(
+                        hidden_states=hidden_states,
+                        encoder_hidden_states=encoder_hidden_states,
+                        temb=emb,
+                        image_rotary_emb=image_rotary_emb,
+                    )
+
+            if not self.config.use_rotary_positional_embeddings:
+                # CogVideoX-2B
+                hidden_states = self.norm_final(hidden_states)
+            else:
+                # CogVideoX-5B
+                hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+                hidden_states = self.norm_final(hidden_states)
+                hidden_states = hidden_states[:, text_seq_length:]
+
+            # 4. Final block
+            hidden_states = self.norm_out(hidden_states, temb=emb)
+            hidden_states = self.proj_out(hidden_states)
+
+            # 5. Unpatchify
+            p = self.config.patch_size
+            p_t = self.config.patch_size_t
+
+            if p_t is None:
+                output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, -1, p, p)
+                output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
+            else:
+                output = hidden_states.reshape(
+                    batch_size, (num_frames + p_t - 1) // p_t, height // p, width // p, -1, p_t, p, p
+                )
+                output = output.permute(0, 1, 5, 4, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(1, 2)
+
+            if USE_PEFT_BACKEND:
+                # remove `lora_scale` from each PEFT layer
+                unscale_lora_layers(self, lora_scale)
+
+            if not return_dict:
+                return (output,)
+            return Transformer2DModelOutput(sample=output)
+
+
+        return forward
+
+    model.forward = custom_forward(model)
+

CogVideoX/CogVideoX_test.py

@@ -0,0 +1,26 @@
+import torch
+from diffusers import CogVideoXImageToVideoPipeline
+from diffusers.utils import export_to_video, load_image
+
+prompt = "A dragon is flipping its wings"
+image = load_image(image="/hpc2hdd/home/lwang592/projects/CogVideo/sat/configs/i2v/Dragon.jpg")
+pipe = CogVideoXImageToVideoPipeline.from_pretrained(
+    "THUDM/CogVideoX-5b-I2V",
+    torch_dtype=torch.bfloat16
+)
+
+pipe.enable_sequential_cpu_offload()
+pipe.vae.enable_tiling()
+pipe.vae.enable_slicing()
+
+video = pipe(
+    prompt=prompt,
+    image=image,
+    num_videos_per_prompt=1,
+    num_inference_steps=50,
+    num_frames=13,
+    guidance_scale=6,
+    generator=torch.Generator(device="cuda").manual_seed(42),
+).frames[0]
+
+export_to_video(video, "output.mp4", fps=8)