diff --git a/README.md b/README.md index 278593c..5720ee5 100755 --- a/README.md +++ b/README.md @@ -545,7 +545,16 @@ CogVideoX-Fun can be found in [Readme Train](scripts/cogvideox_fun/README_TRAIN. # Model zoo -## 1. Wan2.1-Fun +## 1. Wan2.2 + +| Name | Hugging Face | Model Scope | Description | +|--|--|--|--| +| Wan2.2-TI2V-5B | [🤗Link](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.2-TI2V-5B) | Wan2.1-1.3B Text-to-Video Weights | +| Wan2.2-T2V-14B | [🤗Link](https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.2-T2V-A14B) | Wan2.1-14B Text-to-Video Weights | +| Wan2.2-I2V-14B-480P | [🤗Link](https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.2-I2V-A14B) | Wan2.1-14B-480P Image-to-Video Weights | + + +## 2. Wan2.1-Fun V1.1: | Name | Storage Size | Hugging Face | Model Scope | Description | @@ -565,16 +574,16 @@ V1.0: | Wan2.1-Fun-1.3B-Control | 19.0 GB | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-1.3B-Control) | [😄Link](https://modelscope.cn/models/PAI/Wan2.1-Fun-1.3B-Control) | Wan2.1-Fun-1.3B video control weights, supporting various control conditions such as Canny, Depth, Pose, MLSD, etc., and trajectory control. Supports multi-resolution (512, 768, 1024) video prediction at 81 frames, trained at 16 frames per second, with multilingual prediction support. | | Wan2.1-Fun-14B-Control | 47.0 GB | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-14B-Control) | [😄Link](https://modelscope.cn/models/PAI/Wan2.1-Fun-14B-Control) | Wan2.1-Fun-14B video control weights, supporting various control conditions such as Canny, Depth, Pose, MLSD, etc., and trajectory control. Supports multi-resolution (512, 768, 1024) video prediction at 81 frames, trained at 16 frames per second, with multilingual prediction support. | -## 2. Wan2.1 +## 3. Wan2.1 | Name | Hugging Face | Model Scope | Description | |--|--|--|--| -| Wan2.1-T2V-1.3B | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-1.3B-InP) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-T2V-1.3B) | Wanxiang 2.1-1.3B text-to-video weights | -| Wan2.1-T2V-14B | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-1.3B-InP) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-T2V-14B) | Wanxiang 2.1-14B text-to-video weights | -| Wan2.1-I2V-14B-480P | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-14B-InP) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-I2V-14B-480P) | Wanxiang 2.1-14B-480P image-to-video weights | -| Wan2.1-I2V-14B-720P| [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-14B-InP) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-I2V-14B-480P) | Wanxiang 2.1-14B-720P image-to-video weights | +| Wan2.1-T2V-1.3B | [🤗Link](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-T2V-1.3B) | Wanxiang 2.1-1.3B text-to-video weights | +| Wan2.1-T2V-14B | [🤗Link](https://huggingface.co/Wan-AI/Wan2.1-T2V-14B) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-T2V-14B) | Wanxiang 2.1-14B text-to-video weights | +| Wan2.1-I2V-14B-480P | [🤗Link](https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-480P) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-I2V-14B-480P) | Wanxiang 2.1-14B-480P image-to-video weights | +| Wan2.1-I2V-14B-720P| [🤗Link](https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-720P) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-I2V-14B-720P) | Wanxiang 2.1-14B-720P image-to-video weights | -## 3. CogVideoX-Fun +## 4. CogVideoX-Fun V1.5: @@ -608,6 +617,7 @@ V1.1: - CogVideo: https://github.com/THUDM/CogVideo/ - EasyAnimate: https://github.com/aigc-apps/EasyAnimate - Wan2.1: https://github.com/Wan-Video/Wan2.1/ +- Wan2.2: https://github.com/Wan-Video/Wan2.2/ - ComfyUI-KJNodes: https://github.com/kijai/ComfyUI-KJNodes - ComfyUI-EasyAnimateWrapper: https://github.com/kijai/ComfyUI-EasyAnimateWrapper - ComfyUI-CameraCtrl-Wrapper: https://github.com/chaojie/ComfyUI-CameraCtrl-Wrapper diff --git a/README_ja-JP.md b/README_ja-JP.md index 3bef24a..2dca49a 100755 --- a/README_ja-JP.md +++ b/README_ja-JP.md @@ -543,7 +543,17 @@ Wan2.1は[Readme Train](scripts/wan2.1/README_TRAIN.md)と[Readme Lora](scripts/ CogVideoX-Funは[Readme Train](scripts/cogvideox_fun/README_TRAIN.md)と[Readme Lora](scripts/cogvideox_fun/README_TRAIN_LORA.md)を参照してください。 # モデルの場所 -## 1. Wan2.1-Fun + +## 1. Wan2.2 + +| モデル名 | Hugging Face | Model Scope | 説明 | +|--|--|--|--| +| Wan2.2-TI2V-5B | [🤗リンク](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B) | [😄リンク](https://www.modelscope.cn/models/Wan-AI/Wan2.2-TI2V-5B) | 万象2.1-1.3B テキストから動画生成重み | +| Wan2.2-T2V-14B | [🤗リンク](https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B) | [😄リンク](https://www.modelscope.cn/models/Wan-AI/Wan2.2-T2V-A14B) | 万象2.1-14B テキストから動画生成重み | +| Wan2.2-I2V-14B-480P | [🤗リンク](https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B) | [😄リンク](https://www.modelscope.cn/models/Wan-AI/Wan2.2-I2V-A14B) | 万象2.1-14B-480P 画像から動画生成重み | + + +## 2. Wan2.1-Fun V1.1: | 名称 | ストレージ容量 | Hugging Face | Model Scope | 説明 | @@ -564,16 +574,16 @@ V1.0: | Wan2.1-Fun-1.3B-Control | 19.0 GB | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-1.3B-Control) | [😄Link](https://modelscope.cn/models/PAI/Wan2.1-Fun-1.3B-Control) | Wan2.1-Fun-1.3Bのビデオ制御ウェイト。Canny、Depth、Pose、MLSDなどの異なる制御条件をサポートし、トラジェクトリ制御も利用可能。512、768、1024のマルチ解像度でのビデオ予測をサポートし、81フレーム（1秒間に16フレーム）でトレーニング済みで、多言語予測にも対応しています。 | | Wan2.1-Fun-14B-Control | 47.0 GB | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-14B-Control) | [😄Link](https://modelscope.cn/models/PAI/Wan2.1-Fun-14B-Control) | Wan2.1-Fun-14Bのビデオ制御ウェイト。Canny、Depth、Pose、MLSDなどの異なる制御条件をサポートし、トラジェクトリ制御も利用可能。512、768、1024のマルチ解像度でのビデオ予測をサポートし、81フレーム（1秒間に16フレーム）でトレーニング済みで、多言語予測にも対応しています。 | -## 2. Wan2.1 +## 3. Wan2.1 | 名称 | Hugging Face | Model Scope | 説明 | |--|--|--|--| -| Wan2.1-T2V-1.3B | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-1.3B-InP) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-T2V-1.3B) | 万象2.1-1.3Bのテキストから動画生成する重み | -| Wan2.1-T2V-14B | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-1.3B-InP) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-T2V-14B) | 万象2.1-14Bのテキストから動画生成する重み | -| Wan2.1-I2V-14B-480P | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-14B-InP) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-I2V-14B-480P) | 万象2.1-14B-480Pの画像から動画生成する重み | -| Wan2.1-I2V-14B-720P| [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-14B-InP) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-I2V-14B-480P) | 万象2.1-14B-720Pの画像から動画生成する重み | +| Wan2.1-T2V-1.3B | [🤗Link](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-T2V-1.3B) | 万象2.1-1.3Bのテキストから動画生成する重み | +| Wan2.1-T2V-14B | [🤗Link](https://huggingface.co/Wan-AI/Wan2.1-T2V-14B) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-T2V-14B) | 万象2.1-14Bのテキストから動画生成する重み | +| Wan2.1-I2V-14B-480P | [🤗Link](https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-480P) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-I2V-14B-480P) | 万象2.1-14B-480Pの画像から動画生成する重み | +| Wan2.1-I2V-14B-720P| [🤗Link](https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-720P) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-I2V-14B-720P) | 万象2.1-14B-720Pの画像から動画生成する重み | -## 3. CogVideoX-Fun +## 4. CogVideoX-Fun V1.5: @@ -610,6 +620,7 @@ V1.1: - CogVideo: https://github.com/THUDM/CogVideo/ - EasyAnimate: https://github.com/aigc-apps/EasyAnimate - Wan2.1: https://github.com/Wan-Video/Wan2.1/ +- Wan2.2: https://github.com/Wan-Video/Wan2.2/ # ライセンス このプロジェクトは[Apache License (Version 2.0)](https://github.com/modelscope/modelscope/blob/master/LICENSE)の下でライセンスされています。 diff --git a/README_zh-CN.md b/README_zh-CN.md index e77ae28..7e59955 100755 --- a/README_zh-CN.md +++ b/README_zh-CN.md @@ -534,7 +534,15 @@ CogVideoX-Fun可以查看[Readme Train](scripts/cogvideox_fun/README_TRAIN.md) # 模型地址 -## 1. Wan2.1-Fun +## 1. Wan2.2 + +| 名称 | Hugging Face | Model Scope | 描述 | +|--|--|--|--| +| Wan2.2-TI2V-5B | [🤗Link](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.2-TI2V-5B) | 万象2.1-1.3B文生视频权重 | +| Wan2.2-T2V-14B | [🤗Link](https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.2-T2V-A14B) | 万象2.1-14B文生视频权重 | +| Wan2.2-I2V-14B-480P | [🤗Link](https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.2-I2V-A14B) | 万象2.1-14B-480P图生视频权重 | + +## 2. Wan2.1-Fun V1.1: | 名称 | 存储空间 | Hugging Face | Model Scope | 描述 | @@ -554,16 +562,16 @@ V1.0: | Wan2.1-Fun-1.3B-Control | 19.0 GB | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-1.3B-Control) | [😄Link](https://modelscope.cn/models/PAI/Wan2.1-Fun-1.3B-Control)| Wan2.1-Fun-1.3B视频控制权重，支持不同的控制条件，如Canny、Depth、Pose、MLSD等，同时支持使用轨迹控制。支持多分辨率（512，768，1024）的视频预测，支持多分辨率（512，768，1024）的视频预测，以81帧、每秒16帧进行训练，支持多语言预测 | | Wan2.1-Fun-14B-Control | 47.0 GB | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-14B-Control) | [😄Link](https://modelscope.cn/models/PAI/Wan2.1-Fun-14B-Control)| Wan2.1-Fun-14B视频控制权重，支持不同的控制条件，如Canny、Depth、Pose、MLSD等，同时支持使用轨迹控制。支持多分辨率（512，768，1024）的视频预测，支持多分辨率（512，768，1024）的视频预测，以81帧、每秒16帧进行训练，支持多语言预测 | -## 2. Wan2.1 +## 3. Wan2.1 | 名称 | Hugging Face | Model Scope | 描述 | |--|--|--|--| -| Wan2.1-T2V-1.3B | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-1.3B-InP) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-T2V-1.3B) | 万象2.1-1.3B文生视频权重 | -| Wan2.1-T2V-14B | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-1.3B-InP) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-T2V-14B) | 万象2.1-14B文生视频权重 | -| Wan2.1-I2V-14B-480P | [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-14B-InP) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-I2V-14B-480P) | 万象2.1-14B-480P图生视频权重 | -| Wan2.1-I2V-14B-720P| [🤗Link](https://huggingface.co/alibaba-pai/Wan2.1-Fun-14B-InP) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-I2V-14B-480P) | 万象2.1-14B-720P图生视频权重 | +| Wan2.1-T2V-1.3B | [🤗Link](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-T2V-1.3B) | 万象2.1-1.3B文生视频权重 | +| Wan2.1-T2V-14B | [🤗Link](https://huggingface.co/Wan-AI/Wan2.1-T2V-14B) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-T2V-14B) | 万象2.1-14B文生视频权重 | +| Wan2.1-I2V-14B-480P | [🤗Link](https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-480P) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-I2V-14B-480P) | 万象2.1-14B-480P图生视频权重 | +| Wan2.1-I2V-14B-720P| [🤗Link](https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-720P) | [😄Link](https://www.modelscope.cn/models/Wan-AI/Wan2.1-I2V-14B-720P) | 万象2.1-14B-720P图生视频权重 | -## 3. CogVideoX-Fun +## 4. CogVideoX-Fun V1.5: @@ -598,6 +606,7 @@ V1.1: - CogVideo: https://github.com/THUDM/CogVideo/ - EasyAnimate: https://github.com/aigc-apps/EasyAnimate - Wan2.1: https://github.com/Wan-Video/Wan2.1/ +- Wan2.2: https://github.com/Wan-Video/Wan2.2/ - ComfyUI-KJNodes: https://github.com/kijai/ComfyUI-KJNodes - ComfyUI-EasyAnimateWrapper: https://github.com/kijai/ComfyUI-EasyAnimateWrapper - ComfyUI-CameraCtrl-Wrapper: https://github.com/chaojie/ComfyUI-CameraCtrl-Wrapper diff --git a/examples/wan2.2/predict_i2v.py b/examples/wan2.2/predict_i2v.py new file mode 100644 index 0000000..de3484d --- /dev/null +++ b/examples/wan2.2/predict_i2v.py @@ -0,0 +1,318 @@ +import os +import sys + +import numpy as np +import torch +from diffusers import FlowMatchEulerDiscreteScheduler +from omegaconf import OmegaConf +from PIL import Image + +current_file_path = os.path.abspath(__file__) +project_roots = [os.path.dirname(current_file_path), os.path.dirname(os.path.dirname(current_file_path)), os.path.dirname(os.path.dirname(os.path.dirname(current_file_path)))] +for project_root in project_roots: + sys.path.insert(0, project_root) if project_root not in sys.path else None + +from videox_fun.dist import set_multi_gpus_devices, shard_model +from videox_fun.models import (AutoencoderKLWan, AutoTokenizer, CLIPModel, + WanT5EncoderModel, Wan2_2Transformer3DModel) +from videox_fun.models.cache_utils import get_teacache_coefficients +from videox_fun.pipeline import Wan2_2I2VPipeline +from videox_fun.utils.fp8_optimization import (convert_model_weight_to_float8, replace_parameters_by_name, + convert_weight_dtype_wrapper) +from videox_fun.utils.lora_utils import merge_lora, unmerge_lora +from videox_fun.utils.utils import (filter_kwargs, get_image_to_video_latent, + save_videos_grid) +from videox_fun.utils.fm_solvers import FlowDPMSolverMultistepScheduler +from videox_fun.utils.fm_solvers_unipc import FlowUniPCMultistepScheduler + +# GPU memory mode, which can be choosen in [model_full_load, model_full_load_and_qfloat8, model_cpu_offload, model_cpu_offload_and_qfloat8, sequential_cpu_offload]. +# model_full_load means that the entire model will be moved to the GPU. +# +# model_full_load_and_qfloat8 means that the entire model will be moved to the GPU, +# and the transformer model has been quantized to float8, which can save more GPU memory. +# +# model_cpu_offload means that the entire model will be moved to the CPU after use, which can save some GPU memory. +# +# model_cpu_offload_and_qfloat8 indicates that the entire model will be moved to the CPU after use, +# and the transformer model has been quantized to float8, which can save more GPU memory. +# +# sequential_cpu_offload means that each layer of the model will be moved to the CPU after use, +# resulting in slower speeds but saving a large amount of GPU memory. +GPU_memory_mode = "sequential_cpu_offload" +# Multi GPUs config +# Please ensure that the product of ulysses_degree and ring_degree equals the number of GPUs used. +# For example, if you are using 8 GPUs, you can set ulysses_degree = 2 and ring_degree = 4. +# If you are using 1 GPU, you can set ulysses_degree = 1 and ring_degree = 1. +ulysses_degree = 1 +ring_degree = 1 +# Use FSDP to save more GPU memory in multi gpus. +fsdp_dit = False +fsdp_text_encoder = True +# Compile will give a speedup in fixed resolution and need a little GPU memory. +# The compile_dit is not compatible with the fsdp_dit and sequential_cpu_offload. +compile_dit = False + +# TeaCache config +enable_teacache = True +# Recommended to be set between 0.05 and 0.30. A larger threshold can cache more steps, speeding up the inference process, +# but it may cause slight differences between the generated content and the original content. +# # --------------------------------------------------------------------------------------------------- # +# | Model Name | threshold | Model Name | threshold | Model Name | threshold | +# | Wan2.1-T2V-1.3B | 0.05~0.10 | Wan2.1-T2V-14B | 0.10~0.15 | Wan2.1-I2V-14B-720P | 0.20~0.30 | +# | Wan2.1-I2V-14B-480P | 0.20~0.25 | Wan2.1-Fun-*-1.3B-* | 0.05~0.10 | Wan2.1-Fun-*-14B-* | 0.20~0.30 | +# # --------------------------------------------------------------------------------------------------- # +teacache_threshold = 0.10 +# The number of steps to skip TeaCache at the beginning of the inference process, which can +# reduce the impact of TeaCache on generated video quality. +num_skip_start_steps = 5 +# Whether to offload TeaCache tensors to cpu to save a little bit of GPU memory. +teacache_offload = False + +# Skip some cfg steps in inference +# Recommended to be set between 0.00 and 0.25 +cfg_skip_ratio = 0 + +# Riflex config +enable_riflex = False +# Index of intrinsic frequency +riflex_k = 6 + +# Config and model path +config_path = "config/wan2.2/wan_civitai_i2v.yaml" +# model path +model_name = "models/Diffusion_Transformer/Wan2.2-I2V-A14B" + +# Choose the sampler in "Flow", "Flow_Unipc", "Flow_DPM++" +sampler_name = "Flow_Unipc" +# [NOTE]: Noise schedule shift parameter. Affects temporal dynamics. +# Used when the sampler is in "Flow_Unipc", "Flow_DPM++". +# If you want to generate a 480p video, it is recommended to set the shift value to 3.0. +# If you want to generate a 720p video, it is recommended to set the shift value to 5.0. +shift = 5 + +# Load pretrained model if need +transformer_path = None +vae_path = None +lora_path = None + +# Other params +sample_size = [480, 832] +video_length = 81 +fps = 16 + +# Use torch.float16 if GPU does not support torch.bfloat16 +# ome graphics cards, such as v100, 2080ti, do not support torch.bfloat16 +weight_dtype = torch.bfloat16 +# If you want to generate from text, please set the validation_image_start = None and validation_image_end = None +validation_image_start = "asset/1.png" + +# prompts +prompt = "一只棕色的狗摇着头，坐在舒适房间里的浅色沙发上。在狗的后面，架子上有一幅镶框的画，周围是粉红色的花朵。房间里柔和温暖的灯光营造出舒适的氛围。" +negative_prompt = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" +guidance_scale = 6.0 +seed = 43 +num_inference_steps = 50 +lora_weight = 0.55 +save_path = "samples/wan-videos-i2v" + +device = set_multi_gpus_devices(ulysses_degree, ring_degree) +config = OmegaConf.load(config_path) +boundary = config['transformer_additional_kwargs'].get('boundary', 0.900) + +transformer = Wan2_2Transformer3DModel.from_pretrained( + os.path.join(model_name, config['transformer_additional_kwargs'].get('transformer_low_noise_model_subpath', 'transformer')), + transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']), + low_cpu_mem_usage=True, + torch_dtype=weight_dtype, +) + +transformer_2 = Wan2_2Transformer3DModel.from_pretrained( + os.path.join(model_name, config['transformer_additional_kwargs'].get('transformer_high_noise_model_subpath', 'transformer')), + transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']), + low_cpu_mem_usage=True, + torch_dtype=weight_dtype, +) + +if transformer_path is not None: + print(f"From checkpoint: {transformer_path}") + if transformer_path.endswith("safetensors"): + from safetensors.torch import load_file, safe_open + state_dict = load_file(transformer_path) + else: + state_dict = torch.load(transformer_path, map_location="cpu") + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = transformer.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + +# Get Vae +vae = AutoencoderKLWan.from_pretrained( + os.path.join(model_name, config['vae_kwargs'].get('vae_subpath', 'vae')), + additional_kwargs=OmegaConf.to_container(config['vae_kwargs']), +).to(weight_dtype) + +if vae_path is not None: + print(f"From checkpoint: {vae_path}") + if vae_path.endswith("safetensors"): + from safetensors.torch import load_file, safe_open + state_dict = load_file(vae_path) + else: + state_dict = torch.load(vae_path, map_location="cpu") + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = vae.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + +# Get Tokenizer +tokenizer = AutoTokenizer.from_pretrained( + os.path.join(model_name, config['text_encoder_kwargs'].get('tokenizer_subpath', 'tokenizer')), +) + +# Get Text encoder +text_encoder = WanT5EncoderModel.from_pretrained( + os.path.join(model_name, config['text_encoder_kwargs'].get('text_encoder_subpath', 'text_encoder')), + additional_kwargs=OmegaConf.to_container(config['text_encoder_kwargs']), + low_cpu_mem_usage=True, + torch_dtype=weight_dtype, +) +text_encoder = text_encoder.eval() + +# Get Scheduler +Choosen_Scheduler = scheduler_dict = { + "Flow": FlowMatchEulerDiscreteScheduler, + "Flow_Unipc": FlowUniPCMultistepScheduler, + "Flow_DPM++": FlowDPMSolverMultistepScheduler, +}[sampler_name] +if sampler_name == "Flow_Unipc" or sampler_name == "Flow_DPM++": + config['scheduler_kwargs']['shift'] = 1 +scheduler = Choosen_Scheduler( + **filter_kwargs(Choosen_Scheduler, OmegaConf.to_container(config['scheduler_kwargs'])) +) + +# Get Pipeline +pipeline = Wan2_2I2VPipeline( + transformer=transformer, + transformer_2=transformer_2, + vae=vae, + tokenizer=tokenizer, + text_encoder=text_encoder, + scheduler=scheduler, +) +if ulysses_degree > 1 or ring_degree > 1: + from functools import partial + transformer.enable_multi_gpus_inference() + transformer_2.enable_multi_gpus_inference() + if fsdp_dit: + shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype) + pipeline.transformer = shard_fn(pipeline.transformer) + pipeline.transformer_2 = shard_fn(pipeline.transformer_2) + print("Add FSDP DIT") + if fsdp_text_encoder: + shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype) + pipeline.text_encoder = shard_fn(pipeline.text_encoder) + print("Add FSDP TEXT ENCODER") + +if compile_dit: + for i in range(len(pipeline.transformer.blocks)): + pipeline.transformer.blocks[i] = torch.compile(pipeline.transformer.blocks[i]) + for i in range(len(pipeline.transformer_2.blocks)): + pipeline.transformer_2.blocks[i] = torch.compile(pipeline.transformer_2.blocks[i]) + print("Add Compile") + +if GPU_memory_mode == "sequential_cpu_offload": + replace_parameters_by_name(transformer, ["modulation",], device=device) + replace_parameters_by_name(transformer_2, ["modulation",], device=device) + transformer.freqs = transformer.freqs.to(device=device) + transformer_2.freqs = transformer_2.freqs.to(device=device) + pipeline.enable_sequential_cpu_offload(device=device) +elif GPU_memory_mode == "model_cpu_offload_and_qfloat8": + convert_model_weight_to_float8(transformer, exclude_module_name=["modulation",], device=device) + convert_model_weight_to_float8(transformer_2, exclude_module_name=["modulation",], device=device) + convert_weight_dtype_wrapper(transformer, weight_dtype) + convert_weight_dtype_wrapper(transformer_2, weight_dtype) + pipeline.enable_model_cpu_offload(device=device) +elif GPU_memory_mode == "model_cpu_offload": + pipeline.enable_model_cpu_offload(device=device) +elif GPU_memory_mode == "model_full_load_and_qfloat8": + convert_model_weight_to_float8(transformer, exclude_module_name=["modulation",], device=device) + convert_model_weight_to_float8(transformer_2, exclude_module_name=["modulation",], device=device) + convert_weight_dtype_wrapper(transformer, weight_dtype) + convert_weight_dtype_wrapper(transformer_2, weight_dtype) + pipeline.to(device=device) +else: + pipeline.to(device=device) +transformer_2 = transformer_2.to(device=device) + +coefficients = get_teacache_coefficients(model_name) if enable_teacache else None +if coefficients is not None: + print(f"Enable TeaCache with threshold {teacache_threshold} and skip the first {num_skip_start_steps} steps.") + pipeline.transformer.enable_teacache( + coefficients, num_inference_steps, teacache_threshold, num_skip_start_steps=num_skip_start_steps, offload=teacache_offload + ) + pipeline.transformer_2.enable_teacache( + coefficients, num_inference_steps, teacache_threshold, num_skip_start_steps=num_skip_start_steps, offload=teacache_offload + ) + +if cfg_skip_ratio is not None: + print(f"Enable cfg_skip_ratio {cfg_skip_ratio}.") + pipeline.transformer.enable_cfg_skip(cfg_skip_ratio, num_inference_steps) + pipeline.transformer_2.enable_cfg_skip(cfg_skip_ratio, num_inference_steps) + +generator = torch.Generator(device=device).manual_seed(seed) + +if lora_path is not None: + pipeline = merge_lora(pipeline, lora_path, lora_weight, device=device) + +with torch.no_grad(): + video_length = int((video_length - 1) // vae.config.temporal_compression_ratio * vae.config.temporal_compression_ratio) + 1 if video_length != 1 else 1 + latent_frames = (video_length - 1) // vae.config.temporal_compression_ratio + 1 + + if enable_riflex: + pipeline.transformer.enable_riflex(k = riflex_k, L_test = latent_frames) + pipeline.transformer_2.enable_riflex(k = riflex_k, L_test = latent_frames) + + input_video, input_video_mask, clip_image = get_image_to_video_latent(validation_image_start, None, video_length=video_length, sample_size=sample_size) + + sample = pipeline( + prompt, + num_frames = video_length, + negative_prompt = negative_prompt, + height = sample_size[0], + width = sample_size[1], + generator = generator, + guidance_scale = guidance_scale, + num_inference_steps = num_inference_steps, + boundary = boundary, + + video = input_video, + mask_video = input_video_mask, + shift = shift, + ).videos + +if lora_path is not None: + pipeline = unmerge_lora(pipeline, lora_path, lora_weight, device=device) + +def save_results(): + if not os.path.exists(save_path): + os.makedirs(save_path, exist_ok=True) + + index = len([path for path in os.listdir(save_path)]) + 1 + prefix = str(index).zfill(8) + if video_length == 1: + video_path = os.path.join(save_path, prefix + ".png") + + image = sample[0, :, 0] + image = image.transpose(0, 1).transpose(1, 2) + image = (image * 255).numpy().astype(np.uint8) + image = Image.fromarray(image) + image.save(video_path) + else: + video_path = os.path.join(save_path, prefix + ".mp4") + save_videos_grid(sample, video_path, fps=fps) + +if ulysses_degree * ring_degree > 1: + import torch.distributed as dist + if dist.get_rank() == 0: + save_results() +else: + save_results() \ No newline at end of file diff --git a/examples/wan2.2/predict_t2v.py b/examples/wan2.2/predict_t2v.py new file mode 100755 index 0000000..4a94598 --- /dev/null +++ b/examples/wan2.2/predict_t2v.py @@ -0,0 +1,307 @@ +import os +import sys + +import numpy as np +import torch +from diffusers import FlowMatchEulerDiscreteScheduler +from omegaconf import OmegaConf +from PIL import Image + +current_file_path = os.path.abspath(__file__) +project_roots = [os.path.dirname(current_file_path), os.path.dirname(os.path.dirname(current_file_path)), os.path.dirname(os.path.dirname(os.path.dirname(current_file_path)))] +for project_root in project_roots: + sys.path.insert(0, project_root) if project_root not in sys.path else None + +from videox_fun.dist import set_multi_gpus_devices, shard_model +from videox_fun.models import (AutoencoderKLWan, WanT5EncoderModel, AutoTokenizer, + Wan2_2Transformer3DModel) +from videox_fun.models.cache_utils import get_teacache_coefficients +from videox_fun.pipeline import Wan2_2Pipeline +from videox_fun.utils.fp8_optimization import (convert_model_weight_to_float8, replace_parameters_by_name, + convert_weight_dtype_wrapper) +from videox_fun.utils.lora_utils import merge_lora, unmerge_lora +from videox_fun.utils.utils import (filter_kwargs, get_image_to_video_latent, + save_videos_grid) +from videox_fun.utils.fm_solvers import FlowDPMSolverMultistepScheduler +from videox_fun.utils.fm_solvers_unipc import FlowUniPCMultistepScheduler + +# GPU memory mode, which can be choosen in [model_full_load, model_full_load_and_qfloat8, model_cpu_offload, model_cpu_offload_and_qfloat8, sequential_cpu_offload]. +# model_full_load means that the entire model will be moved to the GPU. +# +# model_full_load_and_qfloat8 means that the entire model will be moved to the GPU, +# and the transformer model has been quantized to float8, which can save more GPU memory. +# +# model_cpu_offload means that the entire model will be moved to the CPU after use, which can save some GPU memory. +# +# model_cpu_offload_and_qfloat8 indicates that the entire model will be moved to the CPU after use, +# and the transformer model has been quantized to float8, which can save more GPU memory. +# +# sequential_cpu_offload means that each layer of the model will be moved to the CPU after use, +# resulting in slower speeds but saving a large amount of GPU memory. +GPU_memory_mode = "sequential_cpu_offload" +# Multi GPUs config +# Please ensure that the product of ulysses_degree and ring_degree equals the number of GPUs used. +# For example, if you are using 8 GPUs, you can set ulysses_degree = 2 and ring_degree = 4. +# If you are using 1 GPU, you can set ulysses_degree = 1 and ring_degree = 1. +ulysses_degree = 1 +ring_degree = 1 +# Use FSDP to save more GPU memory in multi gpus. +fsdp_dit = False +fsdp_text_encoder = True +# Compile will give a speedup in fixed resolution and need a little GPU memory. +# The compile_dit is not compatible with the fsdp_dit and sequential_cpu_offload. +compile_dit = False + +# TeaCache config +enable_teacache = True +# Recommended to be set between 0.05 and 0.30. A larger threshold can cache more steps, speeding up the inference process, +# but it may cause slight differences between the generated content and the original content. +# # --------------------------------------------------------------------------------------------------- # +# | Model Name | threshold | Model Name | threshold | Model Name | threshold | +# | Wan2.1-T2V-1.3B | 0.05~0.10 | Wan2.1-T2V-14B | 0.10~0.15 | Wan2.1-I2V-14B-720P | 0.20~0.30 | +# | Wan2.1-I2V-14B-480P | 0.20~0.25 | Wan2.1-Fun-*-1.3B-* | 0.05~0.10 | Wan2.1-Fun-*-14B-* | 0.20~0.30 | +# # --------------------------------------------------------------------------------------------------- # +teacache_threshold = 0.10 +# The number of steps to skip TeaCache at the beginning of the inference process, which can +# reduce the impact of TeaCache on generated video quality. +num_skip_start_steps = 5 +# Whether to offload TeaCache tensors to cpu to save a little bit of GPU memory. +teacache_offload = False + +# Skip some cfg steps in inference +# Recommended to be set between 0.00 and 0.25 +cfg_skip_ratio = 0 + +# Riflex config +enable_riflex = False +# Index of intrinsic frequency +riflex_k = 6 + +# Config and model path +config_path = "config/wan2.2/wan_civitai_t2v.yaml" +# model path +model_name = "models/Diffusion_Transformer/Wan2.2-T2V-A14B" + +# Choose the sampler in "Flow", "Flow_Unipc", "Flow_DPM++" +sampler_name = "Flow_Unipc" +# [NOTE]: Noise schedule shift parameter. Affects temporal dynamics. +# Used when the sampler is in "Flow_Unipc", "Flow_DPM++". +# If you want to generate a 480p video, it is recommended to set the shift value to 3.0. +# If you want to generate a 720p video, it is recommended to set the shift value to 5.0. +shift = 12 + +# Load pretrained model if need +transformer_path = None +vae_path = None +lora_path = None + +# Other params +sample_size = [480, 832] +video_length = 81 +fps = 16 + +# Use torch.float16 if GPU does not support torch.bfloat16 +# ome graphics cards, such as v100, 2080ti, do not support torch.bfloat16 +weight_dtype = torch.bfloat16 +prompt = "一只棕色的狗摇着头，坐在舒适房间里的浅色沙发上。在狗的后面，架子上有一幅镶框的画，周围是粉红色的花朵。房间里柔和温暖的灯光营造出舒适的氛围。" +negative_prompt = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" +guidance_scale = 6.0 +seed = 43 +num_inference_steps = 50 +lora_weight = 0.55 +save_path = "samples/wan-videos-t2v" + +device = set_multi_gpus_devices(ulysses_degree, ring_degree) +config = OmegaConf.load(config_path) +boundary = config['transformer_additional_kwargs'].get('boundary', 0.875) + +transformer = Wan2_2Transformer3DModel.from_pretrained( + os.path.join(model_name, config['transformer_additional_kwargs'].get('transformer_low_noise_model_subpath', 'transformer')), + transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']), + low_cpu_mem_usage=True, + torch_dtype=weight_dtype, +) + +transformer_2 = Wan2_2Transformer3DModel.from_pretrained( + os.path.join(model_name, config['transformer_additional_kwargs'].get('transformer_high_noise_model_subpath', 'transformer')), + transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']), + low_cpu_mem_usage=True, + torch_dtype=weight_dtype, +) + +if transformer_path is not None: + print(f"From checkpoint: {transformer_path}") + if transformer_path.endswith("safetensors"): + from safetensors.torch import load_file, safe_open + state_dict = load_file(transformer_path) + else: + state_dict = torch.load(transformer_path, map_location="cpu") + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = transformer.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + +# Get Vae +vae = AutoencoderKLWan.from_pretrained( + os.path.join(model_name, config['vae_kwargs'].get('vae_subpath', 'vae')), + additional_kwargs=OmegaConf.to_container(config['vae_kwargs']), +).to(weight_dtype) + +if vae_path is not None: + print(f"From checkpoint: {vae_path}") + if vae_path.endswith("safetensors"): + from safetensors.torch import load_file, safe_open + state_dict = load_file(vae_path) + else: + state_dict = torch.load(vae_path, map_location="cpu") + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = vae.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + +# Get Tokenizer +tokenizer = AutoTokenizer.from_pretrained( + os.path.join(model_name, config['text_encoder_kwargs'].get('tokenizer_subpath', 'tokenizer')), +) + +# Get Text encoder +text_encoder = WanT5EncoderModel.from_pretrained( + os.path.join(model_name, config['text_encoder_kwargs'].get('text_encoder_subpath', 'text_encoder')), + additional_kwargs=OmegaConf.to_container(config['text_encoder_kwargs']), + low_cpu_mem_usage=True, + torch_dtype=weight_dtype, +) + +# Get Scheduler +Choosen_Scheduler = scheduler_dict = { + "Flow": FlowMatchEulerDiscreteScheduler, + "Flow_Unipc": FlowUniPCMultistepScheduler, + "Flow_DPM++": FlowDPMSolverMultistepScheduler, +}[sampler_name] +if sampler_name == "Flow_Unipc" or sampler_name == "Flow_DPM++": + config['scheduler_kwargs']['shift'] = 1 +scheduler = Choosen_Scheduler( + **filter_kwargs(Choosen_Scheduler, OmegaConf.to_container(config['scheduler_kwargs'])) +) + +# Get Pipeline +pipeline = Wan2_2Pipeline( + transformer=transformer, + transformer_2=transformer_2, + vae=vae, + tokenizer=tokenizer, + text_encoder=text_encoder, + scheduler=scheduler, +) +if ulysses_degree > 1 or ring_degree > 1: + from functools import partial + transformer.enable_multi_gpus_inference() + transformer_2.enable_multi_gpus_inference() + if fsdp_dit: + shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype) + pipeline.transformer = shard_fn(pipeline.transformer) + pipeline.transformer_2 = shard_fn(pipeline.transformer_2) + print("Add FSDP DIT") + if fsdp_text_encoder: + shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype) + pipeline.text_encoder = shard_fn(pipeline.text_encoder) + print("Add FSDP TEXT ENCODER") + +if compile_dit: + for i in range(len(pipeline.transformer.blocks)): + pipeline.transformer.blocks[i] = torch.compile(pipeline.transformer.blocks[i]) + for i in range(len(pipeline.transformer_2.blocks)): + pipeline.transformer_2.blocks[i] = torch.compile(pipeline.transformer_2.blocks[i]) + print("Add Compile") + +if GPU_memory_mode == "sequential_cpu_offload": + replace_parameters_by_name(transformer, ["modulation",], device=device) + replace_parameters_by_name(transformer_2, ["modulation",], device=device) + transformer.freqs = transformer.freqs.to(device=device) + transformer_2.freqs = transformer_2.freqs.to(device=device) + pipeline.enable_sequential_cpu_offload(device=device) +elif GPU_memory_mode == "model_cpu_offload_and_qfloat8": + convert_model_weight_to_float8(transformer, exclude_module_name=["modulation",], device=device) + convert_model_weight_to_float8(transformer_2, exclude_module_name=["modulation",], device=device) + convert_weight_dtype_wrapper(transformer, weight_dtype) + convert_weight_dtype_wrapper(transformer_2, weight_dtype) + pipeline.enable_model_cpu_offload(device=device) +elif GPU_memory_mode == "model_cpu_offload": + pipeline.enable_model_cpu_offload(device=device) +elif GPU_memory_mode == "model_full_load_and_qfloat8": + convert_model_weight_to_float8(transformer, exclude_module_name=["modulation",], device=device) + convert_model_weight_to_float8(transformer_2, exclude_module_name=["modulation",], device=device) + convert_weight_dtype_wrapper(transformer, weight_dtype) + convert_weight_dtype_wrapper(transformer_2, weight_dtype) + pipeline.to(device=device) +else: + pipeline.to(device=device) + +coefficients = get_teacache_coefficients(model_name) if enable_teacache else None +if coefficients is not None: + print(f"Enable TeaCache with threshold {teacache_threshold} and skip the first {num_skip_start_steps} steps.") + pipeline.transformer.enable_teacache( + coefficients, num_inference_steps, teacache_threshold, num_skip_start_steps=num_skip_start_steps, offload=teacache_offload + ) + pipeline.transformer_2.enable_teacache( + coefficients, num_inference_steps, teacache_threshold, num_skip_start_steps=num_skip_start_steps, offload=teacache_offload + ) + +if cfg_skip_ratio is not None: + print(f"Enable cfg_skip_ratio {cfg_skip_ratio}.") + pipeline.transformer.enable_cfg_skip(cfg_skip_ratio, num_inference_steps) + pipeline.transformer_2.enable_cfg_skip(cfg_skip_ratio, num_inference_steps) + +generator = torch.Generator(device=device).manual_seed(seed) + +if lora_path is not None: + pipeline = merge_lora(pipeline, lora_path, lora_weight, device=device) + +with torch.no_grad(): + video_length = int((video_length - 1) // vae.config.temporal_compression_ratio * vae.config.temporal_compression_ratio) + 1 if video_length != 1 else 1 + latent_frames = (video_length - 1) // vae.config.temporal_compression_ratio + 1 + + if enable_riflex: + pipeline.transformer.enable_riflex(k = riflex_k, L_test = latent_frames) + pipeline.transformer_2.enable_riflex(k = riflex_k, L_test = latent_frames) + + sample = pipeline( + prompt, + num_frames = video_length, + negative_prompt = negative_prompt, + height = sample_size[0], + width = sample_size[1], + generator = generator, + guidance_scale = guidance_scale, + num_inference_steps = num_inference_steps, + boundary = boundary, + shift = shift, + ).videos + +if lora_path is not None: + pipeline = unmerge_lora(pipeline, lora_path, lora_weight, device=device) + +def save_results(): + if not os.path.exists(save_path): + os.makedirs(save_path, exist_ok=True) + + index = len([path for path in os.listdir(save_path)]) + 1 + prefix = str(index).zfill(8) + if video_length == 1: + video_path = os.path.join(save_path, prefix + ".png") + + image = sample[0, :, 0] + image = image.transpose(0, 1).transpose(1, 2) + image = (image * 255).numpy().astype(np.uint8) + image = Image.fromarray(image) + image.save(video_path) + else: + video_path = os.path.join(save_path, prefix + ".mp4") + save_videos_grid(sample, video_path, fps=fps) + +if ulysses_degree * ring_degree > 1: + import torch.distributed as dist + if dist.get_rank() == 0: + save_results() +else: + save_results() \ No newline at end of file diff --git a/scripts/wan2.2/README_TRAIN_LORA.md b/scripts/wan2.2/README_TRAIN_LORA.md new file mode 100755 index 0000000..2c1412a --- /dev/null +++ b/scripts/wan2.2/README_TRAIN_LORA.md @@ -0,0 +1,216 @@ +## Lora Training Code + +We can choose whether to use deep speed in Wan, which can save a lot of video memory. + +Some parameters in the sh file can be confusing, and they are explained in this document: + +- `enable_bucket` is used to enable bucket training. When enabled, the model does not crop the images and videos at the center, but instead, it trains the entire images and videos after grouping them into buckets based on resolution. +- `random_frame_crop` is used for random cropping on video frames to simulate videos with different frame counts. +- `random_hw_adapt` is used to enable automatic height and width scaling for images and videos. When `random_hw_adapt` is enabled, the training images will have their height and width set to `image_sample_size` as the maximum and `min(video_sample_size, 512)` as the minimum. For training videos, the height and width will be set to `image_sample_size` as the maximum and `min(video_sample_size, 512)` as the minimum. + - For example, when `random_hw_adapt` is enabled, with `video_sample_n_frames=49`, `video_sample_size=1024`, and `image_sample_size=1024`, the resolution of image inputs for training is `512x512` to `1024x1024`, and the resolution of video inputs for training is `512x512x49` to `1024x1024x49`. + - For example, when `random_hw_adapt` is enabled, with `video_sample_n_frames=49`, `video_sample_size=1024`, and `image_sample_size=256`, the resolution of image inputs for training is `256x256` to `1024x1024`, and the resolution of video inputs for training is `256x256x49`. +- `training_with_video_token_length` specifies training the model according to token length. For training images and videos, the height and width will be set to `image_sample_size` as the maximum and `video_sample_size` as the minimum. + - For example, when `training_with_video_token_length` is enabled, with `video_sample_n_frames=49`, `token_sample_size=1024`, `video_sample_size=1024`, and `image_sample_size=256`, the resolution of image inputs for training is `256x256` to `1024x1024`, and the resolution of video inputs for training is `256x256x49` to `1024x1024x49`. + - For example, when `training_with_video_token_length` is enabled, with `video_sample_n_frames=49`, `token_sample_size=512`, `video_sample_size=1024`, and `image_sample_size=256`, the resolution of image inputs for training is `256x256` to `1024x1024`, and the resolution of video inputs for training is `256x256x49` to `1024x1024x9`. + - The token length for a video with dimensions 512x512 and 49 frames is 13,312. We need to set the `token_sample_size = 512`. + - At 512x512 resolution, the number of video frames is 49 (~= 512 * 512 * 49 / 512 / 512). + - At 768x768 resolution, the number of video frames is 21 (~= 512 * 512 * 49 / 768 / 768). + - At 1024x1024 resolution, the number of video frames is 9 (~= 512 * 512 * 49 / 1024 / 1024). + - These resolutions combined with their corresponding lengths allow the model to generate videos of different sizes. +- `train_mode` is used to specify the training mode, which can be either normal or i2v. Since Wan uses the inpaint model to achieve image-to-video generation, the default is set to inpaint mode. If you only wish to achieve text-to-video generation, you can remove this line, and it will default to the text-to-video mode. +- `resume_from_checkpoint` is used to set the training should be resumed from a previous checkpoint. Use a path or `"latest"` to automatically select the last available checkpoint. +- The Wan2.2 series includes two distinct models that handle different noise levels, specified via the `boundary_type` parameter. `low`: Corresponds to the **low noise model** (low_noise_model). `high`: Corresponds to the **high noise model**. (high_noise_model) + + +Wan2.2 T2V without deepspeed: + +Training 14B Wan2.2 without DeepSpeed may result in insufficient GPU memory. +```sh +export MODEL_NAME="models/Diffusion_Transformer/Wan2.2-T2V-A14B" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --mixed_precision="bf16" scripts/wan2.2/train_lora.py \ + --config_path="config/wan2.2/wan_civitai_t2v.yaml" \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --image_sample_size=1024 \ + --video_sample_size=256 \ + --token_sample_size=512 \ + --video_sample_stride=2 \ + --video_sample_n_frames=81 \ + --train_batch_size=1 \ + --video_repeat=1 \ + --gradient_accumulation_steps=1 \ + --dataloader_num_workers=8 \ + --num_train_epochs=100 \ + --checkpointing_steps=50 \ + --learning_rate=1e-04 \ + --seed=42 \ + --output_dir="output_dir" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --random_hw_adapt \ + --training_with_video_token_length \ + --enable_bucket \ + --uniform_sampling \ + --boundary_type="low" \ + --train_mode="normal" \ + --low_vram +``` + +Wan T2V with deepspeed zero-2: + +Wan with DeepSpeed Zero-2 is suitable for training 1.3B Wan and 14B Wan at low resolutions, but training 14B Wan at high resolutions may still result in insufficient GPU memory. + +```sh +export MODEL_NAME="models/Diffusion_Transformer/Wan2.2-T2V-A14B" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --use_deepspeed --deepspeed_config_file config/zero_stage2_config.json --deepspeed_multinode_launcher standard scripts/wan2.2/train_lora.py \ + --config_path="config/wan2.2/wan_civitai_t2v.yaml" \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --image_sample_size=1024 \ + --video_sample_size=256 \ + --token_sample_size=512 \ + --video_sample_stride=2 \ + --video_sample_n_frames=81 \ + --train_batch_size=1 \ + --video_repeat=1 \ + --gradient_accumulation_steps=1 \ + --dataloader_num_workers=8 \ + --num_train_epochs=100 \ + --checkpointing_steps=50 \ + --learning_rate=1e-04 \ + --seed=42 \ + --output_dir="output_dir" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --random_hw_adapt \ + --training_with_video_token_length \ + --enable_bucket \ + --uniform_sampling \ + --boundary_type="low" \ + --train_mode="normal" \ + --use_deepspeed \ + --low_vram +``` + +Wan T2V with deepspeed zero-3: + +Wan with DeepSpeed Zero-3 is suitable for 14B Wan at high resolutions. You must set save_state to True to save the model. After training, you can use the following command to get the final model: +```sh +python scripts/zero_to_bf16.py output_dir/checkpoint-{our-num-steps} output_dir/checkpoint-{your-num-steps}-outputs --max_shard_size 80GB --safe_serialization +``` + +Training shell command is as follows: +```sh +export MODEL_NAME="models/Diffusion_Transformer/Wan2.2-T2V-A14B" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --zero_stage 3 --zero3_save_16bit_model true --zero3_init_flag true --use_deepspeed --deepspeed_config_file config/zero_stage3_config.json --deepspeed_multinode_launcher standard scripts/wan2.2/train_lora.py \ + --config_path="config/wan2.2/wan_civitai_t2v.yaml" \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --image_sample_size=1024 \ + --video_sample_size=256 \ + --token_sample_size=512 \ + --video_sample_stride=2 \ + --video_sample_n_frames=81 \ + --train_batch_size=1 \ + --video_repeat=1 \ + --gradient_accumulation_steps=1 \ + --dataloader_num_workers=8 \ + --num_train_epochs=100 \ + --checkpointing_steps=50 \ + --learning_rate=1e-04 \ + --seed=42 \ + --output_dir="output_dir" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --random_hw_adapt \ + --training_with_video_token_length \ + --enable_bucket \ + --uniform_sampling \ + --boundary_type="low" \ + --train_mode="normal" \ + --use_deepspeed \ + --low_vram +``` + +Wan T2V with FSDP: + +Wan with FSDP is suitable for 14B Wan at high resolutions. Training shell command is as follows: +```sh +export MODEL_NAME="models/Diffusion_Transformer/Wan2.2-T2V-A14B" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --mixed_precision="bf16" --use_fsdp --fsdp_auto_wrap_policy TRANSFORMER_BASED_WRAP --fsdp_transformer_layer_cls_to_wrap=WanAttentionBlock --fsdp_sharding_strategy "FULL_SHARD" --fsdp_state_dict_type=SHARDED_STATE_DICT --fsdp_backward_prefetch "BACKWARD_PRE" --fsdp_cpu_ram_efficient_loading False scripts/wan2.2/train_lora.py \ + --config_path="config/wan2.2/wan_civitai_t2v.yaml" \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --image_sample_size=1024 \ + --video_sample_size=256 \ + --token_sample_size=512 \ + --video_sample_stride=2 \ + --video_sample_n_frames=81 \ + --train_batch_size=1 \ + --video_repeat=1 \ + --gradient_accumulation_steps=1 \ + --dataloader_num_workers=8 \ + --num_train_epochs=100 \ + --checkpointing_steps=50 \ + --learning_rate=1e-04 \ + --seed=42 \ + --output_dir="output_dir" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --random_hw_adapt \ + --training_with_video_token_length \ + --enable_bucket \ + --uniform_sampling \ + --boundary_type="low" \ + --train_mode="normal" \ + --use_deepspeed \ + --low_vram +``` \ No newline at end of file diff --git a/scripts/wan2.2/train_lora.py b/scripts/wan2.2/train_lora.py new file mode 100755 index 0000000..c8b00b0 --- /dev/null +++ b/scripts/wan2.2/train_lora.py @@ -0,0 +1,1892 @@ +"""Modified from https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image.py +""" +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 The HuggingFace Inc. 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 + +import argparse +import gc +import logging +import math +import os +import pickle +import shutil +import sys + +import accelerate +import diffusers +import numpy as np +import torch +import torch.nn.functional as F +import torch.utils.checkpoint +import torchvision.transforms.functional as TF +import transformers +from accelerate import Accelerator +from accelerate.logging import get_logger +from accelerate.state import AcceleratorState +from accelerate.utils import ProjectConfiguration, set_seed +from diffusers import DDIMScheduler, FlowMatchEulerDiscreteScheduler +from diffusers.optimization import get_scheduler +from diffusers.training_utils import (EMAModel, + compute_density_for_timestep_sampling, + compute_loss_weighting_for_sd3) +from diffusers.utils import check_min_version, deprecate, is_wandb_available +from diffusers.utils.torch_utils import is_compiled_module +from einops import rearrange +from omegaconf import OmegaConf +from packaging import version +from PIL import Image +from torch.utils.data import RandomSampler +from torch.utils.tensorboard import SummaryWriter +from torchvision import transforms +from tqdm.auto import tqdm +from transformers import AutoTokenizer +from transformers.utils import ContextManagers + +import datasets + +current_file_path = os.path.abspath(__file__) +project_roots = [os.path.dirname(current_file_path), os.path.dirname(os.path.dirname(current_file_path)), os.path.dirname(os.path.dirname(os.path.dirname(current_file_path)))] +for project_root in project_roots: + sys.path.insert(0, project_root) if project_root not in sys.path else None +from videox_fun.data.bucket_sampler import (ASPECT_RATIO_512, + ASPECT_RATIO_RANDOM_CROP_512, + ASPECT_RATIO_RANDOM_CROP_PROB, + AspectRatioBatchImageVideoSampler, + RandomSampler, get_closest_ratio) +from videox_fun.data.dataset_image_video import (ImageVideoDataset, + ImageVideoSampler, + get_random_mask) +from videox_fun.models import (AutoencoderKLWan, CLIPModel, WanT5EncoderModel, + Wan2_2Transformer3DModel) +from videox_fun.pipeline import Wan2_2Pipeline, Wan2_2I2VPipeline +from videox_fun.utils.discrete_sampler import DiscreteSampling +from videox_fun.utils.lora_utils import create_network, merge_lora, unmerge_lora +from videox_fun.utils.utils import get_image_to_video_latent, save_videos_grid + +if is_wandb_available(): + import wandb + + +def filter_kwargs(cls, kwargs): + import inspect + sig = inspect.signature(cls.__init__) + valid_params = set(sig.parameters.keys()) - {'self', 'cls'} + filtered_kwargs = {k: v for k, v in kwargs.items() if k in valid_params} + return filtered_kwargs + +def get_random_downsample_ratio(sample_size, image_ratio=[], + all_choices=False, rng=None): + def _create_special_list(length): + if length == 1: + return [1.0] + if length >= 2: + first_element = 0.75 + remaining_sum = 1.0 - first_element + other_elements_value = remaining_sum / (length - 1) + special_list = [first_element] + [other_elements_value] * (length - 1) + return special_list + + if sample_size >= 1536: + number_list = [1, 1.25, 1.5, 2, 2.5, 3] + image_ratio + elif sample_size >= 1024: + number_list = [1, 1.25, 1.5, 2] + image_ratio + elif sample_size >= 768: + number_list = [1, 1.25, 1.5] + image_ratio + elif sample_size >= 512: + number_list = [1] + image_ratio + else: + number_list = [1] + + if all_choices: + return number_list + + number_list_prob = np.array(_create_special_list(len(number_list))) + if rng is None: + return np.random.choice(number_list, p = number_list_prob) + else: + return rng.choice(number_list, p = number_list_prob) + +def resize_mask(mask, latent, process_first_frame_only=True): + latent_size = latent.size() + batch_size, channels, num_frames, height, width = mask.shape + + if process_first_frame_only: + target_size = list(latent_size[2:]) + target_size[0] = 1 + first_frame_resized = F.interpolate( + mask[:, :, 0:1, :, :], + size=target_size, + mode='trilinear', + align_corners=False + ) + + target_size = list(latent_size[2:]) + target_size[0] = target_size[0] - 1 + if target_size[0] != 0: + remaining_frames_resized = F.interpolate( + mask[:, :, 1:, :, :], + size=target_size, + mode='trilinear', + align_corners=False + ) + resized_mask = torch.cat([first_frame_resized, remaining_frames_resized], dim=2) + else: + resized_mask = first_frame_resized + else: + target_size = list(latent_size[2:]) + resized_mask = F.interpolate( + mask, + size=target_size, + mode='trilinear', + align_corners=False + ) + return resized_mask + +# Will error if the minimal version of diffusers is not installed. Remove at your own risks. +check_min_version("0.18.0.dev0") + +logger = get_logger(__name__, log_level="INFO") + +def log_validation(vae, text_encoder, tokenizer, clip_image_encoder, transformer3d, network, config, args, accelerator, weight_dtype, global_step): + try: + logger.info("Running validation... ") + + transformer3d_val = Wan2_2Transformer3DModel.from_pretrained( + os.path.join(args.pretrained_model_name_or_path, config['transformer_additional_kwargs'].get('transformer_subpath', 'transformer')), + transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']), + ).to(weight_dtype) + transformer3d_val.load_state_dict(accelerator.unwrap_model(transformer3d).state_dict()) + scheduler = FlowMatchEulerDiscreteScheduler( + **filter_kwargs(FlowMatchEulerDiscreteScheduler, OmegaConf.to_container(config['scheduler_kwargs'])) + ) + + if args.train_mode != "normal": + pipeline = Wan2_2I2VPipeline( + vae=accelerator.unwrap_model(vae).to(weight_dtype), + text_encoder=accelerator.unwrap_model(text_encoder), + tokenizer=tokenizer, + transformer=transformer3d_val, + scheduler=scheduler, + clip_image_encoder=clip_image_encoder, + ) + else: + pipeline = Wan2_2Pipeline( + vae=accelerator.unwrap_model(vae).to(weight_dtype), + text_encoder=accelerator.unwrap_model(text_encoder), + tokenizer=tokenizer, + transformer=transformer3d_val, + scheduler=scheduler, + ) + pipeline = pipeline.to(accelerator.device) + + pipeline = merge_lora( + pipeline, None, 1, accelerator.device, state_dict=accelerator.unwrap_model(network).state_dict(), transformer_only=True + ) + + if args.seed is None: + generator = None + else: + generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) + + for i in range(len(args.validation_prompts)): + with torch.no_grad(): + if args.train_mode != "normal": + with torch.autocast("cuda", dtype=weight_dtype): + video_length = int((args.video_sample_n_frames - 1) // vae.config.temporal_compression_ratio * vae.config.temporal_compression_ratio) + 1 if args.video_sample_n_frames != 1 else 1 + input_video, input_video_mask, _ = get_image_to_video_latent(None, None, video_length=video_length, sample_size=[args.video_sample_size, args.video_sample_size]) + sample = pipeline( + args.validation_prompts[i], + num_frames = video_length, + negative_prompt = "bad detailed", + height = args.video_sample_size, + width = args.video_sample_size, + guidance_scale = 6.0, + generator = generator, + + video = input_video, + mask_video = input_video_mask, + ).videos + os.makedirs(os.path.join(args.output_dir, "sample"), exist_ok=True) + save_videos_grid(sample, os.path.join(args.output_dir, f"sample/sample-{global_step}-{i}.gif")) + + video_length = 1 + input_video, input_video_mask, _ = get_image_to_video_latent(None, None, video_length=video_length, sample_size=[args.video_sample_size, args.video_sample_size]) + sample = pipeline( + args.validation_prompts[i], + num_frames = video_length, + negative_prompt = "bad detailed", + height = args.video_sample_size, + width = args.video_sample_size, + guidance_scale = 6.0, + generator = generator, + + video = input_video, + mask_video = input_video_mask, + ).videos + os.makedirs(os.path.join(args.output_dir, "sample"), exist_ok=True) + save_videos_grid(sample, os.path.join(args.output_dir, f"sample/sample-{global_step}-image-{i}.gif")) + else: + with torch.autocast("cuda", dtype=weight_dtype): + sample = pipeline( + args.validation_prompts[i], + num_frames = args.video_sample_n_frames, + negative_prompt = "bad detailed", + height = args.video_sample_size, + width = args.video_sample_size, + generator = generator + ).videos + os.makedirs(os.path.join(args.output_dir, "sample"), exist_ok=True) + save_videos_grid(sample, os.path.join(args.output_dir, f"sample/sample-{global_step}-{i}.gif")) + + sample = pipeline( + args.validation_prompts[i], + num_frames = 1, + negative_prompt = "bad detailed", + height = args.video_sample_size, + width = args.video_sample_size, + generator = generator + ).videos + os.makedirs(os.path.join(args.output_dir, "sample"), exist_ok=True) + save_videos_grid(sample, os.path.join(args.output_dir, f"sample/sample-{global_step}-image-{i}.gif")) + + del pipeline + del transformer3d_val + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() + except Exception as e: + gc.collect() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() + print(f"Eval error with info {e}") + return None + +def linear_decay(initial_value, final_value, total_steps, current_step): + if current_step >= total_steps: + return final_value + current_step = max(0, current_step) + step_size = (final_value - initial_value) / total_steps + current_value = initial_value + step_size * current_step + return current_value + +def generate_timestep_with_lognorm(low, high, shape, device="cpu", generator=None): + u = torch.normal(mean=0.0, std=1.0, size=shape, device=device, generator=generator) + t = 1 / (1 + torch.exp(-u)) * (high - low) + low + return torch.clip(t.to(torch.int32), low, high - 1) + +def parse_args(): + parser = argparse.ArgumentParser(description="Simple example of a training script.") + parser.add_argument( + "--input_perturbation", type=float, default=0, help="The scale of input perturbation. Recommended 0.1." + ) + parser.add_argument( + "--pretrained_model_name_or_path", + type=str, + default=None, + required=True, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--revision", + type=str, + default=None, + required=False, + help="Revision of pretrained model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--variant", + type=str, + default=None, + help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16", + ) + parser.add_argument( + "--train_data_dir", + type=str, + default=None, + help=( + "A folder containing the training data. " + ), + ) + parser.add_argument( + "--train_data_meta", + type=str, + default=None, + help=( + "A csv containing the training data. " + ), + ) + parser.add_argument( + "--max_train_samples", + type=int, + default=None, + help=( + "For debugging purposes or quicker training, truncate the number of training examples to this " + "value if set." + ), + ) + parser.add_argument( + "--validation_prompts", + type=str, + default=None, + nargs="+", + help=("A set of prompts evaluated every `--validation_epochs` and logged to `--report_to`."), + ) + parser.add_argument( + "--output_dir", + type=str, + default="sd-model-finetuned", + help="The output directory where the model predictions and checkpoints will be written.", + ) + parser.add_argument( + "--cache_dir", + type=str, + default=None, + help="The directory where the downloaded models and datasets will be stored.", + ) + parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") + parser.add_argument( + "--random_flip", + action="store_true", + help="whether to randomly flip images horizontally", + ) + parser.add_argument( + "--use_came", + action="store_true", + help="whether to use came", + ) + parser.add_argument( + "--multi_stream", + action="store_true", + help="whether to use cuda multi-stream", + ) + parser.add_argument( + "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader." + ) + parser.add_argument( + "--vae_mini_batch", type=int, default=32, help="mini batch size for vae." + ) + parser.add_argument("--num_train_epochs", type=int, default=100) + parser.add_argument( + "--max_train_steps", + type=int, + default=None, + help="Total number of training steps to perform. If provided, overrides num_train_epochs.", + ) + parser.add_argument( + "--gradient_accumulation_steps", + type=int, + default=1, + help="Number of updates steps to accumulate before performing a backward/update pass.", + ) + parser.add_argument( + "--gradient_checkpointing", + action="store_true", + help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.", + ) + parser.add_argument( + "--learning_rate", + type=float, + default=1e-4, + help="Initial learning rate (after the potential warmup period) to use.", + ) + parser.add_argument( + "--scale_lr", + action="store_true", + default=False, + help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.", + ) + parser.add_argument( + "--lr_scheduler", + type=str, + default="constant", + help=( + 'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",' + ' "constant", "constant_with_warmup"]' + ), + ) + parser.add_argument( + "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler." + ) + parser.add_argument( + "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes." + ) + parser.add_argument( + "--allow_tf32", + action="store_true", + help=( + "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see" + " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices" + ), + ) + parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.") + parser.add_argument( + "--non_ema_revision", + type=str, + default=None, + required=False, + help=( + "Revision of pretrained non-ema model identifier. Must be a branch, tag or git identifier of the local or" + " remote repository specified with --pretrained_model_name_or_path." + ), + ) + parser.add_argument( + "--dataloader_num_workers", + type=int, + default=0, + help=( + "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process." + ), + ) + parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.") + parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.") + parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.") + parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer") + parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.") + parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.") + parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.") + parser.add_argument( + "--prediction_type", + type=str, + default=None, + help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediciton_type` is chosen.", + ) + parser.add_argument( + "--hub_model_id", + type=str, + default=None, + help="The name of the repository to keep in sync with the local `output_dir`.", + ) + parser.add_argument( + "--logging_dir", + type=str, + default="logs", + help=( + "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to" + " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***." + ), + ) + parser.add_argument( + "--mixed_precision", + type=str, + default=None, + choices=["no", "fp16", "bf16"], + help=( + "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >=" + " 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the" + " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config." + ), + ) + parser.add_argument( + "--report_to", + type=str, + default="tensorboard", + help=( + 'The integration to report the results and logs to. Supported platforms are `"tensorboard"`' + ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.' + ), + ) + parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") + parser.add_argument( + "--checkpointing_steps", + type=int, + default=500, + help=( + "Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming" + " training using `--resume_from_checkpoint`." + ), + ) + parser.add_argument( + "--checkpoints_total_limit", + type=int, + default=None, + help=("Max number of checkpoints to store."), + ) + parser.add_argument( + "--resume_from_checkpoint", + type=str, + default=None, + help=( + "Whether training should be resumed from a previous checkpoint. Use a path saved by" + ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.' + ), + ) + parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.") + parser.add_argument( + "--validation_epochs", + type=int, + default=5, + help="Run validation every X epochs.", + ) + parser.add_argument( + "--validation_steps", + type=int, + default=2000, + help="Run validation every X steps.", + ) + parser.add_argument( + "--tracker_project_name", + type=str, + default="text2image-fine-tune", + help=( + "The `project_name` argument passed to Accelerator.init_trackers for" + " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator" + ), + ) + + parser.add_argument( + "--rank", + type=int, + default=128, + help=("The dimension of the LoRA update matrices."), + ) + parser.add_argument( + "--network_alpha", + type=int, + default=64, + help=("The dimension of the LoRA update matrices."), + ) + parser.add_argument( + "--train_text_encoder", + action="store_true", + help="Whether to train the text encoder. If set, the text encoder should be float32 precision.", + ) + parser.add_argument( + "--snr_loss", action="store_true", help="Whether or not to use snr_loss." + ) + parser.add_argument( + "--uniform_sampling", action="store_true", help="Whether or not to use uniform_sampling." + ) + parser.add_argument( + "--enable_text_encoder_in_dataloader", action="store_true", help="Whether or not to use text encoder in dataloader." + ) + parser.add_argument( + "--enable_bucket", action="store_true", help="Whether enable bucket sample in datasets." + ) + parser.add_argument( + "--random_ratio_crop", action="store_true", help="Whether enable random ratio crop sample in datasets." + ) + parser.add_argument( + "--random_frame_crop", action="store_true", help="Whether enable random frame crop sample in datasets." + ) + parser.add_argument( + "--random_hw_adapt", action="store_true", help="Whether enable random adapt height and width in datasets." + ) + parser.add_argument( + "--training_with_video_token_length", action="store_true", help="The training stage of the model in training.", + ) + parser.add_argument( + "--noise_share_in_frames", action="store_true", help="Whether enable noise share in frames." + ) + parser.add_argument( + "--noise_share_in_frames_ratio", type=float, default=0.5, help="Noise share ratio.", + ) + parser.add_argument( + "--motion_sub_loss", action="store_true", help="Whether enable motion sub loss." + ) + parser.add_argument( + "--motion_sub_loss_ratio", type=float, default=0.25, help="The ratio of motion sub loss." + ) + parser.add_argument( + "--keep_all_node_same_token_length", + action="store_true", + help="Reference of the length token.", + ) + parser.add_argument( + "--train_sampling_steps", + type=int, + default=1000, + help="Run train_sampling_steps.", + ) + parser.add_argument( + "--token_sample_size", + type=int, + default=512, + help="Sample size of the token.", + ) + parser.add_argument( + "--video_sample_size", + type=int, + default=512, + help="Sample size of the video.", + ) + parser.add_argument( + "--image_sample_size", + type=int, + default=512, + help="Sample size of the image.", + ) + parser.add_argument( + "--fix_sample_size", + nargs=2, type=int, default=None, + help="Fix Sample size [height, width] when using bucket and collate_fn." + ) + parser.add_argument( + "--video_sample_stride", + type=int, + default=4, + help="Sample stride of the video.", + ) + parser.add_argument( + "--video_sample_n_frames", + type=int, + default=17, + help="Num frame of video.", + ) + parser.add_argument( + "--video_repeat", + type=int, + default=0, + help="Num of repeat video.", + ) + parser.add_argument( + "--config_path", + type=str, + default=None, + help=( + "The config of the model in training." + ), + ) + parser.add_argument( + "--transformer_path", + type=str, + default=None, + help=("If you want to load the weight from other transformers, input its path."), + ) + parser.add_argument( + "--vae_path", + type=str, + default=None, + help=("If you want to load the weight from other vaes, input its path."), + ) + parser.add_argument("--save_state", action="store_true", help="Whether or not to save state.") + + parser.add_argument( + '--tokenizer_max_length', + type=int, + default=512, + help='Max length of tokenizer' + ) + parser.add_argument( + "--use_deepspeed", action="store_true", help="Whether or not to use deepspeed." + ) + parser.add_argument( + "--use_fsdp", action="store_true", help="Whether or not to use fsdp." + ) + parser.add_argument( + "--low_vram", action="store_true", help="Whether enable low_vram mode." + ) + parser.add_argument( + "--boundary_type", + type=str, + default="low", + help=( + 'The format of training data. Support `"low"` and `"high"`' + ), + ) + parser.add_argument( + "--train_mode", + type=str, + default="normal", + help=( + 'The format of training data. Support `"normal"`' + ' (default), `"i2v"`.' + ), + ) + parser.add_argument( + "--weighting_scheme", + type=str, + default="none", + choices=["sigma_sqrt", "logit_normal", "mode", "cosmap", "none"], + help=('We default to the "none" weighting scheme for uniform sampling and uniform loss'), + ) + parser.add_argument( + "--logit_mean", type=float, default=0.0, help="mean to use when using the `'logit_normal'` weighting scheme." + ) + parser.add_argument( + "--logit_std", type=float, default=1.0, help="std to use when using the `'logit_normal'` weighting scheme." + ) + parser.add_argument( + "--mode_scale", + type=float, + default=1.29, + help="Scale of mode weighting scheme. Only effective when using the `'mode'` as the `weighting_scheme`.", + ) + parser.add_argument( + "--lora_skip_name", + type=str, + default=None, + help=("The module is not trained in loras. "), + ) + + args = parser.parse_args() + env_local_rank = int(os.environ.get("LOCAL_RANK", -1)) + if env_local_rank != -1 and env_local_rank != args.local_rank: + args.local_rank = env_local_rank + + # default to using the same revision for the non-ema model if not specified + if args.non_ema_revision is None: + args.non_ema_revision = args.revision + + return args + + +def main(): + args = parse_args() + + if args.report_to == "wandb" and args.hub_token is not None: + raise ValueError( + "You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token." + " Please use `huggingface-cli login` to authenticate with the Hub." + ) + + if args.non_ema_revision is not None: + deprecate( + "non_ema_revision!=None", + "0.15.0", + message=( + "Downloading 'non_ema' weights from revision branches of the Hub is deprecated. Please make sure to" + " use `--variant=non_ema` instead." + ), + ) + logging_dir = os.path.join(args.output_dir, args.logging_dir) + + config = OmegaConf.load(args.config_path) + accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir) + + accelerator = Accelerator( + gradient_accumulation_steps=args.gradient_accumulation_steps, + mixed_precision=args.mixed_precision, + log_with=args.report_to, + project_config=accelerator_project_config, + ) + + deepspeed_plugin = accelerator.state.deepspeed_plugin if hasattr(accelerator.state, "deepspeed_plugin") else None + fsdp_plugin = accelerator.state.fsdp_plugin if hasattr(accelerator.state, "fsdp_plugin") else None + if deepspeed_plugin is not None: + zero_stage = int(deepspeed_plugin.zero_stage) + fsdp_stage = 0 + print(f"Using DeepSpeed Zero stage: {zero_stage}") + + args.use_deepspeed = True + if zero_stage == 3: + print(f"Auto set save_state to True because zero_stage == 3") + args.save_state = True + elif fsdp_plugin is not None: + from torch.distributed.fsdp import ShardingStrategy + zero_stage = 0 + if fsdp_plugin.sharding_strategy is ShardingStrategy.FULL_SHARD: + fsdp_stage = 3 + elif fsdp_plugin.sharding_strategy is None: # The fsdp_plugin.sharding_strategy is None in FSDP 2. + fsdp_stage = 3 + elif fsdp_plugin.sharding_strategy is ShardingStrategy.SHARD_GRAD_OP: + fsdp_stage = 2 + else: + fsdp_stage = 0 + print(f"Using FSDP stage: {fsdp_stage}") + + args.use_fsdp = True + if fsdp_stage == 3: + print(f"Auto set save_state to True because fsdp_stage == 3") + args.save_state = True + else: + zero_stage = 0 + fsdp_stage = 0 + print("DeepSpeed is not enabled.") + + if accelerator.is_main_process: + writer = SummaryWriter(log_dir=logging_dir) + + # Make one log on every process with the configuration for debugging. + logging.basicConfig( + format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", + datefmt="%m/%d/%Y %H:%M:%S", + level=logging.INFO, + ) + logger.info(accelerator.state, main_process_only=False) + if accelerator.is_local_main_process: + datasets.utils.logging.set_verbosity_warning() + transformers.utils.logging.set_verbosity_warning() + diffusers.utils.logging.set_verbosity_info() + else: + datasets.utils.logging.set_verbosity_error() + transformers.utils.logging.set_verbosity_error() + diffusers.utils.logging.set_verbosity_error() + + # If passed along, set the training seed now. + if args.seed is not None: + set_seed(args.seed) + rng = np.random.default_rng(np.random.PCG64(args.seed + accelerator.process_index)) + torch_rng = torch.Generator(accelerator.device).manual_seed(args.seed + accelerator.process_index) + else: + rng = None + torch_rng = None + index_rng = np.random.default_rng(np.random.PCG64(43)) + print(f"Init rng with seed {args.seed + accelerator.process_index}. Process_index is {accelerator.process_index}") + + # Handle the repository creation + if accelerator.is_main_process: + if args.output_dir is not None: + os.makedirs(args.output_dir, exist_ok=True) + + # For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora transformer3d) to half-precision + # as these weights are only used for inference, keeping weights in full precision is not required. + weight_dtype = torch.float32 + if accelerator.mixed_precision == "fp16": + weight_dtype = torch.float16 + args.mixed_precision = accelerator.mixed_precision + elif accelerator.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + args.mixed_precision = accelerator.mixed_precision + + # Load scheduler, tokenizer and models. + noise_scheduler = FlowMatchEulerDiscreteScheduler( + **filter_kwargs(FlowMatchEulerDiscreteScheduler, OmegaConf.to_container(config['scheduler_kwargs'])) + ) + + # Get Tokenizer + tokenizer = AutoTokenizer.from_pretrained( + os.path.join(args.pretrained_model_name_or_path, config['text_encoder_kwargs'].get('tokenizer_subpath', 'tokenizer')), + ) + + def deepspeed_zero_init_disabled_context_manager(): + """ + returns either a context list that includes one that will disable zero.Init or an empty context list + """ + deepspeed_plugin = AcceleratorState().deepspeed_plugin if accelerate.state.is_initialized() else None + if deepspeed_plugin is None: + return [] + + return [deepspeed_plugin.zero3_init_context_manager(enable=False)] + + # Currently Accelerate doesn't know how to handle multiple models under Deepspeed ZeRO stage 3. + # For this to work properly all models must be run through `accelerate.prepare`. But accelerate + # will try to assign the same optimizer with the same weights to all models during + # `deepspeed.initialize`, which of course doesn't work. + # + # For now the following workaround will partially support Deepspeed ZeRO-3, by excluding the 2 + # frozen models from being partitioned during `zero.Init` which gets called during + # `from_pretrained` So CLIPTextModel and AutoencoderKL will not enjoy the parameter sharding + # across multiple gpus and only UNet2DConditionModel will get ZeRO sharded. + with ContextManagers(deepspeed_zero_init_disabled_context_manager()): + # Get Text encoder + text_encoder = WanT5EncoderModel.from_pretrained( + os.path.join(args.pretrained_model_name_or_path, config['text_encoder_kwargs'].get('text_encoder_subpath', 'text_encoder')), + additional_kwargs=OmegaConf.to_container(config['text_encoder_kwargs']), + low_cpu_mem_usage=True, + torch_dtype=weight_dtype, + ) + text_encoder = text_encoder.eval() + # Get Vae + vae = AutoencoderKLWan.from_pretrained( + os.path.join(args.pretrained_model_name_or_path, config['vae_kwargs'].get('vae_subpath', 'vae')), + additional_kwargs=OmegaConf.to_container(config['vae_kwargs']), + ) + vae.eval() + # Get Clip Image Encoder + if args.train_mode != "normal": + clip_image_encoder = CLIPModel.from_pretrained( + os.path.join(args.pretrained_model_name_or_path, config['image_encoder_kwargs'].get('image_encoder_subpath', 'image_encoder')), + ) + clip_image_encoder = clip_image_encoder.eval() + + # Get Transformer + sub_path = config['transformer_additional_kwargs'].get('transformer_low_noise_model_subpath', 'transformer') \ + if args.boundary_type == "low" else config['transformer_additional_kwargs'].get('transformer_high_noise_model_subpath', 'transformer') + transformer3d = Wan2_2Transformer3DModel.from_pretrained( + os.path.join(args.pretrained_model_name_or_path, sub_path), + transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']), + ).to(weight_dtype) + + # Freeze vae and text_encoder and set transformer3d to trainable + vae.requires_grad_(False) + text_encoder.requires_grad_(False) + transformer3d.requires_grad_(False) + if args.train_mode != "normal": + clip_image_encoder.requires_grad_(False) + + # Lora will work with this... + network = create_network( + 1.0, + args.rank, + args.network_alpha, + text_encoder, + transformer3d, + neuron_dropout=None, + skip_name=args.lora_skip_name, + ) + network.apply_to(text_encoder, transformer3d, args.train_text_encoder and not args.training_with_video_token_length, True) + + if args.transformer_path is not None: + print(f"From checkpoint: {args.transformer_path}") + if args.transformer_path.endswith("safetensors"): + from safetensors.torch import load_file, safe_open + state_dict = load_file(args.transformer_path) + else: + state_dict = torch.load(args.transformer_path, map_location="cpu") + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = transformer3d.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + assert len(u) == 0 + + if args.vae_path is not None: + print(f"From checkpoint: {args.vae_path}") + if args.vae_path.endswith("safetensors"): + from safetensors.torch import load_file, safe_open + state_dict = load_file(args.vae_path) + else: + state_dict = torch.load(args.vae_path, map_location="cpu") + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = vae.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + assert len(u) == 0 + + # `accelerate` 0.16.0 will have better support for customized saving + if version.parse(accelerate.__version__) >= version.parse("0.16.0"): + # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format + if fsdp_stage != 0: + def save_model_hook(models, weights, output_dir): + accelerate_state_dict = accelerator.get_state_dict(models[-1], unwrap=True) + if accelerator.is_main_process: + from safetensors.torch import save_file + + safetensor_save_path = os.path.join(output_dir, f"lora_diffusion_pytorch_model.safetensors") + network_state_dict = {} + for key in accelerate_state_dict: + if "network" in key: + network_state_dict[key.replace("network.", "")] = accelerate_state_dict[key] + + save_file(network_state_dict, safetensor_save_path, metadata={"format": "pt"}) + + with open(os.path.join(output_dir, "sampler_pos_start.pkl"), 'wb') as file: + pickle.dump([batch_sampler.sampler._pos_start, first_epoch], file) + + def load_model_hook(models, input_dir): + pkl_path = os.path.join(input_dir, "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + loaded_number, _ = pickle.load(file) + batch_sampler.sampler._pos_start = max(loaded_number - args.dataloader_num_workers * accelerator.num_processes * 2, 0) + print(f"Load pkl from {pkl_path}. Get loaded_number = {loaded_number}.") + + elif zero_stage == 3: + def save_model_hook(models, weights, output_dir): + if accelerator.is_main_process: + with open(os.path.join(output_dir, "sampler_pos_start.pkl"), 'wb') as file: + pickle.dump([batch_sampler.sampler._pos_start, first_epoch], file) + + def load_model_hook(models, input_dir): + pkl_path = os.path.join(input_dir, "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + loaded_number, _ = pickle.load(file) + batch_sampler.sampler._pos_start = max(loaded_number - args.dataloader_num_workers * accelerator.num_processes * 2, 0) + print(f"Load pkl from {pkl_path}. Get loaded_number = {loaded_number}.") + else: + def save_model_hook(models, weights, output_dir): + if accelerator.is_main_process: + safetensor_save_path = os.path.join(output_dir, f"lora_diffusion_pytorch_model.safetensors") + save_model(safetensor_save_path, accelerator.unwrap_model(models[-1])) + if not args.use_deepspeed: + for _ in range(len(weights)): + weights.pop() + + with open(os.path.join(output_dir, "sampler_pos_start.pkl"), 'wb') as file: + pickle.dump([batch_sampler.sampler._pos_start, first_epoch], file) + + def load_model_hook(models, input_dir): + pkl_path = os.path.join(input_dir, "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + loaded_number, _ = pickle.load(file) + batch_sampler.sampler._pos_start = max(loaded_number - args.dataloader_num_workers * accelerator.num_processes * 2, 0) + print(f"Load pkl from {pkl_path}. Get loaded_number = {loaded_number}.") + + accelerator.register_save_state_pre_hook(save_model_hook) + accelerator.register_load_state_pre_hook(load_model_hook) + + if args.gradient_checkpointing: + transformer3d.enable_gradient_checkpointing() + + # Enable TF32 for faster training on Ampere GPUs, + # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices + if args.allow_tf32: + torch.backends.cuda.matmul.allow_tf32 = True + + if args.scale_lr: + args.learning_rate = ( + args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes + ) + + # Initialize the optimizer + if args.use_8bit_adam: + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError( + "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`" + ) + + optimizer_cls = bnb.optim.AdamW8bit + elif args.use_came: + try: + from came_pytorch import CAME + except: + raise ImportError( + "Please install came_pytorch to use CAME. You can do so by running `pip install came_pytorch`" + ) + + optimizer_cls = CAME + else: + optimizer_cls = torch.optim.AdamW + + logging.info("Add network parameters") + trainable_params = list(filter(lambda p: p.requires_grad, network.parameters())) + trainable_params_optim = network.prepare_optimizer_params(args.learning_rate / 2, args.learning_rate, args.learning_rate) + + if args.use_came: + optimizer = optimizer_cls( + trainable_params_optim, + lr=args.learning_rate, + # weight_decay=args.adam_weight_decay, + betas=(0.9, 0.999, 0.9999), + eps=(1e-30, 1e-16) + ) + else: + optimizer = optimizer_cls( + trainable_params_optim, + lr=args.learning_rate, + betas=(args.adam_beta1, args.adam_beta2), + weight_decay=args.adam_weight_decay, + eps=args.adam_epsilon, + ) + + # Get the training dataset + sample_n_frames_bucket_interval = vae.config.temporal_compression_ratio + + if args.fix_sample_size is not None and args.enable_bucket: + args.video_sample_size = max(max(args.fix_sample_size), args.video_sample_size) + args.image_sample_size = max(max(args.fix_sample_size), args.image_sample_size) + args.training_with_video_token_length = False + args.random_hw_adapt = False + + # Get the dataset + train_dataset = ImageVideoDataset( + args.train_data_meta, args.train_data_dir, + video_sample_size=args.video_sample_size, video_sample_stride=args.video_sample_stride, video_sample_n_frames=args.video_sample_n_frames, + video_repeat=args.video_repeat, + image_sample_size=args.image_sample_size, + enable_bucket=args.enable_bucket, + enable_inpaint=True if args.train_mode != "normal" else False, + ) + + if args.enable_bucket: + aspect_ratio_sample_size = {key : [x / 512 * args.video_sample_size for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()} + batch_sampler_generator = torch.Generator().manual_seed(args.seed) + batch_sampler = AspectRatioBatchImageVideoSampler( + sampler=RandomSampler(train_dataset, generator=batch_sampler_generator), dataset=train_dataset.dataset, + batch_size=args.train_batch_size, train_folder = args.train_data_dir, drop_last=True, + aspect_ratios=aspect_ratio_sample_size, + ) + + def get_length_to_frame_num(token_length): + if args.image_sample_size > args.video_sample_size: + sample_sizes = list(range(args.video_sample_size, args.image_sample_size + 1, 128)) + + if sample_sizes[-1] != args.image_sample_size: + sample_sizes.append(args.image_sample_size) + else: + sample_sizes = [args.image_sample_size] + + length_to_frame_num = { + sample_size: min(token_length / sample_size / sample_size, args.video_sample_n_frames) // sample_n_frames_bucket_interval * sample_n_frames_bucket_interval + 1 for sample_size in sample_sizes + } + + return length_to_frame_num + + def collate_fn(examples): + # Get token length + target_token_length = args.video_sample_n_frames * args.token_sample_size * args.token_sample_size + length_to_frame_num = get_length_to_frame_num(target_token_length) + + # Create new output + new_examples = {} + new_examples["target_token_length"] = target_token_length + new_examples["pixel_values"] = [] + new_examples["text"] = [] + # Used in Inpaint mode + if args.train_mode != "normal": + new_examples["mask_pixel_values"] = [] + new_examples["mask"] = [] + new_examples["clip_pixel_values"] = [] + + # Get downsample ratio in image and videos + pixel_value = examples[0]["pixel_values"] + data_type = examples[0]["data_type"] + f, h, w, c = np.shape(pixel_value) + if data_type == 'image': + random_downsample_ratio = 1 if not args.random_hw_adapt else get_random_downsample_ratio(args.image_sample_size, image_ratio=[args.image_sample_size / args.video_sample_size], rng=rng) + + aspect_ratio_sample_size = {key : [x / 512 * args.image_sample_size / random_downsample_ratio for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()} + aspect_ratio_random_crop_sample_size = {key : [x / 512 * args.image_sample_size / random_downsample_ratio for x in ASPECT_RATIO_RANDOM_CROP_512[key]] for key in ASPECT_RATIO_RANDOM_CROP_512.keys()} + + batch_video_length = args.video_sample_n_frames + sample_n_frames_bucket_interval + else: + if args.random_hw_adapt: + if args.training_with_video_token_length: + local_min_size = np.min(np.array([np.mean(np.array([np.shape(example["pixel_values"])[1], np.shape(example["pixel_values"])[2]])) for example in examples])) + # The video will be resized to a lower resolution than its own. + choice_list = [length for length in list(length_to_frame_num.keys()) if length < local_min_size * 1.25] + if len(choice_list) == 0: + choice_list = list(length_to_frame_num.keys()) + if rng is None: + local_video_sample_size = np.random.choice(choice_list) + else: + local_video_sample_size = rng.choice(choice_list) + batch_video_length = length_to_frame_num[local_video_sample_size] + random_downsample_ratio = args.video_sample_size / local_video_sample_size + else: + random_downsample_ratio = get_random_downsample_ratio( + args.video_sample_size, rng=rng) + batch_video_length = args.video_sample_n_frames + sample_n_frames_bucket_interval + else: + random_downsample_ratio = 1 + batch_video_length = args.video_sample_n_frames + sample_n_frames_bucket_interval + + aspect_ratio_sample_size = {key : [x / 512 * args.video_sample_size / random_downsample_ratio for x in ASPECT_RATIO_512[key]] for key in ASPECT_RATIO_512.keys()} + aspect_ratio_random_crop_sample_size = {key : [x / 512 * args.video_sample_size / random_downsample_ratio for x in ASPECT_RATIO_RANDOM_CROP_512[key]] for key in ASPECT_RATIO_RANDOM_CROP_512.keys()} + + if args.fix_sample_size is not None: + fix_sample_size = [int(x / 16) * 16 for x in args.fix_sample_size] + elif args.random_ratio_crop: + if rng is None: + random_sample_size = aspect_ratio_random_crop_sample_size[ + np.random.choice(list(aspect_ratio_random_crop_sample_size.keys()), p = ASPECT_RATIO_RANDOM_CROP_PROB) + ] + else: + random_sample_size = aspect_ratio_random_crop_sample_size[ + rng.choice(list(aspect_ratio_random_crop_sample_size.keys()), p = ASPECT_RATIO_RANDOM_CROP_PROB) + ] + random_sample_size = [int(x / 16) * 16 for x in random_sample_size] + else: + closest_size, closest_ratio = get_closest_ratio(h, w, ratios=aspect_ratio_sample_size) + closest_size = [int(x / 16) * 16 for x in closest_size] + + for example in examples: + if args.fix_sample_size is not None: + # To 0~1 + pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous() + pixel_values = pixel_values / 255. + + # Get adapt hw for resize + fix_sample_size = list(map(lambda x: int(x), fix_sample_size)) + transform = transforms.Compose([ + transforms.Resize(fix_sample_size, interpolation=transforms.InterpolationMode.BILINEAR), # Image.BICUBIC + transforms.CenterCrop(fix_sample_size), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ]) + elif args.random_ratio_crop: + # To 0~1 + pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous() + pixel_values = pixel_values / 255. + + # Get adapt hw for resize + b, c, h, w = pixel_values.size() + th, tw = random_sample_size + if th / tw > h / w: + nh = int(th) + nw = int(w / h * nh) + else: + nw = int(tw) + nh = int(h / w * nw) + + transform = transforms.Compose([ + transforms.Resize([nh, nw]), + transforms.CenterCrop([int(x) for x in random_sample_size]), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ]) + else: + # To 0~1 + pixel_values = torch.from_numpy(example["pixel_values"]).permute(0, 3, 1, 2).contiguous() + pixel_values = pixel_values / 255. + + # Get adapt hw for resize + closest_size = list(map(lambda x: int(x), closest_size)) + if closest_size[0] / h > closest_size[1] / w: + resize_size = closest_size[0], int(w * closest_size[0] / h) + else: + resize_size = int(h * closest_size[1] / w), closest_size[1] + + transform = transforms.Compose([ + transforms.Resize(resize_size, interpolation=transforms.InterpolationMode.BILINEAR), # Image.BICUBIC + transforms.CenterCrop(closest_size), + transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True), + ]) + new_examples["pixel_values"].append(transform(pixel_values)) + new_examples["text"].append(example["text"]) + + batch_video_length = int(min(batch_video_length, len(pixel_values))) + + # Magvae needs the number of frames to be 4n + 1. + batch_video_length = (batch_video_length - 1) // sample_n_frames_bucket_interval * sample_n_frames_bucket_interval + 1 + + if batch_video_length <= 0: + batch_video_length = 1 + + if args.train_mode != "normal": + mask = get_random_mask(new_examples["pixel_values"][-1].size(), image_start_only=True) + mask_pixel_values = new_examples["pixel_values"][-1] * (1 - mask) + # Wan 2.1 use 0 for masked pixels + # + torch.ones_like(new_examples["pixel_values"][-1]) * -1 * mask + new_examples["mask_pixel_values"].append(mask_pixel_values) + new_examples["mask"].append(mask) + + clip_pixel_values = new_examples["pixel_values"][-1][0].permute(1, 2, 0).contiguous() + clip_pixel_values = (clip_pixel_values * 0.5 + 0.5) * 255 + new_examples["clip_pixel_values"].append(clip_pixel_values) + + # Limit the number of frames to the same + new_examples["pixel_values"] = torch.stack([example[:batch_video_length] for example in new_examples["pixel_values"]]) + if args.train_mode != "normal": + new_examples["mask_pixel_values"] = torch.stack([example[:batch_video_length] for example in new_examples["mask_pixel_values"]]) + new_examples["mask"] = torch.stack([example[:batch_video_length] for example in new_examples["mask"]]) + new_examples["clip_pixel_values"] = torch.stack([example for example in new_examples["clip_pixel_values"]]) + + # Encode prompts when enable_text_encoder_in_dataloader=True + if args.enable_text_encoder_in_dataloader: + prompt_ids = tokenizer( + new_examples['text'], + max_length=args.tokenizer_max_length, + padding="max_length", + add_special_tokens=True, + truncation=True, + return_tensors="pt" + ) + encoder_hidden_states = text_encoder( + prompt_ids.input_ids + )[0] + new_examples['encoder_attention_mask'] = prompt_ids.attention_mask + new_examples['encoder_hidden_states'] = encoder_hidden_states + + return new_examples + + # DataLoaders creation: + train_dataloader = torch.utils.data.DataLoader( + train_dataset, + batch_sampler=batch_sampler, + collate_fn=collate_fn, + persistent_workers=True if args.dataloader_num_workers != 0 else False, + num_workers=args.dataloader_num_workers, + ) + else: + # DataLoaders creation: + batch_sampler_generator = torch.Generator().manual_seed(args.seed) + batch_sampler = ImageVideoSampler(RandomSampler(train_dataset, generator=batch_sampler_generator), train_dataset, args.train_batch_size) + train_dataloader = torch.utils.data.DataLoader( + train_dataset, + batch_sampler=batch_sampler, + persistent_workers=True if args.dataloader_num_workers != 0 else False, + num_workers=args.dataloader_num_workers, + ) + + # Scheduler and math around the number of training steps. + overrode_max_train_steps = False + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if args.max_train_steps is None: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + overrode_max_train_steps = True + + lr_scheduler = get_scheduler( + args.lr_scheduler, + optimizer=optimizer, + num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes, + num_training_steps=args.max_train_steps * accelerator.num_processes, + ) + + # Prepare everything with our `accelerator`. + if fsdp_stage != 0: + transformer3d.network = network + transformer3d = transformer3d.to(weight_dtype) + transformer3d, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + transformer3d, optimizer, train_dataloader, lr_scheduler + ) + else: + network, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + network, optimizer, train_dataloader, lr_scheduler + ) + + if zero_stage == 3: + from functools import partial + from videox_fun.dist import set_multi_gpus_devices, shard_model + shard_fn = partial(shard_model, device_id=accelerator.device, param_dtype=weight_dtype) + transformer3d = shard_fn(transformer3d) + + if fsdp_stage != 0: + from functools import partial + from videox_fun.dist import set_multi_gpus_devices, shard_model + shard_fn = partial(shard_model, device_id=accelerator.device, param_dtype=weight_dtype) + text_encoder = shard_fn(text_encoder) + + # Move text_encode and vae to gpu and cast to weight_dtype + vae.to(accelerator.device, dtype=weight_dtype) + transformer3d.to(accelerator.device, dtype=weight_dtype) + if not args.enable_text_encoder_in_dataloader: + text_encoder.to(accelerator.device) + if args.train_mode != "normal": + clip_image_encoder.to(accelerator.device, dtype=weight_dtype) + + # We need to recalculate our total training steps as the size of the training dataloader may have changed. + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if overrode_max_train_steps: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + # Afterwards we recalculate our number of training epochs + args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) + + # We need to initialize the trackers we use, and also store our configuration. + # The trackers initializes automatically on the main process. + if accelerator.is_main_process: + tracker_config = dict(vars(args)) + tracker_config.pop("validation_prompts") + tracker_config.pop("fix_sample_size") + accelerator.init_trackers(args.tracker_project_name, tracker_config) + + # Function for unwrapping if model was compiled with `torch.compile`. + def unwrap_model(model): + model = accelerator.unwrap_model(model) + model = model._orig_mod if is_compiled_module(model) else model + return model + + # Train! + total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps + + logger.info("***** Running training *****") + logger.info(f" Num examples = {len(train_dataset)}") + logger.info(f" Num Epochs = {args.num_train_epochs}") + logger.info(f" Instantaneous batch size per device = {args.train_batch_size}") + logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") + logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") + logger.info(f" Total optimization steps = {args.max_train_steps}") + global_step = 0 + first_epoch = 0 + + # Potentially load in the weights and states from a previous save + if args.resume_from_checkpoint: + if args.resume_from_checkpoint != "latest": + path = os.path.basename(args.resume_from_checkpoint) + else: + # Get the most recent checkpoint + dirs = os.listdir(args.output_dir) + dirs = [d for d in dirs if d.startswith("checkpoint")] + dirs = sorted(dirs, key=lambda x: int(x.split("-")[1])) + path = dirs[-1] if len(dirs) > 0 else None + + if path is None: + accelerator.print( + f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run." + ) + args.resume_from_checkpoint = None + initial_global_step = 0 + else: + global_step = int(path.split("-")[1]) + + initial_global_step = global_step + + checkpoint_folder_path = os.path.join(args.output_dir, path) + pkl_path = os.path.join(checkpoint_folder_path, "sampler_pos_start.pkl") + if os.path.exists(pkl_path): + with open(pkl_path, 'rb') as file: + _, first_epoch = pickle.load(file) + else: + first_epoch = global_step // num_update_steps_per_epoch + print(f"Load pkl from {pkl_path}. Get first_epoch = {first_epoch}.") + + if zero_stage != 3 and not args.use_fsdp: + from safetensors.torch import load_file + state_dict = load_file(os.path.join(checkpoint_folder_path, "lora_diffusion_pytorch_model.safetensors"), device=str(accelerator.device)) + m, u = accelerator.unwrap_model(network).load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + + optimizer_file_pt = os.path.join(checkpoint_folder_path, "optimizer.pt") + optimizer_file_bin = os.path.join(checkpoint_folder_path, "optimizer.bin") + optimizer_file_to_load = None + + if os.path.exists(optimizer_file_pt): + optimizer_file_to_load = optimizer_file_pt + elif os.path.exists(optimizer_file_bin): + optimizer_file_to_load = optimizer_file_bin + + if optimizer_file_to_load: + try: + accelerator.print(f"Loading optimizer state from {optimizer_file_to_load}") + optimizer_state = torch.load(optimizer_file_to_load, map_location=accelerator.device) + optimizer.load_state_dict(optimizer_state) + accelerator.print("Optimizer state loaded successfully.") + except Exception as e: + accelerator.print(f"Failed to load optimizer state from {optimizer_file_to_load}: {e}") + + scheduler_file_pt = os.path.join(checkpoint_folder_path, "scheduler.pt") + scheduler_file_bin = os.path.join(checkpoint_folder_path, "scheduler.bin") + scheduler_file_to_load = None + + if os.path.exists(scheduler_file_pt): + scheduler_file_to_load = scheduler_file_pt + elif os.path.exists(scheduler_file_bin): + scheduler_file_to_load = scheduler_file_bin + + if scheduler_file_to_load: + try: + accelerator.print(f"Loading scheduler state from {scheduler_file_to_load}") + scheduler_state = torch.load(scheduler_file_to_load, map_location=accelerator.device) + lr_scheduler.load_state_dict(scheduler_state) + accelerator.print("Scheduler state loaded successfully.") + except Exception as e: + accelerator.print(f"Failed to load scheduler state from {scheduler_file_to_load}: {e}") + + if hasattr(accelerator, 'scaler') and accelerator.scaler is not None: + scaler_file = os.path.join(checkpoint_folder_path, "scaler.pt") + if os.path.exists(scaler_file): + try: + accelerator.print(f"Loading GradScaler state from {scaler_file}") + scaler_state = torch.load(scaler_file, map_location=accelerator.device) + accelerator.scaler.load_state_dict(scaler_state) + accelerator.print("GradScaler state loaded successfully.") + except Exception as e: + accelerator.print(f"Failed to load GradScaler state: {e}") + + else: + accelerator.load_state(checkpoint_folder_path) + accelerator.print("accelerator.load_state() completed for zero_stage 3.") + + else: + initial_global_step = 0 + + # function for saving/removing + def save_model(ckpt_file, unwrapped_nw): + os.makedirs(args.output_dir, exist_ok=True) + accelerator.print(f"\nsaving checkpoint: {ckpt_file}") + unwrapped_nw.save_weights(ckpt_file, weight_dtype, None) + + progress_bar = tqdm( + range(0, args.max_train_steps), + initial=initial_global_step, + desc="Steps", + # Only show the progress bar once on each machine. + disable=not accelerator.is_local_main_process, + ) + + if args.multi_stream and args.train_mode != "normal": + # create extra cuda streams to speedup inpaint vae computation + vae_stream_1 = torch.cuda.Stream() + vae_stream_2 = torch.cuda.Stream() + else: + vae_stream_1 = None + vae_stream_2 = None + + idx_sampling = DiscreteSampling(args.train_sampling_steps, uniform_sampling=args.uniform_sampling) + + for epoch in range(first_epoch, args.num_train_epochs): + train_loss = 0.0 + batch_sampler.sampler.generator = torch.Generator().manual_seed(args.seed + epoch) + for step, batch in enumerate(train_dataloader): + if epoch == first_epoch and step == 0: + pixel_values, texts = batch['pixel_values'].cpu(), batch['text'] + pixel_values = rearrange(pixel_values, "b f c h w -> b c f h w") + os.makedirs(os.path.join(args.output_dir, "sanity_check"), exist_ok=True) + for idx, (pixel_value, text) in enumerate(zip(pixel_values, texts)): + pixel_value = pixel_value[None, ...] + gif_name = '-'.join(text.replace('/', '').split()[:10]) if not text == '' else f'{global_step}-{idx}' + save_videos_grid(pixel_value, f"{args.output_dir}/sanity_check/{gif_name[:10]}.gif", rescale=True) + if args.train_mode != "normal": + clip_pixel_values, mask_pixel_values, texts = batch['clip_pixel_values'].cpu(), batch['mask_pixel_values'].cpu(), batch['text'] + mask_pixel_values = rearrange(mask_pixel_values, "b f c h w -> b c f h w") + for idx, (clip_pixel_value, pixel_value, text) in enumerate(zip(clip_pixel_values, mask_pixel_values, texts)): + pixel_value = pixel_value[None, ...] + Image.fromarray(np.uint8(clip_pixel_value)).save(f"{args.output_dir}/sanity_check/clip_{gif_name[:10] if not text == '' else f'{global_step}-{idx}'}.png") + save_videos_grid(pixel_value, f"{args.output_dir}/sanity_check/mask_{gif_name[:10] if not text == '' else f'{global_step}-{idx}'}.gif", rescale=True) + + with accelerator.accumulate(transformer3d): + # Convert images to latent space + pixel_values = batch["pixel_values"].to(weight_dtype) + + # Increase the batch size when the length of the latent sequence of the current sample is small + if args.training_with_video_token_length and zero_stage != 3: + if args.video_sample_n_frames * args.token_sample_size * args.token_sample_size // 16 >= pixel_values.size()[1] * pixel_values.size()[3] * pixel_values.size()[4]: + pixel_values = torch.tile(pixel_values, (4, 1, 1, 1, 1)) + if args.enable_text_encoder_in_dataloader: + batch['encoder_hidden_states'] = torch.tile(batch['encoder_hidden_states'], (4, 1, 1)) + batch['encoder_attention_mask'] = torch.tile(batch['encoder_attention_mask'], (4, 1)) + else: + batch['text'] = batch['text'] * 4 + elif args.video_sample_n_frames * args.token_sample_size * args.token_sample_size // 4 >= pixel_values.size()[1] * pixel_values.size()[3] * pixel_values.size()[4]: + pixel_values = torch.tile(pixel_values, (2, 1, 1, 1, 1)) + if args.enable_text_encoder_in_dataloader: + batch['encoder_hidden_states'] = torch.tile(batch['encoder_hidden_states'], (2, 1, 1)) + batch['encoder_attention_mask'] = torch.tile(batch['encoder_attention_mask'], (2, 1)) + else: + batch['text'] = batch['text'] * 2 + + if args.train_mode != "normal": + clip_pixel_values = batch["clip_pixel_values"].to(weight_dtype) + mask_pixel_values = batch["mask_pixel_values"].to(weight_dtype) + mask = batch["mask"].to(weight_dtype) + # Increase the batch size when the length of the latent sequence of the current sample is small + if args.training_with_video_token_length and zero_stage != 3: + if args.video_sample_n_frames * args.token_sample_size * args.token_sample_size // 16 >= pixel_values.size()[1] * pixel_values.size()[3] * pixel_values.size()[4]: + clip_pixel_values = torch.tile(clip_pixel_values, (4, 1, 1, 1)) + mask_pixel_values = torch.tile(mask_pixel_values, (4, 1, 1, 1, 1)) + mask = torch.tile(mask, (4, 1, 1, 1, 1)) + elif args.video_sample_n_frames * args.token_sample_size * args.token_sample_size // 4 >= pixel_values.size()[1] * pixel_values.size()[3] * pixel_values.size()[4]: + clip_pixel_values = torch.tile(clip_pixel_values, (2, 1, 1, 1)) + mask_pixel_values = torch.tile(mask_pixel_values, (2, 1, 1, 1, 1)) + mask = torch.tile(mask, (2, 1, 1, 1, 1)) + + if args.random_frame_crop: + def _create_special_list(length): + if length == 1: + return [1.0] + if length >= 2: + last_element = 0.90 + remaining_sum = 1.0 - last_element + other_elements_value = remaining_sum / (length - 1) + special_list = [other_elements_value] * (length - 1) + [last_element] + return special_list + select_frames = [_tmp for _tmp in list(range(sample_n_frames_bucket_interval + 1, args.video_sample_n_frames + sample_n_frames_bucket_interval, sample_n_frames_bucket_interval))] + select_frames_prob = np.array(_create_special_list(len(select_frames))) + + if len(select_frames) != 0: + if rng is None: + temp_n_frames = np.random.choice(select_frames, p = select_frames_prob) + else: + temp_n_frames = rng.choice(select_frames, p = select_frames_prob) + else: + temp_n_frames = 1 + + # Magvae needs the number of frames to be 4n + 1. + temp_n_frames = (temp_n_frames - 1) // sample_n_frames_bucket_interval + 1 + + pixel_values = pixel_values[:, :temp_n_frames, :, :] + + if args.train_mode != "normal": + mask_pixel_values = mask_pixel_values[:, :temp_n_frames, :, :] + mask = mask[:, :temp_n_frames, :, :] + + # Keep all node same token length to accelerate the traning when resolution grows. + if args.keep_all_node_same_token_length: + if args.token_sample_size > 256: + numbers_list = list(range(256, args.token_sample_size + 1, 128)) + + if numbers_list[-1] != args.token_sample_size: + numbers_list.append(args.token_sample_size) + else: + numbers_list = [256] + numbers_list = [_number * _number * args.video_sample_n_frames for _number in numbers_list] + + actual_token_length = index_rng.choice(numbers_list) + actual_video_length = (min( + actual_token_length / pixel_values.size()[-1] / pixel_values.size()[-2], args.video_sample_n_frames + ) - 1) // sample_n_frames_bucket_interval * sample_n_frames_bucket_interval + 1 + actual_video_length = int(max(actual_video_length, 1)) + + # Magvae needs the number of frames to be 4n + 1. + actual_video_length = (actual_video_length - 1) // sample_n_frames_bucket_interval + 1 + + pixel_values = pixel_values[:, :actual_video_length, :, :] + if args.train_mode != "normal": + mask_pixel_values = mask_pixel_values[:, :actual_video_length, :, :] + mask = mask[:, :actual_video_length, :, :] + + # Make the inpaint latents to be zeros. + if args.train_mode != "normal": + t2v_flag = [(_mask == 1).all() for _mask in mask] + new_t2v_flag = [] + for _mask in t2v_flag: + if _mask and np.random.rand() < 0.90: + new_t2v_flag.append(0) + else: + new_t2v_flag.append(1) + t2v_flag = torch.from_numpy(np.array(new_t2v_flag)).to(accelerator.device, dtype=weight_dtype) + + if args.low_vram: + torch.cuda.empty_cache() + vae.to(accelerator.device) + if args.train_mode != "normal": + clip_image_encoder.to(accelerator.device) + if not args.enable_text_encoder_in_dataloader: + text_encoder.to("cpu") + + with torch.no_grad(): + # This way is quicker when batch grows up + def _batch_encode_vae(pixel_values): + pixel_values = rearrange(pixel_values, "b f c h w -> b c f h w") + bs = args.vae_mini_batch + new_pixel_values = [] + for i in range(0, pixel_values.shape[0], bs): + pixel_values_bs = pixel_values[i : i + bs] + pixel_values_bs = vae.encode(pixel_values_bs)[0] + pixel_values_bs = pixel_values_bs.sample() + new_pixel_values.append(pixel_values_bs) + return torch.cat(new_pixel_values, dim = 0) + if vae_stream_1 is not None: + vae_stream_1.wait_stream(torch.cuda.current_stream()) + with torch.cuda.stream(vae_stream_1): + latents = _batch_encode_vae(pixel_values) + else: + latents = _batch_encode_vae(pixel_values) + + if args.train_mode != "normal": + mask = rearrange(mask, "b f c h w -> b c f h w") + mask = torch.concat( + [ + torch.repeat_interleave(mask[:, :, 0:1], repeats=4, dim=2), + mask[:, :, 1:] + ], dim=2 + ) + mask = mask.view(mask.shape[0], mask.shape[2] // 4, 4, mask.shape[3], mask.shape[4]) + mask = mask.transpose(1, 2) + mask = resize_mask(1 - mask, latents) + + # Encode inpaint latents. + mask_latents = _batch_encode_vae(mask_pixel_values) + if vae_stream_2 is not None: + torch.cuda.current_stream().wait_stream(vae_stream_2) + + inpaint_latents = torch.concat([mask, mask_latents], dim=1) + inpaint_latents = t2v_flag[:, None, None, None, None] * inpaint_latents + + clip_context = [] + for clip_pixel_value in clip_pixel_values: + clip_image = Image.fromarray(np.uint8(clip_pixel_value.float().cpu().numpy())) + clip_image = TF.to_tensor(clip_image).sub_(0.5).div_(0.5).to(clip_image_encoder.device, weight_dtype) + _clip_context = clip_image_encoder([clip_image[:, None, :, :]]) + clip_context.append(_clip_context) + clip_context = torch.cat(clip_context) + + # wait for latents = vae.encode(pixel_values) to complete + if vae_stream_1 is not None: + torch.cuda.current_stream().wait_stream(vae_stream_1) + + if args.low_vram: + vae.to('cpu') + if args.train_mode != "normal": + clip_image_encoder.to('cpu') + torch.cuda.empty_cache() + if not args.enable_text_encoder_in_dataloader: + text_encoder.to(accelerator.device) + + if args.enable_text_encoder_in_dataloader: + prompt_embeds = batch['encoder_hidden_states'].to(device=latents.device) + else: + with torch.no_grad(): + prompt_ids = tokenizer( + batch['text'], + padding="max_length", + max_length=args.tokenizer_max_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt" + ) + text_input_ids = prompt_ids.input_ids + prompt_attention_mask = prompt_ids.attention_mask + + seq_lens = prompt_attention_mask.gt(0).sum(dim=1).long() + prompt_embeds = text_encoder(text_input_ids.to(latents.device), attention_mask=prompt_attention_mask.to(latents.device))[0] + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + + if args.low_vram and not args.enable_text_encoder_in_dataloader: + text_encoder.to('cpu') + torch.cuda.empty_cache() + + bsz, channel, num_frames, height, width = latents.size() + noise = torch.randn(latents.size(), device=latents.device, generator=torch_rng, dtype=weight_dtype) + + if not args.uniform_sampling: + u = compute_density_for_timestep_sampling( + weighting_scheme=args.weighting_scheme, + batch_size=bsz, + logit_mean=args.logit_mean, + logit_std=args.logit_std, + mode_scale=args.mode_scale, + ) + indices = (u * noise_scheduler.config.num_train_timesteps).long() + else: + # Sample a random timestep for each image + # timesteps = generate_timestep_with_lognorm(0, args.train_sampling_steps, (bsz,), device=latents.device, generator=torch_rng) + # timesteps = torch.randint(0, args.train_sampling_steps, (bsz,), device=latents.device, generator=torch_rng) + indices = idx_sampling(bsz, generator=torch_rng, device=latents.device) + indices = indices.long().cpu() + timesteps = noise_scheduler.timesteps[indices].to(device=latents.device) + + def get_sigmas(timesteps, n_dim=4, dtype=torch.float32): + sigmas = noise_scheduler.sigmas.to(device=accelerator.device, dtype=dtype) + schedule_timesteps = noise_scheduler.timesteps.to(accelerator.device) + timesteps = timesteps.to(accelerator.device) + step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < n_dim: + sigma = sigma.unsqueeze(-1) + return sigma + + # Add noise according to flow matching. + # zt = (1 - texp) * x + texp * z1 + sigmas = get_sigmas(timesteps, n_dim=latents.ndim, dtype=latents.dtype) + noisy_latents = (1.0 - sigmas) * latents + sigmas * noise + + # Add noise + target = noise - latents + + target_shape = (vae.latent_channels, num_frames, width, height) + seq_len = math.ceil( + (target_shape[2] * target_shape[3]) / + (accelerator.unwrap_model(transformer3d).config.patch_size[1] * accelerator.unwrap_model(transformer3d).config.patch_size[2]) * + target_shape[1] + ) + # Predict the noise residual + with torch.cuda.amp.autocast(dtype=weight_dtype), torch.cuda.device(device=accelerator.device): + noise_pred = transformer3d( + x=noisy_latents, + context=prompt_embeds, + t=timesteps, + seq_len=seq_len, + y=inpaint_latents if args.train_mode != "normal" else None, + clip_fea=clip_context if args.train_mode != "normal" else None, + ) + + def custom_mse_loss(noise_pred, target, weighting=None, threshold=50): + noise_pred = noise_pred.float() + target = target.float() + diff = noise_pred - target + mse_loss = F.mse_loss(noise_pred, target, reduction='none') + mask = (diff.abs() <= threshold).float() + masked_loss = mse_loss * mask + if weighting is not None: + masked_loss = masked_loss * weighting + final_loss = masked_loss.mean() + return final_loss + + weighting = compute_loss_weighting_for_sd3(weighting_scheme=args.weighting_scheme, sigmas=sigmas) + loss = custom_mse_loss(noise_pred.float(), target.float(), weighting.float()) + loss = loss.mean() + + if args.motion_sub_loss and noise_pred.size()[1] > 2: + gt_sub_noise = noise_pred[:, 1:, :].float() - noise_pred[:, :-1, :].float() + pre_sub_noise = target[:, 1:, :].float() - target[:, :-1, :].float() + sub_loss = F.mse_loss(gt_sub_noise, pre_sub_noise, reduction="mean") + loss = loss * (1 - args.motion_sub_loss_ratio) + sub_loss * args.motion_sub_loss_ratio + + # Gather the losses across all processes for logging (if we use distributed training). + avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean() + train_loss += avg_loss.item() / args.gradient_accumulation_steps + + # Backpropagate + accelerator.backward(loss) + if accelerator.sync_gradients: + accelerator.clip_grad_norm_(trainable_params, args.max_grad_norm) + optimizer.step() + lr_scheduler.step() + optimizer.zero_grad() + + # Checks if the accelerator has performed an optimization step behind the scenes + if accelerator.sync_gradients: + progress_bar.update(1) + global_step += 1 + accelerator.log({"train_loss": train_loss}, step=global_step) + train_loss = 0.0 + + if global_step % args.checkpointing_steps == 0: + if args.use_deepspeed or args.use_fsdp or accelerator.is_main_process: + # _before_ saving state, check if this save would set us over the `checkpoints_total_limit` + if args.checkpoints_total_limit is not None: + checkpoints = os.listdir(args.output_dir) + checkpoints = [d for d in checkpoints if d.startswith("checkpoint")] + checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1])) + + # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints + if len(checkpoints) >= args.checkpoints_total_limit: + num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1 + removing_checkpoints = checkpoints[0:num_to_remove] + + logger.info( + f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints" + ) + logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}") + + for removing_checkpoint in removing_checkpoints: + removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint) + shutil.rmtree(removing_checkpoint) + if not args.save_state: + safetensor_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}.safetensors") + save_model(safetensor_save_path, accelerator.unwrap_model(network)) + logger.info(f"Saved safetensor to {safetensor_save_path}") + else: + accelerator_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(accelerator_save_path) + logger.info(f"Saved state to {accelerator_save_path}") + + if accelerator.is_main_process: + if args.validation_prompts is not None and global_step % args.validation_steps == 0: + log_validation( + vae, + text_encoder, + tokenizer, + clip_image_encoder, + transformer3d, + network, + config, + args, + accelerator, + weight_dtype, + global_step, + ) + + logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]} + progress_bar.set_postfix(**logs) + + if global_step >= args.max_train_steps: + break + + if accelerator.is_main_process: + if args.validation_prompts is not None and epoch % args.validation_epochs == 0: + log_validation( + vae, + text_encoder, + tokenizer, + clip_image_encoder, + transformer3d, + network, + config, + args, + accelerator, + weight_dtype, + global_step, + ) + + # Create the pipeline using the trained modules and save it. + accelerator.wait_for_everyone() + if args.use_deepspeed or args.use_fsdp or accelerator.is_main_process: + if not args.save_state: + safetensor_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}.safetensors") + save_model(safetensor_save_path, accelerator.unwrap_model(network)) + else: + accelerator_save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(accelerator_save_path) + logger.info(f"Saved state to {accelerator_save_path}") + + accelerator.end_training() + + +if __name__ == "__main__": + main() diff --git a/scripts/wan2.2/train_lora_i2v.sh b/scripts/wan2.2/train_lora_i2v.sh new file mode 100755 index 0000000..facb05e --- /dev/null +++ b/scripts/wan2.2/train_lora_i2v.sh @@ -0,0 +1,40 @@ +export MODEL_NAME="models/Diffusion_Transformer/Wan2.2-I2V-A14B" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --mixed_precision="bf16" scripts/wan2.2/train_lora.py \ + --config_path="config/wan2.2/wan_civitai_i2v.yaml" \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --image_sample_size=1024 \ + --video_sample_size=256 \ + --token_sample_size=512 \ + --video_sample_stride=2 \ + --video_sample_n_frames=81 \ + --train_batch_size=1 \ + --video_repeat=1 \ + --gradient_accumulation_steps=1 \ + --dataloader_num_workers=8 \ + --num_train_epochs=100 \ + --checkpointing_steps=50 \ + --learning_rate=1e-04 \ + --seed=42 \ + --output_dir="output_dir" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --random_hw_adapt \ + --training_with_video_token_length \ + --enable_bucket \ + --uniform_sampling \ + --boundary_type="low" \ + --train_mode="i2v" \ + --low_vram diff --git a/scripts/wan2.2/train_lora_t2v.sh b/scripts/wan2.2/train_lora_t2v.sh new file mode 100755 index 0000000..8a1853c --- /dev/null +++ b/scripts/wan2.2/train_lora_t2v.sh @@ -0,0 +1,40 @@ +export MODEL_NAME="models/Diffusion_Transformer/Wan2.2-T2V-A14B" +export DATASET_NAME="datasets/internal_datasets/" +export DATASET_META_NAME="datasets/internal_datasets/metadata.json" +# NCCL_IB_DISABLE=1 and NCCL_P2P_DISABLE=1 are used in multi nodes without RDMA. +# export NCCL_IB_DISABLE=1 +# export NCCL_P2P_DISABLE=1 +NCCL_DEBUG=INFO + +accelerate launch --mixed_precision="bf16" scripts/wan2.2/train_lora.py \ + --config_path="config/wan2.2/wan_civitai_t2v.yaml" \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATASET_NAME \ + --train_data_meta=$DATASET_META_NAME \ + --image_sample_size=1024 \ + --video_sample_size=256 \ + --token_sample_size=512 \ + --video_sample_stride=2 \ + --video_sample_n_frames=81 \ + --train_batch_size=1 \ + --video_repeat=1 \ + --gradient_accumulation_steps=1 \ + --dataloader_num_workers=8 \ + --num_train_epochs=100 \ + --checkpointing_steps=50 \ + --learning_rate=1e-04 \ + --seed=42 \ + --output_dir="output_dir" \ + --gradient_checkpointing \ + --mixed_precision="bf16" \ + --adam_weight_decay=3e-2 \ + --adam_epsilon=1e-10 \ + --vae_mini_batch=1 \ + --max_grad_norm=0.05 \ + --random_hw_adapt \ + --training_with_video_token_length \ + --enable_bucket \ + --uniform_sampling \ + --boundary_type="low" \ + --train_mode="normal" \ + --low_vram diff --git a/videox_fun/data/dataset_image_video.py b/videox_fun/data/dataset_image_video.py index 3ea5b75..dfaded0 100755 --- a/videox_fun/data/dataset_image_video.py +++ b/videox_fun/data/dataset_image_video.py @@ -20,6 +20,7 @@ from einops import rearrange from func_timeout import FunctionTimedOut, func_timeout from packaging import version as pver from PIL import Image +from safetensors.torch import load_file from torch.utils.data import BatchSampler, Sampler from torch.utils.data.dataset import Dataset @@ -344,6 +345,7 @@ class ImageVideoDataset(Dataset): video_length_drop_start=0.0, video_length_drop_end=1.0, enable_inpaint=False, + return_file_name=False, ): # Loading annotations from files print(f"loading annotations from {ann_path} ...") @@ -375,7 +377,8 @@ class ImageVideoDataset(Dataset): # TODO: enable bucket training self.enable_bucket = enable_bucket self.text_drop_ratio = text_drop_ratio - self.enable_inpaint = enable_inpaint + self.enable_inpaint = enable_inpaint + self.return_file_name = return_file_name self.video_length_drop_start = video_length_drop_start self.video_length_drop_end = video_length_drop_end @@ -456,7 +459,7 @@ class ImageVideoDataset(Dataset): # Random use no text generation if random.random() < self.text_drop_ratio: text = '' - return pixel_values, text, 'video' + return pixel_values, text, 'video', video_dir else: image_path, text = data_info['file_path'], data_info['text'] if self.data_root is not None: @@ -468,7 +471,7 @@ class ImageVideoDataset(Dataset): image = np.expand_dims(np.array(image), 0) if random.random() < self.text_drop_ratio: text = '' - return image, text, 'image' + return image, text, 'image', image_path def __len__(self): return self.length @@ -484,11 +487,13 @@ class ImageVideoDataset(Dataset): if data_type_local != data_type: raise ValueError("data_type_local != data_type") - pixel_values, name, data_type = self.get_batch(idx) + pixel_values, name, data_type, file_path = self.get_batch(idx) sample["pixel_values"] = pixel_values sample["text"] = name sample["data_type"] = data_type sample["idx"] = idx + if self.return_file_name: + sample["file_name"] = os.path.basename(file_path) if len(sample) > 0: break @@ -498,7 +503,7 @@ class ImageVideoDataset(Dataset): if self.enable_inpaint and not self.enable_bucket: mask = get_random_mask(pixel_values.size()) - mask_pixel_values = pixel_values * (1 - mask) + torch.ones_like(pixel_values) * mask + mask_pixel_values = pixel_values * (1 - mask) + torch.ones_like(pixel_values) * -1 * mask sample["mask_pixel_values"] = mask_pixel_values sample["mask"] = mask @@ -521,6 +526,7 @@ class ImageVideoControlDataset(Dataset): video_length_drop_end=0.9, enable_inpaint=False, enable_camera_info=False, + return_file_name=False, ): # Loading annotations from files print(f"loading annotations from {ann_path} ...") @@ -552,8 +558,9 @@ class ImageVideoControlDataset(Dataset): # TODO: enable bucket training self.enable_bucket = enable_bucket self.text_drop_ratio = text_drop_ratio - self.enable_inpaint = enable_inpaint + self.enable_inpaint = enable_inpaint self.enable_camera_info = enable_camera_info + self.return_file_name = return_file_name self.video_length_drop_start = video_length_drop_start self.video_length_drop_end = video_length_drop_end @@ -700,7 +707,8 @@ class ImageVideoControlDataset(Dataset): control_pixel_values = self.video_transforms(control_pixel_values) control_camera_values = None - return pixel_values, control_pixel_values, control_camera_values, text, "video" + return pixel_values, control_pixel_values, control_camera_values, text, "video", video_dir + else: image_path, text = data_info['file_path'], data_info['text'] if self.data_root is not None: @@ -726,7 +734,8 @@ class ImageVideoControlDataset(Dataset): control_image = self.image_transforms(control_image).unsqueeze(0) else: control_image = np.expand_dims(np.array(control_image), 0) - return image, control_image, None, text, 'image' + return image, control_image, None, text, 'image', image_path + def __len__(self): return self.length @@ -741,7 +750,7 @@ class ImageVideoControlDataset(Dataset): if data_type_local != data_type: raise ValueError("data_type_local != data_type") - pixel_values, control_pixel_values, control_camera_values, name, data_type = self.get_batch(idx) + pixel_values, control_pixel_values, control_camera_values, name, data_type, file_path = self.get_batch(idx) sample["pixel_values"] = pixel_values sample["control_pixel_values"] = control_pixel_values @@ -752,6 +761,9 @@ class ImageVideoControlDataset(Dataset): if self.enable_camera_info: sample["control_camera_values"] = control_camera_values + if self.return_file_name: + sample["file_name"] = os.path.basename(file_path) + if len(sample) > 0: break except Exception as e: @@ -769,3 +781,24 @@ class ImageVideoControlDataset(Dataset): sample["clip_pixel_values"] = clip_pixel_values return sample + +class ImageVideoSafetensorsDataset(Dataset): + def __init__( + self, + ann_path + ): + # Loading annotations from files + print(f"loading annotations from {ann_path} ...") + if ann_path.endswith('.json'): + dataset = json.load(open(ann_path)) + + self.dataset = dataset + self.length = len(self.dataset) + print(f"data scale: {self.length}") + + def __len__(self): + return self.length + + def __getitem__(self, idx): + state_dict = load_file(self.dataset[idx]["file_path"]) + return state_dict \ No newline at end of file diff --git a/videox_fun/models/__init__.py b/videox_fun/models/__init__.py index 52ee607..70de90c 100755 --- a/videox_fun/models/__init__.py +++ b/videox_fun/models/__init__.py @@ -6,7 +6,8 @@ from .cogvideox_transformer3d import CogVideoXTransformer3DModel from .cogvideox_vae import AutoencoderKLCogVideoX from .wan_image_encoder import CLIPModel from .wan_text_encoder import WanT5EncoderModel -from .wan_transformer3d import WanSelfAttention, WanTransformer3DModel +from .wan_transformer3d import (Wan2_2Transformer3DModel, WanSelfAttention, + WanTransformer3DModel) from .wan_vae import AutoencoderKLWan, AutoencoderKLWan_ # The pai_fuser is an internally developed acceleration package, which can be used on PAI. diff --git a/videox_fun/models/wan_transformer3d.py b/videox_fun/models/wan_transformer3d.py index d58ce97..041cd3d 100755 --- a/videox_fun/models/wan_transformer3d.py +++ b/videox_fun/models/wan_transformer3d.py @@ -554,9 +554,31 @@ class WanI2VCrossAttention(WanSelfAttention): return x +class WanCrossAttention(WanSelfAttention): + def forward(self, x, context, context_lens, dtype, t): + r""" + Args: + x(Tensor): Shape [B, L1, C] + context(Tensor): Shape [B, L2, C] + context_lens(Tensor): Shape [B] + """ + b, n, d = x.size(0), self.num_heads, self.head_dim + # compute query, key, value + q = self.norm_q(self.q(x)).view(b, -1, n, d) + k = self.norm_k(self.k(context)).view(b, -1, n, d) + v = self.v(context).view(b, -1, n, d) + # compute attention + x = attention(q, k, v, k_lens=context_lens) + # output + x = x.flatten(2) + x = self.o(x) + return x + + WAN_CROSSATTENTION_CLASSES = { 't2v_cross_attn': WanT2VCrossAttention, 'i2v_cross_attn': WanI2VCrossAttention, + 'cross_attn': WanCrossAttention, } @@ -841,6 +863,7 @@ class WanTransformer3DModel(ModelMixin, ConfigMixin, FromOriginalModelMixin): self.gradient_checkpointing = False self.sp_world_size = 1 self.sp_world_rank = 0 + self.init_weights() def _set_gradient_checkpointing(self, *args, **kwargs): if "value" in kwargs: @@ -952,8 +975,9 @@ class WanTransformer3DModel(ModelMixin, ConfigMixin, FromOriginalModelMixin): List[Tensor]: List of denoised video tensors with original input shapes [C_out, F, H / 8, W / 8] """ - if self.model_type == 'i2v': - assert clip_fea is not None and y is not None + # Wan2.2 don't need a clip. + # if self.model_type == 'i2v': + # assert clip_fea is not None and y is not None # params device = self.patch_embedding.weight.device dtype = x.dtype @@ -1352,4 +1376,112 @@ class WanTransformer3DModel(ModelMixin, ConfigMixin, FromOriginalModelMixin): print(f"### attn1 Parameters: {sum(params) / 1e6} M") model = model.to(torch_dtype) - return model \ No newline at end of file + return model + + +class Wan2_2Transformer3DModel(WanTransformer3DModel): + r""" + Wan diffusion backbone supporting both text-to-video and image-to-video. + """ + + # ignore_for_config = [ + # 'patch_size', 'cross_attn_norm', 'qk_norm', 'text_dim', 'window_size' + # ] + # _no_split_modules = ['WanAttentionBlock'] + _supports_gradient_checkpointing = True + + @register_to_config + def __init__(self, + model_type='t2v', + patch_size=(1, 2, 2), + text_len=512, + in_dim=16, + dim=2048, + ffn_dim=8192, + freq_dim=256, + text_dim=4096, + out_dim=16, + num_heads=16, + num_layers=32, + window_size=(-1, -1), + qk_norm=True, + cross_attn_norm=True, + eps=1e-6): + r""" + Initialize the diffusion model backbone. + Args: + model_type (`str`, *optional*, defaults to 't2v'): + Model variant - 't2v' (text-to-video) or 'i2v' (image-to-video) + patch_size (`tuple`, *optional*, defaults to (1, 2, 2)): + 3D patch dimensions for video embedding (t_patch, h_patch, w_patch) + text_len (`int`, *optional*, defaults to 512): + Fixed length for text embeddings + in_dim (`int`, *optional*, defaults to 16): + Input video channels (C_in) + dim (`int`, *optional*, defaults to 2048): + Hidden dimension of the transformer + ffn_dim (`int`, *optional*, defaults to 8192): + Intermediate dimension in feed-forward network + freq_dim (`int`, *optional*, defaults to 256): + Dimension for sinusoidal time embeddings + text_dim (`int`, *optional*, defaults to 4096): + Input dimension for text embeddings + out_dim (`int`, *optional*, defaults to 16): + Output video channels (C_out) + num_heads (`int`, *optional*, defaults to 16): + Number of attention heads + num_layers (`int`, *optional*, defaults to 32): + Number of transformer blocks + window_size (`tuple`, *optional*, defaults to (-1, -1)): + Window size for local attention (-1 indicates global attention) + qk_norm (`bool`, *optional*, defaults to True): + Enable query/key normalization + cross_attn_norm (`bool`, *optional*, defaults to False): + Enable cross-attention normalization + eps (`float`, *optional*, defaults to 1e-6): + Epsilon value for normalization layers + """ + super().__init__() + assert model_type in ['t2v', 'i2v', 'ti2v'] + self.model_type = model_type + self.patch_size = patch_size + self.text_len = text_len + self.in_dim = in_dim + self.dim = dim + self.ffn_dim = ffn_dim + self.freq_dim = freq_dim + self.text_dim = text_dim + self.out_dim = out_dim + self.num_heads = num_heads + self.num_layers = num_layers + self.window_size = window_size + self.qk_norm = qk_norm + self.cross_attn_norm = cross_attn_norm + self.eps = eps + # embeddings + self.patch_embedding = nn.Conv3d( + in_dim, dim, kernel_size=patch_size, stride=patch_size) + self.text_embedding = nn.Sequential( + nn.Linear(text_dim, dim), nn.GELU(approximate='tanh'), + nn.Linear(dim, dim)) + self.time_embedding = nn.Sequential( + nn.Linear(freq_dim, dim), nn.SiLU(), nn.Linear(dim, dim)) + self.time_projection = nn.Sequential(nn.SiLU(), nn.Linear(dim, dim * 6)) + # blocks + self.blocks = nn.ModuleList([ + WanAttentionBlock("cross_attn", dim, ffn_dim, num_heads, window_size, qk_norm, + cross_attn_norm, eps) for _ in range(num_layers) + ]) + # head + self.head = Head(dim, out_dim, patch_size, eps) + # buffers (don't use register_buffer otherwise dtype will be changed in to()) + assert (dim % num_heads) == 0 and (dim // num_heads) % 2 == 0 + d = dim // num_heads + self.freqs = torch.cat([ + rope_params(1024, d - 4 * (d // 6)), + rope_params(1024, 2 * (d // 6)), + rope_params(1024, 2 * (d // 6)) + ], + dim=1) + # initialize weights + self.init_weights() \ No newline at end of file diff --git a/videox_fun/pipeline/__init__.py b/videox_fun/pipeline/__init__.py index 2b01503..786ed8d 100755 --- a/videox_fun/pipeline/__init__.py +++ b/videox_fun/pipeline/__init__.py @@ -5,6 +5,8 @@ from .pipeline_wan_fun import WanFunPipeline from .pipeline_wan_fun_inpaint import WanFunInpaintPipeline from .pipeline_wan_fun_control import WanFunControlPipeline from .pipeline_wan_phantom import WanFunPhantomPipeline +from .pipeline_wan2_2 import Wan2_2Pipeline +from .pipeline_wan2_2_i2v import Wan2_2I2VPipeline WanPipeline = WanFunPipeline WanI2VPipeline = WanFunInpaintPipeline @@ -19,4 +21,6 @@ if importlib.util.find_spec("pai_fuser") is not None: WanFunControlPipeline.__call__ = sparse_reset(WanFunControlPipeline.__call__) WanI2VPipeline.__call__ = sparse_reset(WanI2VPipeline.__call__) WanPipeline.__call__ = sparse_reset(WanPipeline.__call__) - WanFunPhantomPipeline.__call__ = sparse_reset(WanFunPhantomPipeline.__call__) \ No newline at end of file + WanFunPhantomPipeline.__call__ = sparse_reset(WanFunPhantomPipeline.__call__) + Wan2_2Pipeline.__call__ = sparse_reset(Wan2_2Pipeline.__call__) + Wan2_2I2VPipeline.__call__ = sparse_reset(Wan2_2I2VPipeline.__call__) \ No newline at end of file diff --git a/videox_fun/pipeline/pipeline_wan2_2.py b/videox_fun/pipeline/pipeline_wan2_2.py new file mode 100755 index 0000000..b6d975f --- /dev/null +++ b/videox_fun/pipeline/pipeline_wan2_2.py @@ -0,0 +1,591 @@ +import inspect +import math +from dataclasses import dataclass +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import numpy as np +import torch +from diffusers import FlowMatchEulerDiscreteScheduler +from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.utils import BaseOutput, logging, replace_example_docstring +from diffusers.utils.torch_utils import randn_tensor +from diffusers.video_processor import VideoProcessor + +from ..models import (AutoencoderKLWan, AutoTokenizer, + WanT5EncoderModel, Wan2_2Transformer3DModel) +from ..utils.fm_solvers import (FlowDPMSolverMultistepScheduler, + get_sampling_sigmas) +from ..utils.fm_solvers_unipc import FlowUniPCMultistepScheduler + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + pass + ``` +""" + + +# 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, +): + """ + 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 + + +@dataclass +class WanPipelineOutput(BaseOutput): + r""" + Output class for CogVideo pipelines. + + Args: + video (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]): + List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)`. + """ + + videos: torch.Tensor + + +class Wan2_2Pipeline(DiffusionPipeline): + r""" + Pipeline for text-to-video generation using Wan. + + 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.) + """ + + _optional_components = ["transformer_2"] + model_cpu_offload_seq = "text_encoder->transformer->transformer_2->vae" + + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + ] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: WanT5EncoderModel, + vae: AutoencoderKLWan, + transformer: Wan2_2Transformer3DModel, + transformer_2: Wan2_2Transformer3DModel = None, + scheduler: FlowMatchEulerDiscreteScheduler = None, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, + transformer_2=transformer_2, scheduler=scheduler + ) + self.video_processor = VideoProcessor(vae_scale_factor=self.vae.spacial_compression_ratio) + + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 512, + 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 + prompt_attention_mask = text_inputs.attention_mask + 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}" + ) + + seq_lens = prompt_attention_mask.gt(0).sum(dim=1).long() + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask.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 [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + + 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 = 512, + 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_channels_latents, + (num_frames - 1) // self.vae.temporal_compression_ratio + 1, + height // self.vae.spacial_compression_ratio, + width // self.vae.spacial_compression_ratio, + ) + + 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 + if hasattr(self.scheduler, "init_noise_sigma"): + latents = latents * self.scheduler.init_noise_sigma + return latents + + def decode_latents(self, latents: torch.Tensor) -> torch.Tensor: + frames = self.vae.decode(latents.to(self.vae.dtype)).sample + frames = (frames / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16 + frames = frames.cpu().float().numpy() + 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}." + ) + + @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, + 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 = "numpy", + return_dict: bool = False, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 512, + boundary: float = 0.875, + comfyui_progressbar: bool = False, + shift: int = 5, + ) -> Union[WanPipelineOutput, Tuple]: + """ + Function invoked when calling the pipeline for generation. + Args: + + Examples: + + Returns: + + """ + + 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 + weight_dtype = self.text_encoder.dtype + + # 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: + in_prompt_embeds = negative_prompt_embeds + prompt_embeds + else: + in_prompt_embeds = prompt_embeds + + # 4. Prepare timesteps + if isinstance(self.scheduler, FlowMatchEulerDiscreteScheduler): + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps, mu=1) + elif isinstance(self.scheduler, FlowUniPCMultistepScheduler): + self.scheduler.set_timesteps(num_inference_steps, device=device, shift=shift) + timesteps = self.scheduler.timesteps + elif isinstance(self.scheduler, FlowDPMSolverMultistepScheduler): + sampling_sigmas = get_sampling_sigmas(num_inference_steps, shift) + timesteps, _ = retrieve_timesteps( + self.scheduler, + device=device, + sigmas=sampling_sigmas) + else: + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps) + self._num_timesteps = len(timesteps) + if comfyui_progressbar: + from comfy.utils import ProgressBar + pbar = ProgressBar(num_inference_steps + 1) + + # 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, + weight_dtype, + device, + generator, + latents, + ) + if comfyui_progressbar: + pbar.update(1) + + # 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) + + target_shape = (self.vae.latent_channels, (num_frames - 1) // self.vae.temporal_compression_ratio + 1, width // self.vae.spacial_compression_ratio, height // self.vae.spacial_compression_ratio) + seq_len = math.ceil((target_shape[2] * target_shape[3]) / (self.transformer.config.patch_size[1] * self.transformer.config.patch_size[2]) * target_shape[1]) + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self.transformer.num_inference_steps = num_inference_steps + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + self.transformer.current_steps = i + + if self.interrupt: + continue + + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + if hasattr(self.scheduler, "scale_model_input"): + 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]) + + if self.transformer_2 is not None: + if t >= boundary * self.scheduler.config.num_train_timesteps: + local_transformer = self.transformer_2 + else: + local_transformer = self.transformer + else: + local_transformer = self.transformer + + # predict noise model_output + with torch.cuda.amp.autocast(dtype=weight_dtype), torch.cuda.device(device=device): + noise_pred = local_transformer( + x=latent_model_input, + context=in_prompt_embeds, + t=timestep, + seq_len=seq_len, + ) + + # perform guidance + if do_classifier_free_guidance: + if self.transformer_2 is not None and (isinstance(self.guidance_scale, (list, tuple))): + sample_guide_scale = self.guidance_scale[1] if t >= self.transformer_2.config.boundary * self.scheduler.config.num_train_timesteps else self.guidance_scale[0] + else: + sample_guide_scale = self.guidance_scale + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + sample_guide_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + 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 comfyui_progressbar: + pbar.update(1) + + if output_type == "numpy": + video = self.decode_latents(latents) + elif 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: + video = torch.from_numpy(video) + + return WanPipelineOutput(videos=video) diff --git a/videox_fun/pipeline/pipeline_wan2_2_i2v.py b/videox_fun/pipeline/pipeline_wan2_2_i2v.py new file mode 100644 index 0000000..dbf9026 --- /dev/null +++ b/videox_fun/pipeline/pipeline_wan2_2_i2v.py @@ -0,0 +1,733 @@ +import inspect +import math +from dataclasses import dataclass +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import numpy as np +import torch +import torch.nn.functional as F +import torchvision.transforms.functional as TF +from diffusers import FlowMatchEulerDiscreteScheduler +from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback +from diffusers.image_processor import VaeImageProcessor +from diffusers.models.embeddings import get_1d_rotary_pos_embed +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.utils import BaseOutput, logging, replace_example_docstring +from diffusers.utils.torch_utils import randn_tensor +from diffusers.video_processor import VideoProcessor +from einops import rearrange +from PIL import Image +from transformers import T5Tokenizer + +from ..models import (AutoencoderKLWan, AutoTokenizer, CLIPModel, + WanT5EncoderModel, Wan2_2Transformer3DModel) +from ..utils.fm_solvers import (FlowDPMSolverMultistepScheduler, + get_sampling_sigmas) +from ..utils.fm_solvers_unipc import FlowUniPCMultistepScheduler + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + pass + ``` +""" + + +# 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, +): + """ + 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 + + +def resize_mask(mask, latent, process_first_frame_only=True): + latent_size = latent.size() + batch_size, channels, num_frames, height, width = mask.shape + + if process_first_frame_only: + target_size = list(latent_size[2:]) + target_size[0] = 1 + first_frame_resized = F.interpolate( + mask[:, :, 0:1, :, :], + size=target_size, + mode='trilinear', + align_corners=False + ) + + target_size = list(latent_size[2:]) + target_size[0] = target_size[0] - 1 + if target_size[0] != 0: + remaining_frames_resized = F.interpolate( + mask[:, :, 1:, :, :], + size=target_size, + mode='trilinear', + align_corners=False + ) + resized_mask = torch.cat([first_frame_resized, remaining_frames_resized], dim=2) + else: + resized_mask = first_frame_resized + else: + target_size = list(latent_size[2:]) + resized_mask = F.interpolate( + mask, + size=target_size, + mode='trilinear', + align_corners=False + ) + return resized_mask + + +@dataclass +class WanPipelineOutput(BaseOutput): + r""" + Output class for CogVideo pipelines. + + Args: + video (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]): + List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)`. + """ + + videos: torch.Tensor + + +class Wan2_2I2VPipeline(DiffusionPipeline): + r""" + Pipeline for text-to-video generation using Wan. + + 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.) + """ + + _optional_components = ["transformer_2"] + model_cpu_offload_seq = "text_encoder->transformer->transformer_2->vae" + + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + ] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: WanT5EncoderModel, + vae: AutoencoderKLWan, + transformer: Wan2_2Transformer3DModel, + transformer_2: Wan2_2Transformer3DModel = None, + scheduler: FlowMatchEulerDiscreteScheduler = None, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, + transformer_2=transformer_2, scheduler=scheduler + ) + self.video_processor = VideoProcessor(vae_scale_factor=self.vae.spacial_compression_ratio) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae.spacial_compression_ratio) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae.spacial_compression_ratio, do_normalize=False, do_binarize=True, do_convert_grayscale=True + ) + + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 512, + 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 + prompt_attention_mask = text_inputs.attention_mask + 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}" + ) + + seq_lens = prompt_attention_mask.gt(0).sum(dim=1).long() + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask.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 [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + + 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 = 512, + 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_channels_latents, + (num_frames - 1) // self.vae.temporal_compression_ratio + 1, + height // self.vae.spacial_compression_ratio, + width // self.vae.spacial_compression_ratio, + ) + + 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 + if hasattr(self.scheduler, "init_noise_sigma"): + latents = latents * self.scheduler.init_noise_sigma + return latents + + + def prepare_mask_latents( + self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance, noise_aug_strength + ): + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + + if mask is not None: + mask = mask.to(device=device, dtype=self.vae.dtype) + bs = 1 + new_mask = [] + for i in range(0, mask.shape[0], bs): + mask_bs = mask[i : i + bs] + mask_bs = self.vae.encode(mask_bs)[0] + mask_bs = mask_bs.mode() + new_mask.append(mask_bs) + mask = torch.cat(new_mask, dim = 0) + # mask = mask * self.vae.config.scaling_factor + + if masked_image is not None: + masked_image = masked_image.to(device=device, dtype=self.vae.dtype) + bs = 1 + new_mask_pixel_values = [] + for i in range(0, masked_image.shape[0], bs): + mask_pixel_values_bs = masked_image[i : i + bs] + mask_pixel_values_bs = self.vae.encode(mask_pixel_values_bs)[0] + mask_pixel_values_bs = mask_pixel_values_bs.mode() + new_mask_pixel_values.append(mask_pixel_values_bs) + masked_image_latents = torch.cat(new_mask_pixel_values, dim = 0) + # masked_image_latents = masked_image_latents * self.vae.config.scaling_factor + else: + masked_image_latents = None + + return mask, masked_image_latents + + def decode_latents(self, latents: torch.Tensor) -> torch.Tensor: + frames = self.vae.decode(latents.to(self.vae.dtype)).sample + frames = (frames / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16 + frames = frames.cpu().float().numpy() + 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}." + ) + + @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, + video: Union[torch.FloatTensor] = None, + mask_video: Union[torch.FloatTensor] = None, + num_frames: int = 49, + num_inference_steps: int = 50, + timesteps: Optional[List[int]] = None, + guidance_scale: float = 6, + 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 = "numpy", + return_dict: bool = False, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 512, + boundary: float = 0.875, + comfyui_progressbar: bool = False, + shift: int = 5, + ) -> Union[WanPipelineOutput, Tuple]: + """ + Function invoked when calling the pipeline for generation. + Args: + + Examples: + + Returns: + + """ + 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 + weight_dtype = self.text_encoder.dtype + + # 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: + in_prompt_embeds = negative_prompt_embeds + prompt_embeds + else: + in_prompt_embeds = prompt_embeds + + # 4. Prepare timesteps + if isinstance(self.scheduler, FlowMatchEulerDiscreteScheduler): + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps, mu=1) + elif isinstance(self.scheduler, FlowUniPCMultistepScheduler): + self.scheduler.set_timesteps(num_inference_steps, device=device, shift=shift) + timesteps = self.scheduler.timesteps + elif isinstance(self.scheduler, FlowDPMSolverMultistepScheduler): + sampling_sigmas = get_sampling_sigmas(num_inference_steps, shift) + timesteps, _ = retrieve_timesteps( + self.scheduler, + device=device, + sigmas=sampling_sigmas) + else: + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps) + self._num_timesteps = len(timesteps) + if comfyui_progressbar: + from comfy.utils import ProgressBar + pbar = ProgressBar(num_inference_steps + 2) + + # 5. Prepare latents. + if video is not None: + video_length = video.shape[2] + init_video = self.image_processor.preprocess(rearrange(video, "b c f h w -> (b f) c h w"), height=height, width=width) + init_video = init_video.to(dtype=torch.float32) + init_video = rearrange(init_video, "(b f) c h w -> b c f h w", f=video_length) + else: + init_video = None + + latent_channels = self.vae.config.latent_channels + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + latent_channels, + num_frames, + height, + width, + weight_dtype, + device, + generator, + latents, + ) + if comfyui_progressbar: + pbar.update(1) + + # Prepare mask latent variables + if init_video is not None: + if (mask_video == 255).all(): + mask_latents = torch.tile( + torch.zeros_like(latents)[:, :1].to(device, weight_dtype), [1, 4, 1, 1, 1] + ) + masked_video_latents = torch.zeros_like(latents).to(device, weight_dtype) + else: + bs, _, video_length, height, width = video.size() + mask_condition = self.mask_processor.preprocess(rearrange(mask_video, "b c f h w -> (b f) c h w"), height=height, width=width) + mask_condition = mask_condition.to(dtype=torch.float32) + mask_condition = rearrange(mask_condition, "(b f) c h w -> b c f h w", f=video_length) + + masked_video = init_video * (torch.tile(mask_condition, [1, 3, 1, 1, 1]) < 0.5) + _, masked_video_latents = self.prepare_mask_latents( + None, + masked_video, + batch_size, + height, + width, + weight_dtype, + device, + generator, + do_classifier_free_guidance, + noise_aug_strength=None, + ) + + mask_condition = torch.concat( + [ + torch.repeat_interleave(mask_condition[:, :, 0:1], repeats=4, dim=2), + mask_condition[:, :, 1:] + ], dim=2 + ) + mask_condition = mask_condition.view(bs, mask_condition.shape[2] // 4, 4, height, width) + mask_condition = mask_condition.transpose(1, 2) + mask_latents = resize_mask(1 - mask_condition, masked_video_latents, True).to(device, weight_dtype) + + if comfyui_progressbar: + pbar.update(1) + + # 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) + + target_shape = (self.vae.latent_channels, (num_frames - 1) // self.vae.temporal_compression_ratio + 1, width // self.vae.spacial_compression_ratio, height // self.vae.spacial_compression_ratio) + seq_len = math.ceil((target_shape[2] * target_shape[3]) / (self.transformer.config.patch_size[1] * self.transformer.config.patch_size[2]) * target_shape[1]) + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self.transformer.num_inference_steps = num_inference_steps + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + self.transformer.current_steps = i + + if self.interrupt: + continue + + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + if hasattr(self.scheduler, "scale_model_input"): + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + if init_video is not None: + mask_input = torch.cat([mask_latents] * 2) if do_classifier_free_guidance else mask_latents + masked_video_latents_input = ( + torch.cat([masked_video_latents] * 2) if do_classifier_free_guidance else masked_video_latents + ) + y = torch.cat([mask_input, masked_video_latents_input], dim=1).to(device, weight_dtype) + + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + if self.transformer_2 is not None: + if t >= boundary * self.scheduler.config.num_train_timesteps: + local_transformer = self.transformer_2 + else: + local_transformer = self.transformer + else: + local_transformer = self.transformer + + # predict noise model_output + with torch.cuda.amp.autocast(dtype=weight_dtype), torch.cuda.device(device=device): + noise_pred = local_transformer( + x=latent_model_input, + context=in_prompt_embeds, + t=timestep, + seq_len=seq_len, + y=y, + ) + + # perform guidance + if do_classifier_free_guidance: + if self.transformer_2 is not None and (isinstance(self.guidance_scale, (list, tuple))): + sample_guide_scale = self.guidance_scale[1] if t >= self.transformer_2.config.boundary * self.scheduler.config.num_train_timesteps else self.guidance_scale[0] + else: + sample_guide_scale = self.guidance_scale + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + sample_guide_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + 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 comfyui_progressbar: + pbar.update(1) + + if output_type == "numpy": + video = self.decode_latents(latents) + elif 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: + video = torch.from_numpy(video) + + return WanPipelineOutput(videos=video) diff --git a/videox_fun/utils/lora_utils.py b/videox_fun/utils/lora_utils.py index 6ff52ba..c66a406 100755 --- a/videox_fun/utils/lora_utils.py +++ b/videox_fun/utils/lora_utils.py @@ -156,7 +156,7 @@ def precalculate_safetensors_hashes(tensors, metadata): class LoRANetwork(torch.nn.Module): - TRANSFORMER_TARGET_REPLACE_MODULE = ["CogVideoXTransformer3DModel", "WanTransformer3DModel"] + TRANSFORMER_TARGET_REPLACE_MODULE = ["CogVideoXTransformer3DModel", "WanTransformer3DModel", "Wan2_2Transformer3DModel"] TEXT_ENCODER_TARGET_REPLACE_MODULE = ["T5LayerSelfAttention", "T5LayerFF", "BertEncoder", "T5SelfAttention", "T5CrossAttention"] LORA_PREFIX_TRANSFORMER = "lora_unet" LORA_PREFIX_TEXT_ENCODER = "lora_te"