train.py

import torch, os, argparse, accelerate, warnings
from diffsynth.core import UnifiedDataset
from diffsynth.core.data.operators import LoadVideo, LoadAudio, ImageCropAndResize, ToAbsolutePath
from diffsynth.pipelines.wan_video import WanVideoPipeline, ModelConfig
from diffsynth.diffusion import *
os.environ["TOKENIZERS_PARALLELISM"] = "false"
from time import time
from accelerate import Accelerator, DeepSpeedPlugin
from accelerate.utils import DistributedDataParallelKwargs
import json

class WanTrainingModule(DiffusionTrainingModule):
    def __init__(
        self,
        wan_version="t2v-1.3B",
        resume_ckpt=None,
        load_path_json=None,
        trainable_models=None,
        lora_base_model=None, lora_target_modules="", lora_rank=32, lora_checkpoint=None,
        preset_lora_path=None, preset_lora_model=None,
        use_gradient_checkpointing=True,
        use_gradient_checkpointing_offload=False,
        extra_inputs=None,
        fp8_models=None,
        offload_models=None,
        device="cpu",
        task="sft",
        max_timestep_boundary=1.0,
        min_timestep_boundary=0.0,
    ):
        super().__init__()
        # Warning
        if not use_gradient_checkpointing:
            warnings.warn("Gradient checkpointing is detected as disabled. To prevent out-of-memory errors, the training framework will forcibly enable gradient checkpointing.")
            use_gradient_checkpointing = True
        
        # Load models
        with open(load_path_json) as user_file:
            model_paths = json.load(user_file)

        if resume_ckpt == "None":
            load_path = model_paths["dit"]
        else:
            load_path = resume_ckpt


        print(f"resume from {load_path}")
        # Load models
        if wan_version == "t2v-1.3B":
            self.pipe = WanVideoPipeline.from_pretrained(
                torch_dtype=torch.bfloat16,
                device=device,
                model_configs=[
                    ModelConfig(path=load_path),
                    ModelConfig(path=model_paths["t5"]),
                    ModelConfig(path=model_paths["vae"]),      
                ],
                tokenizer_config=ModelConfig(path=model_paths["tokenizer"]),
            )     

        # Load models
        elif wan_version == "t2v-14B":
            self.pipe = WanVideoPipeline.from_pretrained(
                torch_dtype=torch.bfloat16,
                device=device,
                model_configs=[
                    ModelConfig(path=load_path),
                    ModelConfig(path=model_paths["t5"]),
                    ModelConfig(path=model_paths["vae"]),  
                ],            
                tokenizer_config=ModelConfig(path=model_paths["tokenizer"]),
            )
        self.pipe = self.split_pipeline_units(task, self.pipe, trainable_models, lora_base_model)

        if lora_checkpoint == "None":
            lora_checkpoint = None
            
        # Training mode
        self.switch_pipe_to_training_mode(
            self.pipe, trainable_models,
            lora_base_model, lora_target_modules, lora_rank, lora_checkpoint,
            preset_lora_path, preset_lora_model,
            task=task,
        )
        
        # Store other configs
        self.use_gradient_checkpointing = use_gradient_checkpointing
        self.use_gradient_checkpointing_offload = use_gradient_checkpointing_offload
        self.extra_inputs = extra_inputs.split(",") if extra_inputs is not None else []
        self.fp8_models = fp8_models
        self.task = task
        self.task_to_loss = {
            "sft:data_process": lambda pipe, *args: args,
            "direct_distill:data_process": lambda pipe, *args: args,
            "sft": lambda pipe, inputs_shared, inputs_posi, inputs_nega: FlowMatchSFTLoss(pipe, **inputs_shared, **inputs_posi),
            "sft:train": lambda pipe, inputs_shared, inputs_posi, inputs_nega: FlowMatchSFTLoss(pipe, **inputs_shared, **inputs_posi),
            "direct_distill": lambda pipe, inputs_shared, inputs_posi, inputs_nega: DirectDistillLoss(pipe, **inputs_shared, **inputs_posi),
            "direct_distill:train": lambda pipe, inputs_shared, inputs_posi, inputs_nega: DirectDistillLoss(pipe, **inputs_shared, **inputs_posi),
        }
        self.max_timestep_boundary = max_timestep_boundary
        self.min_timestep_boundary = min_timestep_boundary
        
    def parse_extra_inputs(self, data, extra_inputs, inputs_shared):
        for extra_input in extra_inputs:
            if extra_input == "input_image":
                inputs_shared["input_image"] = data["video"][0]
            elif extra_input == "end_image":
                inputs_shared["end_image"] = data["video"][-1]
            elif extra_input == "reference_image" or extra_input == "vace_reference_image":
                inputs_shared[extra_input] = data[extra_input][0]
            else:
                inputs_shared[extra_input] = data[extra_input]      
        return inputs_shared
    
    def get_pipeline_inputs(self, data):
        shot_groups = data["shot_groups"]
        latent_shot_groups = data["latent_shot_groups"]

        inputs_posi = {"prompt": data["prompt"]}

        inputs_nega = {}
        inputs_shared = {
            # Assume you are using this pipeline for inference,
            # please fill in the input parameters.
            "input_video": data["video"],
            "shot_groups": shot_groups, 
            "latent_shot_groups": latent_shot_groups,
            "height": data["video"][0].shape[1],
            "width": data["video"][0].shape[2],
            "num_frames": len(data["video"]),
            # Please do not modify the following parameters
            # unless you clearly know what this will cause.
            "cfg_scale": 1,
            "tiled": False,
            "rand_device": self.pipe.device,
            "use_gradient_checkpointing": self.use_gradient_checkpointing,
            "use_gradient_checkpointing_offload": self.use_gradient_checkpointing_offload,
            "cfg_merge": False,
            "vace_scale": 1,
            "max_timestep_boundary": self.max_timestep_boundary,
            "min_timestep_boundary": self.min_timestep_boundary,
        }
        inputs_shared = self.parse_extra_inputs(data, self.extra_inputs, inputs_shared)
        return inputs_shared, inputs_posi, inputs_nega
    
    def forward(self, data, inputs=None):
        if inputs is None: inputs = self.get_pipeline_inputs(data)      
        inputs = self.transfer_data_to_device(inputs, self.pipe.device, self.pipe.torch_dtype)
        for unit in self.pipe.units:                                    
            inputs = self.pipe.unit_runner(unit, self.pipe, *inputs)    
        loss = self.task_to_loss[self.task](self.pipe, *inputs)         
        return loss


def wan_parser():
    parser = argparse.ArgumentParser(description="Simple example of a training script.")
    parser = add_general_config(parser)
    parser = add_video_size_config(parser)
    parser.add_argument("--wan_version", type=str, default="t2v-1.3B", help="wan_version.")
    parser.add_argument("--resume_ckpt", type=str, default=None, help="resume_ckpt.")
    parser.add_argument("--load_path_json", type=str, default=None, help="load_path_json.")

    parser.add_argument("--max_timestep_boundary", type=float, default=1.0, help="Max timestep boundary (for mixed models, e.g., Wan-AI/Wan2.2-I2V-A14B).")
    parser.add_argument("--min_timestep_boundary", type=float, default=0.0, help="Min timestep boundary (for mixed models, e.g., Wan-AI/Wan2.2-I2V-A14B).")
    parser.add_argument("--initialize_model_on_cpu", default=False, action="store_true", help="Whether to initialize models on CPU.")
    return parser


if __name__ == "__main__":
    parser = wan_parser()
    args = parser.parse_args()

    ds_plugin = DeepSpeedPlugin(
        zero_stage=2,
        offload_optimizer_device="cpu",
        offload_param_device="cpu",
        gradient_clipping=1.0
    )

    accelerator = accelerate.Accelerator(
        mixed_precision="bf16",
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        kwargs_handlers=[accelerate.DistributedDataParallelKwargs(find_unused_parameters=args.find_unused_parameters)],
        deepspeed_plugin=ds_plugin
    )
    
    dataset = UnifiedDataset(
        metadata_path=args.dataset_metadata_path,
        repeat=args.dataset_repeat,
        max_pixels=args.max_pixels,
        height=args.height,
        width=args.width,
        height_division_factor=16,
        width_division_factor=16,
        time_division_factor=4
    )
    model = WanTrainingModule(
        wan_version=args.wan_version,
        resume_ckpt=args.resume_ckpt,
        load_path_json=args.load_path_json,
        trainable_models=args.trainable_models,
        lora_base_model=args.lora_base_model,
        lora_target_modules=args.lora_target_modules,
        lora_rank=args.lora_rank,
        lora_checkpoint=args.lora_checkpoint,
        preset_lora_path=args.preset_lora_path,
        preset_lora_model=args.preset_lora_model,
        use_gradient_checkpointing=args.use_gradient_checkpointing,
        use_gradient_checkpointing_offload=args.use_gradient_checkpointing_offload,
        extra_inputs=args.extra_inputs,
        fp8_models=args.fp8_models,
        offload_models=args.offload_models,
        task=args.task,
        device="cpu" if args.initialize_model_on_cpu else accelerator.device,
        max_timestep_boundary=args.max_timestep_boundary,
        min_timestep_boundary=args.min_timestep_boundary,
    )
    model_logger = ModelLogger(
        args.output_path,
        remove_prefix_in_ckpt=args.remove_prefix_in_ckpt,
    )
    launcher_map = {
        "sft:data_process": launch_data_process_task,
        "direct_distill:data_process": launch_data_process_task,
        "sft": launch_training_task,
        "sft:train": launch_training_task,
        "direct_distill": launch_training_task,
        "direct_distill:train": launch_training_task,
    }
    launcher_map[args.task](accelerator, dataset, model, model_logger, args=args)