"""LingBot-Video ti2v full-parameter training script. Flow-matching training of the LingBot-Video single-stream joint DiT, mirroring the ti2v inference path of examples/lingbot_video/predict_i2v.py: - The first frame is used twice, exactly like at inference time: 1. as visual input of the frozen Qwen3-VL text encoder (<|vision_start|>...<|image_pad|>...<|vision_end|> tokens); 2. as a clean latent written into the temporal prefix of the noisy latent (inpainting-style clamping, see LingBotVideoPipeline._apply_inpainting). - Loss is computed on the non-conditioned latent frames only. - Sigma is sampled from the same shifted FlowUniPC schedule used at inference (sigma' = shift * sigma / (1 + (shift - 1) * sigma)), and the transformer receives timestep = sigma * 1000. Modified from scripts/wan2.2/train.py / scripts/qwenimage/train.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 transformers from accelerate import Accelerator from accelerate.logging import get_logger from accelerate.utils import ProjectConfiguration, set_seed 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 packaging import version from PIL import Image from torch.utils.data import RandomSampler as TorchRandomSampler from torch.utils.tensorboard import SummaryWriter from torchvision import transforms from transformers import AutoProcessor 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 import (ASPECT_RATIO_512, ASPECT_RATIO_RANDOM_CROP_512, ASPECT_RATIO_RANDOM_CROP_PROB, AspectRatioBatchImageVideoSampler, ImageVideoDataset, ImageVideoSampler, RandomSampler, get_closest_ratio) from videox_fun.models import (AutoencoderKLQwenImage, LingBotVideoTransformer3DModel, Qwen3VLForConditionalGeneration) from videox_fun.pipeline import LingBotVideoI2VPipeline from videox_fun.pipeline.pipeline_lingbot_video_i2v import (SPATIAL_MERGE_SIZE, smart_resize) from videox_fun.utils.discrete_sampler import DiscreteSampling from videox_fun.utils.fm_solvers_unipc import FlowUniPCMultistepScheduler from videox_fun.utils.fsdp_ema import FSDPEMA from videox_fun.utils.tqdm_bar import PauseAwareTqdm from videox_fun.utils.utils import calculate_dimensions, 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 build_sigma_table(sigma_max, sigma_min, num_steps, shift): """The shifted sigma grid used by FlowUniPC at inference (see FlowUniPCMultistepScheduler.set_timesteps and compute_refiner_sigmas).""" sigmas = np.linspace(float(sigma_max), float(sigma_min), int(num_steps) + 1)[:-1] sigmas = shift * sigmas / (1.0 + (shift - 1.0) * sigmas) return torch.from_numpy(sigmas.astype(np.float32)) def vae_latent_to_dit(vae, latents): """Normalize raw VAE latents with the per-channel mean/std, matching LingBotVideoPipeline._vae_latent_to_dit.""" device = latents.device mean = torch.tensor(vae.config.latents_mean, device=device, dtype=torch.float32).view(1, -1, 1, 1, 1) std_inv = (1.0 / torch.tensor(vae.config.latents_std, device=device, dtype=torch.float32)).view(1, -1, 1, 1, 1) return (latents.float() - mean) * std_inv def frame_to_vlm_image(frame): """frame: (C, H, W) tensor in [-1, 1]. Returns the PIL image fed to Qwen3-VL, resized exactly like LingBotVideoI2VPipeline._vlm_image.""" frame = (frame.detach().cpu().clamp(-1, 1) * 0.5 + 0.5) array = frame.permute(1, 2, 0).mul(255).byte().numpy() image = Image.fromarray(array, mode="RGB") return image def resize_vlm_image(image, patch_factor): resized_height, resized_width = smart_resize(image.height, image.width, factor=patch_factor) return image.resize((resized_width, resized_height)) # 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, processor, transformer3d, args, accelerator, weight_dtype, global_step): try: with torch.no_grad(), torch.cuda.amp.autocast(dtype=weight_dtype), torch.cuda.device(device=accelerator.device): logger.info("Running validation... ") scheduler = FlowUniPCMultistepScheduler.from_pretrained( args.pretrained_model_name_or_path, subfolder="scheduler", ) # Under FSDP the transformer3d passed here is the wrapped (sharded) # module; it must be used as-is because its forward is collective. pipeline = LingBotVideoI2VPipeline( transformer=transformer3d, vae=vae, text_encoder=text_encoder, processor=processor, scheduler=scheduler, ) if not args.use_fsdp: pipeline = pipeline.to(accelerator.device) if args.seed is None: generator = None else: rank_seed = args.seed + accelerator.process_index generator = torch.Generator(device=accelerator.device).manual_seed(rank_seed) logger.info(f"Rank {accelerator.process_index} using seed: {rank_seed}") video_length = int((args.video_sample_n_frames - 1) // pipeline.vae_scale_factor_temporal * pipeline.vae_scale_factor_temporal) + 1 if args.video_sample_n_frames != 1 else 1 for i in range(len(args.validation_prompts)): image = Image.open(args.validation_paths[i]).convert("RGB") width, height = calculate_dimensions( args.image_sample_size * args.image_sample_size, image.width / image.height) width, height = int(width), int(height) sample = pipeline( args.validation_prompts[i], image = image, num_frames = video_length, height = height, width = width, generator = generator, guidance_scale = 3.0, shift = args.train_shift, num_inference_steps = 25, ).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}-rank{accelerator.process_index}-image-{i}.mp4" ), fps=24, ) del pipeline gc.collect() torch.cuda.empty_cache() torch.cuda.ipc_collect() vae.to(accelerator.device if not args.low_vram else "cpu") text_encoder.to(accelerator.device if not args.low_vram else "cpu") except Exception as e: gc.collect() torch.cuda.empty_cache() torch.cuda.ipc_collect() print(f"Eval error on rank {accelerator.process_index} with info {e}") vae.to(accelerator.device if not args.low_vram else "cpu") text_encoder.to(accelerator.device if not args.low_vram else "cpu") 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 parse_args(): parser = argparse.ArgumentParser(description="LingBot-Video ti2v training script.") parser.add_argument( "--pretrained_model_name_or_path", type=str, default=None, required=True, help="Path to the LingBot-Video model root (contains transformer/ vae/ text_encoder/ processor/ scheduler/).", ) 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 json 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_steps` and logged to `--report_to`. " "Must be structured JSON captions (rewriter schema), same as the training data.", ) parser.add_argument( "--validation_paths", type=str, default=None, nargs="+", help="A set of condition (first-frame) images evaluated every `--validation_steps`, paired with `--validation_prompts`.", ) parser.add_argument( "--output_dir", type=str, default="lingbot-video-finetuned", help="The output directory where the model predictions and checkpoints will be written.", ) parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") parser.add_argument( "--train_batch_size", type=int, default=1, help="Batch size (per device) for the training dataloader." ) parser.add_argument( "--vae_mini_batch", type=int, default=1, 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-5, 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( "--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( "--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( "--report_model_info", action="store_true", help="Whether or not to report more info about model (such as norm, grad)." ) 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( "--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="lingbot-video-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( "--uniform_sampling", action="store_true", help="Whether or not to use uniform_sampling." ) 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_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="Keep the video token length constant while adapting height and width.", ) parser.add_argument( "--train_sampling_steps", type=int, default=1000, help="Number of discrete sigma levels used for training-time sigma sampling.", ) parser.add_argument( "--train_shift", type=float, default=3.0, help="Shift of the flow-matching sigma schedule, same as the inference `shift` (3.0 for LingBot-Video).", ) 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=1, help="Sample stride of the video.", ) parser.add_argument( "--video_sample_n_frames", type=int, default=81, help="Num frame of video.", ) parser.add_argument( "--video_repeat", type=int, default=0, help="Num of repeat video.", ) 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( '--trainable_modules', nargs='+', default=["."], help='Enter a list of trainable modules (substring match).' ) parser.add_argument( '--trainable_modules_low_learning_rate', nargs='+', default=[], help='Enter a list of trainable modules with lower learning rate' ) 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( "--abnormal_norm_clip_start", type=int, default=1000, help=('When do we start doing additional processing on abnormal gradients.'), ) parser.add_argument( "--initial_grad_norm_ratio", type=int, default=5, help=('The initial gradient is relative to the multiple of the max_grad_norm.'), ) 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`.", ) 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 if args.validation_prompts is not None and args.validation_paths is not None: assert len(args.validation_prompts) == len(args.validation_paths), \ "validation_prompts and validation_paths must be paired one-to-one." return args def main(): args = parse_args() logging_dir = os.path.join(args.output_dir, args.logging_dir) 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, ) fsdp_plugin = accelerator.state.fsdp_plugin if hasattr(accelerator.state, "fsdp_plugin") else None if fsdp_plugin is not None: from torch.distributed.fsdp import ShardingStrategy zero_stage = 0 if fsdp_plugin.sharding_strategy in (ShardingStrategy.FULL_SHARD, 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("FSDP 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 print(f"Init rng with seed {args.seed + accelerator.process_index if args.seed is not None else None}. 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 weights (vae, text_encoder) # 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 the sigma scheduler (only used to build the training sigma table). noise_scheduler = FlowUniPCMultistepScheduler.from_pretrained( args.pretrained_model_name_or_path, subfolder="scheduler", ) # Get Processor (Qwen3-VL tokenizer + image processor) processor = AutoProcessor.from_pretrained( os.path.join(args.pretrained_model_name_or_path, "processor"), ) vision_patch_size = getattr(getattr(processor, "image_processor", None), "config", None) vision_patch_size = getattr(vision_patch_size, "patch_size", 16) if vision_patch_size is not None else 16 vlm_patch_factor = int(vision_patch_size) * SPATIAL_MERGE_SIZE def deepspeed_zero_init_disabled_context_manager(): """ returns a list of context managers to disable deepspeed zero init """ deepspeed_plugin = accelerate.state.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)] with ContextManagers(deepspeed_zero_init_disabled_context_manager()): # Get Text encoder (frozen Qwen3-VL) text_encoder = Qwen3VLForConditionalGeneration.from_pretrained( os.path.join(args.pretrained_model_name_or_path, "text_encoder"), low_cpu_mem_usage=True, torch_dtype=weight_dtype, ) text_encoder = text_encoder.eval() # Get Vae vae = AutoencoderKLQwenImage.from_pretrained( os.path.join(args.pretrained_model_name_or_path, "vae"), ) vae.eval() # Get Transformer. `.to(weight_dtype)` keeps the fp32-sensitive modules # (norms / router / modulation / scale_shift_table) in fp32 automatically. transformer3d = LingBotVideoTransformer3DModel.from_pretrained( os.path.join(args.pretrained_model_name_or_path, "transformer"), low_cpu_mem_usage=True, torch_dtype=weight_dtype, ).to(weight_dtype) # FSDP flattens parameters of one wrap unit into a single flat buffer and # requires a uniform dtype. The custom `.to()` above keeps fp32-sensitive # modules (norms / router / modulation / scale_shift_table) in fp32, so under # FSDP we force-cast everything to weight_dtype (matches bf16 inference). if fsdp_stage != 0: for _name, _param in transformer3d.named_parameters(): _param.data = _param.data.to(weight_dtype) for _name, _buffer in transformer3d.named_buffers(): _buffer.data = _buffer.data.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.transformer_path is not None: print(f"From checkpoint: {args.transformer_path}") if args.transformer_path.endswith("safetensors"): from safetensors.torch import load_file 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 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 transformer3d.train() if accelerator.is_main_process: accelerator.print( f"Trainable modules '{args.trainable_modules}'." ) for name, param in transformer3d.named_parameters(): for trainable_module_name in args.trainable_modules + args.trainable_modules_low_learning_rate: if trainable_module_name in name: param.requires_grad = True break # Create EMA for the transformer3d. if args.use_ema: if fsdp_stage == 3: raise NotImplementedError("DeepSpeed Zero-3 does not support EMA.") ema_module = LingBotVideoTransformer3DModel.from_pretrained( os.path.join(args.pretrained_model_name_or_path, "transformer"), low_cpu_mem_usage=True, torch_dtype=weight_dtype, ) if args.use_fsdp: # The EMA copy gets the same FSDP wrap as the live model so that # every local shard of the copy pairs 1:1 with the live shard. ema_transformer3d = FSDPEMA(ema_module, source=transformer3d, accelerator=accelerator, fsdp_plugin=fsdp_plugin) else: ema_module = ema_module.to(weight_dtype) ema_transformer3d = EMAModel(ema_module.parameters(), model_cls=LingBotVideoTransformer3DModel, model_config=ema_module.config) # `accelerate` 0.16.0 will have better support for customized saving if version.parse(accelerate.__version__) >= version.parse("0.16.0"): 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"diffusion_pytorch_model.safetensors") accelerate_state_dict = {k: v.to(dtype=weight_dtype) for k, v in accelerate_state_dict.items()} save_file(accelerate_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) if args.use_ema: # Every rank joins the FULL_STATE_DICT all-gather inside. ema_transformer3d.save_pretrained(os.path.join(output_dir, "transformer_ema")) def load_model_hook(models, input_dir): if args.use_ema: ema_transformer3d.load_pretrained(os.path.join(input_dir, "transformer_ema")) 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: # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format def save_model_hook(models, weights, output_dir): if accelerator.is_main_process: if args.use_ema: ema_transformer3d.save_pretrained(os.path.join(output_dir, "transformer_ema")) models[0].save_pretrained(os.path.join(output_dir, "transformer")) 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): if args.use_ema: ema_path = os.path.join(input_dir, "transformer_ema") _, ema_kwargs = LingBotVideoTransformer3DModel.load_config(ema_path, return_unused_kwargs=True) load_model = LingBotVideoTransformer3DModel.from_pretrained( input_dir, subfolder="transformer_ema", low_cpu_mem_usage=True, ) load_model = EMAModel(load_model.parameters(), model_cls=LingBotVideoTransformer3DModel, model_config=load_model.config) load_model.load_state_dict(ema_kwargs) ema_transformer3d.load_state_dict(load_model.state_dict()) ema_transformer3d.to(accelerator.device) del load_model for i in range(len(models)): # pop models so that they are not loaded again model = models.pop() # load diffusers style into model load_model = LingBotVideoTransformer3DModel.from_pretrained( input_dir, subfolder="transformer", low_cpu_mem_usage=True, ) model.register_to_config(**load_model.config) model.load_state_dict(load_model.state_dict()) del load_model 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 else: optimizer_cls = torch.optim.AdamW trainable_params = list(filter(lambda p: p.requires_grad, transformer3d.parameters())) trainable_params_optim = [ {'params': [], 'lr': args.learning_rate}, {'params': [], 'lr': args.learning_rate / 2}, ] in_already = [] for name, param in transformer3d.named_parameters(): high_lr_flag = False if name in in_already: continue for trainable_module_name in args.trainable_modules: if trainable_module_name in name: in_already.append(name) high_lr_flag = True trainable_params_optim[0]['params'].append(param) if accelerator.is_main_process: print(f"Set {name} to lr : {args.learning_rate}") break if high_lr_flag: continue for trainable_module_name in args.trainable_modules_low_learning_rate: if trainable_module_name in name: in_already.append(name) trainable_params_optim[1]['params'].append(param) if accelerator.is_main_process: print(f"Set {name} to lr : {args.learning_rate / 2}") break 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 # AutoencoderKLQwenImage's config has no compression-ratio fields; the # pipeline constants are temporal=4 / spatial=8. sample_n_frames_bucket_interval = getattr(vae.config, "temporal_compression_ratio", 4) spatial_compression_ratio = getattr(vae.config, "spatial_compression_ratio", 8) 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=False, ) # The DiT was trained on structured JSON captions only; plain natural # language captions are out-of-distribution and degrade fine-tuning. from videox_fun.models.lingbot_video_rewriter import is_valid_caption n_plain = sum(1 for d in train_dataset.dataset if not is_valid_caption(d.get("text", ""))) if n_plain: logger.warning( f"{n_plain}/{len(train_dataset.dataset)} dataset captions are NOT structured JSON captions. " "LingBot-Video expects rewriter-style JSON captions; training on natural-language text " "feeds out-of-distribution prompts to the DiT. All captions must go through the rewriter " "(no hand-written captions) — convert the dataset with " "scripts/lingbot_video/prepare_captions.py (batch) or " "videox_fun.models.lingbot_video_rewriter.ensure_json_caption (single)." ) 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"] = [] # 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 / spatial_compression_ratio / 2) * spatial_compression_ratio * 2 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 / spatial_compression_ratio / 2) * spatial_compression_ratio * 2 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 / spatial_compression_ratio / 2) * spatial_compression_ratio * 2 for x in closest_size] min_example_length = min( [example["pixel_values"].shape[0] for example in examples] ) batch_video_length = int(min(batch_video_length, min_example_length)) # VAE 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 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)[:batch_video_length]) new_examples["text"].append(example["text"]) # Limit the number of frames to the same new_examples["pixel_values"] = torch.stack([example for example in new_examples["pixel_values"]]) 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`. transformer3d, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( transformer3d, optimizer, train_dataloader, lr_scheduler ) if fsdp_stage != 0: from functools import partial from videox_fun.dist import shard_model # Qwen3-VL has no `.blocks`; wrap the text decoder layers and vision # blocks explicitly via the transformer-class policy. shard_fn = partial( shard_model, device_id=accelerator.device, param_dtype=weight_dtype, transformer_layer_cls_to_wrap=["Qwen3VLTextDecoderLayer", "Qwen3VLVisionBlock"], ) text_encoder = shard_fn(text_encoder) if args.use_ema: ema_transformer3d.to(accelerator.device) # Move text_encoder and vae to gpu and cast to weight_dtype vae.to(accelerator.device if not args.low_vram else "cpu", dtype=weight_dtype) text_encoder.to(accelerator.device if not args.low_vram else "cpu", 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)) keys_to_pop = [k for k, v in tracker_config.items() if isinstance(v, list)] for k in keys_to_pop: tracker_config.pop(k) print(f"Removed tracker_config['{k}']") 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 pkl_path = os.path.join(os.path.join(args.output_dir, 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}.") accelerator.print(f"Resuming from checkpoint {path}") accelerator.load_state(os.path.join(args.output_dir, path)) else: initial_global_step = 0 progress_bar = PauseAwareTqdm( 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, ) # Training-time sigma table: the same shifted FlowUniPC grid used at inference. sigma_table = build_sigma_table( noise_scheduler.sigma_max, noise_scheduler.sigma_min, args.train_sampling_steps, args.train_shift, ) idx_sampling = DiscreteSampling(args.train_sampling_steps, uniform_sampling=args.uniform_sampling) # A throw-away pipeline used only to reuse the inference-time prompt encoding # (chat template + image tokens + system-prefix cropping). encode_pipeline = LingBotVideoI2VPipeline( transformer=None, vae=vae, text_encoder=text_encoder, processor=processor, scheduler=None, ) 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): # Data batch sanity check 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]}.mp4", rescale=True) with accelerator.accumulate(transformer3d): # Convert videos to latent space pixel_values = batch["pixel_values"].to(weight_dtype) pixel_values = rearrange(pixel_values, "b f c h w -> b c f h w") if args.low_vram: torch.cuda.empty_cache() vae.to(accelerator.device) with torch.no_grad(): # This way is quicker when batch grows up def _batch_encode_vae(pixel_values): 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.to(vae.dtype)).latent_dist.sample() new_pixel_values.append(pixel_values_bs) return torch.cat(new_pixel_values, dim=0) latents = vae_latent_to_dit(vae, _batch_encode_vae(pixel_values)) # ti2v condition: the first frame encoded on its own, exactly # like LingBotVideoI2VPipeline.encode_image_latent at inference. cond_latent = vae_latent_to_dit(vae, _batch_encode_vae(pixel_values[:, :, :1])) if args.low_vram: vae.to('cpu') torch.cuda.empty_cache() # Text (+ first-frame image) encoding with the frozen Qwen3-VL. with torch.no_grad(): if args.low_vram: text_encoder.to(accelerator.device) vlm_images = [ resize_vlm_image(frame_to_vlm_image(frame), vlm_patch_factor) for frame in pixel_values[:, :, 0] ] prompt_embeds, prompt_mask = encode_pipeline.encode_prompt( batch['text'], images=vlm_images, device=accelerator.device) prompt_embeds = prompt_embeds.to(device=latents.device) prompt_mask = prompt_mask.to(device=latents.device) if args.low_vram: 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) # Sample sigmas from the shifted table. 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 * args.train_sampling_steps).long() else: indices = idx_sampling(bsz, generator=torch_rng, device=latents.device) indices = indices.long().cpu() sigmas = sigma_table[indices].to(device=latents.device) # Add noise according to flow matching. # zt = (1 - sigma) * x0 + sigma * noise sigmas = sigmas.view(-1, 1, 1, 1, 1).to(latents.dtype) noisy_latents = (1.0 - sigmas) * latents + sigmas * noise # Clamp the condition frame(s) back to the clean latent, matching # LingBotVideoPipeline._apply_inpainting at inference. The loss is # masked on these frames accordingly. cond_t = cond_latent.shape[2] frame_mask = torch.ones_like(latents[:, :, :1].expand(-1, -1, num_frames, -1, -1)) if num_frames > cond_t: noisy_latents[:, :, :cond_t] = cond_latent.to(noisy_latents.dtype) frame_mask[:, :, :cond_t] = 0.0 else: # Degenerate single-latent-frame sample: nothing to generate. frame_mask[:, :, :] = 0.0 # Predict the noise residual timesteps = (sigmas.reshape(-1) * 1000.0).float() with torch.cuda.amp.autocast(dtype=weight_dtype), torch.cuda.device(device=accelerator.device): noise_pred = transformer3d( noisy_latents, timesteps, prompt_embeds.to(weight_dtype), encoder_attention_mask=prompt_mask, return_dict=False, )[0] target = noise - latents weighting = compute_loss_weighting_for_sd3(weighting_scheme=args.weighting_scheme, sigmas=sigmas.squeeze()) weighting = weighting.view(-1, 1, 1, 1, 1).float() mse_loss = F.mse_loss(noise_pred.float(), target.float(), reduction='none') masked_loss = mse_loss * frame_mask * weighting denom = frame_mask.sum() * channel * height * width loss = masked_loss.sum() / denom.clamp(min=1.0) loss = loss.mean() # 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: if not args.use_fsdp: trainable_params_grads = [p.grad for p in trainable_params if p.grad is not None] trainable_params_total_norm = torch.norm(torch.stack([torch.norm(g.detach(), 2) for g in trainable_params_grads]), 2) max_grad_norm = linear_decay(args.max_grad_norm * args.initial_grad_norm_ratio, args.max_grad_norm, args.abnormal_norm_clip_start, global_step) if trainable_params_total_norm / max_grad_norm > 5 and global_step > args.abnormal_norm_clip_start: actual_max_grad_norm = max_grad_norm / min((trainable_params_total_norm / max_grad_norm), 10) else: actual_max_grad_norm = max_grad_norm else: actual_max_grad_norm = args.max_grad_norm if not args.use_fsdp and args.report_model_info and accelerator.is_main_process: if trainable_params_total_norm > 1 and global_step > args.abnormal_norm_clip_start: for name, param in transformer3d.named_parameters(): if param.requires_grad: writer.add_scalar(f'gradients/before_clip_norm/{name}', param.grad.norm(), global_step=global_step) norm_sum = accelerator.clip_grad_norm_(trainable_params, actual_max_grad_norm) if not args.use_fsdp and args.report_model_info and accelerator.is_main_process: writer.add_scalar(f'gradients/norm_sum', norm_sum, global_step=global_step) writer.add_scalar(f'gradients/actual_max_grad_norm', actual_max_grad_norm, global_step=global_step) optimizer.step() lr_scheduler.step() optimizer.zero_grad() # Checks if the accelerator has performed an optimization step behind the scenes if accelerator.sync_gradients: if args.use_ema: ema_transformer3d.step(transformer3d.parameters()) 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_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) gc.collect() torch.cuda.empty_cache() torch.cuda.ipc_collect() save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") # Keep the checkpoint out of the progress bar rate: a minute-long save would # otherwise land in the next step's interval and be shown as a slow step. The # save also stages the whole state in host RAM (safetensors materializes every # tensor as bytes) and leaves the freed blocks in the allocator caches, so the # cache flushes run inside the same window. with progress_bar.paused(): accelerator.save_state(save_path) gc.collect() torch.cuda.empty_cache() torch.cuda.ipc_collect() logger.info(f"Saved state to {save_path}") if args.validation_prompts is not None and global_step % args.validation_steps == 0: with progress_bar.paused(): if args.use_ema: # Store the transformer parameters temporarily and load the EMA parameters to perform inference. ema_transformer3d.store(transformer3d.parameters()) ema_transformer3d.copy_to(transformer3d.parameters()) log_validation( vae, text_encoder, processor, transformer3d, args, accelerator, weight_dtype, global_step, ) if args.use_ema: # Switch back to the original transformer3d parameters. ema_transformer3d.restore(transformer3d.parameters()) 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 args.validation_prompts is not None and epoch % args.validation_epochs == 0: with progress_bar.paused(): if args.use_ema: # Store the transformer parameters temporarily and load the EMA parameters to perform inference. ema_transformer3d.store(transformer3d.parameters()) ema_transformer3d.copy_to(transformer3d.parameters()) log_validation( vae, text_encoder, processor, transformer3d, args, accelerator, weight_dtype, global_step, ) if args.use_ema: # Switch back to the original transformer3d parameters. ema_transformer3d.restore(transformer3d.parameters()) # Close the bar before the end-of-run checkpoint: tqdm keeps redrawing a live bar whenever # something else writes to the console. PauseAwareTqdm.close() rebases the closing line onto # the smoothed rate, so the worker warm-up and the first dataloader fetch do not dilute it. progress_bar.close() # Create the pipeline using the trained modules and save it. accelerator.wait_for_everyone() if args.use_ema and args.use_fsdp: # Under FSDP every rank must write its own shards, and the shards only # exist while the model is still wrapped, so this runs before the # `unwrap_model` below. ema_transformer3d.copy_to(transformer3d.parameters()) if accelerator.is_main_process: transformer3d = unwrap_model(transformer3d) if args.use_ema and not args.use_fsdp: ema_transformer3d.copy_to(transformer3d.parameters()) if args.use_fsdp or accelerator.is_main_process: gc.collect() torch.cuda.empty_cache() torch.cuda.ipc_collect() save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") accelerator.save_state(save_path) logger.info(f"Saved state to {save_path}") accelerator.end_training() if __name__ == "__main__": main()