diff --git a/fastvideo/dataset/__init__.py b/fastvideo/dataset/__init__.py index b82c653..f76077b 100644 --- a/fastvideo/dataset/__init__.py +++ b/fastvideo/dataset/__init__.py @@ -4,6 +4,8 @@ from torchvision.transforms import Lambda from fastvideo.dataset.parquet_dataset_map_style import ( build_parquet_map_style_dataloader) +from fastvideo.dataset.pants_latent_dataset import ( + build_pants_latent_dataloader, is_pants_latent_path) from fastvideo.dataset.ltx2_precomputed_dataset import ( build_ltx2_precomputed_dataloader, LTX2PrecomputedDataset) from fastvideo.dataset.preprocessing_datasets import VideoCaptionMergedDataset, TextDataset @@ -46,6 +48,8 @@ def gettextdataset(args) -> TextDataset: __all__ = [ "build_parquet_map_style_dataloader", + "build_pants_latent_dataloader", + "is_pants_latent_path", "build_ltx2_precomputed_dataloader", "LTX2PrecomputedDataset", "ValidationDataset", diff --git a/fastvideo/training/training_pipeline.py b/fastvideo/training/training_pipeline.py index 575d6dc..140bb31 100644 --- a/fastvideo/training/training_pipeline.py +++ b/fastvideo/training/training_pipeline.py @@ -28,7 +28,11 @@ try: except Exception: pass from fastvideo.api.sampling_param import SamplingParam -from fastvideo.dataset import build_parquet_map_style_dataloader +from fastvideo.dataset import ( + build_pants_latent_dataloader, + build_parquet_map_style_dataloader, + is_pants_latent_path, +) from fastvideo.dataset.dataloader.schema import pyarrow_schema_t2v from fastvideo.dataset.validation_dataset import ValidationDataset from fastvideo.distributed import (cleanup_dist_env_and_memory, get_local_torch_device, get_sp_group, get_world_group) @@ -42,7 +46,8 @@ from fastvideo.training.activation_checkpoint import (apply_activation_checkpoin from fastvideo.training.trackers import (DummyTracker, TrackerType, initialize_trackers, Trackers) from fastvideo.training.training_utils import (clip_grad_norm_while_handling_failing_dtensor_cases, compute_density_for_timestep_sampling, count_trainable, get_scheduler, - get_sigmas, load_checkpoint, normalize_dit_input, save_checkpoint) + get_sigmas, load_checkpoint, normalize_dit_input, save_checkpoint, + EMA_FSDP, gather_state_dict_on_cpu_rank0, custom_to_hf_state_dict) from fastvideo.utils import (is_vmoba_available, is_vsa_available, set_random_seed, shallow_asdict) try: @@ -82,6 +87,7 @@ class TrainingPipeline(LoRAPipeline, ABC): super().__init__(model_path, fastvideo_args, required_config_modules, loaded_modules) # type: ignore self.tracker = DummyTracker() self.validation_ref_videos_logged = False + self.generator_ema: EMA_FSDP | None = None def create_pipeline_stages(self, fastvideo_args: FastVideoArgs): raise RuntimeError("create_pipeline_stages should not be called for training pipeline") @@ -167,16 +173,27 @@ class TrainingPipeline(LoRAPipeline, ABC): last_epoch=self.init_steps - 1, ) - self.train_dataset, self.train_dataloader = build_parquet_map_style_dataloader( - training_args.data_path, - training_args.train_batch_size, - parquet_schema=self.train_dataset_schema, - num_data_workers=training_args.dataloader_num_workers, - cfg_rate=training_args.training_cfg_rate, - drop_last=True, - text_padding_length=training_args.pipeline_config.text_encoder_configs[0].arch_config. - text_len, # type: ignore[attr-defined] - seed=self.seed) + text_padding_length = training_args.pipeline_config.text_encoder_configs[0].arch_config.text_len # type: ignore[attr-defined] + if is_pants_latent_path(training_args.data_path): + self.train_dataset, self.train_dataloader = build_pants_latent_dataloader( + training_args.data_path, + training_args.train_batch_size, + num_data_workers=training_args.dataloader_num_workers, + cfg_rate=training_args.training_cfg_rate, + drop_last=True, + text_padding_length=text_padding_length, + seed=self.seed, + ) + else: + self.train_dataset, self.train_dataloader = build_parquet_map_style_dataloader( + training_args.data_path, + training_args.train_batch_size, + parquet_schema=self.train_dataset_schema, + num_data_workers=training_args.dataloader_num_workers, + cfg_rate=training_args.training_cfg_rate, + drop_last=True, + text_padding_length=text_padding_length, + seed=self.seed) self.noise_scheduler = noise_scheduler if self.training_args.boundary_ratio is not None: @@ -460,6 +477,43 @@ class TrainingPipeline(LoRAPipeline, ABC): training_batch.grad_norm = grad_norm return training_batch + def _maybe_init_ema(self, step: int) -> None: + if not self.training_args.use_ema: + return + if self.generator_ema is not None: + return + if step < self.training_args.ema_start_step: + return + if self.training_args.ema_decay <= 0: + return + self.generator_ema = EMA_FSDP(self.transformer, decay=self.training_args.ema_decay) + logger.info("Created generator EMA at step %s with decay=%s", step, self.training_args.ema_decay) + + def _maybe_update_ema(self, step: int) -> None: + self._maybe_init_ema(step) + if self.generator_ema is not None: + self.generator_ema.update(self.transformer) + + def _save_ema_weights(self, step: int) -> None: + if not self.training_args.use_ema or self.generator_ema is None: + return + ema_dir = os.path.join(self.training_args.output_dir, f"ema_checkpoint-{step}") + os.makedirs(ema_dir, exist_ok=True) + with self.generator_ema.apply_to_model(self.transformer): + cpu_state = gather_state_dict_on_cpu_rank0(self.transformer, device=None) + if self.global_rank == 0: + from safetensors.torch import save_file + + diffusers_state_dict = custom_to_hf_state_dict( + cpu_state, + self.transformer.reverse_param_names_mapping, + ) + save_file( + diffusers_state_dict, + os.path.join(ema_dir, "diffusion_pytorch_model.safetensors"), + ) + logger.info("Saved EMA transformer weights to %s", ema_dir) + @profile_region("profiler_region_training_train_one_step") def train_one_step(self, training_batch: TrainingBatch) -> TrainingBatch: training_batch = self._prepare_training(training_batch) @@ -571,6 +625,7 @@ class TrainingPipeline(LoRAPipeline, ABC): training_batch.current_timestep = step training_batch.current_vsa_sparsity = current_vsa_sparsity training_batch = self.train_one_step(training_batch) + self._maybe_update_ema(step) loss = float(training_batch.total_loss) grad_norm = training_batch.grad_norm @@ -594,6 +649,9 @@ class TrainingPipeline(LoRAPipeline, ABC): "grad_norm": grad_norm, "vsa_sparsity": current_vsa_sparsity, } + if self.training_args.use_ema: + metrics["ema_enabled"] = self.generator_ema is not None + metrics["ema_decay"] = self.training_args.ema_decay try: metrics["batch_size"] = int(training_batch.raw_latent_shape[0]) @@ -622,6 +680,7 @@ class TrainingPipeline(LoRAPipeline, ABC): save_checkpoint(self.transformer, self.global_rank, self.training_args.output_dir, step, self.optimizer, self.train_dataloader, self.lr_scheduler, self.noise_random_generator) + self._save_ema_weights(step) self.transformer.train() self.sp_group.barrier() @@ -637,9 +696,13 @@ class TrainingPipeline(LoRAPipeline, ABC): trainable_params) self.tracker.finish() - save_checkpoint(self.transformer, self.global_rank, self.training_args.output_dir, - self.training_args.max_train_steps, self.optimizer, self.train_dataloader, self.lr_scheduler, - self.noise_random_generator) + if os.environ.get("FASTVIDEO_SKIP_FINAL_CHECKPOINT", "0") == "1": + logger.info("Skipping final checkpoint because FASTVIDEO_SKIP_FINAL_CHECKPOINT=1") + else: + save_checkpoint(self.transformer, self.global_rank, self.training_args.output_dir, + self.training_args.max_train_steps, self.optimizer, self.train_dataloader, + self.lr_scheduler, self.noise_random_generator) + self._save_ema_weights(self.training_args.max_train_steps) if envs.FASTVIDEO_TORCH_PROFILER_DIR: logger.info("Stopping profiler...")