[misc]: realign blank lines left by the comment sweep

2026-11-10 01:32:33 +00:00
parent 054149ee3f
commit 74e3061646
29 changed files with 0 additions and 68 deletions
@@ -50,8 +50,6 @@ class AttentionBackend(ABC):
    def get_metadata_cls() -> type["AttentionMetadata"]:
        raise NotImplementedError

-
-
    @staticmethod
    @abstractmethod
    def get_builder_cls() -> type["AttentionMetadataBuilder"]:
@@ -44,7 +44,6 @@ logger = init_logger(__name__)
 BSA_TILE_SIZE = (4, 4, 4)


-
@functools.lru_cache(maxsize=10)
 def get_tile_partition_indices(
    dit_seq_shape: tuple[int, int, int],
@@ -77,8 +76,6 @@ def get_reverse_tile_partition_indices(
    return torch.argsort(get_tile_partition_indices(dit_seq_shape, tile_size, device))


-
-
 def _prune_queries(
    q_blocks: torch.Tensor,
    keep_ratio: float,
@@ -530,8 +527,6 @@ def _reconstruct_pruned(
    return full_output


-
-
 class BSAAttentionBackend(AttentionBackend):

    accept_output_buffer: bool = False
@@ -27,7 +27,6 @@ class SageAttentionBackend(AttentionBackend):
        return SageAttentionImpl


-
 class SageAttentionImpl(AttentionImpl):

    def __init__(
@@ -35,7 +35,6 @@ class SageAttention3Backend(AttentionBackend):
        raise NotImplementedError


-
 class SageAttention3Impl(AttentionImpl):

    def __init__(
@@ -33,7 +33,6 @@ class SDPABackend(AttentionBackend):
        return SDPAImpl


-
@dataclass
 class SDPAMetadata(AttentionMetadata):
    current_timestep: int
@@ -35,7 +35,6 @@ from fastvideo.logger import init_logger
 logger = init_logger(__name__)


-
@triton.jit
 def compress_kernel(
    X,
@@ -107,8 +106,6 @@ def get_block_map(
    return sparse_map, lut, topk


-
-
 class SLAAttentionBackend(AttentionBackend):
    """Sparse-Linear Attention backend."""

@@ -36,6 +36,5 @@ __all__ = [
    "get_tp_group",
    "get_tp_rank",
    "get_tp_world_size",
-
    "get_local_torch_device",
 ]
@@ -22,7 +22,6 @@
 # changing the environment variable `FASTVIDEO_NCCL_SO_PATH`, or the `so_file`
 # variable in the code.

-
 import ctypes
 import platform
 from dataclasses import dataclass
@@ -262,8 +262,6 @@ def samples_from_manifest(
    return samples


-
-
 def _expand(spec: PathSpec, exts: tuple[str, ...]) -> list[Path]:
    """Resolve a path-style input into a sorted list of file paths.

@@ -937,8 +937,6 @@ class MiniMaxH3SchedulerState:
        return timestep * sample + (1.0 - timestep) * noise


-
-
 def _as_mx_indices(indices):
    """NumPy/python index arrays are not accepted by MLX indexing; convert."""
    import mlx.core as mx
@@ -1728,8 +1726,6 @@ class MLXMiniMaxH3DiT:
    __call__ = forward


-
-
 def _load_array(array, dtype):
    import mlx.core as mx

@@ -235,8 +235,6 @@ def _activation1d(x, alpha, beta, up_filter, down_filter, ratio: int = ACTIVATIO
    return _low_pass(x, down_filter, ratio)


-
-
 class MLXMiniMaxH3AudioVAE:
    """DAC-style posterior encoder plus BigVGAN waveform decoder."""

@@ -12,7 +12,6 @@ from fastvideo.pipelines.composed_pipeline_base import ComposedPipelineBase
 from fastvideo.pipelines.stages import (ConditioningStage, DecodingStage, DenoisingStage, InputValidationStage,
                                        LatentPreparationStage, TextEncodingStage, TimestepPreparationStage)

-
 logger = init_logger(__name__)


@@ -16,7 +16,6 @@ from fastvideo.pipelines.stages import (ConditioningStage, DecodingStage, Denois
                                        Hy15ImageEncodingStage, SRDenoisingStage)
 from fastvideo.distributed.parallel_state import get_local_torch_device

-
 logger = init_logger(__name__)


@@ -13,7 +13,6 @@ from fastvideo.pipelines.stages import (ConditioningStage, DecodingStage, Denois
                                        LatentPreparationStage, TextEncodingStage, TimestepPreparationStage,
                                        Hy15ImageEncodingStage)

-
 logger = init_logger(__name__)


@@ -13,7 +13,6 @@ from fastvideo.pipelines.stages import (ConditioningStage, DecodingStage, Denois
                                        LatentPreparationStage, TextEncodingStage, TimestepPreparationStage,
                                        Hy15ImageEncodingStage)

-
 logger = init_logger(__name__)


@@ -16,7 +16,6 @@ from fastvideo.pipelines.stages import (ConditioningStage, DecodingStage, Denois
                                        Hy15ImageEncodingStage, SRDenoisingStage)
 from fastvideo.distributed.parallel_state import get_local_torch_device

-
 logger = init_logger(__name__)


@@ -104,7 +104,6 @@ class PreprocessPipeline_MatrixGame2(BasePreprocessPipeline):
        # mask_lat_size = mask_lat_size.transpose(1, 2)
        # mask_lat_size = mask_lat_size.to(latent_condition.device)

-
        features["first_frame_latent"] = latent_condition

        if "action_path" in valid_data and valid_data["action_path"]:
@@ -236,7 +236,6 @@ class ValidationCallback(Callback):
            return value.strip().lower() in {"1", "true", "yes", "on"}
        return bool(value)

-
    def _adopt_training_denoising_ladder(
        self,
        method: TrainingMethod,
@@ -319,7 +318,6 @@ class ValidationCallback(Callback):

        self._run_validation(method, iteration)

-
    def _run_validation(
        self,
        method: TrainingMethod,
@@ -839,7 +837,6 @@ class ValidationCallback(Callback):
                step,
            )

-
    def _metric_device(self) -> str:
        device = self.metrics_config.device
        if device == "cuda" and torch.cuda.is_available():
@@ -1170,7 +1167,6 @@ class ValidationCallback(Callback):
            return -1
        return int(self.metrics_config.mouse_pitch_sign)

-
    def _get_sampling_param(self) -> SamplingParam:
        if self._sampling_param is None:
            self._sampling_param = (SamplingParam.from_pretrained(self._pipeline_model_path()))
@@ -1329,7 +1325,6 @@ class ValidationCallback(Callback):
        self._pipeline_key = key
        return self._pipeline

-
    def _prepare_validation_batch(
        self,
        sampling_param: SamplingParam,
@@ -1450,7 +1445,6 @@ class ValidationCallback(Callback):
                tensor,
            )

-
    def _student_sampling_pipeline_config(
        self,
        pipeline_config: Any,
@@ -1655,7 +1649,6 @@ class ValidationCallback(Callback):
            action=action,
        )

-
    def state_dict(self) -> dict[str, Any]:
        state: dict[str, Any] = {}
        if self.validation_random_generator is not None:
@@ -77,7 +77,6 @@ class TrainingMethod(torch.nn.Module, ABC):
            if mods:
                self.role_modules[role] = torch.nn.ModuleDict(mods)

-
    def set_tracker(self, tracker: Any) -> None:
        self.tracker = tracker

@@ -58,7 +58,6 @@ class CausalConsistencyDistillationMethod(TrainingMethod):
        )
        self._init_optimizers_and_schedulers()

-
    @property
    def _optimizer_dict(self) -> dict[str, Any]:
        return {"student": self._student_optimizer}
@@ -84,7 +83,6 @@ class CausalConsistencyDistillationMethod(TrainingMethod):
        states["roles.ema.transformer"] = _FullModelState(self.ema_model.transformer)
        return states

-
    def single_train_step(
        self,
        batch: dict[str, Any],
@@ -166,7 +164,6 @@ class CausalConsistencyDistillationMethod(TrainingMethod):
        metrics: dict[str, LogScalar] = {}
        return loss_map, outputs, metrics

-
    def backward(
        self,
        loss_map: dict[str, torch.Tensor],
@@ -186,7 +183,6 @@ class CausalConsistencyDistillationMethod(TrainingMethod):
        if iteration >= self._ema_start_step:
            self._update_ema()

-
    def _predict_flow(
        self,
        model: ModelBase,
@@ -61,7 +61,6 @@ from fastvideo.train.utils.optimizer import build_optimizer_and_scheduler

 logger = logging.getLogger(__name__)

-
 _COMPLETE_SENTINEL = "COMPLETE"
 _METADATA_FILE = "metadata.json"

@@ -153,8 +152,6 @@ class _KDPathCache:
            dist.barrier()


-
-
 class _KDPathDataset(Dataset):
    """Reads per-sample ``.pt`` files from a completed KD cache.

@@ -181,8 +178,6 @@ class _KDPathDataset(Dataset):
        return torch.load(path, weights_only=True, map_location="cpu")


-
-
 class _KDDataLoaderWrapper:
    """Wraps the parquet dataloader initially; activates to KD cache after generation.

@@ -250,8 +245,6 @@ def _kd_collate(samples: list[dict[str, torch.Tensor]]) -> dict[str, torch.Tenso
    return {k: torch.stack([s[k] for s in samples]) for k in keys}


-
-
 class KDMethod(TrainingMethod):
    """Knowledge Distillation training method.

@@ -335,7 +328,6 @@ class KDMethod(TrainingMethod):
        # --- Student optimizer / scheduler (same as FineTuneMethod) ---
        self._init_optimizers_and_schedulers()

-
    @property
    def _optimizer_dict(self) -> dict[str, Any]:
        return {"student": self._student_optimizer}
@@ -352,7 +344,6 @@ class KDMethod(TrainingMethod):
        del iteration
        return [self._student_lr_scheduler]

-
    def on_train_start(self) -> None:
        super().on_train_start()  # sets CUDA RNG, calls student.on_train_start()

@@ -440,7 +431,6 @@ class KDMethod(TrainingMethod):
             for t in self._t_list[:-1]},
        )

-
    def _generate_paths(
        self,
        source_dataset: Any,
@@ -602,7 +592,6 @@ class KDMethod(TrainingMethod):
        idx = torch.argmin((timesteps - t).abs())
        return sigmas[idx].expand(numel).reshape(numel, 1, 1, 1)

-
    def single_train_step(
        self,
        batch: dict[str, Any],
@@ -711,7 +700,6 @@ class KDMethod(TrainingMethod):
            grad_accum_rounds=grad_accum_rounds,
        )

-
    def _init_optimizers_and_schedulers(self) -> None:
        tc = self.training_config
        student_lr = float(tc.optimizer.learning_rate)
@@ -732,8 +720,6 @@ class KDMethod(TrainingMethod):
        )


-
-
 class KDCausalMethod(KDMethod):
    """KD for causal Wan: per-frame block-quantized timestep sampling.

@@ -71,7 +71,6 @@ class CosmosModel(WanModel):
            transformer_override_safetensor=(transformer_override_safetensor),
        )

-
    def prepare_batch(
        self,
        raw_batch: dict[str, Any],
@@ -65,7 +65,6 @@ class HunyuanModel(WanModel):
            lora=lora,
        )

-
    def prepare_batch(
        self,
        raw_batch: dict[str, Any],
@@ -199,7 +199,6 @@ class MiniMaxH3DMDModel(MiniMaxH3Model):
    def num_train_timesteps(self) -> int:
        return _DMD_TIMESTEP_SCALE

-
    def _modality_shapes(self) -> tuple[tuple[int, int, int, int, int], tuple[int, int, int, int]]:
        """Return the ``[1, T, C, H, W]`` video and ``[1, 2, 32, Ta]`` audio shapes."""
        data = self.training_config.data
@@ -360,7 +359,6 @@ class MiniMaxH3DMDModel(MiniMaxH3Model):
            ),
        )

-
    def set_requires_negative_conditioning(self, requires: bool) -> None:
        """Fail fast: H3 cannot encode negative prompts at training time."""
        if requires:
@@ -725,7 +723,6 @@ class MiniMaxH3DMDModel(MiniMaxH3Model):
        sigma_fp64 = sigma.to(device=noisy.device, dtype=torch.float64)
        return (noisy_fp64 - sigma_fp64 * pred_noise_fp64).to(original_dtype)

-
    def _load_vis_vae(self) -> Any:
        """Lazily load the H3 video VAE for visualization decodes.

@@ -73,7 +73,6 @@ class MiniMaxH3PDDMixin(_PDDMixinHost):
        """Only this role widens its heads; the shared pipeline config stays single-head."""
        return {"pdd_steps": self.pdd_steps}

-
    def pdd_fine_grid(self, device: torch.device) -> torch.Tensor:
        return pdd_fine_grid(self.pdd_steps, max_t=PDD_GRID_MAX_T, device=device)

@@ -99,7 +98,6 @@ class MiniMaxH3PDDMixin(_PDDMixinHost):
            raise ValueError("MiniMaxH3PDDModel packs one document per batch; every sample must share a base time")
        return base * _DMD_TIMESTEP_SCALE

-
    def predict_pdd_heads(
        self,
        state: torch.Tensor,
@@ -83,8 +83,6 @@ def make_inference_args(
    return args


-
-
 def load_module_from_path(
    *,
    model_path: str,
@@ -60,7 +60,6 @@ class Cosmos25TrainingPipeline(TrainingPipeline):
        )
        self.validation_pipeline = validation_pipeline

-
    def _normalize_dit_input(self, training_batch: TrainingBatch) -> TrainingBatch:
        """Skip: Cosmos25WanVAEWrapper normalises latents during encoding.

@@ -126,7 +125,6 @@ class Cosmos25TrainingPipeline(TrainingPipeline):
        return training_batch


-
 if __name__ == "__main__":
    logger.warning("\n"
                   "================================================================================\n"
@@ -998,7 +998,6 @@ class DistillationPipeline(TrainingPipeline):

        sampling_param = SamplingParam.from_pretrained(training_args.model_path)

-
        logger.info("Using validation seed: %s", self.seed)

        logger.info('rank: %s: fastvideo_args.validation_dataset_file: %s',
@@ -69,7 +69,6 @@ class Worker:
        self.rank = rank
        self.distributed_init_method = distributed_init_method

-
    def init_device(self) -> None:
        """Initialize the device for the worker."""