feat: Implement self-contained specialist managers

common/cache.py

@@ -0,0 +1,47 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from typing import Callable
+
+
+class Cache:
+    """Caching reusable args for faster inference"""
+
+    def __init__(self, disable=False, prefix="", cache=None):
+        self.cache = cache if cache is not None else {}
+        self.disable = disable
+        self.prefix = prefix
+
+    def __call__(self, key: str, fn: Callable):
+        if self.disable:
+            return fn()
+
+        key = self.prefix + key
+        try:
+            result = self.cache[key]
+        except KeyError:
+            result = fn()
+            self.cache[key] = result
+        return result
+
+    def namespace(self, namespace: str):
+        return Cache(
+            disable=self.disable,
+            prefix=self.prefix + namespace + ".",
+            cache=self.cache,
+        )
+
+    def get(self, key: str):
+        key = self.prefix + key
+        return self.cache[key]

common/config.py

@@ -0,0 +1,110 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Configuration utility functions
+"""
+
+import importlib
+from typing import Any, Callable, List, Union
+from omegaconf import DictConfig, ListConfig, OmegaConf
+
+OmegaConf.register_new_resolver("eval", eval)
+
+
+def load_config(path: str, argv: List[str] = None) -> Union[DictConfig, ListConfig]:
+    """
+    Load a configuration. Will resolve inheritance.
+    """
+    config = OmegaConf.load(path)
+    if argv is not None:
+        config_argv = OmegaConf.from_dotlist(argv)
+        config = OmegaConf.merge(config, config_argv)
+    config = resolve_recursive(config, resolve_inheritance)
+    return config
+
+
+def resolve_recursive(
+    config: Any,
+    resolver: Callable[[Union[DictConfig, ListConfig]], Union[DictConfig, ListConfig]],
+) -> Any:
+    config = resolver(config)
+    if isinstance(config, DictConfig):
+        for k in config.keys():
+            v = config.get(k)
+            if isinstance(v, (DictConfig, ListConfig)):
+                config[k] = resolve_recursive(v, resolver)
+    if isinstance(config, ListConfig):
+        for i in range(len(config)):
+            v = config.get(i)
+            if isinstance(v, (DictConfig, ListConfig)):
+                config[i] = resolve_recursive(v, resolver)
+    return config
+
+
+def resolve_inheritance(config: Union[DictConfig, ListConfig]) -> Any:
+    """
+    Recursively resolve inheritance if the config contains:
+    __inherit__: path/to/parent.yaml or a ListConfig of such paths.
+    """
+    if isinstance(config, DictConfig):
+        inherit = config.pop("__inherit__", None)
+
+        if inherit:
+            inherit_list = inherit if isinstance(inherit, ListConfig) else [inherit]
+
+            parent_config = None
+            for parent_path in inherit_list:
+                assert isinstance(parent_path, str)
+                parent_config = (
+                    load_config(parent_path)
+                    if parent_config is None
+                    else OmegaConf.merge(parent_config, load_config(parent_path))
+                )
+
+            if len(config.keys()) > 0:
+                config = OmegaConf.merge(parent_config, config)
+            else:
+                config = parent_config
+    return config
+
+
+def import_item(path: str, name: str) -> Any:
+    """
+    Import a python item. Example: import_item("path.to.file", "MyClass") -> MyClass
+    """
+    return getattr(importlib.import_module(path), name)
+
+
+def create_object(config: DictConfig) -> Any:
+    """
+    Create an object from config.
+    The config is expected to contains the following:
+    __object__:
+      path: path.to.module
+      name: MyClass
+      args: as_config | as_params (default to as_config)
+    """
+    item = import_item(
+        path=config.__object__.path,
+        name=config.__object__.name,
+    )
+    args = config.__object__.get("args", "as_config")
+    if args == "as_config":
+        return item(config)
+    if args == "as_params":
+        config = OmegaConf.to_object(config)
+        config.pop("__object__")
+        return item(**config)
+    raise NotImplementedError(f"Unknown args type: {args}")

common/decorators.py

@@ -0,0 +1,147 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Decorators.
+"""
+
+import functools
+import threading
+import time
+from typing import Callable
+import torch
+
+from common.distributed import barrier_if_distributed, get_global_rank, get_local_rank
+from common.logger import get_logger
+
+logger = get_logger(__name__)
+
+
+def log_on_entry(func: Callable) -> Callable:
+    """
+    Functions with this decorator will log the function name at entry.
+    When using multiple decorators, this must be applied innermost to properly capture the name.
+    """
+
+    def log_on_entry_wrapper(*args, **kwargs):
+        logger.info(f"Entering {func.__name__}")
+        return func(*args, **kwargs)
+
+    return log_on_entry_wrapper
+
+
+def barrier_on_entry(func: Callable) -> Callable:
+    """
+    Functions with this decorator will start executing when all ranks are ready to enter.
+    """
+
+    def barrier_on_entry_wrapper(*args, **kwargs):
+        barrier_if_distributed()
+        return func(*args, **kwargs)
+
+    return barrier_on_entry_wrapper
+
+
+def _conditional_execute_wrapper_factory(execute: bool, func: Callable) -> Callable:
+    """
+    Helper function for local_rank_zero_only and global_rank_zero_only.
+    """
+
+    def conditional_execute_wrapper(*args, **kwargs):
+        # Only execute if needed.
+        result = func(*args, **kwargs) if execute else None
+        # All GPUs must wait.
+        barrier_if_distributed()
+        # Return results.
+        return result
+
+    return conditional_execute_wrapper
+
+
+def _asserted_wrapper_factory(condition: bool, func: Callable, err_msg: str = "") -> Callable:
+    """
+    Helper function for some functions with special constraints,
+    especially functions called by other global_rank_zero_only / local_rank_zero_only ones,
+    in case they are wrongly invoked in other scenarios.
+    """
+
+    def asserted_execute_wrapper(*args, **kwargs):
+        assert condition, err_msg
+        result = func(*args, **kwargs)
+        return result
+
+    return asserted_execute_wrapper
+
+
+def local_rank_zero_only(func: Callable) -> Callable:
+    """
+    Functions with this decorator will only execute on local rank zero.
+    """
+    return _conditional_execute_wrapper_factory(get_local_rank() == 0, func)
+
+
+def global_rank_zero_only(func: Callable) -> Callable:
+    """
+    Functions with this decorator will only execute on global rank zero.
+    """
+    return _conditional_execute_wrapper_factory(get_global_rank() == 0, func)
+
+
+def assert_only_global_rank_zero(func: Callable) -> Callable:
+    """
+    Functions with this decorator are only accessible to processes with global rank zero.
+    """
+    return _asserted_wrapper_factory(
+        get_global_rank() == 0, func, err_msg="Not accessible to processes with global_rank != 0"
+    )
+
+
+def assert_only_local_rank_zero(func: Callable) -> Callable:
+    """
+    Functions with this decorator are only accessible to processes with local rank zero.
+    """
+    return _asserted_wrapper_factory(
+        get_local_rank() == 0, func, err_msg="Not accessible to processes with local_rank != 0"
+    )
+
+
+def new_thread(func: Callable) -> Callable:
+    """
+    Functions with this decorator will run in a new thread.
+    The function will return the thread, which can be joined to wait for completion.
+    """
+
+    def new_thread_wrapper(*args, **kwargs):
+        thread = threading.Thread(target=func, args=args, kwargs=kwargs)
+        thread.start()
+        return thread
+
+    return new_thread_wrapper
+
+
+def log_runtime(func: Callable) -> Callable:
+    """
+    Functions with this decorator will logging the runtime.
+    """
+
+    @functools.wraps(func)
+    def wrapped(*args, **kwargs):
+        torch.distributed.barrier()
+        start = time.perf_counter()
+        result = func(*args, **kwargs)
+        torch.distributed.barrier()
+        logger.info(f"Completed {func.__name__} in {time.perf_counter() - start:.3f} seconds.")
+        return result
+
+    return wrapped

common/diffusion/__init__.py

@@ -0,0 +1,56 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Diffusion package.
+"""
+
+from .config import (
+    create_sampler_from_config,
+    create_sampling_timesteps_from_config,
+    create_schedule_from_config,
+)
+from .samplers.base import Sampler
+from .samplers.euler import EulerSampler
+from .schedules.base import Schedule
+from .schedules.lerp import LinearInterpolationSchedule
+from .timesteps.base import SamplingTimesteps, Timesteps
+from .timesteps.sampling.trailing import UniformTrailingSamplingTimesteps
+from .types import PredictionType, SamplingDirection
+from .utils import classifier_free_guidance, classifier_free_guidance_dispatcher, expand_dims
+
+__all__ = [
+    # Configs
+    "create_sampler_from_config",
+    "create_sampling_timesteps_from_config",
+    "create_schedule_from_config",
+    # Schedules
+    "Schedule",
+    "DiscreteVariancePreservingSchedule",
+    "LinearInterpolationSchedule",
+    # Samplers
+    "Sampler",
+    "EulerSampler",
+    # Timesteps
+    "Timesteps",
+    "SamplingTimesteps",
+    # Types
+    "PredictionType",
+    "SamplingDirection",
+    "UniformTrailingSamplingTimesteps",
+    # Utils
+    "classifier_free_guidance",
+    "classifier_free_guidance_dispatcher",
+    "expand_dims",
+]

common/diffusion/config.py

@@ -0,0 +1,74 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Utility functions for creating schedules and samplers from config.
+"""
+
+import torch
+from omegaconf import DictConfig
+
+from .samplers.base import Sampler
+from .samplers.euler import EulerSampler
+from .schedules.base import Schedule
+from .schedules.lerp import LinearInterpolationSchedule
+from .timesteps.base import SamplingTimesteps
+from .timesteps.sampling.trailing import UniformTrailingSamplingTimesteps
+
+
+def create_schedule_from_config(
+    config: DictConfig,
+    device: torch.device,
+    dtype: torch.dtype = torch.float32,
+) -> Schedule:
+    """
+    Create a schedule from configuration.
+    """
+    if config.type == "lerp":
+        return LinearInterpolationSchedule(T=config.get("T", 1.0))
+
+    raise NotImplementedError
+
+
+def create_sampler_from_config(
+    config: DictConfig,
+    schedule: Schedule,
+    timesteps: SamplingTimesteps,
+) -> Sampler:
+    """
+    Create a sampler from configuration.
+    """
+    if config.type == "euler":
+        return EulerSampler(
+            schedule=schedule,
+            timesteps=timesteps,
+            prediction_type=config.prediction_type,
+        )
+    raise NotImplementedError
+
+
+def create_sampling_timesteps_from_config(
+    config: DictConfig,
+    schedule: Schedule,
+    device: torch.device,
+    dtype: torch.dtype = torch.float32,
+) -> SamplingTimesteps:
+    if config.type == "uniform_trailing":
+        return UniformTrailingSamplingTimesteps(
+            T=schedule.T,
+            steps=config.steps,
+            shift=config.get("shift", 1.0),
+            device=device,
+        )
+    raise NotImplementedError

common/diffusion/samplers/base.py

@@ -0,0 +1,108 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Sampler base class.
+"""
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from typing import Callable
+import torch
+from tqdm import tqdm
+
+from ..schedules.base import Schedule
+from ..timesteps.base import SamplingTimesteps
+from ..types import PredictionType, SamplingDirection
+from ..utils import assert_schedule_timesteps_compatible
+
+
+@dataclass
+class SamplerModelArgs:
+    x_t: torch.Tensor
+    t: torch.Tensor
+    i: int
+
+
+class Sampler(ABC):
+    """
+    Samplers are ODE/SDE solvers.
+    """
+
+    def __init__(
+        self,
+        schedule: Schedule,
+        timesteps: SamplingTimesteps,
+        prediction_type: PredictionType,
+        return_endpoint: bool = True,
+    ):
+        assert_schedule_timesteps_compatible(
+            schedule=schedule,
+            timesteps=timesteps,
+        )
+        self.schedule = schedule
+        self.timesteps = timesteps
+        self.prediction_type = prediction_type
+        self.return_endpoint = return_endpoint
+
+    @abstractmethod
+    def sample(
+        self,
+        x: torch.Tensor,
+        f: Callable[[SamplerModelArgs], torch.Tensor],
+    ) -> torch.Tensor:
+        """
+        Generate a new sample given the the intial sample x and score function f.
+        """
+
+    def get_next_timestep(
+        self,
+        t: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Get the next sample timestep.
+        Support multiple different timesteps t in a batch.
+        If no more steps, return out of bound value -1 or T+1.
+        """
+        T = self.timesteps.T
+        steps = len(self.timesteps)
+        curr_idx = self.timesteps.index(t)
+        next_idx = curr_idx + 1
+        bound = -1 if self.timesteps.direction == SamplingDirection.backward else T + 1
+
+        s = self.timesteps[next_idx.clamp_max(steps - 1)]
+        s = s.where(next_idx < steps, bound)
+        return s
+
+    def get_endpoint(
+        self,
+        pred: torch.Tensor,
+        x_t: torch.Tensor,
+        t: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Get to the endpoint of the probability flow.
+        """
+        x_0, x_T = self.schedule.convert_from_pred(pred, self.prediction_type, x_t, t)
+        return x_0 if self.timesteps.direction == SamplingDirection.backward else x_T
+
+    def get_progress_bar(self):
+        """
+        Get progress bar for sampling.
+        """
+        return tqdm(
+            iterable=range(len(self.timesteps) - (0 if self.return_endpoint else 1)),
+            dynamic_ncols=True,
+            desc=self.__class__.__name__,
+        )

common/diffusion/samplers/euler.py

@@ -0,0 +1,89 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+
+"""
+Euler ODE solver.
+"""
+
+from typing import Callable
+import torch
+from einops import rearrange
+from torch.nn import functional as F
+
+from models.dit_v2 import na
+
+from ..types import PredictionType
+from ..utils import expand_dims
+from .base import Sampler, SamplerModelArgs
+
+
+class EulerSampler(Sampler):
+    """
+    The Euler method is the simplest ODE solver.
+    <https://en.wikipedia.org/wiki/Euler_method>
+    """
+
+    def sample(
+        self,
+        x: torch.Tensor,
+        f: Callable[[SamplerModelArgs], torch.Tensor],
+    ) -> torch.Tensor:
+        timesteps = self.timesteps.timesteps
+        progress = self.get_progress_bar()
+        i = 0
+        for t, s in zip(timesteps[:-1], timesteps[1:]):
+            pred = f(SamplerModelArgs(x, t, i))
+            x = self.step_to(pred, x, t, s)
+            i += 1
+            progress.update()
+
+        if self.return_endpoint:
+            t = timesteps[-1]
+            pred = f(SamplerModelArgs(x, t, i))
+            x = self.get_endpoint(pred, x, t)
+            progress.update()
+        return x
+
+    def step(
+        self,
+        pred: torch.Tensor,
+        x_t: torch.Tensor,
+        t: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Step to the next timestep.
+        """
+        return self.step_to(pred, x_t, t, self.get_next_timestep(t))
+
+    def step_to(
+        self,
+        pred: torch.Tensor,
+        x_t: torch.Tensor,
+        t: torch.Tensor,
+        s: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Steps from x_t at timestep t to x_s at timestep s. Returns x_s.
+        """
+        t = expand_dims(t, x_t.ndim)
+        s = expand_dims(s, x_t.ndim)
+        T = self.schedule.T
+        # Step from x_t to x_s.
+        pred_x_0, pred_x_T = self.schedule.convert_from_pred(pred, self.prediction_type, x_t, t)
+        pred_x_s = self.schedule.forward(pred_x_0, pred_x_T, s.clamp(0, T))
+        # Clamp x_s to x_0 and x_T if s is out of bound.
+        pred_x_s = pred_x_s.where(s >= 0, pred_x_0)
+        pred_x_s = pred_x_s.where(s <= T, pred_x_T)
+        return pred_x_s

common/diffusion/schedules/base.py

@@ -0,0 +1,131 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Schedule base class.
+"""
+
+from abc import ABC, abstractmethod, abstractproperty
+from typing import Tuple, Union
+import torch
+
+from ..types import PredictionType
+from ..utils import expand_dims
+
+
+class Schedule(ABC):
+    """
+    Diffusion schedules are uniquely defined by T, A, B:
+
+        x_t = A(t) * x_0 + B(t) * x_T, where t in [0, T]
+
+    Schedules can be continuous or discrete.
+    """
+
+    @abstractproperty
+    def T(self) -> Union[int, float]:
+        """
+        Maximum timestep inclusive.
+        Schedule is continuous if float, discrete if int.
+        """
+
+    @abstractmethod
+    def A(self, t: torch.Tensor) -> torch.Tensor:
+        """
+        Interpolation coefficient A.
+        Returns tensor with the same shape as t.
+        """
+
+    @abstractmethod
+    def B(self, t: torch.Tensor) -> torch.Tensor:
+        """
+        Interpolation coefficient B.
+        Returns tensor with the same shape as t.
+        """
+
+    # ----------------------------------------------------
+
+    def snr(self, t: torch.Tensor) -> torch.Tensor:
+        """
+        Signal to noise ratio.
+        Returns tensor with the same shape as t.
+        """
+        return (self.A(t) ** 2) / (self.B(t) ** 2)
+
+    def isnr(self, snr: torch.Tensor) -> torch.Tensor:
+        """
+        Inverse signal to noise ratio.
+        Returns tensor with the same shape as snr.
+        Subclass may implement.
+        """
+        raise NotImplementedError
+
+    # ----------------------------------------------------
+
+    def is_continuous(self) -> bool:
+        """
+        Whether the schedule is continuous.
+        """
+        return isinstance(self.T, float)
+
+    def forward(self, x_0: torch.Tensor, x_T: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
+        """
+        Diffusion forward function.
+        """
+        t = expand_dims(t, x_0.ndim)
+        return self.A(t) * x_0 + self.B(t) * x_T
+
+    def convert_from_pred(
+        self, pred: torch.Tensor, pred_type: PredictionType, x_t: torch.Tensor, t: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Convert from prediction. Return predicted x_0 and x_T.
+        """
+        t = expand_dims(t, x_t.ndim)
+        A_t = self.A(t)
+        B_t = self.B(t)
+
+        if pred_type == PredictionType.x_T:
+            pred_x_T = pred
+            pred_x_0 = (x_t - B_t * pred_x_T) / A_t
+        elif pred_type == PredictionType.x_0:
+            pred_x_0 = pred
+            pred_x_T = (x_t - A_t * pred_x_0) / B_t
+        elif pred_type == PredictionType.v_cos:
+            pred_x_0 = A_t * x_t - B_t * pred
+            pred_x_T = A_t * pred + B_t * x_t
+        elif pred_type == PredictionType.v_lerp:
+            pred_x_0 = (x_t - B_t * pred) / (A_t + B_t)
+            pred_x_T = (x_t + A_t * pred) / (A_t + B_t)
+        else:
+            raise NotImplementedError
+
+        return pred_x_0, pred_x_T
+
+    def convert_to_pred(
+        self, x_0: torch.Tensor, x_T: torch.Tensor, t: torch.Tensor, pred_type: PredictionType
+    ) -> torch.FloatTensor:
+        """
+        Convert to prediction target given x_0 and x_T.
+        """
+        if pred_type == PredictionType.x_T:
+            return x_T
+        if pred_type == PredictionType.x_0:
+            return x_0
+        if pred_type == PredictionType.v_cos:
+            t = expand_dims(t, x_0.ndim)
+            return self.A(t) * x_T - self.B(t) * x_0
+        if pred_type == PredictionType.v_lerp:
+            return x_T - x_0
+        raise NotImplementedError

common/diffusion/schedules/lerp.py

@@ -0,0 +1,55 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Linear interpolation schedule (lerp).
+"""
+
+from typing import Union
+import torch
+
+from .base import Schedule
+
+
+class LinearInterpolationSchedule(Schedule):
+    """
+    Linear interpolation schedule (lerp) is proposed by flow matching and rectified flow.
+    It leads to straighter probability flow theoretically. It is also used by Stable Diffusion 3.
+    <https://arxiv.org/abs/2209.03003>
+    <https://arxiv.org/abs/2210.02747>
+
+        x_t = (1 - t) * x_0 + t * x_T
+
+    Can be either continuous or discrete.
+    """
+
+    def __init__(self, T: Union[int, float] = 1.0):
+        self._T = T
+
+    @property
+    def T(self) -> Union[int, float]:
+        return self._T
+
+    def A(self, t: torch.Tensor) -> torch.Tensor:
+        return 1 - (t / self.T)
+
+    def B(self, t: torch.Tensor) -> torch.Tensor:
+        return t / self.T
+
+    # ----------------------------------------------------
+
+    def isnr(self, snr: torch.Tensor) -> torch.Tensor:
+        t = self.T / (1 + snr**0.5)
+        t = t if self.is_continuous() else t.round().int()
+        return t

common/diffusion/timesteps/base.py

@@ -0,0 +1,72 @@
+from abc import ABC, abstractmethod
+from typing import Sequence, Union
+import torch
+
+from ..types import SamplingDirection
+
+
+class Timesteps(ABC):
+    """
+    Timesteps base class.
+    """
+
+    def __init__(self, T: Union[int, float]):
+        assert T > 0
+        self._T = T
+
+    @property
+    def T(self) -> Union[int, float]:
+        """
+        Maximum timestep inclusive.
+        int if discrete, float if continuous.
+        """
+        return self._T
+
+    def is_continuous(self) -> bool:
+        """
+        Whether the schedule is continuous.
+        """
+        return isinstance(self.T, float)
+
+
+class SamplingTimesteps(Timesteps):
+    """
+    Sampling timesteps.
+    It defines the discretization of sampling steps.
+    """
+
+    def __init__(
+        self,
+        T: Union[int, float],
+        timesteps: torch.Tensor,
+        direction: SamplingDirection,
+    ):
+        assert timesteps.ndim == 1
+        super().__init__(T)
+        self.timesteps = timesteps
+        self.direction = direction
+
+    def __len__(self) -> int:
+        """
+        Number of sampling steps.
+        """
+        return len(self.timesteps)
+
+    def __getitem__(self, idx: Union[int, torch.IntTensor]) -> torch.Tensor:
+        """
+        The timestep at the sampling step.
+        Returns a scalar tensor if idx is int,
+        or tensor of the same size if idx is a tensor.
+        """
+        return self.timesteps[idx]
+
+    def index(self, t: torch.Tensor) -> torch.Tensor:
+        """
+        Find index by t.
+        Return index of the same shape as t.
+        Index is -1 if t not found in timesteps.
+        """
+        i, j = t.reshape(-1, 1).eq(self.timesteps).nonzero(as_tuple=True)
+        idx = torch.full_like(t, fill_value=-1, dtype=torch.int)
+        idx.view(-1)[i] = j.int()
+        return idx

common/diffusion/timesteps/sampling/trailing.py

@@ -0,0 +1,49 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+import torch
+
+from ...types import SamplingDirection
+from ..base import SamplingTimesteps
+
+
+class UniformTrailingSamplingTimesteps(SamplingTimesteps):
+    """
+    Uniform trailing sampling timesteps.
+    Defined in (https://arxiv.org/abs/2305.08891)
+
+    Shift is proposed in SD3 for RF schedule.
+    Defined in (https://arxiv.org/pdf/2403.03206) eq.23
+    """
+
+    def __init__(
+        self,
+        T: int,
+        steps: int,
+        shift: float = 1.0,
+        device: torch.device = "cpu",
+    ):
+        # Create trailing timesteps.
+        timesteps = torch.arange(1.0, 0.0, -1.0 / steps, device=device)
+
+        # Shift timesteps.
+        timesteps = shift * timesteps / (1 + (shift - 1) * timesteps)
+
+        # Scale to T range.
+        if isinstance(T, float):
+            timesteps = timesteps * T
+        else:
+            timesteps = timesteps.mul(T + 1).sub(1).round().int()
+
+        super().__init__(T=T, timesteps=timesteps, direction=SamplingDirection.backward)

common/diffusion/types.py

@@ -0,0 +1,59 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Type definitions.
+"""
+
+from enum import Enum
+
+
+class PredictionType(str, Enum):
+    """
+    x_0:
+        Predict data sample.
+    x_T:
+        Predict noise sample.
+        Proposed by DDPM (https://arxiv.org/abs/2006.11239)
+        Proved problematic by zsnr paper (https://arxiv.org/abs/2305.08891)
+    v_cos:
+        Predict velocity dx/dt based on the cosine schedule (A_t * x_T - B_t * x_0).
+        Proposed by progressive distillation (https://arxiv.org/abs/2202.00512)
+    v_lerp:
+        Predict velocity dx/dt based on the lerp schedule (x_T - x_0).
+        Proposed by rectified flow (https://arxiv.org/abs/2209.03003)
+    """
+
+    x_0 = "x_0"
+    x_T = "x_T"
+    v_cos = "v_cos"
+    v_lerp = "v_lerp"
+
+
+class SamplingDirection(str, Enum):
+    """
+    backward: Sample from x_T to x_0 for data generation.
+    forward:  Sample from x_0 to x_T for noise inversion.
+    """
+
+    backward = "backward"
+    forward = "forward"
+
+    @staticmethod
+    def reverse(direction):
+        if direction == SamplingDirection.backward:
+            return SamplingDirection.forward
+        if direction == SamplingDirection.forward:
+            return SamplingDirection.backward
+        raise NotImplementedError

common/diffusion/utils.py

@@ -0,0 +1,84 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Utility functions.
+"""
+
+from typing import Callable
+import torch
+
+
+def expand_dims(tensor: torch.Tensor, ndim: int):
+    """
+    Expand tensor to target ndim. New dims are added to the right.
+    For example, if the tensor shape was (8,), target ndim is 4, return (8, 1, 1, 1).
+    """
+    shape = tensor.shape + (1,) * (ndim - tensor.ndim)
+    return tensor.reshape(shape)
+
+
+def assert_schedule_timesteps_compatible(schedule, timesteps):
+    """
+    Check if schedule and timesteps are compatible.
+    """
+    if schedule.T != timesteps.T:
+        raise ValueError("Schedule and timesteps must have the same T.")
+    if schedule.is_continuous() != timesteps.is_continuous():
+        raise ValueError("Schedule and timesteps must have the same continuity.")
+
+
+def classifier_free_guidance(
+    pos: torch.Tensor,
+    neg: torch.Tensor,
+    scale: float,
+    rescale: float = 0.0,
+):
+    """
+    Apply classifier-free guidance.
+    """
+    # Classifier-free guidance (https://arxiv.org/abs/2207.12598)
+    cfg = neg + scale * (pos - neg)
+
+    # Classifier-free guidance rescale (https://arxiv.org/pdf/2305.08891.pdf)
+    if rescale != 0.0:
+        pos_std = pos.std(dim=list(range(1, pos.ndim)), keepdim=True)
+        cfg_std = cfg.std(dim=list(range(1, cfg.ndim)), keepdim=True)
+        factor = pos_std / cfg_std
+        factor = rescale * factor + (1 - rescale)
+        cfg *= factor
+
+    return cfg
+
+
+def classifier_free_guidance_dispatcher(
+    pos: Callable,
+    neg: Callable,
+    scale: float,
+    rescale: float = 0.0,
+):
+    """
+    Optionally execute models depending on classifer-free guidance scale.
+    """
+    # If scale is 1, no need to execute neg model.
+    if scale == 1.0:
+        return pos()
+
+    # Otherwise, execute both pos nad neg models and apply cfg.
+    return classifier_free_guidance(
+        pos=pos(),
+        neg=neg(),
+        scale=scale,
+        rescale=rescale,
+    )

common/distributed/__init__.py

@@ -0,0 +1,37 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Distributed package.
+"""
+
+from .basic import (
+    barrier_if_distributed,
+    convert_to_ddp,
+    get_device,
+    get_global_rank,
+    get_local_rank,
+    get_world_size,
+    init_torch,
+)
+
+__all__ = [
+    "barrier_if_distributed",
+    "convert_to_ddp",
+    "get_device",
+    "get_global_rank",
+    "get_local_rank",
+    "get_world_size",
+    "init_torch",
+]

common/distributed/advanced.py

@@ -0,0 +1,208 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Advanced distributed functions for sequence parallel.
+"""
+
+from typing import Optional, List
+import torch
+import torch.distributed as dist
+from torch.distributed.device_mesh import DeviceMesh, init_device_mesh
+from torch.distributed.fsdp import ShardingStrategy
+
+from .basic import get_global_rank, get_world_size
+
+
+_DATA_PARALLEL_GROUP = None
+_SEQUENCE_PARALLEL_GROUP = None
+_SEQUENCE_PARALLEL_CPU_GROUP = None
+_MODEL_SHARD_CPU_INTER_GROUP = None
+_MODEL_SHARD_CPU_INTRA_GROUP = None
+_MODEL_SHARD_INTER_GROUP = None
+_MODEL_SHARD_INTRA_GROUP = None
+_SEQUENCE_PARALLEL_GLOBAL_RANKS = None
+
+
+def get_data_parallel_group() -> Optional[dist.ProcessGroup]:
+    """
+    Get data parallel process group.
+    """
+    return _DATA_PARALLEL_GROUP
+
+
+def get_sequence_parallel_group() -> Optional[dist.ProcessGroup]:
+    """
+    Get sequence parallel process group.
+    """
+    return _SEQUENCE_PARALLEL_GROUP
+
+
+def get_sequence_parallel_cpu_group() -> Optional[dist.ProcessGroup]:
+    """
+    Get sequence parallel CPU process group.
+    """
+    return _SEQUENCE_PARALLEL_CPU_GROUP
+
+
+def get_data_parallel_rank() -> int:
+    """
+    Get data parallel rank.
+    """
+    group = get_data_parallel_group()
+    return dist.get_rank(group) if group else get_global_rank()
+
+
+def get_data_parallel_world_size() -> int:
+    """
+    Get data parallel world size.
+    """
+    group = get_data_parallel_group()
+    return dist.get_world_size(group) if group else get_world_size()
+
+
+def get_sequence_parallel_rank() -> int:
+    """
+    Get sequence parallel rank.
+    """
+    group = get_sequence_parallel_group()
+    return dist.get_rank(group) if group else 0
+
+
+def get_sequence_parallel_world_size() -> int:
+    """
+    Get sequence parallel world size.
+    """
+    group = get_sequence_parallel_group()
+    return dist.get_world_size(group) if group else 1
+
+
+def get_model_shard_cpu_intra_group() -> Optional[dist.ProcessGroup]:
+    """
+    Get the CPU intra process group of model sharding.
+    """
+    return _MODEL_SHARD_CPU_INTRA_GROUP
+
+
+def get_model_shard_cpu_inter_group() -> Optional[dist.ProcessGroup]:
+    """
+    Get the CPU inter process group of model sharding.
+    """
+    return _MODEL_SHARD_CPU_INTER_GROUP
+
+
+def get_model_shard_intra_group() -> Optional[dist.ProcessGroup]:
+    """
+    Get the GPU intra process group of model sharding.
+    """
+    return _MODEL_SHARD_INTRA_GROUP
+
+
+def get_model_shard_inter_group() -> Optional[dist.ProcessGroup]:
+    """
+    Get the GPU inter process group of model sharding.
+    """
+    return _MODEL_SHARD_INTER_GROUP
+
+
+def init_sequence_parallel(sequence_parallel_size: int):
+    """
+    Initialize sequence parallel.
+    """
+    global _DATA_PARALLEL_GROUP
+    global _SEQUENCE_PARALLEL_GROUP
+    global _SEQUENCE_PARALLEL_CPU_GROUP
+    global _SEQUENCE_PARALLEL_GLOBAL_RANKS
+    assert dist.is_initialized()
+    world_size = dist.get_world_size()
+    rank = dist.get_rank()
+    data_parallel_size = world_size // sequence_parallel_size
+    for i in range(data_parallel_size):
+        start_rank = i * sequence_parallel_size
+        end_rank = (i + 1) * sequence_parallel_size
+        ranks = range(start_rank, end_rank)
+        group = dist.new_group(ranks)
+        cpu_group = dist.new_group(ranks, backend="gloo")
+        if rank in ranks:
+            _SEQUENCE_PARALLEL_GROUP = group
+            _SEQUENCE_PARALLEL_CPU_GROUP = cpu_group
+            _SEQUENCE_PARALLEL_GLOBAL_RANKS = list(ranks)
+
+
+def init_model_shard_group(
+    *,
+    sharding_strategy: ShardingStrategy,
+    device_mesh: Optional[DeviceMesh] = None,
+):
+    """
+    Initialize process group of model sharding.
+    """
+    global _MODEL_SHARD_INTER_GROUP
+    global _MODEL_SHARD_INTRA_GROUP
+    global _MODEL_SHARD_CPU_INTER_GROUP
+    global _MODEL_SHARD_CPU_INTRA_GROUP
+    assert dist.is_initialized()
+    world_size = dist.get_world_size()
+    if device_mesh is not None:
+        num_shards_per_group = device_mesh.shape[1]
+    elif sharding_strategy == ShardingStrategy.NO_SHARD:
+        num_shards_per_group = 1
+    elif sharding_strategy in [
+        ShardingStrategy.HYBRID_SHARD,
+        ShardingStrategy._HYBRID_SHARD_ZERO2,
+    ]:
+        num_shards_per_group = torch.cuda.device_count()
+    else:
+        num_shards_per_group = world_size
+    num_groups = world_size // num_shards_per_group
+    device_mesh = (num_groups, num_shards_per_group)
+
+    gpu_mesh_2d = init_device_mesh("cuda", device_mesh, mesh_dim_names=("inter", "intra"))
+    cpu_mesh_2d = init_device_mesh("cpu", device_mesh, mesh_dim_names=("inter", "intra"))
+
+    _MODEL_SHARD_INTER_GROUP = gpu_mesh_2d.get_group("inter")
+    _MODEL_SHARD_INTRA_GROUP = gpu_mesh_2d.get_group("intra")
+    _MODEL_SHARD_CPU_INTER_GROUP = cpu_mesh_2d.get_group("inter")
+    _MODEL_SHARD_CPU_INTRA_GROUP = cpu_mesh_2d.get_group("intra")
+
+def get_sequence_parallel_global_ranks() -> List[int]:
+    """
+    Get all global ranks of the sequence parallel process group
+    that the caller rank belongs to.
+    """
+    if _SEQUENCE_PARALLEL_GLOBAL_RANKS is None:
+        return [dist.get_rank()]
+    return _SEQUENCE_PARALLEL_GLOBAL_RANKS
+
+
+def get_next_sequence_parallel_rank() -> int:
+    """
+    Get the next global rank of the sequence parallel process group
+    that the caller rank belongs to.
+    """
+    sp_global_ranks = get_sequence_parallel_global_ranks()
+    sp_rank = get_sequence_parallel_rank()
+    sp_size = get_sequence_parallel_world_size()
+    return sp_global_ranks[(sp_rank + 1) % sp_size]
+
+
+def get_prev_sequence_parallel_rank() -> int:
+    """
+    Get the previous global rank of the sequence parallel process group
+    that the caller rank belongs to.
+    """
+    sp_global_ranks = get_sequence_parallel_global_ranks()
+    sp_rank = get_sequence_parallel_rank()
+    sp_size = get_sequence_parallel_world_size()
+    return sp_global_ranks[(sp_rank + sp_size - 1) % sp_size]

common/distributed/basic.py

@@ -0,0 +1,84 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Distributed basic functions.
+"""
+
+import os
+from datetime import timedelta
+import torch
+import torch.distributed as dist
+from torch.nn.parallel import DistributedDataParallel
+
+
+def get_global_rank() -> int:
+    """
+    Get the global rank, the global index of the GPU.
+    """
+    return int(os.environ.get("RANK", "0"))
+
+
+def get_local_rank() -> int:
+    """
+    Get the local rank, the local index of the GPU.
+    """
+    return int(os.environ.get("LOCAL_RANK", "0"))
+
+
+def get_world_size() -> int:
+    """
+    Get the world size, the total amount of GPUs.
+    """
+    return int(os.environ.get("WORLD_SIZE", "1"))
+
+
+def get_device() -> torch.device:
+    """
+    Get current rank device.
+    """
+    return torch.device("cuda", get_local_rank())
+
+
+def barrier_if_distributed(*args, **kwargs):
+    """
+    Synchronizes all processes if under distributed context.
+    """
+    if dist.is_initialized():
+        return dist.barrier(*args, **kwargs)
+
+
+def init_torch(cudnn_benchmark=True, timeout=timedelta(seconds=600)):
+    """
+    Common PyTorch initialization configuration.
+    """
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.backends.cudnn.allow_tf32 = True
+    torch.backends.cudnn.benchmark = cudnn_benchmark
+    torch.cuda.set_device(get_local_rank())
+    dist.init_process_group(
+        backend="nccl",
+        rank=get_global_rank(),
+        world_size=get_world_size(),
+        timeout=timeout,
+    )
+
+
+def convert_to_ddp(module: torch.nn.Module, **kwargs) -> DistributedDataParallel:
+    return DistributedDataParallel(
+        module=module,
+        device_ids=[get_local_rank()],
+        output_device=get_local_rank(),
+        **kwargs,
+    )

common/distributed/meta_init_utils.py

@@ -0,0 +1,41 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+import torch
+from rotary_embedding_torch import RotaryEmbedding
+from torch import nn
+from torch.distributed.fsdp._common_utils import _is_fsdp_flattened
+
+__all__ = ["meta_non_persistent_buffer_init_fn"]
+
+
+def meta_non_persistent_buffer_init_fn(module: nn.Module) -> nn.Module:
+    """
+    Used for materializing `non-persistent tensor buffers` while model resuming.
+
+    Since non-persistent tensor buffers are not saved in state_dict,
+    when initializing model with meta device, user should materialize those buffers manually.
+
+    Currently, only `rope.dummy` is this special case.
+    """
+    with torch.no_grad():
+        for submodule in module.modules():
+            if not isinstance(submodule, RotaryEmbedding):
+                continue
+            for buffer_name, buffer in submodule.named_buffers(recurse=False):
+                if buffer.is_meta and "dummy" in buffer_name:
+                    materialized_buffer = torch.zeros_like(buffer, device="cpu")
+                    setattr(submodule, buffer_name, materialized_buffer)
+    assert not any(b.is_meta for n, b in module.named_buffers())
+    return module

common/distributed/ops.py

@@ -0,0 +1,494 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Distributed ops for supporting sequence parallel.
+"""
+
+from collections import defaultdict
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+import torch
+import torch.distributed as dist
+from torch import Tensor
+
+from common.cache import Cache
+from common.distributed.advanced import (
+    get_sequence_parallel_group,
+    get_sequence_parallel_rank,
+    get_sequence_parallel_world_size,
+)
+
+from .basic import get_device
+
+_SEQ_DATA_BUF = defaultdict(lambda: [None, None, None])
+_SEQ_DATA_META_SHAPES = defaultdict()
+_SEQ_DATA_META_DTYPES = defaultdict()
+_SEQ_DATA_ASYNC_COMMS = defaultdict(list)
+_SYNC_BUFFER = defaultdict(dict)
+
+
+def single_all_to_all(
+    local_input: Tensor,
+    scatter_dim: int,
+    gather_dim: int,
+    group: dist.ProcessGroup,
+    async_op: bool = False,
+):
+    """
+    A function to do all-to-all on a tensor
+    """
+    seq_world_size = dist.get_world_size(group)
+    prev_scatter_dim = scatter_dim
+    if scatter_dim != 0:
+        local_input = local_input.transpose(0, scatter_dim)
+        if gather_dim == 0:
+            gather_dim = scatter_dim
+        scatter_dim = 0
+
+    inp_shape = list(local_input.shape)
+    inp_shape[scatter_dim] = inp_shape[scatter_dim] // seq_world_size
+    input_t = local_input.reshape(
+        [seq_world_size, inp_shape[scatter_dim]] + inp_shape[scatter_dim + 1 :]
+    ).contiguous()
+    output = torch.empty_like(input_t)
+    comm = dist.all_to_all_single(output, input_t, group=group, async_op=async_op)
+    if async_op:
+        # let user's code transpose & reshape
+        return output, comm, prev_scatter_dim
+
+    # first dim is seq_world_size, so we can split it directly
+    output = torch.cat(output.split(1), dim=gather_dim + 1).squeeze(0)
+    if prev_scatter_dim:
+        output = output.transpose(0, prev_scatter_dim).contiguous()
+    return output
+
+
+def _all_to_all(
+    local_input: Tensor,
+    scatter_dim: int,
+    gather_dim: int,
+    group: dist.ProcessGroup,
+):
+    seq_world_size = dist.get_world_size(group)
+    input_list = [
+        t.contiguous() for t in torch.tensor_split(local_input, seq_world_size, scatter_dim)
+    ]
+    output_list = [torch.empty_like(input_list[0]) for _ in range(seq_world_size)]
+    dist.all_to_all(output_list, input_list, group=group)
+    return torch.cat(output_list, dim=gather_dim).contiguous()
+
+
+class SeqAllToAll(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx: Any,
+        group: dist.ProcessGroup,
+        local_input: Tensor,
+        scatter_dim: int,
+        gather_dim: int,
+        async_op: bool,
+    ) -> Tensor:
+        ctx.group = group
+        ctx.scatter_dim = scatter_dim
+        ctx.gather_dim = gather_dim
+        ctx.async_op = async_op
+        if async_op:
+            output, comm, prev_scatter_dim = single_all_to_all(
+                local_input, scatter_dim, gather_dim, group, async_op=async_op
+            )
+            ctx.prev_scatter_dim = prev_scatter_dim
+            return output, comm
+
+        return _all_to_all(local_input, scatter_dim, gather_dim, group)
+
+    @staticmethod
+    def backward(ctx: Any, *grad_output: Tensor) -> Tuple[None, Tensor, None, None]:
+        if ctx.async_op:
+            input_t = torch.cat(grad_output[0].split(1), dim=ctx.gather_dim + 1).squeeze(0)
+            if ctx.prev_scatter_dim:
+                input_t = input_t.transpose(0, ctx.prev_scatter_dim)
+        else:
+            input_t = grad_output[0]
+        return (
+            None,
+            _all_to_all(input_t, ctx.gather_dim, ctx.scatter_dim, ctx.group),
+            None,
+            None,
+            None,
+        )
+
+
+class Slice(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx: Any, group: dist.ProcessGroup, local_input: Tensor, dim: int) -> Tensor:
+        ctx.group = group
+        ctx.rank = dist.get_rank(group)
+        seq_world_size = dist.get_world_size(group)
+        ctx.seq_world_size = seq_world_size
+        ctx.dim = dim
+        dim_size = local_input.shape[dim]
+        return local_input.split(dim_size // seq_world_size, dim=dim)[ctx.rank].contiguous()
+
+    @staticmethod
+    def backward(ctx: Any, grad_output: Tensor) -> Tuple[None, Tensor, None]:
+        dim_size = list(grad_output.size())
+        split_size = dim_size[0]
+        dim_size[0] = dim_size[0] * ctx.seq_world_size
+        output = torch.empty(dim_size, dtype=grad_output.dtype, device=torch.cuda.current_device())
+        dist._all_gather_base(output, grad_output, group=ctx.group)
+        return (None, torch.cat(output.split(split_size), dim=ctx.dim), None)
+
+
+class Gather(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx: Any,
+        group: dist.ProcessGroup,
+        local_input: Tensor,
+        dim: int,
+        grad_scale: Optional[bool] = False,
+    ) -> Tensor:
+        ctx.group = group
+        ctx.rank = dist.get_rank(group)
+        ctx.dim = dim
+        ctx.grad_scale = grad_scale
+        seq_world_size = dist.get_world_size(group)
+        ctx.seq_world_size = seq_world_size
+        dim_size = list(local_input.size())
+        split_size = dim_size[0]
+        ctx.part_size = dim_size[dim]
+        dim_size[0] = dim_size[0] * seq_world_size
+        output = torch.empty(dim_size, dtype=local_input.dtype, device=torch.cuda.current_device())
+        dist._all_gather_base(output, local_input.contiguous(), group=ctx.group)
+        return torch.cat(output.split(split_size), dim=dim)
+
+    @staticmethod
+    def backward(ctx: Any, grad_output: Tensor) -> Tuple[None, Tensor]:
+        if ctx.grad_scale:
+            grad_output = grad_output * ctx.seq_world_size
+        return (
+            None,
+            grad_output.split(ctx.part_size, dim=ctx.dim)[ctx.rank].contiguous(),
+            None,
+            None,
+        )
+
+
+def gather_seq_scatter_heads_qkv(
+    qkv_tensor: Tensor,
+    *,
+    seq_dim: int,
+    qkv_shape: Optional[Tensor] = None,
+    cache: Cache = Cache(disable=True),
+    restore_shape: bool = True,
+):
+    """
+    A func to sync splited qkv tensor
+    qkv_tensor: the tensor we want to do alltoall with. The last dim must
+        be the projection_idx, which we will split into 3 part. After
+        spliting, the gather idx will be projecttion_idx + 1
+    seq_dim: gather_dim for all2all comm
+    restore_shape: if True, output will has the same shape length as input
+    """
+    group = get_sequence_parallel_group()
+    if not group:
+        return qkv_tensor
+    world = get_sequence_parallel_world_size()
+    orig_shape = qkv_tensor.shape
+    scatter_dim = qkv_tensor.dim()
+    bef_all2all_shape = list(orig_shape)
+    qkv_proj_dim = bef_all2all_shape[-1]
+    bef_all2all_shape = bef_all2all_shape[:-1] + [3, qkv_proj_dim // 3]
+    qkv_tensor = qkv_tensor.view(bef_all2all_shape)
+    qkv_tensor = SeqAllToAll.apply(group, qkv_tensor, scatter_dim, seq_dim, False)
+    if restore_shape:
+        out_shape = list(orig_shape)
+        out_shape[seq_dim] *= world
+        out_shape[-1] = qkv_proj_dim // world
+        qkv_tensor = qkv_tensor.view(out_shape)
+
+    # remove padding
+    if qkv_shape is not None:
+        unpad_dim_size = cache(
+            "unpad_dim_size", lambda: torch.sum(torch.prod(qkv_shape, dim=-1)).item()
+        )
+        if unpad_dim_size % world != 0:
+            padding_size = qkv_tensor.size(seq_dim) - unpad_dim_size
+            qkv_tensor = _unpad_tensor(qkv_tensor, seq_dim, padding_size)
+    return qkv_tensor
+
+
+def slice_inputs(x: Tensor, dim: int, padding: bool = True):
+    """
+    A func to slice the input sequence in sequence parallel
+    """
+    group = get_sequence_parallel_group()
+    if group is None:
+        return x
+    sp_rank = get_sequence_parallel_rank()
+    sp_world = get_sequence_parallel_world_size()
+    dim_size = x.shape[dim]
+    unit = (dim_size + sp_world - 1) // sp_world
+    if padding and dim_size % sp_world:
+        padding_size = sp_world - (dim_size % sp_world)
+        x = _pad_tensor(x, dim, padding_size)
+    slc = [slice(None)] * len(x.shape)
+    slc[dim] = slice(unit * sp_rank, unit * (sp_rank + 1))
+    return x[slc]
+
+
+def remove_seqeunce_parallel_padding(x: Tensor, dim: int, unpad_dim_size: int):
+    """
+    A func to remove the padding part of the tensor based on its original shape
+    """
+    group = get_sequence_parallel_group()
+    if group is None:
+        return x
+    sp_world = get_sequence_parallel_world_size()
+    if unpad_dim_size % sp_world == 0:
+        return x
+    padding_size = sp_world - (unpad_dim_size % sp_world)
+    assert (padding_size + unpad_dim_size) % sp_world == 0
+    return _unpad_tensor(x, dim=dim, padding_size=padding_size)
+
+
+def gather_heads_scatter_seq(x: Tensor, head_dim: int, seq_dim: int) -> Tensor:
+    """
+    A func to sync attention result with alltoall in sequence parallel
+    """
+    group = get_sequence_parallel_group()
+    if not group:
+        return x
+    dim_size = x.size(seq_dim)
+    sp_world = get_sequence_parallel_world_size()
+    if dim_size % sp_world != 0:
+        padding_size = sp_world - (dim_size % sp_world)
+        x = _pad_tensor(x, seq_dim, padding_size)
+    return SeqAllToAll.apply(group, x, seq_dim, head_dim, False)
+
+
+def gather_seq_scatter_heads(x: Tensor, seq_dim: int, head_dim: int) -> Tensor:
+    """
+    A func to sync embedding input with alltoall in sequence parallel
+    """
+    group = get_sequence_parallel_group()
+    if not group:
+        return x
+    return SeqAllToAll.apply(group, x, head_dim, seq_dim, False)
+
+
+def scatter_heads(x: Tensor, dim: int) -> Tensor:
+    """
+    A func to split heads before attention in sequence parallel
+    """
+    group = get_sequence_parallel_group()
+    if not group:
+        return x
+    return Slice.apply(group, x, dim)
+
+
+def gather_heads(x: Tensor, dim: int, grad_scale: Optional[bool] = False) -> Tensor:
+    """
+    A func to gather heads for the attention result in sequence parallel
+    """
+    group = get_sequence_parallel_group()
+    if not group:
+        return x
+    return Gather.apply(group, x, dim, grad_scale)
+
+
+def gather_outputs(
+    x: Tensor,
+    *,
+    gather_dim: int,
+    padding_dim: Optional[int] = None,
+    unpad_shape: Optional[Tensor] = None,
+    cache: Cache = Cache(disable=True),
+    scale_grad=True,
+):
+    """
+    A func to gather the outputs for the model result in sequence parallel
+    """
+    group = get_sequence_parallel_group()
+    if not group:
+        return x
+    x = Gather.apply(group, x, gather_dim, scale_grad)
+    if padding_dim is not None:
+        unpad_dim_size = cache(
+            "unpad_dim_size", lambda: torch.sum(torch.prod(unpad_shape, dim=1)).item()
+        )
+        x = remove_seqeunce_parallel_padding(x, padding_dim, unpad_dim_size)
+    return x
+
+
+def _pad_tensor(x: Tensor, dim: int, padding_size: int):
+    shape = list(x.shape)
+    shape[dim] = padding_size
+    pad = torch.zeros(shape, dtype=x.dtype, device=x.device)
+    return torch.cat([x, pad], dim=dim)
+
+
+def _unpad_tensor(x: Tensor, dim: int, padding_size):
+    slc = [slice(None)] * len(x.shape)
+    slc[dim] = slice(0, -padding_size)
+    return x[slc]
+
+
+def _broadcast_data(data, shape, dtype, src, group, async_op):
+    comms = []
+    if isinstance(data, (list, tuple)):
+        for i, sub_shape in enumerate(shape):
+            comms += _broadcast_data(data[i], sub_shape, dtype[i], src, group, async_op)
+    elif isinstance(data, dict):
+        for key, sub_data in data.items():
+            comms += _broadcast_data(sub_data, shape[key], dtype[key], src, group, async_op)
+    elif isinstance(data, Tensor):
+        comms.append(dist.broadcast(data, src=src, group=group, async_op=async_op))
+    return comms
+
+
+def _traverse(data: Any, op: Callable) -> Union[None, List, Dict, Any]:
+    if isinstance(data, (list, tuple)):
+        return [_traverse(sub_data, op) for sub_data in data]
+    elif isinstance(data, dict):
+        return {key: _traverse(sub_data, op) for key, sub_data in data.items()}
+    elif isinstance(data, Tensor):
+        return op(data)
+    else:
+        return None
+
+
+def _get_shapes(data):
+    return _traverse(data, op=lambda x: x.shape)
+
+
+def _get_dtypes(data):
+    return _traverse(data, op=lambda x: x.dtype)
+
+
+def _construct_broadcast_buffer(shapes, dtypes, device):
+    if isinstance(shapes, torch.Size):
+        return torch.empty(shapes, dtype=dtypes, device=device)
+
+    if isinstance(shapes, (list, tuple)):
+        buffer = []
+        for i, sub_shape in enumerate(shapes):
+            buffer.append(_construct_broadcast_buffer(sub_shape, dtypes[i], device))
+    elif isinstance(shapes, dict):
+        buffer = {}
+        for key, sub_shape in shapes.items():
+            buffer[key] = _construct_broadcast_buffer(sub_shape, dtypes[key], device)
+    else:
+        return None
+    return buffer
+
+
+class SPDistForward:
+    """A forward tool to sync different result across sp group
+
+    Args:
+        module: a function or module to process users input
+        sp_step: current training step to judge which rank to broadcast its result to all
+        name: a distinct str to save meta and async comm
+        comm_shape: if different ranks have different shape, mark this arg to True
+        device: the device for current rank, can be empty
+    """
+
+    def __init__(
+        self,
+        name: str,
+        comm_shape: bool,
+        device: torch.device = None,
+    ):
+        self.name = name
+        self.comm_shape = comm_shape
+        if device:
+            self.device = device
+        else:
+            self.device = get_device()
+
+    def __call__(self, inputs) -> Any:
+        group = get_sequence_parallel_group()
+        if not group:
+            yield inputs
+        else:
+            device = self.device
+            sp_world = get_sequence_parallel_world_size()
+            sp_rank = get_sequence_parallel_rank()
+            for local_step in range(sp_world):
+                src_rank = dist.get_global_rank(group, local_step)
+                is_src = sp_rank == local_step
+                local_shapes = []
+                local_dtypes = []
+                if local_step == 0:
+                    local_result = inputs
+                    _SEQ_DATA_BUF[self.name][-1] = local_result
+                    local_shapes = _get_shapes(local_result)
+                    local_dtypes = _get_dtypes(local_result)
+                    if self.comm_shape:
+                        group_shapes_lists = [None] * sp_world
+                        dist.all_gather_object(group_shapes_lists, local_shapes, group=group)
+                        _SEQ_DATA_META_SHAPES[self.name] = group_shapes_lists
+                    else:
+                        _SEQ_DATA_META_SHAPES[self.name] = [local_shapes] * sp_world
+                    _SEQ_DATA_META_DTYPES[self.name] = local_dtypes
+                shapes = _SEQ_DATA_META_SHAPES[self.name][local_step]
+                dtypes = _SEQ_DATA_META_DTYPES[self.name]
+                buf_id = local_step % 2
+                if local_step == 0:
+                    sync_data = (
+                        local_result
+                        if is_src
+                        else _construct_broadcast_buffer(shapes, dtypes, device)
+                    )
+                    _broadcast_data(sync_data, shapes, dtypes, src_rank, group, False)
+                    _SEQ_DATA_BUF[self.name][buf_id] = sync_data
+
+                # wait for async comm ops
+                if _SEQ_DATA_ASYNC_COMMS[self.name]:
+                    for comm in _SEQ_DATA_ASYNC_COMMS[self.name]:
+                        comm.wait()
+                # before return the sync result, do async broadcast for next batch
+                if local_step < sp_world - 1:
+                    next_buf_id = 1 - buf_id
+                    shapes = _SEQ_DATA_META_SHAPES[self.name][local_step + 1]
+                    src_rank = dist.get_global_rank(group, local_step + 1)
+                    is_src = sp_rank == local_step + 1
+                    next_sync_data = (
+                        _SEQ_DATA_BUF[self.name][-1]
+                        if is_src
+                        else _construct_broadcast_buffer(shapes, dtypes, device)
+                    )
+                    _SEQ_DATA_ASYNC_COMMS[self.name] = _broadcast_data(
+                        next_sync_data, shapes, dtypes, src_rank, group, True
+                    )
+                    _SEQ_DATA_BUF[self.name][next_buf_id] = next_sync_data
+                yield _SEQ_DATA_BUF[self.name][buf_id]
+
+
+sync_inputs = SPDistForward(name="bef_fwd", comm_shape=True)
+
+
+def sync_data(data, sp_idx, name="tmp"):
+    group = get_sequence_parallel_group()
+    if group is None:
+        return data
+    # if sp_idx in _SYNC_BUFFER[name]:
+    #     return _SYNC_BUFFER[name][sp_idx]
+    sp_rank = get_sequence_parallel_rank()
+    src_rank = dist.get_global_rank(group, sp_idx)
+    objects = [data] if sp_rank == sp_idx else [None]
+    dist.broadcast_object_list(objects, src=src_rank, group=group)
+    # _SYNC_BUFFER[name] = {sp_idx: objects[0]}
+    return objects[0]

common/logger.py

@@ -0,0 +1,44 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Logging utility functions.
+"""
+
+import logging
+import sys
+from typing import Optional
+
+from common.distributed import get_global_rank, get_local_rank, get_world_size
+
+_default_handler = logging.StreamHandler(sys.stdout)
+_default_handler.setFormatter(
+    logging.Formatter(
+        "%(asctime)s "
+        + (f"[Rank:{get_global_rank()}]" if get_world_size() > 1 else "")
+        + (f"[LocalRank:{get_local_rank()}]" if get_world_size() > 1 else "")
+        + "[%(threadName).12s][%(name)s][%(levelname).5s] "
+        + "%(message)s"
+    )
+)
+
+
+def get_logger(name: Optional[str] = None) -> logging.Logger:
+    """
+    Get a logger.
+    """
+    logger = logging.getLogger(name)
+    logger.addHandler(_default_handler)
+    logger.setLevel(logging.INFO)
+    return logger

common/partition.py

@@ -0,0 +1,59 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+"""
+Partition utility functions.
+"""
+
+from typing import Any, List
+
+
+def partition_by_size(data: List[Any], size: int) -> List[List[Any]]:
+    """
+    Partition a list by size.
+    When indivisible, the last group contains fewer items than the target size.
+
+    Examples:
+        - data: [1,2,3,4,5]
+        - size: 2
+        - return: [[1,2], [3,4], [5]]
+    """
+    assert size > 0
+    return [data[i : (i + size)] for i in range(0, len(data), size)]
+
+
+def partition_by_groups(data: List[Any], groups: int) -> List[List[Any]]:
+    """
+    Partition a list by groups.
+    When indivisible, some groups may have more items than others.
+
+    Examples:
+        - data: [1,2,3,4,5]
+        - groups: 2
+        - return: [[1,3,5], [2,4]]
+    """
+    assert groups > 0
+    return [data[i::groups] for i in range(groups)]
+
+
+def shift_list(data: List[Any], n: int) -> List[Any]:
+    """
+    Rotate a list by n elements.
+
+    Examples:
+        - data: [1,2,3,4,5]
+        - n: 3
+        - return: [4,5,1,2,3]
+    """
+    return data[(n % len(data)) :] + data[: (n % len(data))]

common/seed.py

@@ -0,0 +1,30 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+import random
+from typing import Optional
+import numpy as np
+import torch
+
+from common.distributed import get_global_rank
+
+
+def set_seed(seed: Optional[int], same_across_ranks: bool = False):
+    """Function that sets the seed for pseudo-random number generators."""
+    if seed is not None:
+        seed += get_global_rank() if not same_across_ranks else 0
+        random.seed(seed)
+        np.random.seed(seed)
+        torch.manual_seed(seed)
+

configs_3b/main.yaml

@@ -0,0 +1,88 @@
+__object__:
+  path: projects.video_diffusion_sr.train
+  name: VideoDiffusionTrainer
+
+dit:
+  model:
+    __object__:
+      path: models.dit_v2.nadit
+      name: NaDiT
+      args: as_params
+    vid_in_channels: 33
+    vid_out_channels: 16
+    vid_dim: 2560
+    vid_out_norm: fusedrms
+    txt_in_dim: 5120
+    txt_in_norm: fusedln
+    txt_dim: ${.vid_dim}
+    emb_dim: ${eval:'6 * ${.vid_dim}'}
+    heads: 20
+    head_dim: 128 # llm-like
+    expand_ratio: 4
+    norm: fusedrms
+    norm_eps: 1.0e-05
+    ada: single
+    qk_bias: False
+    qk_norm: fusedrms
+    patch_size: [ 1,2,2 ]
+    num_layers: 32 # llm-like
+    mm_layers: 10
+    mlp_type: swiglu
+    msa_type: None
+    block_type: ${eval:'${.num_layers} * ["mmdit_sr"]'} # space-full
+    window: ${eval:'${.num_layers} * [(4,3,3)]'} # space-full
+    window_method: ${eval:'${.num_layers} // 2 * ["720pwin_by_size_bysize","720pswin_by_size_bysize"]'} # space-full
+    rope_type: mmrope3d
+    rope_dim: 128
+  compile: False
+  gradient_checkpoint: True
+  fsdp:
+    sharding_strategy: _HYBRID_SHARD_ZERO2
+
+ema:
+  decay: 0.9998
+
+vae:
+  model:
+    __inherit__: models/video_vae_v3/s8_c16_t4_inflation_sd3.yaml
+    freeze_encoder: False
+    # gradient_checkpoint: True
+  slicing:
+    split_size: 4
+    memory_device: same
+  memory_limit:
+    conv_max_mem: 0.5
+    norm_max_mem: 0.5
+  checkpoint: ./ckpts/ema_vae.pth
+  scaling_factor: 0.9152
+  compile: False
+  grouping: False
+  dtype: bfloat16
+
+diffusion:
+  schedule:
+    type: lerp
+    T: 1000.0
+  sampler:
+    type: euler
+    prediction_type: v_lerp
+  timesteps:
+    training:
+      type: logitnormal
+      loc: 0.0
+      scale: 1.0
+    sampling:
+      type: uniform_trailing
+      steps: 50
+    transform: True
+  loss:
+    type: v_lerp
+  cfg:
+    scale: 7.5
+    rescale: 0
+
+condition:
+  i2v: 0.0
+  v2v: 0.0
+  sr: 1.0
+  noise_scale: 0.25

configs_7b/main.yaml

@@ -0,0 +1,85 @@
+__object__:
+  path: projects.video_diffusion_sr.train
+  name: VideoDiffusionTrainer
+
+dit:
+  model:
+    __object__:
+      path: models.dit.nadit
+      name: NaDiT
+      args: as_params
+    vid_in_channels: 33
+    vid_out_channels: 16
+    vid_dim: 3072
+    txt_in_dim: 5120
+    txt_dim: ${.vid_dim}
+    emb_dim: ${eval:'6 * ${.vid_dim}'}
+    heads: 24
+    head_dim: 128 # llm-like
+    expand_ratio: 4
+    norm: fusedrms
+    norm_eps: 1e-5
+    ada: single
+    qk_bias: False
+    qk_rope: True
+    qk_norm: fusedrms
+    patch_size: [ 1,2,2 ]
+    num_layers: 36 # llm-like
+    shared_mlp: False
+    shared_qkv: False
+    mlp_type: normal
+    block_type: ${eval:'${.num_layers} * ["mmdit_sr"]'} # space-full
+    window: ${eval:'${.num_layers} * [(4,3,3)]'} # space-full
+    window_method: ${eval:'${.num_layers} // 2 * ["720pwin_by_size_bysize","720pswin_by_size_bysize"]'} # space-full
+  compile: False
+  gradient_checkpoint: True
+  fsdp:
+    sharding_strategy: _HYBRID_SHARD_ZERO2
+
+ema:
+  decay: 0.9998
+
+vae:
+  model:
+    __inherit__: models/video_vae_v3/s8_c16_t4_inflation_sd3.yaml
+    freeze_encoder: False
+    # gradient_checkpoint: True
+  slicing:
+    split_size: 4
+    memory_device: same
+  memory_limit:
+    conv_max_mem: 0.5
+    norm_max_mem: 0.5
+  checkpoint: ./ckpts/ema_vae.pth
+  scaling_factor: 0.9152
+  compile: False
+  grouping: False
+  dtype: bfloat16
+
+diffusion:
+  schedule:
+    type: lerp
+    T: 1000.0
+  sampler:
+    type: euler
+    prediction_type: v_lerp
+  timesteps:
+    training:
+      type: logitnormal
+      loc: 0.0
+      scale: 1.0
+    sampling:
+      type: uniform_trailing
+      steps: 50
+    transform: True
+  loss:
+    type: v_lerp
+  cfg:
+    scale: 7.5
+    rescale: 0
+
+condition:
+  i2v: 0.0
+  v2v: 0.0
+  sr: 1.0
+  noise_scale: 0.25

data/image/transforms/area_resize.py

@@ -0,0 +1,135 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+import math
+import random
+from typing import Union
+import torch
+from PIL import Image
+from torchvision.transforms import functional as TVF
+from torchvision.transforms.functional import InterpolationMode
+
+
+class AreaResize:
+    def __init__(
+        self,
+        max_area: float,
+        downsample_only: bool = False,
+        interpolation: InterpolationMode = InterpolationMode.BICUBIC,
+    ):
+        self.max_area = max_area
+        self.downsample_only = downsample_only
+        self.interpolation = interpolation
+
+    def __call__(self, image: Union[torch.Tensor, Image.Image]):
+
+        if isinstance(image, torch.Tensor):
+            height, width = image.shape[-2:]
+        elif isinstance(image, Image.Image):
+            width, height = image.size
+        else:
+            raise NotImplementedError
+
+        scale = math.sqrt(self.max_area / (height * width))
+
+        # keep original height and width for small pictures.
+        scale = 1 if scale >= 1 and self.downsample_only else scale
+
+        resized_height, resized_width = round(height * scale), round(width * scale)
+
+        return TVF.resize(
+            image,
+            size=(resized_height, resized_width),
+            interpolation=self.interpolation,
+        )
+
+
+class AreaRandomCrop:
+    def __init__(
+        self,
+        max_area: float,
+    ):
+        self.max_area = max_area
+
+    def get_params(self, input_size, output_size):
+        """Get parameters for ``crop`` for a random crop.
+
+        Args:
+            img (PIL Image): Image to be cropped.
+            output_size (tuple): Expected output size of the crop.
+
+        Returns:
+            tuple: params (i, j, h, w) to be passed to ``crop`` for random crop.
+        """
+        # w, h = _get_image_size(img)
+        h, w = input_size
+        th, tw = output_size
+        if w <= tw and h <= th:
+            return 0, 0, h, w
+
+        i = random.randint(0, h - th)
+        j = random.randint(0, w - tw)
+        return i, j, th, tw
+
+    def __call__(self, image: Union[torch.Tensor, Image.Image]):
+        if isinstance(image, torch.Tensor):
+            height, width = image.shape[-2:]
+        elif isinstance(image, Image.Image):
+            width, height = image.size
+        else:
+            raise NotImplementedError
+
+        resized_height = math.sqrt(self.max_area / (width / height))
+        resized_width = (width / height) * resized_height
+
+        # print('>>>>>>>>>>>>>>>>>>>>>')
+        # print((height, width))
+        # print( (resized_height, resized_width))
+
+        resized_height, resized_width = round(resized_height), round(resized_width)
+        i, j, h, w = self.get_params((height, width), (resized_height, resized_width))
+        image = TVF.crop(image, i, j, h, w)
+        return image
+
+class ScaleResize:
+    def __init__(
+        self,
+        scale: float,
+    ):
+        self.scale = scale
+
+    def __call__(self, image: Union[torch.Tensor, Image.Image]):
+        if isinstance(image, torch.Tensor):
+            height, width = image.shape[-2:]
+            interpolation_mode = InterpolationMode.BILINEAR
+            antialias = True if image.ndim == 4 else "warn"
+        elif isinstance(image, Image.Image):
+            width, height = image.size
+            interpolation_mode = InterpolationMode.LANCZOS
+            antialias = "warn"
+        else:
+            raise NotImplementedError
+
+        scale = self.scale
+
+        # keep original height and width for small pictures
+
+        resized_height, resized_width = round(height * scale), round(width * scale)
+        image = TVF.resize(
+            image,
+            size=(resized_height, resized_width),
+            interpolation=interpolation_mode,
+            antialias=antialias,
+        )
+        return image

data/image/transforms/divisible_crop.py

@@ -0,0 +1,40 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from typing import Union
+import torch
+from PIL import Image
+from torchvision.transforms import functional as TVF
+
+
+class DivisibleCrop:
+    def __init__(self, factor):
+        if not isinstance(factor, tuple):
+            factor = (factor, factor)
+
+        self.height_factor, self.width_factor = factor[0], factor[1]
+
+    def __call__(self, image: Union[torch.Tensor, Image.Image]):
+        if isinstance(image, torch.Tensor):
+            height, width = image.shape[-2:]
+        elif isinstance(image, Image.Image):
+            width, height = image.size
+        else:
+            raise NotImplementedError
+
+        cropped_height = height - (height % self.height_factor)
+        cropped_width = width - (width % self.width_factor)
+
+        image = TVF.center_crop(img=image, output_size=(cropped_height, cropped_width))
+        return image

data/image/transforms/na_resize.py

@@ -0,0 +1,50 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from typing import Literal
+from torchvision.transforms import CenterCrop, Compose, InterpolationMode, Resize
+
+from .area_resize import AreaResize
+from .side_resize import SideResize
+
+
+def NaResize(
+    resolution: int,
+    mode: Literal["area", "side"],
+    downsample_only: bool,
+    interpolation: InterpolationMode = InterpolationMode.BICUBIC,
+):
+    if mode == "area":
+        return AreaResize(
+            max_area=resolution**2,
+            downsample_only=downsample_only,
+            interpolation=interpolation,
+        )
+    if mode == "side":
+        return SideResize(
+            size=resolution,
+            downsample_only=downsample_only,
+            interpolation=interpolation,
+        )
+    if mode == "square":
+        return Compose(
+            [
+                Resize(
+                    size=resolution,
+                    interpolation=interpolation,
+                ),
+                CenterCrop(resolution),
+            ]
+        )
+    raise ValueError(f"Unknown resize mode: {mode}")

data/image/transforms/side_resize.py

@@ -0,0 +1,54 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from typing import Union
+import torch
+from PIL import Image
+from torchvision.transforms import InterpolationMode
+from torchvision.transforms import functional as TVF
+
+
+class SideResize:
+    def __init__(
+        self,
+        size: int,
+        downsample_only: bool = False,
+        interpolation: InterpolationMode = InterpolationMode.BICUBIC,
+    ):
+        self.size = size
+        self.downsample_only = downsample_only
+        self.interpolation = interpolation
+
+    def __call__(self, image: Union[torch.Tensor, Image.Image]):
+        """
+        Args:
+            image (PIL Image or Tensor): Image to be scaled.
+
+        Returns:
+            PIL Image or Tensor: Rescaled image.
+        """
+        if isinstance(image, torch.Tensor):
+            height, width = image.shape[-2:]
+        elif isinstance(image, Image.Image):
+            width, height = image.size
+        else:
+            raise NotImplementedError
+
+        if self.downsample_only and min(width, height) < self.size:
+            # keep original height and width for small pictures.
+            size = min(width, height)
+        else:
+            size = self.size
+
+        return TVF.resize(image, size, self.interpolation)

data/video/transforms/rearrange.py

@@ -0,0 +1,24 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from einops import rearrange
+
+
+class Rearrange:
+    def __init__(self, pattern: str, **kwargs):
+        self.pattern = pattern
+        self.kwargs = kwargs
+
+    def __call__(self, x):
+        return rearrange(x, self.pattern, **self.kwargs)

models/dit/attention.py

@@ -0,0 +1,46 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+import torch
+import torch.nn.functional as F
+
+from flash_attn import flash_attn_varlen_func
+
+from torch import nn
+
+class TorchAttention(nn.Module):
+    def tflops(self, args, kwargs, output) -> float:
+        assert len(args) == 0 or len(args) > 2, "query, key should both provided by args / kwargs"
+        q = kwargs.get("query") or args[0]
+        k = kwargs.get("key") or args[1]
+        b, h, sq, d = q.shape
+        b, h, sk, d = k.shape
+        return b * h * (4 * d * (sq / 1e6) * (sk / 1e6))
+
+    def forward(self, *args, **kwargs):
+        return F.scaled_dot_product_attention(*args, **kwargs)
+
+
+class FlashAttentionVarlen(nn.Module):
+    def tflops(self, args, kwargs, output) -> float:
+        cu_seqlens_q = kwargs["cu_seqlens_q"]
+        cu_seqlens_k = kwargs["cu_seqlens_k"]
+        _, h, d = output.shape
+        seqlens_q = (cu_seqlens_q[1:] - cu_seqlens_q[:-1]) / 1e6
+        seqlens_k = (cu_seqlens_k[1:] - cu_seqlens_k[:-1]) / 1e6
+        return h * (4 * d * (seqlens_q * seqlens_k).sum())
+
+    def forward(self, *args, **kwargs):
+        kwargs["deterministic"] = torch.are_deterministic_algorithms_enabled()
+        return flash_attn_varlen_func(*args, **kwargs)

models/dit/blocks/__init__.py

@@ -0,0 +1,25 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from .mmdit_window_block import MMWindowTransformerBlock
+
+dit_blocks = {
+    "mmdit_window": MMWindowTransformerBlock,
+}
+
+
+def get_block(block_type: str):
+    if block_type in dit_blocks:
+        return dit_blocks[block_type]
+    raise NotImplementedError(f"{block_type} is not supported")

models/dit/blocks/mmdit_window_block.py

@@ -0,0 +1,233 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from typing import Tuple, Union
+import torch
+from einops import rearrange
+from torch import nn
+from torch.nn import functional as F
+from torch.nn.modules.utils import _triple
+
+from common.distributed.ops import (
+    gather_heads,
+    gather_heads_scatter_seq,
+    gather_seq_scatter_heads_qkv,
+    scatter_heads,
+)
+
+from ..attention import TorchAttention
+from ..mlp import get_mlp
+from ..mm import MMArg, MMModule
+from ..modulation import ada_layer_type
+from ..normalization import norm_layer_type
+from ..rope import RotaryEmbedding3d
+
+
+class MMWindowAttention(nn.Module):
+    def __init__(
+        self,
+        vid_dim: int,
+        txt_dim: int,
+        heads: int,
+        head_dim: int,
+        qk_bias: bool,
+        qk_rope: bool,
+        qk_norm: norm_layer_type,
+        qk_norm_eps: float,
+        window: Union[int, Tuple[int, int, int]],
+        window_method: str,
+        shared_qkv: bool,
+    ):
+        super().__init__()
+        dim = MMArg(vid_dim, txt_dim)
+        inner_dim = heads * head_dim
+        qkv_dim = inner_dim * 3
+
+        self.window = _triple(window)
+        self.window_method = window_method
+        assert all(map(lambda v: isinstance(v, int) and v >= 0, self.window))
+
+        self.head_dim = head_dim
+        self.proj_qkv = MMModule(nn.Linear, dim, qkv_dim, bias=qk_bias, shared_weights=shared_qkv)
+        self.proj_out = MMModule(nn.Linear, inner_dim, dim, shared_weights=shared_qkv)
+        self.norm_q = MMModule(qk_norm, dim=head_dim, eps=qk_norm_eps, elementwise_affine=True)
+        self.norm_k = MMModule(qk_norm, dim=head_dim, eps=qk_norm_eps, elementwise_affine=True)
+        self.rope = RotaryEmbedding3d(dim=head_dim // 2) if qk_rope else None
+        self.attn = TorchAttention()
+
+    def forward(
+        self,
+        vid: torch.FloatTensor,  # b T H W c
+        txt: torch.FloatTensor,  # b L c
+        txt_mask: torch.BoolTensor,  # b L
+    ) -> Tuple[
+        torch.FloatTensor,
+        torch.FloatTensor,
+    ]:
+        # Project q, k, v.
+        vid_qkv, txt_qkv = self.proj_qkv(vid, txt)
+        vid_qkv = gather_seq_scatter_heads_qkv(vid_qkv, seq_dim=2)
+        _, T, H, W, _ = vid_qkv.shape
+        _, L, _ = txt.shape
+
+        if self.window_method == "win":
+            nt, nh, nw = self.window
+            tt, hh, ww = T // nt, H // nh, W // nw
+        elif self.window_method == "win_by_size":
+            tt, hh, ww = self.window
+            tt, hh, ww = (
+                tt if tt > 0 else T,
+                hh if hh > 0 else H,
+                ww if ww > 0 else W,
+            )
+            nt, nh, nw = T // tt, H // hh, W // ww
+        else:
+            raise NotImplementedError
+
+        vid_qkv = rearrange(vid_qkv, "b T H W (o h d) -> o b h (T H W) d", o=3, d=self.head_dim)
+        txt_qkv = rearrange(txt_qkv, "b L (o h d) -> o b h L d", o=3, d=self.head_dim)
+        txt_qkv = scatter_heads(txt_qkv, dim=2)
+
+        vid_q, vid_k, vid_v = vid_qkv.unbind()
+        txt_q, txt_k, txt_v = txt_qkv.unbind()
+
+        vid_q, txt_q = self.norm_q(vid_q, txt_q)
+        vid_k, txt_k = self.norm_k(vid_k, txt_k)
+
+        if self.rope:
+            vid_q, vid_k = self.rope(vid_q, vid_k, (T, H, W))
+
+        def vid_window(v):
+            return rearrange(
+                v,
+                "b h (nt tt nh hh nw ww) d -> b h (nt nh nw) (tt hh ww) d",
+                hh=hh,
+                ww=ww,
+                tt=tt,
+                nh=nh,
+                nw=nw,
+                nt=nt,
+            )
+
+        def txt_window(t):
+            return rearrange(t, "b h L d -> b h 1 L d").expand(-1, -1, nt * nh * nw, -1, -1)
+
+        # Process video attention.
+        vid_msk = F.pad(txt_mask, (tt * hh * ww, 0), value=True)
+        vid_msk = rearrange(vid_msk, "b l -> b 1 1 1 l").expand(-1, 1, 1, tt * hh * ww, -1)
+        vid_out = self.attn(
+            vid_window(vid_q),
+            torch.cat([vid_window(vid_k), txt_window(txt_k)], dim=-2),
+            torch.cat([vid_window(vid_v), txt_window(txt_v)], dim=-2),
+            vid_msk,
+        )
+        vid_out = rearrange(
+            vid_out,
+            "b h (nt nh nw) (tt hh ww) d -> b (nt tt) (nh hh) (nw ww) (h d)",
+            hh=hh,
+            ww=ww,
+            tt=tt,
+            nh=nh,
+            nw=nw,
+        )
+        vid_out = gather_heads_scatter_seq(vid_out, head_dim=4, seq_dim=2)
+
+        # Process text attention.
+        txt_msk = F.pad(txt_mask, (T * H * W, 0), value=True)
+        txt_msk = rearrange(txt_msk, "b l -> b 1 1 l").expand(-1, 1, L, -1)
+        txt_out = self.attn(
+            txt_q,
+            torch.cat([vid_k, txt_k], dim=-2),
+            torch.cat([vid_v, txt_v], dim=-2),
+            txt_msk,
+        )
+        txt_out = rearrange(txt_out, "b h L d -> b L (h d)")
+        txt_out = gather_heads(txt_out, dim=2)
+
+        # Project output.
+        vid_out, txt_out = self.proj_out(vid_out, txt_out)
+        return vid_out, txt_out
+
+
+class MMWindowTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        vid_dim: int,
+        txt_dim: int,
+        emb_dim: int,
+        heads: int,
+        head_dim: int,
+        expand_ratio: int,
+        norm: norm_layer_type,
+        norm_eps: float,
+        ada: ada_layer_type,
+        qk_bias: bool,
+        qk_rope: bool,
+        qk_norm: norm_layer_type,
+        window: Union[int, Tuple[int, int, int]],
+        window_method: str,
+        shared_qkv: bool,
+        shared_mlp: bool,
+        mlp_type: str,
+        **kwargs,
+    ):
+        super().__init__()
+        dim = MMArg(vid_dim, txt_dim)
+        self.attn_norm = MMModule(norm, dim=dim, eps=norm_eps, elementwise_affine=False)
+        self.attn = MMWindowAttention(
+            vid_dim=vid_dim,
+            txt_dim=txt_dim,
+            heads=heads,
+            head_dim=head_dim,
+            qk_bias=qk_bias,
+            qk_rope=qk_rope,
+            qk_norm=qk_norm,
+            qk_norm_eps=norm_eps,
+            window=window,
+            window_method=window_method,
+            shared_qkv=shared_qkv,
+        )
+        self.mlp_norm = MMModule(norm, dim=dim, eps=norm_eps, elementwise_affine=False)
+        self.mlp = MMModule(
+            get_mlp(mlp_type),
+            dim=dim,
+            expand_ratio=expand_ratio,
+            shared_weights=shared_mlp,
+        )
+        self.ada = MMModule(ada, dim=dim, emb_dim=emb_dim, layers=["attn", "mlp"])
+
+    def forward(
+        self,
+        vid: torch.FloatTensor,
+        txt: torch.FloatTensor,
+        txt_mask: torch.BoolTensor,
+        emb: torch.FloatTensor,
+    ) -> Tuple[
+        torch.FloatTensor,
+        torch.FloatTensor,
+    ]:
+        vid_attn, txt_attn = self.attn_norm(vid, txt)
+        vid_attn, txt_attn = self.ada(vid_attn, txt_attn, emb=emb, layer="attn", mode="in")
+        vid_attn, txt_attn = self.attn(vid_attn, txt_attn, txt_mask=txt_mask)
+        vid_attn, txt_attn = self.ada(vid_attn, txt_attn, emb=emb, layer="attn", mode="out")
+        vid_attn, txt_attn = (vid_attn + vid), (txt_attn + txt)
+
+        vid_mlp, txt_mlp = self.mlp_norm(vid_attn, txt_attn)
+        vid_mlp, txt_mlp = self.ada(vid_mlp, txt_mlp, emb=emb, layer="mlp", mode="in")
+        vid_mlp, txt_mlp = self.mlp(vid_mlp, txt_mlp)
+        vid_mlp, txt_mlp = self.ada(vid_mlp, txt_mlp, emb=emb, layer="mlp", mode="out")
+        vid_mlp, txt_mlp = (vid_mlp + vid_attn), (txt_mlp + txt_attn)
+
+        return vid_mlp, txt_mlp

models/dit/embedding.py

@@ -0,0 +1,62 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from typing import Optional, Union
+import torch
+from diffusers.models.embeddings import get_timestep_embedding
+from torch import nn
+
+
+def emb_add(emb1: torch.Tensor, emb2: Optional[torch.Tensor]):
+    return emb1 if emb2 is None else emb1 + emb2
+
+
+class TimeEmbedding(nn.Module):
+    def __init__(
+        self,
+        sinusoidal_dim: int,
+        hidden_dim: int,
+        output_dim: int,
+    ):
+        super().__init__()
+        self.sinusoidal_dim = sinusoidal_dim
+        self.proj_in = nn.Linear(sinusoidal_dim, hidden_dim)
+        self.proj_hid = nn.Linear(hidden_dim, hidden_dim)
+        self.proj_out = nn.Linear(hidden_dim, output_dim)
+        self.act = nn.SiLU()
+
+    def forward(
+        self,
+        timestep: Union[int, float, torch.IntTensor, torch.FloatTensor],
+        device: torch.device,
+        dtype: torch.dtype,
+    ) -> torch.FloatTensor:
+        if not torch.is_tensor(timestep):
+            timestep = torch.tensor([timestep], device=device, dtype=dtype)
+        if timestep.ndim == 0:
+            timestep = timestep[None]
+
+        emb = get_timestep_embedding(
+            timesteps=timestep,
+            embedding_dim=self.sinusoidal_dim,
+            flip_sin_to_cos=False,
+            downscale_freq_shift=0,
+        )
+        emb = emb.to(dtype)
+        emb = self.proj_in(emb)
+        emb = self.act(emb)
+        emb = self.proj_hid(emb)
+        emb = self.act(emb)
+        emb = self.proj_out(emb)
+        return emb

models/dit/mlp.py

@@ -0,0 +1,62 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from typing import Optional
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+def get_mlp(mlp_type: Optional[str] = "normal"):
+    if mlp_type == "normal":
+        return MLP
+    elif mlp_type == "swiglu":
+        return SwiGLUMLP
+
+
+class MLP(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        expand_ratio: int,
+    ):
+        super().__init__()
+        self.proj_in = nn.Linear(dim, dim * expand_ratio)
+        self.act = nn.GELU("tanh")
+        self.proj_out = nn.Linear(dim * expand_ratio, dim)
+
+    def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+        x = self.proj_in(x)
+        x = self.act(x)
+        x = self.proj_out(x)
+        return x
+
+
+class SwiGLUMLP(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        expand_ratio: int,
+        multiple_of: int = 256,
+    ):
+        super().__init__()
+        hidden_dim = int(2 * dim * expand_ratio / 3)
+        hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+        self.proj_in_gate = nn.Linear(dim, hidden_dim, bias=False)
+        self.proj_out = nn.Linear(hidden_dim, dim, bias=False)
+        self.proj_in = nn.Linear(dim, hidden_dim, bias=False)
+
+    def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+        x = self.proj_out(F.silu(self.proj_in_gate(x)) * self.proj_in(x))
+        return x

models/dit/mm.py

@@ -0,0 +1,67 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Tuple
+import torch
+from torch import nn
+
+
+@dataclass
+class MMArg:
+    vid: Any
+    txt: Any
+
+
+def get_args(key: str, args: List[Any]) -> List[Any]:
+    return [getattr(v, key) if isinstance(v, MMArg) else v for v in args]
+
+
+def get_kwargs(key: str, kwargs: Dict[str, Any]) -> Dict[str, Any]:
+    return {k: getattr(v, key) if isinstance(v, MMArg) else v for k, v in kwargs.items()}
+
+
+class MMModule(nn.Module):
+    def __init__(
+        self,
+        module: Callable[..., nn.Module],
+        *args,
+        shared_weights: bool = False,
+        **kwargs,
+    ):
+        super().__init__()
+        self.shared_weights = shared_weights
+        if self.shared_weights:
+            assert get_args("vid", args) == get_args("txt", args)
+            assert get_kwargs("vid", kwargs) == get_kwargs("txt", kwargs)
+            self.all = module(*get_args("vid", args), **get_kwargs("vid", kwargs))
+        else:
+            self.vid = module(*get_args("vid", args), **get_kwargs("vid", kwargs))
+            self.txt = module(*get_args("txt", args), **get_kwargs("txt", kwargs))
+
+    def forward(
+        self,
+        vid: torch.FloatTensor,
+        txt: torch.FloatTensor,
+        *args,
+        **kwargs,
+    ) -> Tuple[
+        torch.FloatTensor,
+        torch.FloatTensor,
+    ]:
+        vid_module = self.vid if not self.shared_weights else self.all
+        txt_module = self.txt if not self.shared_weights else self.all
+        vid = vid_module(vid, *get_args("vid", args), **get_kwargs("vid", kwargs))
+        txt = txt_module(txt, *get_args("txt", args), **get_kwargs("txt", kwargs))
+        return vid, txt

models/dit/modulation.py

@@ -0,0 +1,97 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from typing import Callable, List, Optional
+import torch
+from einops import rearrange
+from torch import nn
+
+from common.cache import Cache
+from common.distributed.ops import slice_inputs
+
+# (dim: int, emb_dim: int)
+ada_layer_type = Callable[[int, int], nn.Module]
+
+
+def get_ada_layer(ada_layer: str) -> ada_layer_type:
+    if ada_layer == "single":
+        return AdaSingle
+    raise NotImplementedError(f"{ada_layer} is not supported")
+
+
+def expand_dims(x: torch.Tensor, dim: int, ndim: int):
+    """
+    Expand tensor "x" to "ndim" by adding empty dims at "dim".
+    Example: x is (b d), target ndim is 5, add dim at 1, return (b 1 1 1 d).
+    """
+    shape = x.shape
+    shape = shape[:dim] + (1,) * (ndim - len(shape)) + shape[dim:]
+    return x.reshape(shape)
+
+
+class AdaSingle(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        emb_dim: int,
+        layers: List[str],
+    ):
+        assert emb_dim == 6 * dim, "AdaSingle requires emb_dim == 6 * dim"
+        super().__init__()
+        self.dim = dim
+        self.emb_dim = emb_dim
+        self.layers = layers
+        for l in layers:
+            self.register_parameter(f"{l}_shift", nn.Parameter(torch.randn(dim) / dim**0.5))
+            self.register_parameter(f"{l}_scale", nn.Parameter(torch.randn(dim) / dim**0.5 + 1))
+            self.register_parameter(f"{l}_gate", nn.Parameter(torch.randn(dim) / dim**0.5))
+
+    def forward(
+        self,
+        hid: torch.FloatTensor,  # b ... c
+        emb: torch.FloatTensor,  # b d
+        layer: str,
+        mode: str,
+        cache: Cache = Cache(disable=True),
+        branch_tag: str = "",
+        hid_len: Optional[torch.LongTensor] = None,  # b
+    ) -> torch.FloatTensor:
+        idx = self.layers.index(layer)
+        emb = rearrange(emb, "b (d l g) -> b d l g", l=len(self.layers), g=3)[..., idx, :]
+        emb = expand_dims(emb, 1, hid.ndim + 1)
+
+        if hid_len is not None:
+            emb = cache(
+                f"emb_repeat_{idx}_{branch_tag}",
+                lambda: slice_inputs(
+                    torch.cat([e.repeat(l, *([1] * e.ndim)) for e, l in zip(emb, hid_len)]),
+                    dim=0,
+                ),
+            )
+
+        shiftA, scaleA, gateA = emb.unbind(-1)
+        shiftB, scaleB, gateB = (
+            getattr(self, f"{layer}_shift"),
+            getattr(self, f"{layer}_scale"),
+            getattr(self, f"{layer}_gate"),
+        )
+
+        if mode == "in":
+            return hid.mul_(scaleA + scaleB).add_(shiftA + shiftB)
+        if mode == "out":
+            return hid.mul_(gateA + gateB)
+        raise NotImplementedError
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, emb_dim={self.emb_dim}, layers={self.layers}"

models/dit/na.py

@@ -0,0 +1,241 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from itertools import chain
+from typing import Callable, Dict, List, Tuple
+import einops
+import torch
+
+
+def flatten(
+    hid: List[torch.FloatTensor],  # List of (*** c)
+) -> Tuple[
+    torch.FloatTensor,  # (L c)
+    torch.LongTensor,  # (b n)
+]:
+    assert len(hid) > 0
+    shape = torch.stack([torch.tensor(x.shape[:-1], device=hid[0].device) for x in hid])
+    hid = torch.cat([x.flatten(0, -2) for x in hid])
+    return hid, shape
+
+
+def unflatten(
+    hid: torch.FloatTensor,  # (L c) or (L ... c)
+    hid_shape: torch.LongTensor,  # (b n)
+) -> List[torch.Tensor]:  # List of (*** c) or (*** ... c)
+    hid_len = hid_shape.prod(-1)
+    hid = hid.split(hid_len.tolist())
+    hid = [x.unflatten(0, s.tolist()) for x, s in zip(hid, hid_shape)]
+    return hid
+
+
+def concat(
+    vid: torch.FloatTensor,  # (VL ... c)
+    txt: torch.FloatTensor,  # (TL ... c)
+    vid_len: torch.LongTensor,  # (b)
+    txt_len: torch.LongTensor,  # (b)
+) -> torch.FloatTensor:  # (L ... c)
+    vid = torch.split(vid, vid_len.tolist())
+    txt = torch.split(txt, txt_len.tolist())
+    return torch.cat(list(chain(*zip(vid, txt))))
+
+
+def concat_idx(
+    vid_len: torch.LongTensor,  # (b)
+    txt_len: torch.LongTensor,  # (b)
+) -> Tuple[
+    Callable,
+    Callable,
+]:
+    device = vid_len.device
+    vid_idx = torch.arange(vid_len.sum(), device=device)
+    txt_idx = torch.arange(len(vid_idx), len(vid_idx) + txt_len.sum(), device=device)
+    tgt_idx = concat(vid_idx, txt_idx, vid_len, txt_len)
+    src_idx = torch.argsort(tgt_idx)
+    return (
+        lambda vid, txt: torch.index_select(torch.cat([vid, txt]), 0, tgt_idx),
+        lambda all: torch.index_select(all, 0, src_idx).split([len(vid_idx), len(txt_idx)]),
+    )
+
+
+def unconcat(
+    all: torch.FloatTensor,  # (L ... c)
+    vid_len: torch.LongTensor,  # (b)
+    txt_len: torch.LongTensor,  # (b)
+) -> Tuple[
+    torch.FloatTensor,  # (VL ... c)
+    torch.FloatTensor,  # (TL ... c)
+]:
+    interleave_len = list(chain(*zip(vid_len.tolist(), txt_len.tolist())))
+    all = all.split(interleave_len)
+    vid = torch.cat(all[0::2])
+    txt = torch.cat(all[1::2])
+    return vid, txt
+
+
+def repeat_concat(
+    vid: torch.FloatTensor,  # (VL ... c)
+    txt: torch.FloatTensor,  # (TL ... c)
+    vid_len: torch.LongTensor,  # (n*b)
+    txt_len: torch.LongTensor,  # (b)
+    txt_repeat: List,  # (n)
+) -> torch.FloatTensor:  # (L ... c)
+    vid = torch.split(vid, vid_len.tolist())
+    txt = torch.split(txt, txt_len.tolist())
+    txt = [[x] * n for x, n in zip(txt, txt_repeat)]
+    txt = list(chain(*txt))
+    return torch.cat(list(chain(*zip(vid, txt))))
+
+
+def repeat_concat_idx(
+    vid_len: torch.LongTensor,  # (n*b)
+    txt_len: torch.LongTensor,  # (b)
+    txt_repeat: torch.LongTensor,  # (n)
+) -> Tuple[
+    Callable,
+    Callable,
+]:
+    device = vid_len.device
+    vid_idx = torch.arange(vid_len.sum(), device=device)
+    txt_idx = torch.arange(len(vid_idx), len(vid_idx) + txt_len.sum(), device=device)
+    txt_repeat_list = txt_repeat.tolist()
+    tgt_idx = repeat_concat(vid_idx, txt_idx, vid_len, txt_len, txt_repeat)
+    src_idx = torch.argsort(tgt_idx)
+    txt_idx_len = len(tgt_idx) - len(vid_idx)
+    repeat_txt_len = (txt_len * txt_repeat).tolist()
+
+    def unconcat_coalesce(all):
+        """
+        Un-concat vid & txt, and coalesce the repeated txt.
+        e.g. vid [0 1 2 3 4 5 6 7 8] -> 3 splits -> [0 1 2] [3 4 5] [6 7 8]
+             txt [9 10]
+             repeat_concat ==> [0 1 2 9 10 3 4 5 9 10 6 7 8 9 10]
+             1. argsort re-index ==> [0 1 2 3 4 5 6 7 8 9 9 9 10 10 10]
+                           split ==> vid_out [0 1 2 3 4 5 6 7 8] txt_out [9 9 9 10 10 10]
+             2. reshape & mean for each sample to coalesce the repeated txt.
+        """
+        vid_out, txt_out = all[src_idx].split([len(vid_idx), txt_idx_len])
+        txt_out_coalesced = []
+        for txt, repeat_time in zip(txt_out.split(repeat_txt_len), txt_repeat_list):
+            txt = txt.reshape(-1, repeat_time, *txt.shape[1:]).mean(1)
+            txt_out_coalesced.append(txt)
+        return vid_out, torch.cat(txt_out_coalesced)
+
+    # Note: Backward of torch.index_select is non-deterministic when existing repeated index,
+    # the difference may cumulative like torch.repeat_interleave, so we use vanilla index here.
+    return (
+        lambda vid, txt: torch.cat([vid, txt])[tgt_idx],
+        lambda all: unconcat_coalesce(all),
+    )
+
+
+def rearrange(
+    hid: torch.FloatTensor,  # (L c)
+    hid_shape: torch.LongTensor,  # (b n)
+    pattern: str,
+    **kwargs: Dict[str, int],
+) -> Tuple[
+    torch.FloatTensor,
+    torch.LongTensor,
+]:
+    return flatten([einops.rearrange(h, pattern, **kwargs) for h in unflatten(hid, hid_shape)])
+
+
+def rearrange_idx(
+    hid_shape: torch.LongTensor,  # (b n)
+    pattern: str,
+    **kwargs: Dict[str, int],
+) -> Tuple[Callable, Callable, torch.LongTensor]:
+    hid_idx = torch.arange(hid_shape.prod(-1).sum(), device=hid_shape.device).unsqueeze(-1)
+    tgt_idx, tgt_shape = rearrange(hid_idx, hid_shape, pattern, **kwargs)
+    tgt_idx = tgt_idx.squeeze(-1)
+    src_idx = torch.argsort(tgt_idx)
+    return (
+        lambda hid: torch.index_select(hid, 0, tgt_idx),
+        lambda hid: torch.index_select(hid, 0, src_idx),
+        tgt_shape,
+    )
+
+
+def repeat(
+    hid: torch.FloatTensor,  # (L c)
+    hid_shape: torch.LongTensor,  # (b n)
+    pattern: str,
+    **kwargs: Dict[str, torch.LongTensor],  # (b)
+) -> Tuple[
+    torch.FloatTensor,
+    torch.LongTensor,
+]:
+    hid = unflatten(hid, hid_shape)
+    kwargs = [{k: v[i].item() for k, v in kwargs.items()} for i in range(len(hid))]
+    return flatten([einops.repeat(h, pattern, **a) for h, a in zip(hid, kwargs)])
+
+
+def pack(
+    samples: List[torch.Tensor],  # List of (h w c).
+) -> Tuple[
+    List[torch.Tensor],  # groups [(b1 h1 w1 c1), (b2 h2 w2 c2)]
+    List[List[int]],  # reversal indices.
+]:
+    batches = {}
+    indices = {}
+    for i, sample in enumerate(samples):
+        shape = sample.shape
+        batches[shape] = batches.get(shape, [])
+        indices[shape] = indices.get(shape, [])
+        batches[shape].append(sample)
+        indices[shape].append(i)
+
+    batches = list(map(torch.stack, batches.values()))
+    indices = list(indices.values())
+    return batches, indices
+
+
+def unpack(
+    batches: List[torch.Tensor],
+    indices: List[List[int]],
+) -> List[torch.Tensor]:
+    samples = [None] * (max(chain(*indices)) + 1)
+    for batch, index in zip(batches, indices):
+        for sample, i in zip(batch.unbind(), index):
+            samples[i] = sample
+    return samples
+
+
+def window(
+    hid: torch.FloatTensor,  # (L c)
+    hid_shape: torch.LongTensor,  # (b n)
+    window_fn: Callable[[torch.Tensor], List[torch.Tensor]],
+):
+    hid = unflatten(hid, hid_shape)
+    hid = list(map(window_fn, hid))
+    hid_windows = torch.tensor(list(map(len, hid)), device=hid_shape.device)
+    hid, hid_shape = flatten(list(chain(*hid)))
+    return hid, hid_shape, hid_windows
+
+
+def window_idx(
+    hid_shape: torch.LongTensor,  # (b n)
+    window_fn: Callable[[torch.Tensor], List[torch.Tensor]],
+):
+    hid_idx = torch.arange(hid_shape.prod(-1).sum(), device=hid_shape.device).unsqueeze(-1)
+    tgt_idx, tgt_shape, tgt_windows = window(hid_idx, hid_shape, window_fn)
+    tgt_idx = tgt_idx.squeeze(-1)
+    src_idx = torch.argsort(tgt_idx)
+    return (
+        lambda hid: torch.index_select(hid, 0, tgt_idx),
+        lambda hid: torch.index_select(hid, 0, src_idx),
+        tgt_shape,
+        tgt_windows,
+    )

models/dit/nablocks/__init__.py

@@ -0,0 +1,25 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from .mmsr_block import NaMMSRTransformerBlock
+
+nadit_blocks = {
+    "mmdit_sr": NaMMSRTransformerBlock,
+}
+
+
+def get_nablock(block_type: str):
+    if block_type in nadit_blocks:
+        return nadit_blocks[block_type]
+    raise NotImplementedError(f"{block_type} is not supported")

models/dit/nablocks/mmsr_block.py

@@ -0,0 +1,248 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from typing import Tuple, Union
+import torch
+from einops import rearrange
+from torch.nn import functional as F
+
+# from ..cache import Cache
+from common.cache import Cache
+from common.distributed.ops import gather_heads_scatter_seq, gather_seq_scatter_heads_qkv
+
+from .. import na
+from ..attention import FlashAttentionVarlen
+from ..blocks.mmdit_window_block import MMWindowAttention, MMWindowTransformerBlock
+from ..mm import MMArg
+from ..modulation import ada_layer_type
+from ..normalization import norm_layer_type
+from ..rope import NaRotaryEmbedding3d
+from ..window import get_window_op
+
+
+class NaSwinAttention(MMWindowAttention):
+    def __init__(
+        self,
+        vid_dim: int,
+        txt_dim: int,
+        heads: int,
+        head_dim: int,
+        qk_bias: bool,
+        qk_rope: bool,
+        qk_norm: norm_layer_type,
+        qk_norm_eps: float,
+        window: Union[int, Tuple[int, int, int]],
+        window_method: str,
+        shared_qkv: bool,
+        **kwargs,
+    ):
+        super().__init__(
+            vid_dim=vid_dim,
+            txt_dim=txt_dim,
+            heads=heads,
+            head_dim=head_dim,
+            qk_bias=qk_bias,
+            qk_rope=qk_rope,
+            qk_norm=qk_norm,
+            qk_norm_eps=qk_norm_eps,
+            window=window,
+            window_method=window_method,
+            shared_qkv=shared_qkv,
+        )
+        self.rope = NaRotaryEmbedding3d(dim=head_dim // 2) if qk_rope else None
+        self.attn = FlashAttentionVarlen()
+        self.window_op = get_window_op(window_method)
+
+    def forward(
+        self,
+        vid: torch.FloatTensor,  # l c
+        txt: torch.FloatTensor,  # l c
+        vid_shape: torch.LongTensor,  # b 3
+        txt_shape: torch.LongTensor,  # b 1
+        cache: Cache,
+    ) -> Tuple[
+        torch.FloatTensor,
+        torch.FloatTensor,
+    ]:
+
+        vid_qkv, txt_qkv = self.proj_qkv(vid, txt)
+        vid_qkv = gather_seq_scatter_heads_qkv(
+            vid_qkv,
+            seq_dim=0,
+            qkv_shape=vid_shape,
+            cache=cache.namespace("vid"),
+        )
+        txt_qkv = gather_seq_scatter_heads_qkv(
+            txt_qkv,
+            seq_dim=0,
+            qkv_shape=txt_shape,
+            cache=cache.namespace("txt"),
+        )
+
+        # re-org the input seq for window attn
+        cache_win = cache.namespace(f"{self.window_method}_{self.window}_sd3")
+
+        def make_window(x: torch.Tensor):
+            t, h, w, _ = x.shape
+            window_slices = self.window_op((t, h, w), self.window)
+            return [x[st, sh, sw] for (st, sh, sw) in window_slices]
+
+        window_partition, window_reverse, window_shape, window_count = cache_win(
+            "win_transform",
+            lambda: na.window_idx(vid_shape, make_window),
+        )
+        vid_qkv_win = window_partition(vid_qkv)
+
+        vid_qkv_win = rearrange(vid_qkv_win, "l (o h d) -> l o h d", o=3, d=self.head_dim)
+        txt_qkv = rearrange(txt_qkv, "l (o h d) -> l o h d", o=3, d=self.head_dim)
+
+        vid_q, vid_k, vid_v = vid_qkv_win.unbind(1)
+        txt_q, txt_k, txt_v = txt_qkv.unbind(1)
+
+        vid_q, txt_q = self.norm_q(vid_q, txt_q)
+        vid_k, txt_k = self.norm_k(vid_k, txt_k)
+
+        txt_len = cache("txt_len", lambda: txt_shape.prod(-1))
+
+        vid_len_win = cache_win("vid_len", lambda: window_shape.prod(-1))
+        txt_len_win = cache_win("txt_len", lambda: txt_len.repeat_interleave(window_count))
+        all_len_win = cache_win("all_len", lambda: vid_len_win + txt_len_win)
+        concat_win, unconcat_win = cache_win(
+            "mm_pnp", lambda: na.repeat_concat_idx(vid_len_win, txt_len, window_count)
+        )
+
+        # window rope
+        if self.rope:
+            vid_q, vid_k = self.rope(vid_q, vid_k, window_shape, cache_win)
+
+        out = self.attn(
+            q=concat_win(vid_q, txt_q).bfloat16(),
+            k=concat_win(vid_k, txt_k).bfloat16(),
+            v=concat_win(vid_v, txt_v).bfloat16(),
+            cu_seqlens_q=cache_win(
+                "vid_seqlens_q", lambda: F.pad(all_len_win.cumsum(0), (1, 0)).int()
+            ),
+            cu_seqlens_k=cache_win(
+                "vid_seqlens_k", lambda: F.pad(all_len_win.cumsum(0), (1, 0)).int()
+            ),
+            max_seqlen_q=cache_win("vid_max_seqlen_q", lambda: all_len_win.max().item()),
+            max_seqlen_k=cache_win("vid_max_seqlen_k", lambda: all_len_win.max().item()),
+        ).type_as(vid_q)
+
+        # text pooling
+        vid_out, txt_out = unconcat_win(out)
+
+        vid_out = rearrange(vid_out, "l h d -> l (h d)")
+        txt_out = rearrange(txt_out, "l h d -> l (h d)")
+        vid_out = window_reverse(vid_out)
+
+        vid_out = gather_heads_scatter_seq(vid_out, head_dim=1, seq_dim=0)
+        txt_out = gather_heads_scatter_seq(txt_out, head_dim=1, seq_dim=0)
+
+        vid_out, txt_out = self.proj_out(vid_out, txt_out)
+
+        return vid_out, txt_out
+
+
+class NaMMSRTransformerBlock(MMWindowTransformerBlock):
+    def __init__(
+        self,
+        *,
+        vid_dim: int,
+        txt_dim: int,
+        emb_dim: int,
+        heads: int,
+        head_dim: int,
+        expand_ratio: int,
+        norm: norm_layer_type,
+        norm_eps: float,
+        ada: ada_layer_type,
+        qk_bias: bool,
+        qk_rope: bool,
+        qk_norm: norm_layer_type,
+        shared_qkv: bool,
+        shared_mlp: bool,
+        mlp_type: str,
+        **kwargs,
+    ):
+        super().__init__(
+            vid_dim=vid_dim,
+            txt_dim=txt_dim,
+            emb_dim=emb_dim,
+            heads=heads,
+            head_dim=head_dim,
+            expand_ratio=expand_ratio,
+            norm=norm,
+            norm_eps=norm_eps,
+            ada=ada,
+            qk_bias=qk_bias,
+            qk_rope=qk_rope,
+            qk_norm=qk_norm,
+            shared_qkv=shared_qkv,
+            shared_mlp=shared_mlp,
+            mlp_type=mlp_type,
+            **kwargs,
+        )
+
+        self.attn = NaSwinAttention(
+            vid_dim=vid_dim,
+            txt_dim=txt_dim,
+            heads=heads,
+            head_dim=head_dim,
+            qk_bias=qk_bias,
+            qk_rope=qk_rope,
+            qk_norm=qk_norm,
+            qk_norm_eps=norm_eps,
+            shared_qkv=shared_qkv,
+            **kwargs,
+        )
+
+    def forward(
+        self,
+        vid: torch.FloatTensor,  # l c
+        txt: torch.FloatTensor,  # l c
+        vid_shape: torch.LongTensor,  # b 3
+        txt_shape: torch.LongTensor,  # b 1
+        emb: torch.FloatTensor,
+        cache: Cache,
+    ) -> Tuple[
+        torch.FloatTensor,
+        torch.FloatTensor,
+        torch.LongTensor,
+        torch.LongTensor,
+    ]:
+        hid_len = MMArg(
+            cache("vid_len", lambda: vid_shape.prod(-1)),
+            cache("txt_len", lambda: txt_shape.prod(-1)),
+        )
+        ada_kwargs = {
+            "emb": emb,
+            "hid_len": hid_len,
+            "cache": cache,
+            "branch_tag": MMArg("vid", "txt"),
+        }
+
+        vid_attn, txt_attn = self.attn_norm(vid, txt)
+        vid_attn, txt_attn = self.ada(vid_attn, txt_attn, layer="attn", mode="in", **ada_kwargs)
+        vid_attn, txt_attn = self.attn(vid_attn, txt_attn, vid_shape, txt_shape, cache)
+        vid_attn, txt_attn = self.ada(vid_attn, txt_attn, layer="attn", mode="out", **ada_kwargs)
+        vid_attn, txt_attn = (vid_attn + vid), (txt_attn + txt)
+
+        vid_mlp, txt_mlp = self.mlp_norm(vid_attn, txt_attn)
+        vid_mlp, txt_mlp = self.ada(vid_mlp, txt_mlp, layer="mlp", mode="in", **ada_kwargs)
+        vid_mlp, txt_mlp = self.mlp(vid_mlp, txt_mlp)
+        vid_mlp, txt_mlp = self.ada(vid_mlp, txt_mlp, layer="mlp", mode="out", **ada_kwargs)
+        vid_mlp, txt_mlp = (vid_mlp + vid_attn), (txt_mlp + txt_attn)
+
+        return vid_mlp, txt_mlp, vid_shape, txt_shape

models/dit/nadit.py

@@ -0,0 +1,350 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union, Callable
+import torch
+from torch import nn
+
+from common.cache import Cache
+from common.distributed.ops import slice_inputs
+
+from . import na
+from .embedding import TimeEmbedding
+from .modulation import get_ada_layer
+from .nablocks import get_nablock
+from .normalization import get_norm_layer
+from .patch import NaPatchIn, NaPatchOut
+
+# Fake func, no checkpointing is required for inference
+def gradient_checkpointing(module: Union[Callable, nn.Module], *args, enabled: bool, **kwargs):
+    return module(*args, **kwargs)
+
+@dataclass
+class NaDiTOutput:
+    vid_sample: torch.Tensor
+
+
+class NaDiT(nn.Module):
+    """
+    Native Resolution Diffusion Transformer (NaDiT)
+    """
+
+    gradient_checkpointing = False
+
+    def __init__(
+        self,
+        vid_in_channels: int,
+        vid_out_channels: int,
+        vid_dim: int,
+        txt_in_dim: Optional[int],
+        txt_dim: Optional[int],
+        emb_dim: int,
+        heads: int,
+        head_dim: int,
+        expand_ratio: int,
+        norm: Optional[str],
+        norm_eps: float,
+        ada: str,
+        qk_bias: bool,
+        qk_rope: bool,
+        qk_norm: Optional[str],
+        patch_size: Union[int, Tuple[int, int, int]],
+        num_layers: int,
+        block_type: Union[str, Tuple[str]],
+        shared_qkv: bool = False,
+        shared_mlp: bool = False,
+        mlp_type: str = "normal",
+        window: Optional[Tuple] = None,
+        window_method: Optional[Tuple[str]] = None,
+        temporal_window_size: int = None,
+        temporal_shifted: bool = False,
+        **kwargs,
+    ):
+        ada = get_ada_layer(ada)
+        norm = get_norm_layer(norm)
+        qk_norm = get_norm_layer(qk_norm)
+        if isinstance(block_type, str):
+            block_type = [block_type] * num_layers
+        elif len(block_type) != num_layers:
+            raise ValueError("The ``block_type`` list should equal to ``num_layers``.")
+        super().__init__()
+        self.vid_in = NaPatchIn(
+            in_channels=vid_in_channels,
+            patch_size=patch_size,
+            dim=vid_dim,
+        )
+        self.txt_in = (
+            nn.Linear(txt_in_dim, txt_dim)
+            if txt_in_dim and txt_in_dim != txt_dim
+            else nn.Identity()
+        )
+        self.emb_in = TimeEmbedding(
+            sinusoidal_dim=256,
+            hidden_dim=max(vid_dim, txt_dim),
+            output_dim=emb_dim,
+        )
+
+        if window is None or isinstance(window[0], int):
+            window = [window] * num_layers
+        if window_method is None or isinstance(window_method, str):
+            window_method = [window_method] * num_layers
+        if temporal_window_size is None or isinstance(temporal_window_size, int):
+            temporal_window_size = [temporal_window_size] * num_layers
+        if temporal_shifted is None or isinstance(temporal_shifted, bool):
+            temporal_shifted = [temporal_shifted] * num_layers
+
+        self.blocks = nn.ModuleList(
+            [
+                get_nablock(block_type[i])(
+                    vid_dim=vid_dim,
+                    txt_dim=txt_dim,
+                    emb_dim=emb_dim,
+                    heads=heads,
+                    head_dim=head_dim,
+                    expand_ratio=expand_ratio,
+                    norm=norm,
+                    norm_eps=norm_eps,
+                    ada=ada,
+                    qk_bias=qk_bias,
+                    qk_rope=qk_rope,
+                    qk_norm=qk_norm,
+                    shared_qkv=shared_qkv,
+                    shared_mlp=shared_mlp,
+                    mlp_type=mlp_type,
+                    window=window[i],
+                    window_method=window_method[i],
+                    temporal_window_size=temporal_window_size[i],
+                    temporal_shifted=temporal_shifted[i],
+                    **kwargs,
+                )
+                for i in range(num_layers)
+            ]
+        )
+        self.vid_out = NaPatchOut(
+            out_channels=vid_out_channels,
+            patch_size=patch_size,
+            dim=vid_dim,
+        )
+
+        self.need_txt_repeat = block_type[0] in [
+            "mmdit_stwin",
+            "mmdit_stwin_spatial",
+            "mmdit_stwin_3d_spatial",
+        ]
+
+    def set_gradient_checkpointing(self, enable: bool):
+        self.gradient_checkpointing = enable
+
+    def forward(
+        self,
+        vid: torch.FloatTensor,  # l c
+        txt: torch.FloatTensor,  # l c
+        vid_shape: torch.LongTensor,  # b 3
+        txt_shape: torch.LongTensor,  # b 1
+        timestep: Union[int, float, torch.IntTensor, torch.FloatTensor],  # b
+        disable_cache: bool = True,  # for test
+    ):
+        # Text input.
+        if txt_shape.size(-1) == 1 and self.need_txt_repeat:
+            txt, txt_shape = na.repeat(txt, txt_shape, "l c -> t l c", t=vid_shape[:, 0])
+        # slice vid after patching in when using sequence parallelism
+        txt = slice_inputs(txt, dim=0)
+        txt = self.txt_in(txt)
+
+        # Video input.
+        # Sequence parallel slicing is done inside patching class.
+        vid, vid_shape = self.vid_in(vid, vid_shape)
+
+        # Embedding input.
+        emb = self.emb_in(timestep, device=vid.device, dtype=vid.dtype)
+
+        # Body
+        cache = Cache(disable=disable_cache)
+        for i, block in enumerate(self.blocks):
+            vid, txt, vid_shape, txt_shape = gradient_checkpointing(
+                enabled=(self.gradient_checkpointing and self.training),
+                module=block,
+                vid=vid,
+                txt=txt,
+                vid_shape=vid_shape,
+                txt_shape=txt_shape,
+                emb=emb,
+                cache=cache,
+            )
+
+        vid, vid_shape = self.vid_out(vid, vid_shape, cache)
+        return NaDiTOutput(vid_sample=vid)
+
+
+class NaDiTUpscaler(nn.Module):
+    """
+    Native Resolution Diffusion Transformer (NaDiT)
+    """
+
+    gradient_checkpointing = False
+
+    def __init__(
+        self,
+        vid_in_channels: int,
+        vid_out_channels: int,
+        vid_dim: int,
+        txt_in_dim: Optional[int],
+        txt_dim: Optional[int],
+        emb_dim: int,
+        heads: int,
+        head_dim: int,
+        expand_ratio: int,
+        norm: Optional[str],
+        norm_eps: float,
+        ada: str,
+        qk_bias: bool,
+        qk_rope: bool,
+        qk_norm: Optional[str],
+        patch_size: Union[int, Tuple[int, int, int]],
+        num_layers: int,
+        block_type: Union[str, Tuple[str]],
+        shared_qkv: bool = False,
+        shared_mlp: bool = False,
+        mlp_type: str = "normal",
+        window: Optional[Tuple] = None,
+        window_method: Optional[Tuple[str]] = None,
+        temporal_window_size: int = None,
+        temporal_shifted: bool = False,
+        **kwargs,
+    ):
+        ada = get_ada_layer(ada)
+        norm = get_norm_layer(norm)
+        qk_norm = get_norm_layer(qk_norm)
+        if isinstance(block_type, str):
+            block_type = [block_type] * num_layers
+        elif len(block_type) != num_layers:
+            raise ValueError("The ``block_type`` list should equal to ``num_layers``.")
+        super().__init__()
+        self.vid_in = NaPatchIn(
+            in_channels=vid_in_channels,
+            patch_size=patch_size,
+            dim=vid_dim,
+        )
+        self.txt_in = (
+            nn.Linear(txt_in_dim, txt_dim)
+            if txt_in_dim and txt_in_dim != txt_dim
+            else nn.Identity()
+        )
+        self.emb_in = TimeEmbedding(
+            sinusoidal_dim=256,
+            hidden_dim=max(vid_dim, txt_dim),
+            output_dim=emb_dim,
+        )
+
+        self.emb_scale = TimeEmbedding(
+            sinusoidal_dim=256,
+            hidden_dim=max(vid_dim, txt_dim),
+            output_dim=emb_dim,
+        )
+
+        if window is None or isinstance(window[0], int):
+            window = [window] * num_layers
+        if window_method is None or isinstance(window_method, str):
+            window_method = [window_method] * num_layers
+        if temporal_window_size is None or isinstance(temporal_window_size, int):
+            temporal_window_size = [temporal_window_size] * num_layers
+        if temporal_shifted is None or isinstance(temporal_shifted, bool):
+            temporal_shifted = [temporal_shifted] * num_layers
+
+        self.blocks = nn.ModuleList(
+            [
+                get_nablock(block_type[i])(
+                    vid_dim=vid_dim,
+                    txt_dim=txt_dim,
+                    emb_dim=emb_dim,
+                    heads=heads,
+                    head_dim=head_dim,
+                    expand_ratio=expand_ratio,
+                    norm=norm,
+                    norm_eps=norm_eps,
+                    ada=ada,
+                    qk_bias=qk_bias,
+                    qk_rope=qk_rope,
+                    qk_norm=qk_norm,
+                    shared_qkv=shared_qkv,
+                    shared_mlp=shared_mlp,
+                    mlp_type=mlp_type,
+                    window=window[i],
+                    window_method=window_method[i],
+                    temporal_window_size=temporal_window_size[i],
+                    temporal_shifted=temporal_shifted[i],
+                    **kwargs,
+                )
+                for i in range(num_layers)
+            ]
+        )
+        self.vid_out = NaPatchOut(
+            out_channels=vid_out_channels,
+            patch_size=patch_size,
+            dim=vid_dim,
+        )
+
+        self.need_txt_repeat = block_type[0] in [
+            "mmdit_stwin",
+            "mmdit_stwin_spatial",
+            "mmdit_stwin_3d_spatial",
+        ]
+
+    def set_gradient_checkpointing(self, enable: bool):
+        self.gradient_checkpointing = enable
+
+    def forward(
+        self,
+        vid: torch.FloatTensor,  # l c
+        txt: torch.FloatTensor,  # l c
+        vid_shape: torch.LongTensor,  # b 3
+        txt_shape: torch.LongTensor,  # b 1
+        timestep: Union[int, float, torch.IntTensor, torch.FloatTensor],  # b
+        downscale: Union[int, float, torch.IntTensor, torch.FloatTensor],  # b
+        disable_cache: bool = False,  # for test
+    ):
+
+        # Text input.
+        if txt_shape.size(-1) == 1 and self.need_txt_repeat:
+            txt, txt_shape = na.repeat(txt, txt_shape, "l c -> t l c", t=vid_shape[:, 0])
+        # slice vid after patching in when using sequence parallelism
+        txt = slice_inputs(txt, dim=0)
+        txt = self.txt_in(txt)
+
+        # Video input.
+        # Sequence parallel slicing is done inside patching class.
+        vid, vid_shape = self.vid_in(vid, vid_shape)
+
+        # Embedding input.
+        emb = self.emb_in(timestep, device=vid.device, dtype=vid.dtype)
+        emb_scale = self.emb_scale(downscale, device=vid.device, dtype=vid.dtype)
+        emb = emb + emb_scale
+
+        # Body
+        cache = Cache(disable=disable_cache)
+        for i, block in enumerate(self.blocks):
+            vid, txt, vid_shape, txt_shape = gradient_checkpointing(
+                enabled=(self.gradient_checkpointing and self.training),
+                module=block,
+                vid=vid,
+                txt=txt,
+                vid_shape=vid_shape,
+                txt_shape=txt_shape,
+                emb=emb,
+                cache=cache,
+            )
+
+        vid, vid_shape = self.vid_out(vid, vid_shape, cache)
+        return NaDiTOutput(vid_sample=vid)

models/dit/normalization.py

@@ -0,0 +1,63 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from typing import Callable, Optional
+from diffusers.models.normalization import RMSNorm
+from torch import nn
+
+# (dim: int, eps: float, elementwise_affine: bool)
+norm_layer_type = Callable[[int, float, bool], nn.Module]
+
+
+def get_norm_layer(norm_type: Optional[str]) -> norm_layer_type:
+
+    def _norm_layer(dim: int, eps: float, elementwise_affine: bool):
+        if norm_type is None:
+            return nn.Identity()
+
+        if norm_type == "layer":
+            return nn.LayerNorm(
+                normalized_shape=dim,
+                eps=eps,
+                elementwise_affine=elementwise_affine,
+            )
+
+        if norm_type == "rms":
+            return RMSNorm(
+                dim=dim,
+                eps=eps,
+                elementwise_affine=elementwise_affine,
+            )
+
+        if norm_type == "fusedln":
+            from apex.normalization import FusedLayerNorm
+
+            return FusedLayerNorm(
+                normalized_shape=dim,
+                elementwise_affine=elementwise_affine,
+                eps=eps,
+            )
+
+        if norm_type == "fusedrms":
+            from apex.normalization import FusedRMSNorm
+
+            return FusedRMSNorm(
+                normalized_shape=dim,
+                elementwise_affine=elementwise_affine,
+                eps=eps,
+            )
+
+        raise NotImplementedError(f"{norm_type} is not supported")
+
+    return _norm_layer

models/dit/patch.py

@@ -0,0 +1,112 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from typing import Tuple, Union
+import torch
+from einops import rearrange
+from torch import nn
+from torch.nn.modules.utils import _triple
+
+from common.cache import Cache
+from common.distributed.ops import gather_outputs, slice_inputs
+
+from . import na
+
+
+class PatchIn(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        patch_size: Union[int, Tuple[int, int, int]],
+        dim: int,
+    ):
+        super().__init__()
+        t, h, w = _triple(patch_size)
+        self.patch_size = t, h, w
+        self.proj = nn.Linear(in_channels * t * h * w, dim)
+
+    def forward(
+        self,
+        vid: torch.Tensor,
+    ) -> torch.Tensor:
+        t, h, w = self.patch_size
+        vid = rearrange(vid, "b c (T t) (H h) (W w) -> b T H W (t h w c)", t=t, h=h, w=w)
+        vid = self.proj(vid)
+        return vid
+
+
+class PatchOut(nn.Module):
+    def __init__(
+        self,
+        out_channels: int,
+        patch_size: Union[int, Tuple[int, int, int]],
+        dim: int,
+    ):
+        super().__init__()
+        t, h, w = _triple(patch_size)
+        self.patch_size = t, h, w
+        self.proj = nn.Linear(dim, out_channels * t * h * w)
+
+    def forward(
+        self,
+        vid: torch.Tensor,
+    ) -> torch.Tensor:
+        t, h, w = self.patch_size
+        vid = self.proj(vid)
+        vid = rearrange(vid, "b T H W (t h w c) -> b c (T t) (H h) (W w)", t=t, h=h, w=w)
+        return vid
+
+
+class NaPatchIn(PatchIn):
+    def forward(
+        self,
+        vid: torch.Tensor,  # l c
+        vid_shape: torch.LongTensor,
+    ) -> torch.Tensor:
+        t, h, w = self.patch_size
+        if not (t == h == w == 1):
+            vid, vid_shape = na.rearrange(
+                vid, vid_shape, "(T t) (H h) (W w) c -> T H W (t h w c)", t=t, h=h, w=w
+            )
+        # slice vid after patching in when using sequence parallelism
+        vid = slice_inputs(vid, dim=0)
+        vid = self.proj(vid)
+        return vid, vid_shape
+
+
+class NaPatchOut(PatchOut):
+    def forward(
+        self,
+        vid: torch.FloatTensor,  # l c
+        vid_shape: torch.LongTensor,
+        cache: Cache = Cache(disable=True),
+    ) -> Tuple[
+        torch.FloatTensor,
+        torch.LongTensor,
+    ]:
+        t, h, w = self.patch_size
+        vid = self.proj(vid)
+        # gather vid before patching out when enabling sequence parallelism
+        vid = gather_outputs(
+            vid,
+            gather_dim=0,
+            padding_dim=0,
+            unpad_shape=vid_shape,
+            cache=cache.namespace("vid"),
+        )
+        if not (t == h == w == 1):
+            vid, vid_shape = na.rearrange(
+                vid, vid_shape, "T H W (t h w c) -> (T t) (H h) (W w) c", t=t, h=h, w=w
+            )
+        return vid, vid_shape

models/dit/rope.py

@@ -0,0 +1,101 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from functools import lru_cache
+from typing import Tuple
+import torch
+from einops import rearrange
+from rotary_embedding_torch import RotaryEmbedding, apply_rotary_emb
+from torch import nn
+
+from common.cache import Cache
+
+
+class RotaryEmbeddingBase(nn.Module):
+    def __init__(self, dim: int, rope_dim: int):
+        super().__init__()
+        self.rope = RotaryEmbedding(
+            dim=dim // rope_dim,
+            freqs_for="pixel",
+            max_freq=256,
+        )
+        # 1. Set model.requires_grad_(True) after model creation will make
+        #    the `requires_grad=False` for rope freqs no longer hold.
+        # 2. Even if we don't set requires_grad_(True) explicitly,
+        #    FSDP is not memory efficient when handling fsdp_wrap
+        #    with mixed requires_grad=True/False.
+        # With above consideration, it is easier just remove the freqs
+        # out of nn.Parameters when `learned_freq=False`
+        freqs = self.rope.freqs
+        del self.rope.freqs
+        self.rope.register_buffer("freqs", freqs.data)
+
+    @lru_cache(maxsize=128)
+    def get_axial_freqs(self, *dims):
+        return self.rope.get_axial_freqs(*dims)
+
+
+class RotaryEmbedding3d(RotaryEmbeddingBase):
+    def __init__(self, dim: int):
+        super().__init__(dim, rope_dim=3)
+
+    def forward(
+        self,
+        q: torch.FloatTensor,  # b h l d
+        k: torch.FloatTensor,  # b h l d
+        size: Tuple[int, int, int],
+    ) -> Tuple[
+        torch.FloatTensor,
+        torch.FloatTensor,
+    ]:
+        T, H, W = size
+        freqs = self.get_axial_freqs(T, H, W)
+        q = rearrange(q, "b h (T H W) d -> b h T H W d", T=T, H=H, W=W)
+        k = rearrange(k, "b h (T H W) d -> b h T H W d", T=T, H=H, W=W)
+        q = apply_rotary_emb(freqs, q)
+        k = apply_rotary_emb(freqs, k)
+        q = rearrange(q, "b h T H W d -> b h (T H W) d")
+        k = rearrange(k, "b h T H W d -> b h (T H W) d")
+        return q, k
+
+
+class NaRotaryEmbedding3d(RotaryEmbedding3d):
+    def forward(
+        self,
+        q: torch.FloatTensor,  # L h d
+        k: torch.FloatTensor,  # L h d
+        shape: torch.LongTensor,
+        cache: Cache,
+    ) -> Tuple[
+        torch.FloatTensor,
+        torch.FloatTensor,
+    ]:
+        freqs = cache("rope_freqs_3d", lambda: self.get_freqs(shape))
+        q = rearrange(q, "L h d -> h L d")
+        k = rearrange(k, "L h d -> h L d")
+        q = apply_rotary_emb(freqs, q.float()).to(q.dtype)
+        k = apply_rotary_emb(freqs, k.float()).to(k.dtype)
+        q = rearrange(q, "h L d -> L h d")
+        k = rearrange(k, "h L d -> L h d")
+        return q, k
+
+    def get_freqs(
+        self,
+        shape: torch.LongTensor,
+    ) -> torch.Tensor:
+        freq_list = []
+        for f, h, w in shape.tolist():
+            freqs = self.get_axial_freqs(f, h, w)
+            freq_list.append(freqs.view(-1, freqs.size(-1)))
+        return torch.cat(freq_list, dim=0)

models/dit/window.py

@@ -0,0 +1,83 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from math import ceil
+from typing import Tuple
+import math
+
+def get_window_op(name: str):
+    if name == "720pwin_by_size_bysize":
+        return make_720Pwindows_bysize
+    if name == "720pswin_by_size_bysize":
+        return make_shifted_720Pwindows_bysize
+    raise ValueError(f"Unknown windowing method: {name}")
+
+
+# -------------------------------- Windowing -------------------------------- #
+def make_720Pwindows_bysize(size: Tuple[int, int, int], num_windows: Tuple[int, int, int]):
+    t, h, w = size
+    resized_nt, resized_nh, resized_nw = num_windows
+    #cal windows under 720p
+    scale = math.sqrt((45 * 80) / (h * w))
+    resized_h, resized_w = round(h * scale), round(w * scale)
+    wh, ww = ceil(resized_h / resized_nh), ceil(resized_w / resized_nw)  # window size.
+    wt = ceil(min(t, 30) / resized_nt)  # window size.
+    nt, nh, nw = ceil(t / wt), ceil(h / wh), ceil(w / ww)  # window size.
+    return [
+        (
+            slice(it * wt, min((it + 1) * wt, t)),
+            slice(ih * wh, min((ih + 1) * wh, h)),
+            slice(iw * ww, min((iw + 1) * ww, w)),
+        )
+        for iw in range(nw)
+        if min((iw + 1) * ww, w) > iw * ww
+        for ih in range(nh)
+        if min((ih + 1) * wh, h) > ih * wh
+        for it in range(nt)
+        if min((it + 1) * wt, t) > it * wt
+    ]
+
+def make_shifted_720Pwindows_bysize(size: Tuple[int, int, int], num_windows: Tuple[int, int, int]):
+    t, h, w = size
+    resized_nt, resized_nh, resized_nw = num_windows
+    #cal windows under 720p
+    scale = math.sqrt((45 * 80) / (h * w))
+    resized_h, resized_w = round(h * scale), round(w * scale)
+    wh, ww = ceil(resized_h / resized_nh), ceil(resized_w / resized_nw)  # window size.
+    wt = ceil(min(t, 30) / resized_nt)  # window size.
+    
+    st, sh, sw = (  # shift size.
+        0.5 if wt < t else 0,
+        0.5 if wh < h else 0,
+        0.5 if ww < w else 0,
+    )
+    nt, nh, nw = ceil((t - st) / wt), ceil((h - sh) / wh), ceil((w - sw) / ww)  # window size.
+    nt, nh, nw = (  # number of window.
+        nt + 1 if st > 0 else 1,
+        nh + 1 if sh > 0 else 1,
+        nw + 1 if sw > 0 else 1,
+    )
+    return [
+        (
+            slice(max(int((it - st) * wt), 0), min(int((it - st + 1) * wt), t)),
+            slice(max(int((ih - sh) * wh), 0), min(int((ih - sh + 1) * wh), h)),
+            slice(max(int((iw - sw) * ww), 0), min(int((iw - sw + 1) * ww), w)),
+        )
+        for iw in range(nw)
+        if min(int((iw - sw + 1) * ww), w) > max(int((iw - sw) * ww), 0)
+        for ih in range(nh)
+        if min(int((ih - sh + 1) * wh), h) > max(int((ih - sh) * wh), 0)
+        for it in range(nt)
+        if min(int((it - st + 1) * wt), t) > max(int((it - st) * wt), 0)
+    ]

models/dit_v2/attention.py

@@ -0,0 +1,46 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+import torch
+import torch.nn.functional as F
+
+from flash_attn import flash_attn_varlen_func
+
+from torch import nn
+
+class TorchAttention(nn.Module):
+    def tflops(self, args, kwargs, output) -> float:
+        assert len(args) == 0 or len(args) > 2, "query, key should both provided by args / kwargs"
+        q = kwargs.get("query") or args[0]
+        k = kwargs.get("key") or args[1]
+        b, h, sq, d = q.shape
+        b, h, sk, d = k.shape
+        return b * h * (4 * d * (sq / 1e6) * (sk / 1e6))
+
+    def forward(self, *args, **kwargs):
+        return F.scaled_dot_product_attention(*args, **kwargs)
+
+
+class FlashAttentionVarlen(nn.Module):
+    def tflops(self, args, kwargs, output) -> float:
+        cu_seqlens_q = kwargs["cu_seqlens_q"]
+        cu_seqlens_k = kwargs["cu_seqlens_k"]
+        _, h, d = output.shape
+        seqlens_q = (cu_seqlens_q[1:] - cu_seqlens_q[:-1]) / 1e6
+        seqlens_k = (cu_seqlens_k[1:] - cu_seqlens_k[:-1]) / 1e6
+        return h * (4 * d * (seqlens_q * seqlens_k).sum())
+
+    def forward(self, *args, **kwargs):
+        kwargs["deterministic"] = torch.are_deterministic_algorithms_enabled()
+        return flash_attn_varlen_func(*args, **kwargs)

models/dit_v2/embedding.py

@@ -0,0 +1,62 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from typing import Optional, Union
+import torch
+from diffusers.models.embeddings import get_timestep_embedding
+from torch import nn
+
+
+def emb_add(emb1: torch.Tensor, emb2: Optional[torch.Tensor]):
+    return emb1 if emb2 is None else emb1 + emb2
+
+
+class TimeEmbedding(nn.Module):
+    def __init__(
+        self,
+        sinusoidal_dim: int,
+        hidden_dim: int,
+        output_dim: int,
+    ):
+        super().__init__()
+        self.sinusoidal_dim = sinusoidal_dim
+        self.proj_in = nn.Linear(sinusoidal_dim, hidden_dim)
+        self.proj_hid = nn.Linear(hidden_dim, hidden_dim)
+        self.proj_out = nn.Linear(hidden_dim, output_dim)
+        self.act = nn.SiLU()
+
+    def forward(
+        self,
+        timestep: Union[int, float, torch.IntTensor, torch.FloatTensor],
+        device: torch.device,
+        dtype: torch.dtype,
+    ) -> torch.FloatTensor:
+        if not torch.is_tensor(timestep):
+            timestep = torch.tensor([timestep], device=device, dtype=dtype)
+        if timestep.ndim == 0:
+            timestep = timestep[None]
+
+        emb = get_timestep_embedding(
+            timesteps=timestep,
+            embedding_dim=self.sinusoidal_dim,
+            flip_sin_to_cos=False,
+            downscale_freq_shift=0,
+        )
+        emb = emb.to(dtype)
+        emb = self.proj_in(emb)
+        emb = self.act(emb)
+        emb = self.proj_hid(emb)
+        emb = self.act(emb)
+        emb = self.proj_out(emb)
+        return emb

models/dit_v2/mlp.py

@@ -0,0 +1,62 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from typing import Optional
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+def get_mlp(mlp_type: Optional[str] = "normal"):
+    if mlp_type == "normal":
+        return MLP
+    elif mlp_type == "swiglu":
+        return SwiGLUMLP
+
+
+class MLP(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        expand_ratio: int,
+    ):
+        super().__init__()
+        self.proj_in = nn.Linear(dim, dim * expand_ratio)
+        self.act = nn.GELU("tanh")
+        self.proj_out = nn.Linear(dim * expand_ratio, dim)
+
+    def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+        x = self.proj_in(x)
+        x = self.act(x)
+        x = self.proj_out(x)
+        return x
+
+
+class SwiGLUMLP(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        expand_ratio: int,
+        multiple_of: int = 256,
+    ):
+        super().__init__()
+        hidden_dim = int(2 * dim * expand_ratio / 3)
+        hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+        self.proj_in_gate = nn.Linear(dim, hidden_dim, bias=False)
+        self.proj_out = nn.Linear(hidden_dim, dim, bias=False)
+        self.proj_in = nn.Linear(dim, hidden_dim, bias=False)
+
+    def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+        x = self.proj_out(F.silu(self.proj_in_gate(x)) * self.proj_in(x))
+        return x

models/dit_v2/mm.py

@@ -0,0 +1,74 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Tuple
+import torch
+from torch import nn
+
+
+@dataclass
+class MMArg:
+    vid: Any
+    txt: Any
+
+
+def get_args(key: str, args: List[Any]) -> List[Any]:
+    return [getattr(v, key) if isinstance(v, MMArg) else v for v in args]
+
+
+def get_kwargs(key: str, kwargs: Dict[str, Any]) -> Dict[str, Any]:
+    return {k: getattr(v, key) if isinstance(v, MMArg) else v for k, v in kwargs.items()}
+
+
+class MMModule(nn.Module):
+    def __init__(
+        self,
+        module: Callable[..., nn.Module],
+        *args,
+        shared_weights: bool = False,
+        vid_only: bool = False,
+        **kwargs,
+    ):
+        super().__init__()
+        self.shared_weights = shared_weights
+        self.vid_only = vid_only
+        if self.shared_weights:
+            assert get_args("vid", args) == get_args("txt", args)
+            assert get_kwargs("vid", kwargs) == get_kwargs("txt", kwargs)
+            self.all = module(*get_args("vid", args), **get_kwargs("vid", kwargs))
+        else:
+            self.vid = module(*get_args("vid", args), **get_kwargs("vid", kwargs))
+            self.txt = (
+                module(*get_args("txt", args), **get_kwargs("txt", kwargs))
+                if not vid_only
+                else None
+            )
+
+    def forward(
+        self,
+        vid: torch.FloatTensor,
+        txt: torch.FloatTensor,
+        *args,
+        **kwargs,
+    ) -> Tuple[
+        torch.FloatTensor,
+        torch.FloatTensor,
+    ]:
+        vid_module = self.vid if not self.shared_weights else self.all
+        vid = vid_module(vid, *get_args("vid", args), **get_kwargs("vid", kwargs))
+        if not self.vid_only:
+            txt_module = self.txt if not self.shared_weights else self.all
+            txt = txt_module(txt, *get_args("txt", args), **get_kwargs("txt", kwargs))
+        return vid, txt

models/dit_v2/modulation.py

@@ -0,0 +1,102 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from typing import Callable, List, Optional
+import torch
+from einops import rearrange
+from torch import nn
+
+from common.cache import Cache
+from common.distributed.ops import slice_inputs
+
+# (dim: int, emb_dim: int)
+ada_layer_type = Callable[[int, int], nn.Module]
+
+
+def get_ada_layer(ada_layer: str) -> ada_layer_type:
+    if ada_layer == "single":
+        return AdaSingle
+    raise NotImplementedError(f"{ada_layer} is not supported")
+
+
+def expand_dims(x: torch.Tensor, dim: int, ndim: int):
+    """
+    Expand tensor "x" to "ndim" by adding empty dims at "dim".
+    Example: x is (b d), target ndim is 5, add dim at 1, return (b 1 1 1 d).
+    """
+    shape = x.shape
+    shape = shape[:dim] + (1,) * (ndim - len(shape)) + shape[dim:]
+    return x.reshape(shape)
+
+
+class AdaSingle(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        emb_dim: int,
+        layers: List[str],
+        modes: List[str] = ["in", "out"],
+    ):
+        assert emb_dim == 6 * dim, "AdaSingle requires emb_dim == 6 * dim"
+        super().__init__()
+        self.dim = dim
+        self.emb_dim = emb_dim
+        self.layers = layers
+        for l in layers:
+            if "in" in modes:
+                self.register_parameter(f"{l}_shift", nn.Parameter(torch.randn(dim) / dim**0.5))
+                self.register_parameter(
+                    f"{l}_scale", nn.Parameter(torch.randn(dim) / dim**0.5 + 1)
+                )
+            if "out" in modes:
+                self.register_parameter(f"{l}_gate", nn.Parameter(torch.randn(dim) / dim**0.5))
+
+    def forward(
+        self,
+        hid: torch.FloatTensor,  # b ... c
+        emb: torch.FloatTensor,  # b d
+        layer: str,
+        mode: str,
+        cache: Cache = Cache(disable=True),
+        branch_tag: str = "",
+        hid_len: Optional[torch.LongTensor] = None,  # b
+    ) -> torch.FloatTensor:
+        idx = self.layers.index(layer)
+        emb = rearrange(emb, "b (d l g) -> b d l g", l=len(self.layers), g=3)[..., idx, :]
+        emb = expand_dims(emb, 1, hid.ndim + 1)
+
+        if hid_len is not None:
+            emb = cache(
+                f"emb_repeat_{idx}_{branch_tag}",
+                lambda: slice_inputs(
+                    torch.cat([e.repeat(l, *([1] * e.ndim)) for e, l in zip(emb, hid_len)]),
+                    dim=0,
+                ),
+            )
+
+        shiftA, scaleA, gateA = emb.unbind(-1)
+        shiftB, scaleB, gateB = (
+            getattr(self, f"{layer}_shift", None),
+            getattr(self, f"{layer}_scale", None),
+            getattr(self, f"{layer}_gate", None),
+        )
+
+        if mode == "in":
+            return hid.mul_(scaleA + scaleB).add_(shiftA + shiftB)
+        if mode == "out":
+            return hid.mul_(gateA + gateB)
+        raise NotImplementedError
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, emb_dim={self.emb_dim}, layers={self.layers}"

models/dit_v2/na.py

@@ -0,0 +1,241 @@
+# // Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
+# //
+# // 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
+# // limitations under the License.
+
+from itertools import chain
+from typing import Callable, Dict, List, Tuple
+import einops
+import torch
+
+
+def flatten(
+    hid: List[torch.FloatTensor],  # List of (*** c)
+) -> Tuple[
+    torch.FloatTensor,  # (L c)
+    torch.LongTensor,  # (b n)
+]:
+    assert len(hid) > 0
+    shape = torch.stack([torch.tensor(x.shape[:-1], device=hid[0].device) for x in hid])
+    hid = torch.cat([x.flatten(0, -2) for x in hid])
+    return hid, shape
+
+
+def unflatten(
+    hid: torch.FloatTensor,  # (L c) or (L ... c)
+    hid_shape: torch.LongTensor,  # (b n)
+) -> List[torch.Tensor]:  # List of (*** c) or (*** ... c)
+    hid_len = hid_shape.prod(-1)
+    hid = hid.split(hid_len.tolist())
+    hid = [x.unflatten(0, s.tolist()) for x, s in zip(hid, hid_shape)]
+    return hid
+
+
+def concat(
+    vid: torch.FloatTensor,  # (VL ... c)
+    txt: torch.FloatTensor,  # (TL ... c)
+    vid_len: torch.LongTensor,  # (b)
+    txt_len: torch.LongTensor,  # (b)
+) -> torch.FloatTensor:  # (L ... c)
+    vid = torch.split(vid, vid_len.tolist())
+    txt = torch.split(txt, txt_len.tolist())
+    return torch.cat(list(chain(*zip(vid, txt))))
+
+
+def concat_idx(
+    vid_len: torch.LongTensor,  # (b)
+    txt_len: torch.LongTensor,  # (b)
+) -> Tuple[
+    Callable,
+    Callable,
+]:
+    device = vid_len.device
+    vid_idx = torch.arange(vid_len.sum(), device=device)
+    txt_idx = torch.arange(len(vid_idx), len(vid_idx) + txt_len.sum(), device=device)
+    tgt_idx = concat(vid_idx, txt_idx, vid_len, txt_len)
+    src_idx = torch.argsort(tgt_idx)
+    return (
+        lambda vid, txt: torch.index_select(torch.cat([vid, txt]), 0, tgt_idx),
+        lambda all: torch.index_select(all, 0, src_idx).split([len(vid_idx), len(txt_idx)]),
+    )
+
+
+def unconcat(
+    all: torch.FloatTensor,  # (L ... c)
+    vid_len: torch.LongTensor,  # (b)
+    txt_len: torch.LongTensor,  # (b)
+) -> Tuple[
+    torch.FloatTensor,  # (VL ... c)
+    torch.FloatTensor,  # (TL ... c)
+]:
+    interleave_len = list(chain(*zip(vid_len.tolist(), txt_len.tolist())))
+    all = all.split(interleave_len)
+    vid = torch.cat(all[0::2])
+    txt = torch.cat(all[1::2])
+    return vid, txt
+
+
+def repeat_concat(
+    vid: torch.FloatTensor,  # (VL ... c)
+    txt: torch.FloatTensor,  # (TL ... c)
+    vid_len: torch.LongTensor,  # (n*b)
+    txt_len: torch.LongTensor,  # (b)
+    txt_repeat: List,  # (n)
+) -> torch.FloatTensor:  # (L ... c)
+    vid = torch.split(vid, vid_len.tolist())
+    txt = torch.split(txt, txt_len.tolist())
+    txt = [[x] * n for x, n in zip(txt, txt_repeat)]
+    txt = list(chain(*txt))
+    return torch.cat(list(chain(*zip(vid, txt))))
+
+
+def repeat_concat_idx(
+    vid_len: torch.LongTensor,  # (n*b)
+    txt_len: torch.LongTensor,  # (b)
+    txt_repeat: torch.LongTensor,  # (n)
+) -> Tuple[
+    Callable,
+    Callable,
+]:
+    device = vid_len.device
+    vid_idx = torch.arange(vid_len.sum(), device=device)
+    txt_idx = torch.arange(len(vid_idx), len(vid_idx) + txt_len.sum(), device=device)
+    txt_repeat_list = txt_repeat.tolist()
+    tgt_idx = repeat_concat(vid_idx, txt_idx, vid_len, txt_len, txt_repeat)
+    src_idx = torch.argsort(tgt_idx)
+    txt_idx_len = len(tgt_idx) - len(vid_idx)
+    repeat_txt_len = (txt_len * txt_repeat).tolist()
+
+    def unconcat_coalesce(all):
+        """
+        Un-concat vid & txt, and coalesce the repeated txt.
+        e.g. vid [0 1 2 3 4 5 6 7 8] -> 3 splits -> [0 1 2] [3 4 5] [6 7 8]
+             txt [9 10]
+             repeat_concat ==> [0 1 2 9 10 3 4 5 9 10 6 7 8 9 10]
+             1. argsort re-index ==> [0 1 2 3 4 5 6 7 8 9 9 9 10 10 10]
+                           split ==> vid_out [0 1 2 3 4 5 6 7 8] txt_out [9 9 9 10 10 10]
+             2. reshape & mean for each sample to coalesce the repeated txt.
+        """
+        vid_out, txt_out = all[src_idx].split([len(vid_idx), txt_idx_len])
+        txt_out_coalesced = []
+        for txt, repeat_time in zip(txt_out.split(repeat_txt_len), txt_repeat_list):
+            txt = txt.reshape(-1, repeat_time, *txt.shape[1:]).mean(1)
+            txt_out_coalesced.append(txt)
+        return vid_out, torch.cat(txt_out_coalesced)
+
+    # Note: Backward of torch.index_select is non-deterministic when existing repeated index,
+    # the difference may cumulative like torch.repeat_interleave, so we use vanilla index here.
+    return (
+        lambda vid, txt: torch.cat([vid, txt])[tgt_idx],
+        lambda all: unconcat_coalesce(all),
+    )
+
+
+def rearrange(
+    hid: torch.FloatTensor,  # (L c)
+    hid_shape: torch.LongTensor,  # (b n)
+    pattern: str,
+    **kwargs: Dict[str, int],
+) -> Tuple[
+    torch.FloatTensor,
+    torch.LongTensor,
+]:
+    return flatten([einops.rearrange(h, pattern, **kwargs) for h in unflatten(hid, hid_shape)])
+
+
+def rearrange_idx(
+    hid_shape: torch.LongTensor,  # (b n)
+    pattern: str,
+    **kwargs: Dict[str, int],
+) -> Tuple[Callable, Callable, torch.LongTensor]:
+    hid_idx = torch.arange(hid_shape.prod(-1).sum(), device=hid_shape.device).unsqueeze(-1)
+    tgt_idx, tgt_shape = rearrange(hid_idx, hid_shape, pattern, **kwargs)
+    tgt_idx = tgt_idx.squeeze(-1)
+    src_idx = torch.argsort(tgt_idx)
+    return (
+        lambda hid: torch.index_select(hid, 0, tgt_idx),
+        lambda hid: torch.index_select(hid, 0, src_idx),
+        tgt_shape,
+    )
+
+
+def repeat(
+    hid: torch.FloatTensor,  # (L c)
+    hid_shape: torch.LongTensor,  # (b n)
+    pattern: str,
+    **kwargs: Dict[str, torch.LongTensor],  # (b)
+) -> Tuple[
+    torch.FloatTensor,
+    torch.LongTensor,
+]:
+    hid = unflatten(hid, hid_shape)
+    kwargs = [{k: v[i].item() for k, v in kwargs.items()} for i in range(len(hid))]
+    return flatten([einops.repeat(h, pattern, **a) for h, a in zip(hid, kwargs)])
+
+
+def pack(
+    samples: List[torch.Tensor],  # List of (h w c).
+) -> Tuple[
+    List[torch.Tensor],  # groups [(b1 h1 w1 c1), (b2 h2 w2 c2)]
+    List[List[int]],  # reversal indices.
+]:
+    batches = {}
+    indices = {}
+    for i, sample in enumerate(samples):
+        shape = sample.shape
+        batches[shape] = batches.get(shape, [])
+        indices[shape] = indices.get(shape, [])
+        batches[shape].append(sample)
+        indices[shape].append(i)
+
+    batches = list(map(torch.stack, batches.values()))
+    indices = list(indices.values())
+    return batches, indices
+
+
+def unpack(
+    batches: List[torch.Tensor],
+    indices: List[List[int]],
+) -> List[torch.Tensor]:
+    samples = [None] * (max(chain(*indices)) + 1)
+    for batch, index in zip(batches, indices):
+        for sample, i in zip(batch.unbind(), index):
+            samples[i] = sample
+    return samples
+
+
+def window(
+    hid: torch.FloatTensor,  # (L c)
+    hid_shape: torch.LongTensor,  # (b n)
+    window_fn: Callable[[torch.Tensor], List[torch.Tensor]],
+):
+    hid = unflatten(hid, hid_shape)
+    hid = list(map(window_fn, hid))
+    hid_windows = torch.tensor(list(map(len, hid)), device=hid_shape.device)
+    hid, hid_shape = flatten(list(chain(*hid)))
+    return hid, hid_shape, hid_windows
+
+
+def window_idx(
+    hid_shape: torch.LongTensor,  # (b n)
+    window_fn: Callable[[torch.Tensor], List[torch.Tensor]],
+):
+    hid_idx = torch.arange(hid_shape.prod(-1).sum(), device=hid_shape.device).unsqueeze(-1)
+    tgt_idx, tgt_shape, tgt_windows = window(hid_idx, hid_shape, window_fn)
+    tgt_idx = tgt_idx.squeeze(-1)
+    src_idx = torch.argsort(tgt_idx)
+    return (
+        lambda hid: torch.index_select(hid, 0, tgt_idx),
+        lambda hid: torch.index_select(hid, 0, src_idx),
+        tgt_shape,
+        tgt_windows,
+    )