Upload folder using huggingface_hub

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "{}"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2025 seed
+
+   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.

README.md

@@ -1,13 +1,21 @@
 ---
-title: Seedvr2 3b
-emoji: 🏢
-colorFrom: yellow
-colorTo: gray
+title: SeedVR2 3B Image Upscaler
+emoji: 🔍
+colorFrom: blue
+colorTo: purple
 sdk: gradio
-sdk_version: 6.24.0
-python_version: '3.13'
+sdk_version: 5.49.1
 app_file: app.py
 pinned: false
+license: apache-2.0
+short_description: One-step diffusion image restoration with SeedVR2-3B
+models:
+  - ByteDance-Seed/SeedVR2-3B
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# SeedVR2-3B Image Upscaler
+
+One-step diffusion restoration for images, from
+[ByteDance-Seed/SeedVR2-3B](https://huggingface.co/ByteDance-Seed/SeedVR2-3B)
+(Apache-2.0). This Space serves the image path only; see the header of `app.py`
+for what was changed from the reference implementation and why.

app.py

@@ -0,0 +1,258 @@
+# SeedVR2-3B image restoration, Upsampler v3 recipe.
+#
+# Ported from ByteDance-Seed/SeedVR2-3B. What changed and why:
+#
+# 1. Image only. The upstream Space serves video and images from one entry
+#    point, carrying the sequence-parallel plumbing, frame cutting and video
+#    writing along with it. This Space backs an image tool, so that is all gone.
+# 2. ONE @spaces.GPU entry. Upstream decorates configure_runner,
+#    generation_step AND generation_loop, so a single request booked three GPU
+#    allocations of 100s each. ZeroGPU checks the requested duration against the
+#    visitor's remaining quota, and an unauthenticated visitor has 120 seconds a
+#    day in total, so the upstream shape cannot serve an anonymous user at all.
+#    Everything now runs inside one call with a measured dynamic duration.
+# 3. No apex. Upstream installs a prebuilt `apex-0.1-cp310-...whl` and selects
+#    `fusedrms` / `fusedln` norms in configs_3b/main.yaml. ZeroGPU runs Python
+#    3.12, where that wheel does not install, so every norm layer then failed.
+#    The config now selects the `rms` / `layer` paths that the same source file
+#    already implements in pure PyTorch, with identical parameter names and
+#    shapes so the checkpoint loads unchanged.
+# 4. No hard flash-attn dependency. See models/dit_v2/attention.py.
+#
+# The model restores at a fixed ~3.7MP working resolution regardless of input
+# size (it was trained at high res and NaResize scales the input to meet it), so
+# GPU cost per request is essentially constant and no tiling is involved.
+
+import gc
+import os
+import mimetypes
+from pathlib import Path
+
+import gradio as gr
+import spaces
+import torch
+import torch.nn.functional as F
+from einops import rearrange
+from omegaconf import OmegaConf
+from PIL import Image
+from huggingface_hub import hf_hub_download
+from torchvision.transforms import Compose, Lambda, Normalize
+import torchvision.transforms as T
+
+from upsampler_theme import UPSAMPLER_CSS, UPSAMPLER_THEME, footer_html, header_html
+
+from data.image.transforms.divisible_crop import DivisibleCrop
+from data.image.transforms.na_resize import NaResize
+from data.video.transforms.rearrange import Rearrange
+from common.config import load_config
+from common.distributed import init_torch
+from common.seed import set_seed
+from projects.video_diffusion_sr.infer import VideoDiffusionInfer
+
+try:
+    from projects.video_diffusion_sr.color_fix import wavelet_reconstruction
+
+    USE_COLOR_FIX = True
+except ImportError:
+    USE_COLOR_FIX = False
+    print("color fix unavailable; output will not be wavelet-reconstructed")
+
+# Weights come from the official ByteDance repo rather than a re-upload. It is
+# the canonical source for these files and is Apache-2.0, so there is no mirror
+# in the chain that could change under us.
+WEIGHTS_REPO = "ByteDance-Seed/SeedVR2-3B"
+CKPT_DIR = Path("./ckpts")
+CKPT_DIR.mkdir(exist_ok=True)
+
+
+def _fetch(filename: str, target: Path) -> str:
+    if target.exists():
+        return str(target)
+    path = hf_hub_download(repo_id=WEIGHTS_REPO, filename=filename)
+    target.symlink_to(path)
+    return str(target)
+
+
+DIT_CKPT = _fetch("seedvr2_ema_3b.pth", CKPT_DIR / "seedvr2_ema_3b.pth")
+VAE_CKPT = _fetch("ema_vae.pth", CKPT_DIR / "ema_vae.pth")
+# The text branch is conditioned by two fixed embeddings shipped with the
+# weights, so this Space runs no text encoder at all.
+POS_EMB = _fetch("pos_emb.pt", Path("./pos_emb.pt"))
+NEG_EMB = _fetch("neg_emb.pt", Path("./neg_emb.pt"))
+
+# The resolution the model restores at, as an area. Upstream hardcodes
+# 2560*1440 for images with `downsample_only=False`, meaning small inputs are
+# scaled UP to it and large inputs down, because the model was only trained at
+# high resolution. Keeping that exact value keeps output quality identical to
+# the reference implementation.
+WORK_AREA = 2560 * 1440
+
+# Single-process "distributed" context. The model code routes every device
+# placement through common.distributed, which reads these.
+os.environ.setdefault("MASTER_ADDR", "127.0.0.1")
+os.environ.setdefault("MASTER_PORT", "12355")
+os.environ.setdefault("RANK", "0")
+os.environ.setdefault("WORLD_SIZE", "1")
+os.environ.setdefault("LOCAL_RANK", "0")
+
+_runner = None
+
+
+def _ensure_runner():
+    """Build the runner once, inside a GPU context.
+
+    Deliberately NOT done at module scope, even though ZeroGPU prefers that for
+    placement: `init_torch` ends in `dist.init_process_group(backend="nccl")`
+    and `torch.cuda.set_device`, which need a real device rather than the CUDA
+    emulation that applies outside `@spaces.GPU`. Memoized because
+    `init_process_group` raises if called twice, and because a warm worker
+    should not reload 3B parameters per request.
+    """
+    global _runner
+    if _runner is not None:
+        return _runner
+
+    if not torch.distributed.is_initialized():
+        init_torch(cudnn_benchmark=False)
+
+    runner = VideoDiffusionInfer(load_config(os.path.join("./configs_3b", "main.yaml")))
+    OmegaConf.set_readonly(runner.config, False)
+    runner.configure_dit_model(device="cuda", checkpoint=DIT_CKPT)
+    runner.configure_vae_model()
+    if hasattr(runner.vae, "set_memory_limit"):
+        runner.vae.set_memory_limit(**runner.config.vae.memory_limit)
+
+    _runner = runner
+    return _runner
+
+
+def _transform():
+    return Compose(
+        [
+            NaResize(resolution=WORK_AREA**0.5, mode="area", downsample_only=False),
+            Lambda(lambda x: torch.clamp(x, 0.0, 1.0)),
+            DivisibleCrop((16, 16)),
+            Normalize(0.5, 0.5),
+            Rearrange("t c h w -> c t h w"),
+        ]
+    )
+
+
+def _duration(image, steps=1, seed=666, progress=None) -> int:
+    """Every request restores at the same ~3.7MP working resolution, so cost
+    tracks the step count and little else. Measured on a cold worker, then
+    given headroom; kept as small as honesty allows, because the request is
+    checked against the visitor's remaining quota and a smaller one also ranks
+    higher in the ZeroGPU queue."""
+    return int(min(120, 35 + 12 * int(steps)))
+
+
+@spaces.GPU(duration=_duration)
+@torch.no_grad()
+def upscale_image(image, steps=1, seed=666, progress=gr.Progress(track_tqdm=True)):
+    if image is None:
+        raise gr.Error("Upload an image first.")
+
+    runner = _ensure_runner()
+
+    runner.config.diffusion.cfg.scale = 1.0
+    runner.config.diffusion.cfg.rescale = 0.0
+    runner.config.diffusion.timesteps.sampling.steps = int(steps)
+    runner.configure_diffusion()
+    set_seed(int(seed) % (2**32), same_across_ranks=True)
+
+    img = Image.open(image).convert("RGB") if isinstance(image, str) else image.convert("RGB")
+    tensor = T.ToTensor()(img).unsqueeze(0)  # (t=1, c, h, w)
+
+    cond = _transform()(tensor.to("cuda"))
+    original = cond
+    latents = runner.vae_encode([cond])
+
+    text_embeds = {
+        "texts_pos": [torch.load(POS_EMB).to("cuda")],
+        "texts_neg": [torch.load(NEG_EMB).to("cuda")],
+    }
+
+    noise = [torch.randn_like(latent) for latent in latents]
+    aug_noise = [torch.randn_like(latent) for latent in latents]
+
+    def _add_noise(x, aug):
+        t = torch.tensor([1000.0], device="cuda") * 0.1
+        shape = torch.tensor(x.shape[1:], device="cuda")[None]
+        return runner.schedule.forward(x, aug, runner.timestep_transform(t, shape))
+
+    conditions = [
+        runner.get_condition(n, task="sr", latent_blur=_add_noise(latent, a))
+        for n, a, latent in zip(noise, aug_noise, latents)
+    ]
+
+    with torch.autocast("cuda", torch.bfloat16, enabled=True):
+        videos = runner.inference(
+            noises=noise, conditions=conditions, dit_offload=False, **text_embeds
+        )
+
+    sample = videos[0]
+    sample = (
+        rearrange(sample[:, None], "c t h w -> t c h w")
+        if sample.ndim == 3
+        else rearrange(sample, "c t h w -> t c h w")
+    )
+    reference = (
+        rearrange(original[:, None], "c t h w -> t c h w")
+        if original.ndim == 3
+        else rearrange(original, "c t h w -> t c h w")
+    )
+    if USE_COLOR_FIX:
+        sample = wavelet_reconstruction(sample.to("cpu"), reference[: sample.size(0)].to("cpu"))
+    else:
+        sample = sample.to("cpu")
+
+    sample = rearrange(sample, "t c h w -> t h w c")
+    sample = sample.clip(-1, 1).mul_(0.5).add_(0.5).mul_(255).round().to(torch.uint8).numpy()
+
+    del latents, conditions, videos
+    gc.collect()
+    torch.cuda.empty_cache()
+
+    return Image.fromarray(sample[0])
+
+
+with gr.Blocks(css=UPSAMPLER_CSS, theme=UPSAMPLER_THEME) as demo:
+    gr.HTML(
+        header_html(
+            "SeedVR2 3B Image Upscaler",
+            "One-step diffusion restoration that rebuilds real detail in blurry, "
+            "compressed, and low-resolution photos.",
+        )
+    )
+
+    with gr.Row():
+        with gr.Column():
+            image_in = gr.Image(label="Image", type="filepath")
+            steps = gr.Slider(1, 4, value=1, step=1, label="Steps")
+            seed = gr.Number(label="Seed", value=666, precision=0)
+            run = gr.Button("Upscale Image", variant="primary")
+        with gr.Column():
+            image_out = gr.Image(label="Result", type="pil")
+
+    # No leading slash: gradio prefixes it, and "/upscale_image" here would
+    # publish the endpoint as "//upscale_image".
+    run.click(
+        upscale_image,
+        inputs=[image_in, steps, seed],
+        outputs=[image_out],
+        api_name="upscale_image",
+    )
+
+    gr.HTML(
+        footer_html(
+            "SeedVR2-3B is ByteDance's one-step diffusion model for image and video "
+            "restoration. It rebuilds genuine texture in photos that are blurry, "
+            "heavily compressed, or simply too small, restoring at high resolution "
+            "rather than smoothing detail away the way a conventional upscaler does.",
+            "https://upsampler.com/free-image-upscaler-no-signup",
+            "free image upscaler",
+        )
+    )
+
+demo.launch(ssr_mode=False, show_error=True)

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: rms
+    txt_in_dim: 5120
+    txt_in_norm: layer
+    txt_dim: ${.vid_dim}
+    emb_dim: ${eval:'6 * ${.vid_dim}'}
+    heads: 20
+    head_dim: 128 # llm-like
+    expand_ratio: 4
+    norm: rms
+    norm_eps: 1.0e-05
+    ada: single
+    qk_bias: False
+    qk_norm: rms
+    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

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_v2/attention.py

@@ -0,0 +1,86 @@
+# // 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 torch import nn
+
+# CHANGED FROM UPSTREAM (Upsampler): flash-attn is imported lazily instead of at
+# module scope, and falls back to PyTorch SDPA when it is unavailable.
+#
+# Upstream ships a prebuilt `apex-0.1-cp310-...whl` and expects a flash-attn
+# built against torch 2.4. ZeroGPU runs Python 3.12 on torch 2.8+, where neither
+# installs, so the module-scope import took the whole Space down at startup.
+try:
+    from flash_attn import flash_attn_varlen_func
+except ImportError:  # pragma: no cover - depends on the runtime image
+    flash_attn_varlen_func = None
+
+
+def _sdpa_varlen(q, k, v, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, **_):
+    """SDPA equivalent of `flash_attn_varlen_func` for this model's usage.
+
+    Only the shape this DiT actually calls is supported, and every one of those
+    constraints is asserted rather than assumed: packed `(total_len, heads,
+    head_dim)` self-attention, identical q/k segmentation, no causal mask, no
+    dropout, and the default softmax scale. Sequences are unpacked from
+    `cu_seqlens`, attended independently, and repacked — with a batch of one
+    (every request this Space serves) that is a single unmasked SDPA call, so
+    the fallback is not a slow path so much as a direct one.
+    """
+    assert torch.equal(cu_seqlens_q, cu_seqlens_k), (
+        "the SDPA fallback assumes self-attention with one segmentation"
+    )
+    lengths = (cu_seqlens_q[1:] - cu_seqlens_q[:-1]).tolist()
+    out = []
+    start = 0
+    for length in lengths:
+        end = start + length
+        # (l h d) -> (1 h l d), attend, then back to (l h d).
+        segment = [
+            x[start:end].transpose(0, 1).unsqueeze(0) for x in (q, k, v)
+        ]
+        attended = F.scaled_dot_product_attention(*segment)
+        out.append(attended.squeeze(0).transpose(0, 1))
+        start = end
+    return torch.cat(out, dim=0)
+
+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):
+        if flash_attn_varlen_func is None:
+            return _sdpa_varlen(*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,
+    )

models/dit_v2/nablocks/__init__.py

@@ -0,0 +1,26 @@
+# // 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_v2/nablocks/attention/__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 .mmattn import NaMMAttention
+
+attns = {
+    "mm_full": NaMMAttention,
+}
+
+
+def get_attn(attn_type: str):
+    if attn_type in attns:
+        return attns[attn_type]
+    raise NotImplementedError(f"{attn_type} is not supported")

models/dit_v2/nablocks/attention/mmattn.py

@@ -0,0 +1,266 @@
+# // 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, 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.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 ...mm import MMArg, MMModule
+from ...normalization import norm_layer_type
+from ...rope import get_na_rope
+from ...window import get_window_op
+from itertools import chain
+
+
+class NaMMAttention(nn.Module):
+    def __init__(
+        self,
+        vid_dim: int,
+        txt_dim: int,
+        heads: int,
+        head_dim: int,
+        qk_bias: bool,
+        qk_norm: norm_layer_type,
+        qk_norm_eps: float,
+        rope_type: Optional[str],
+        rope_dim: int,
+        shared_weights: bool,
+        **kwargs,
+    ):
+        super().__init__()
+        dim = MMArg(vid_dim, txt_dim)
+        inner_dim = heads * head_dim
+        qkv_dim = inner_dim * 3
+        self.head_dim = head_dim
+        self.proj_qkv = MMModule(
+            nn.Linear, dim, qkv_dim, bias=qk_bias, shared_weights=shared_weights
+        )
+        self.proj_out = MMModule(nn.Linear, inner_dim, dim, shared_weights=shared_weights)
+        self.norm_q = MMModule(
+            qk_norm,
+            dim=head_dim,
+            eps=qk_norm_eps,
+            elementwise_affine=True,
+            shared_weights=shared_weights,
+        )
+        self.norm_k = MMModule(
+            qk_norm,
+            dim=head_dim,
+            eps=qk_norm_eps,
+            elementwise_affine=True,
+            shared_weights=shared_weights,
+        )
+
+        self.rope = get_na_rope(rope_type=rope_type, dim=rope_dim)
+        self.attn = FlashAttentionVarlen()
+
+    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"),
+        )
+        vid_qkv = rearrange(vid_qkv, "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.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)
+
+        if self.rope:
+            if self.rope.mm:
+                vid_q, vid_k, txt_q, txt_k = self.rope(
+                    vid_q, vid_k, vid_shape, txt_q, txt_k, txt_shape, cache
+                )
+            else:
+                vid_q, vid_k = self.rope(vid_q, vid_k, vid_shape, cache)
+
+        vid_len = cache("vid_len", lambda: vid_shape.prod(-1))
+        txt_len = cache("txt_len", lambda: txt_shape.prod(-1))
+        all_len = cache("all_len", lambda: vid_len + txt_len)
+
+        concat, unconcat = cache("mm_pnp", lambda: na.concat_idx(vid_len, txt_len))
+
+        attn = self.attn(
+            q=concat(vid_q, txt_q).bfloat16(),
+            k=concat(vid_k, txt_k).bfloat16(),
+            v=concat(vid_v, txt_v).bfloat16(),
+            cu_seqlens_q=cache("mm_seqlens", lambda: F.pad(all_len.cumsum(0), (1, 0)).int()),
+            cu_seqlens_k=cache("mm_seqlens", lambda: F.pad(all_len.cumsum(0), (1, 0)).int()),
+            max_seqlen_q=cache("mm_maxlen", lambda: all_len.max().item()),
+            max_seqlen_k=cache("mm_maxlen", lambda: all_len.max().item()),
+        ).type_as(vid_q)
+
+        attn = rearrange(attn, "l h d -> l (h d)")
+        vid_out, txt_out = unconcat(attn)
+        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 NaSwinAttention(NaMMAttention):
+    def __init__(
+        self,
+        *args,
+        window: Union[int, Tuple[int, int, int]],
+        window_method: str,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        self.window = _triple(window)
+        self.window_method = window_method
+        assert all(map(lambda v: isinstance(v, int) and v >= 0, self.window))
+
+        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:
+            if self.rope.mm:
+                # repeat text q and k for window mmrope
+                _, num_h, _ = txt_q.shape
+                txt_q_repeat = rearrange(txt_q, "l h d -> l (h d)")
+                txt_q_repeat = na.unflatten(txt_q_repeat, txt_shape)
+                txt_q_repeat = [[x] * n for x, n in zip(txt_q_repeat, window_count)]
+                txt_q_repeat = list(chain(*txt_q_repeat))
+                txt_q_repeat, txt_shape_repeat = na.flatten(txt_q_repeat)
+                txt_q_repeat = rearrange(txt_q_repeat, "l (h d) -> l h d", h=num_h)
+
+                txt_k_repeat = rearrange(txt_k, "l h d -> l (h d)")
+                txt_k_repeat = na.unflatten(txt_k_repeat, txt_shape)
+                txt_k_repeat = [[x] * n for x, n in zip(txt_k_repeat, window_count)]
+                txt_k_repeat = list(chain(*txt_k_repeat))
+                txt_k_repeat, _ = na.flatten(txt_k_repeat)
+                txt_k_repeat = rearrange(txt_k_repeat, "l (h d) -> l h d", h=num_h)
+
+                vid_q, vid_k, txt_q, txt_k = self.rope(
+                    vid_q, vid_k, window_shape, txt_q_repeat, txt_k_repeat, txt_shape_repeat, cache_win
+                )
+            else:
+                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

models/dit_v2/nablocks/mmsr_block.py

@@ -0,0 +1,119 @@
+# // 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
+import torch
+import torch.nn as nn
+
+# from ..cache import Cache
+from common.cache import Cache
+
+from .attention.mmattn import NaSwinAttention
+from ..mm import MMArg
+from ..modulation import ada_layer_type
+from ..normalization import norm_layer_type
+from ..mm import MMArg, MMModule
+from ..mlp import get_mlp
+    
+
+class NaMMSRTransformerBlock(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_norm: norm_layer_type,
+        mlp_type: str,
+        shared_weights: bool,
+        rope_type: str,
+        rope_dim: int,
+        is_last_layer: bool,
+        **kwargs,
+    ):
+        super().__init__()
+        dim = MMArg(vid_dim, txt_dim)
+        self.attn_norm = MMModule(norm, dim=dim, eps=norm_eps, elementwise_affine=False, shared_weights=shared_weights,)
+
+        self.attn = NaSwinAttention(
+            vid_dim=vid_dim,
+            txt_dim=txt_dim,
+            heads=heads,
+            head_dim=head_dim,
+            qk_bias=qk_bias,
+            qk_norm=qk_norm,
+            qk_norm_eps=norm_eps,
+            rope_type=rope_type,
+            rope_dim=rope_dim,
+            shared_weights=shared_weights,
+            window=kwargs.pop("window", None),
+            window_method=kwargs.pop("window_method", None),
+        )
+
+        self.mlp_norm = MMModule(norm, dim=dim, eps=norm_eps, elementwise_affine=False, shared_weights=shared_weights, vid_only=is_last_layer)
+        self.mlp = MMModule(
+            get_mlp(mlp_type),
+            dim=dim,
+            expand_ratio=expand_ratio,
+            shared_weights=shared_weights,
+            vid_only=is_last_layer
+        )
+        self.ada = MMModule(ada, dim=dim, emb_dim=emb_dim, layers=["attn", "mlp"], shared_weights=shared_weights, vid_only=is_last_layer)
+        self.is_last_layer = is_last_layer
+
+    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_v2/nadit.py

@@ -0,0 +1,246 @@
+# // 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 List, 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 get_na_patch_layers
+
+# 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: Union[int, List[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_norm: Optional[str],
+        patch_size: Union[int, Tuple[int, int, int]],
+        num_layers: int,
+        block_type: Union[str, Tuple[str]],
+        mm_layers: Union[int, Tuple[bool]],
+        mlp_type: str = "normal",
+        patch_type: str = "v1",
+        rope_type: Optional[str] = "rope3d",
+        rope_dim: Optional[int] = None,
+        window: Optional[Tuple] = None,
+        window_method: Optional[Tuple[str]] = None,
+        msa_type: Optional[Tuple[str]] = None,
+        mca_type: Optional[Tuple[str]] = None,
+        txt_in_norm: Optional[str] = None,
+        txt_in_norm_scale_factor: int = 0.01,
+        txt_proj_type: Optional[str] = "linear",
+        vid_out_norm: Optional[str] = None,
+        **kwargs,
+    ):
+        ada = get_ada_layer(ada)
+        norm = get_norm_layer(norm)
+        qk_norm = get_norm_layer(qk_norm)
+        rope_dim = rope_dim if rope_dim is not None else head_dim // 2
+        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__()
+        NaPatchIn, NaPatchOut = get_na_patch_layers(patch_type)
+        self.vid_in = NaPatchIn(
+            in_channels=vid_in_channels,
+            patch_size=patch_size,
+            dim=vid_dim,
+        )
+        if not isinstance(txt_in_dim, int):
+            self.txt_in = nn.ModuleList([])
+            for in_dim in txt_in_dim:
+                txt_norm_layer = get_norm_layer(txt_in_norm)(txt_dim, norm_eps, True)
+                if txt_proj_type == "linear":
+                    txt_proj_layer = nn.Linear(in_dim, txt_dim)
+                else:
+                    txt_proj_layer = nn.Sequential(
+                        nn.Linear(in_dim, in_dim), nn.GELU("tanh"), nn.Linear(in_dim, txt_dim)
+                    )
+                torch.nn.init.constant_(txt_norm_layer.weight, txt_in_norm_scale_factor)
+                self.txt_in.append(
+                    nn.Sequential(
+                        txt_proj_layer,
+                        txt_norm_layer,
+                    )
+                )
+        else:
+            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 msa_type is None or isinstance(msa_type, str):
+            msa_type = [msa_type] * num_layers
+        if mca_type is None or isinstance(mca_type, str):
+            mca_type = [mca_type] * 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_norm=qk_norm,
+                    shared_weights=not (
+                        (i < mm_layers) if isinstance(mm_layers, int) else mm_layers[i]
+                    ),
+                    mlp_type=mlp_type,
+                    window=window[i],
+                    window_method=window_method[i],
+                    msa_type=msa_type[i],
+                    mca_type=mca_type[i],
+                    rope_type=rope_type,
+                    rope_dim=rope_dim,
+                    is_last_layer=(i == num_layers - 1),
+                    **kwargs,
+                )
+                for i in range(num_layers)
+            ]
+        )
+
+        self.vid_out_norm = None
+        if vid_out_norm is not None:
+            self.vid_out_norm = get_norm_layer(vid_out_norm)(
+                dim=vid_dim,
+                eps=norm_eps,
+                elementwise_affine=True,
+            )
+            self.vid_out_ada = ada(
+                dim=vid_dim,
+                emb_dim=emb_dim,
+                layers=["out"],
+                modes=["in"],
+            )
+
+        self.vid_out = NaPatchOut(
+            out_channels=vid_out_channels,
+            patch_size=patch_size,
+            dim=vid_dim,
+        )
+
+    def set_gradient_checkpointing(self, enable: bool):
+        self.gradient_checkpointing = enable
+
+    def forward(
+        self,
+        vid: torch.FloatTensor,  # l c
+        txt: Union[torch.FloatTensor, List[torch.FloatTensor]],  # l c
+        vid_shape: torch.LongTensor,  # b 3
+        txt_shape: Union[torch.LongTensor, List[torch.LongTensor]],  # b 1
+        timestep: Union[int, float, torch.IntTensor, torch.FloatTensor],  # b
+        disable_cache: bool = False,  # for test
+    ):
+        cache = Cache(disable=disable_cache)
+
+        # slice vid after patching in when using sequence parallelism
+        if isinstance(txt, list):
+            assert isinstance(self.txt_in, nn.ModuleList)
+            txt = [
+                na.unflatten(fc(i), s) for fc, i, s in zip(self.txt_in, txt, txt_shape)
+            ]  # B L D
+            txt, txt_shape = na.flatten([torch.cat(t, dim=0) for t in zip(*txt)])
+            txt = slice_inputs(txt, dim=0)
+        else:
+            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, cache)
+
+        # Embedding input.
+        emb = self.emb_in(timestep, device=vid.device, dtype=vid.dtype)
+
+        # Body
+        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,
+            )
+
+        # Video output norm.
+        if self.vid_out_norm:
+            vid = self.vid_out_norm(vid)
+            vid = self.vid_out_ada(
+                vid,
+                emb=emb,
+                layer="out",
+                mode="in",
+                hid_len=cache("vid_len", lambda: vid_shape.prod(-1)),
+                cache=cache,
+                branch_tag="vid",
+            )
+
+        # Video output.
+        vid, vid_shape = self.vid_out(vid, vid_shape, cache)
+        return NaDiTOutput(vid_sample=vid)

models/dit_v2/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_v2/patch/__init__.py

@@ -0,0 +1,19 @@
+# // 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.
+
+def get_na_patch_layers(patch_type="v1"):
+    assert patch_type in ["v1"]
+    if patch_type == "v1":
+        from .patch_v1 import NaPatchIn, NaPatchOut
+    return NaPatchIn, NaPatchOut

models/dit_v2/patch/patch_v1.py

@@ -0,0 +1,127 @@
+# // 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
+        if t > 1:
+            assert vid.size(2) % t == 1
+            vid = torch.cat([vid[:, :, :1]] * (t - 1) + [vid], dim=2)
+        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)
+        if t > 1:
+            vid = vid[:, :, (t - 1) :]
+        return vid
+
+
+class NaPatchIn(PatchIn):
+    def forward(
+        self,
+        vid: torch.Tensor,  # l c
+        vid_shape: torch.LongTensor,
+        cache: Cache = Cache(disable=True),  # for test
+    ) -> torch.Tensor:
+        cache = cache.namespace("patch")
+        vid_shape_before_patchify = cache("vid_shape_before_patchify", lambda: vid_shape)
+        t, h, w = self.patch_size
+        if not (t == h == w == 1):
+            vid = na.unflatten(vid, vid_shape)
+            for i in range(len(vid)):
+                if t > 1 and vid_shape_before_patchify[i, 0] % t != 0:
+                    vid[i] = torch.cat([vid[i][:1]] * (t - vid[i].size(0) % t) + [vid[i]], dim=0)
+                vid[i] = rearrange(vid[i], "(T t) (H h) (W w) c -> T H W (t h w c)", t=t, h=h, w=w)
+            vid, vid_shape = na.flatten(vid)
+
+        # 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),  # for test
+    ) -> Tuple[
+        torch.FloatTensor,
+        torch.LongTensor,
+    ]:
+        cache = cache.namespace("patch")
+        vid_shape_before_patchify = cache.get("vid_shape_before_patchify")
+
+        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 = na.unflatten(vid, vid_shape)
+            for i in range(len(vid)):
+                vid[i] = rearrange(vid[i], "T H W (t h w c) -> (T t) (H h) (W w) c", t=t, h=h, w=w)
+                if t > 1 and vid_shape_before_patchify[i, 0] % t != 0:
+                    vid[i] = vid[i][(t - vid_shape_before_patchify[i, 0] % t) :]
+            vid, vid_shape = na.flatten(vid)
+
+        return vid, vid_shape

models/dit_v2/rope.py

@@ -0,0 +1,150 @@
+# // 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 Optional, 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)
+        self.mm = False
+
+    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.float()).to(q.dtype)
+        k = apply_rotary_emb(freqs, k.float()).to(k.dtype)
+        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 MMRotaryEmbeddingBase(RotaryEmbeddingBase):
+    def __init__(self, dim: int, rope_dim: int):
+        super().__init__(dim, rope_dim)
+        self.rope = RotaryEmbedding(
+            dim=dim // rope_dim,
+            freqs_for="lang",
+            theta=10000,
+        )
+        freqs = self.rope.freqs
+        del self.rope.freqs
+        self.rope.register_buffer("freqs", freqs.data)
+        self.mm = True
+
+
+class NaMMRotaryEmbedding3d(MMRotaryEmbeddingBase):
+    def __init__(self, dim: int):
+        super().__init__(dim, rope_dim=3)
+
+    def forward(
+        self,
+        vid_q: torch.FloatTensor,  # L h d
+        vid_k: torch.FloatTensor,  # L h d
+        vid_shape: torch.LongTensor,  # B 3
+        txt_q: torch.FloatTensor,  # L h d
+        txt_k: torch.FloatTensor,  # L h d
+        txt_shape: torch.LongTensor,  # B 1
+        cache: Cache,
+    ) -> Tuple[
+        torch.FloatTensor,
+        torch.FloatTensor,
+        torch.FloatTensor,
+        torch.FloatTensor,
+    ]:
+        vid_freqs, txt_freqs = cache(
+            "mmrope_freqs_3d",
+            lambda: self.get_freqs(vid_shape, txt_shape),
+        )
+        vid_q = rearrange(vid_q, "L h d -> h L d")
+        vid_k = rearrange(vid_k, "L h d -> h L d")
+        vid_q = apply_rotary_emb(vid_freqs, vid_q.float()).to(vid_q.dtype)
+        vid_k = apply_rotary_emb(vid_freqs, vid_k.float()).to(vid_k.dtype)
+        vid_q = rearrange(vid_q, "h L d -> L h d")
+        vid_k = rearrange(vid_k, "h L d -> L h d")
+
+        txt_q = rearrange(txt_q, "L h d -> h L d")
+        txt_k = rearrange(txt_k, "L h d -> h L d")
+        txt_q = apply_rotary_emb(txt_freqs, txt_q.float()).to(txt_q.dtype)
+        txt_k = apply_rotary_emb(txt_freqs, txt_k.float()).to(txt_k.dtype)
+        txt_q = rearrange(txt_q, "h L d -> L h d")
+        txt_k = rearrange(txt_k, "h L d -> L h d")
+        return vid_q, vid_k, txt_q, txt_k
+
+    def get_freqs(
+        self,
+        vid_shape: torch.LongTensor,
+        txt_shape: torch.LongTensor,
+    ) -> Tuple[
+        torch.Tensor,
+        torch.Tensor,
+    ]:
+        vid_freqs = self.get_axial_freqs(1024, 128, 128)
+        txt_freqs = self.get_axial_freqs(1024)
+        vid_freq_list, txt_freq_list = [], []
+        for (f, h, w), l in zip(vid_shape.tolist(), txt_shape[:, 0].tolist()):
+            vid_freq = vid_freqs[l : l + f, :h, :w].reshape(-1, vid_freqs.size(-1))
+            txt_freq = txt_freqs[:l].repeat(1, 3).reshape(-1, vid_freqs.size(-1))
+            vid_freq_list.append(vid_freq)
+            txt_freq_list.append(txt_freq)
+        return torch.cat(vid_freq_list, dim=0), torch.cat(txt_freq_list, dim=0)
+
+
+def get_na_rope(rope_type: Optional[str], dim: int):
+    if rope_type is None:
+        return None
+    if rope_type == "mmrope3d":
+        return NaMMRotaryEmbedding3d(dim=dim)
+    raise NotImplementedError(f"{rope_type} is not supported.")

models/dit_v2/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/video_vae_v3/modules/attn_video_vae.py

@@ -0,0 +1,1345 @@
+# Copyright (c) 2023 HuggingFace Team
+# Copyright (c) 2025 ByteDance Ltd. and/or its affiliates.
+# SPDX-License-Identifier: Apache License, Version 2.0 (the "License")
+#
+# This file has been modified by ByteDance Ltd. and/or its affiliates. on 1st June 2025
+#
+# Original file was released under Apache License, Version 2.0 (the "License"), with the full license text
+# available at http://www.apache.org/licenses/LICENSE-2.0.
+#
+# This modified file is released under the same license.
+
+
+from contextlib import nullcontext
+from typing import Literal, Optional, Tuple, Union
+import diffusers
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from diffusers.models.attention_processor import Attention, SpatialNorm
+from diffusers.models.autoencoders.vae import DecoderOutput, DiagonalGaussianDistribution
+from diffusers.models.downsampling import Downsample2D
+from diffusers.models.lora import LoRACompatibleConv
+from diffusers.models.modeling_outputs import AutoencoderKLOutput
+from diffusers.models.resnet import ResnetBlock2D
+from diffusers.models.unets.unet_2d_blocks import DownEncoderBlock2D, UpDecoderBlock2D
+from diffusers.models.upsampling import Upsample2D
+from diffusers.utils import is_torch_version
+from diffusers.utils.accelerate_utils import apply_forward_hook
+from einops import rearrange
+
+from common.distributed.advanced import get_sequence_parallel_world_size
+from common.logger import get_logger
+from models.video_vae_v3.modules.causal_inflation_lib import (
+    InflatedCausalConv3d,
+    causal_norm_wrapper,
+    init_causal_conv3d,
+    remove_head,
+)
+from models.video_vae_v3.modules.context_parallel_lib import (
+    causal_conv_gather_outputs,
+    causal_conv_slice_inputs,
+)
+from models.video_vae_v3.modules.global_config import set_norm_limit
+from models.video_vae_v3.modules.types import (
+    CausalAutoencoderOutput,
+    CausalDecoderOutput,
+    CausalEncoderOutput,
+    MemoryState,
+    _inflation_mode_t,
+    _memory_device_t,
+    _receptive_field_t,
+)
+
+logger = get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class Upsample3D(Upsample2D):
+    """A 3D upsampling layer with an optional convolution."""
+
+    def __init__(
+        self,
+        *args,
+        inflation_mode: _inflation_mode_t = "tail",
+        temporal_up: bool = False,
+        spatial_up: bool = True,
+        slicing: bool = False,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        conv = self.conv if self.name == "conv" else self.Conv2d_0
+
+        assert type(conv) is not nn.ConvTranspose2d
+        # Note: lora_layer is not passed into constructor in the original implementation.
+        # So we make a simplification.
+        conv = init_causal_conv3d(
+            self.channels,
+            self.out_channels,
+            3,
+            padding=1,
+            inflation_mode=inflation_mode,
+        )
+
+        self.temporal_up = temporal_up
+        self.spatial_up = spatial_up
+        self.temporal_ratio = 2 if temporal_up else 1
+        self.spatial_ratio = 2 if spatial_up else 1
+        self.slicing = slicing
+
+        assert not self.interpolate
+        # [Override] MAGViT v2 implementation
+        if not self.interpolate:
+            upscale_ratio = (self.spatial_ratio**2) * self.temporal_ratio
+            self.upscale_conv = nn.Conv3d(
+                self.channels, self.channels * upscale_ratio, kernel_size=1, padding=0
+            )
+            identity = (
+                torch.eye(self.channels)
+                .repeat(upscale_ratio, 1)
+                .reshape_as(self.upscale_conv.weight)
+            )
+            self.upscale_conv.weight.data.copy_(identity)
+            nn.init.zeros_(self.upscale_conv.bias)
+
+        if self.name == "conv":
+            self.conv = conv
+        else:
+            self.Conv2d_0 = conv
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        output_size: Optional[int] = None,
+        memory_state: MemoryState = MemoryState.DISABLED,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        assert hidden_states.shape[1] == self.channels
+
+        if hasattr(self, "norm") and self.norm is not None:
+            # [Overridden] change to causal norm.
+            hidden_states = causal_norm_wrapper(self.norm, hidden_states)
+
+        if self.use_conv_transpose:
+            return self.conv(hidden_states)
+
+        if self.slicing:
+            split_size = hidden_states.size(2) // 2
+            hidden_states = list(
+                hidden_states.split([split_size, hidden_states.size(2) - split_size], dim=2)
+            )
+        else:
+            hidden_states = [hidden_states]
+
+        for i in range(len(hidden_states)):
+            hidden_states[i] = self.upscale_conv(hidden_states[i])
+            hidden_states[i] = rearrange(
+                hidden_states[i],
+                "b (x y z c) f h w -> b c (f z) (h x) (w y)",
+                x=self.spatial_ratio,
+                y=self.spatial_ratio,
+                z=self.temporal_ratio,
+            )
+
+        # [Overridden] For causal temporal conv
+        if self.temporal_up and memory_state != MemoryState.ACTIVE:
+            hidden_states[0] = remove_head(hidden_states[0])
+
+        if not self.slicing:
+            hidden_states = hidden_states[0]
+
+        if self.use_conv:
+            if self.name == "conv":
+                hidden_states = self.conv(hidden_states, memory_state=memory_state)
+            else:
+                hidden_states = self.Conv2d_0(hidden_states, memory_state=memory_state)
+
+        if not self.slicing:
+            return hidden_states
+        else:
+            return torch.cat(hidden_states, dim=2)
+
+
+class Downsample3D(Downsample2D):
+    """A 3D downsampling layer with an optional convolution."""
+
+    def __init__(
+        self,
+        *args,
+        inflation_mode: _inflation_mode_t = "tail",
+        spatial_down: bool = False,
+        temporal_down: bool = False,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        conv = self.conv
+        self.temporal_down = temporal_down
+        self.spatial_down = spatial_down
+
+        self.temporal_ratio = 2 if temporal_down else 1
+        self.spatial_ratio = 2 if spatial_down else 1
+
+        self.temporal_kernel = 3 if temporal_down else 1
+        self.spatial_kernel = 3 if spatial_down else 1
+
+        if type(conv) in [nn.Conv2d, LoRACompatibleConv]:
+            # Note: lora_layer is not passed into constructor in the original implementation.
+            # So we make a simplification.
+            conv = init_causal_conv3d(
+                self.channels,
+                self.out_channels,
+                kernel_size=(self.temporal_kernel, self.spatial_kernel, self.spatial_kernel),
+                stride=(self.temporal_ratio, self.spatial_ratio, self.spatial_ratio),
+                padding=(
+                    1 if self.temporal_down else 0,
+                    self.padding if self.spatial_down else 0,
+                    self.padding if self.spatial_down else 0,
+                ),
+                inflation_mode=inflation_mode,
+            )
+        elif type(conv) is nn.AvgPool2d:
+            assert self.channels == self.out_channels
+            conv = nn.AvgPool3d(
+                kernel_size=(self.temporal_ratio, self.spatial_ratio, self.spatial_ratio),
+                stride=(self.temporal_ratio, self.spatial_ratio, self.spatial_ratio),
+            )
+        else:
+            raise NotImplementedError
+
+        if self.name == "conv":
+            self.Conv2d_0 = conv
+            self.conv = conv
+        else:
+            self.conv = conv
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        memory_state: MemoryState = MemoryState.DISABLED,
+        **kwargs,
+    ) -> torch.FloatTensor:
+
+        assert hidden_states.shape[1] == self.channels
+
+        if hasattr(self, "norm") and self.norm is not None:
+            # [Overridden] change to causal norm.
+            hidden_states = causal_norm_wrapper(self.norm, hidden_states)
+
+        if self.use_conv and self.padding == 0 and self.spatial_down:
+            pad = (0, 1, 0, 1)
+            hidden_states = F.pad(hidden_states, pad, mode="constant", value=0)
+
+        assert hidden_states.shape[1] == self.channels
+
+        hidden_states = self.conv(hidden_states, memory_state=memory_state)
+
+        return hidden_states
+
+
+class ResnetBlock3D(ResnetBlock2D):
+    def __init__(
+        self,
+        *args,
+        inflation_mode: _inflation_mode_t = "tail",
+        time_receptive_field: _receptive_field_t = "half",
+        slicing: bool = False,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        self.conv1 = init_causal_conv3d(
+            self.in_channels,
+            self.out_channels,
+            kernel_size=(1, 3, 3) if time_receptive_field == "half" else (3, 3, 3),
+            stride=1,
+            padding=(0, 1, 1) if time_receptive_field == "half" else (1, 1, 1),
+            inflation_mode=inflation_mode,
+        )
+
+        self.conv2 = init_causal_conv3d(
+            self.out_channels,
+            self.conv2.out_channels,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            inflation_mode=inflation_mode,
+        )
+
+        if self.up:
+            assert type(self.upsample) is Upsample2D
+            self.upsample = Upsample3D(
+                self.in_channels,
+                use_conv=False,
+                inflation_mode=inflation_mode,
+                slicing=slicing,
+            )
+        elif self.down:
+            assert type(self.downsample) is Downsample2D
+            self.downsample = Downsample3D(
+                self.in_channels,
+                use_conv=False,
+                padding=1,
+                name="op",
+                inflation_mode=inflation_mode,
+            )
+
+        if self.use_in_shortcut:
+            self.conv_shortcut = init_causal_conv3d(
+                self.in_channels,
+                self.conv_shortcut.out_channels,
+                kernel_size=1,
+                stride=1,
+                padding=0,
+                bias=(self.conv_shortcut.bias is not None),
+                inflation_mode=inflation_mode,
+            )
+
+    def forward(
+        self, input_tensor, temb, memory_state: MemoryState = MemoryState.DISABLED, **kwargs
+    ):
+        hidden_states = input_tensor
+
+        hidden_states = causal_norm_wrapper(self.norm1, hidden_states)
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        if self.upsample is not None:
+            # upsample_nearest_nhwc fails with large batch sizes.
+            # see https://github.com/huggingface/diffusers/issues/984
+            if hidden_states.shape[0] >= 64:
+                input_tensor = input_tensor.contiguous()
+                hidden_states = hidden_states.contiguous()
+            input_tensor = self.upsample(input_tensor, memory_state=memory_state)
+            hidden_states = self.upsample(hidden_states, memory_state=memory_state)
+        elif self.downsample is not None:
+            input_tensor = self.downsample(input_tensor, memory_state=memory_state)
+            hidden_states = self.downsample(hidden_states, memory_state=memory_state)
+
+        hidden_states = self.conv1(hidden_states, memory_state=memory_state)
+
+        if self.time_emb_proj is not None:
+            if not self.skip_time_act:
+                temb = self.nonlinearity(temb)
+            temb = self.time_emb_proj(temb)[:, :, None, None]
+
+        if temb is not None and self.time_embedding_norm == "default":
+            hidden_states = hidden_states + temb
+
+        hidden_states = causal_norm_wrapper(self.norm2, hidden_states)
+
+        if temb is not None and self.time_embedding_norm == "scale_shift":
+            scale, shift = torch.chunk(temb, 2, dim=1)
+            hidden_states = hidden_states * (1 + scale) + shift
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states, memory_state=memory_state)
+
+        if self.conv_shortcut is not None:
+            input_tensor = self.conv_shortcut(input_tensor, memory_state=memory_state)
+
+        output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
+
+        return output_tensor
+
+
+class DownEncoderBlock3D(DownEncoderBlock2D):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_downsample: bool = True,
+        downsample_padding: int = 1,
+        inflation_mode: _inflation_mode_t = "tail",
+        time_receptive_field: _receptive_field_t = "half",
+        temporal_down: bool = True,
+        spatial_down: bool = True,
+    ):
+        super().__init__(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            dropout=dropout,
+            num_layers=num_layers,
+            resnet_eps=resnet_eps,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_pre_norm=resnet_pre_norm,
+            output_scale_factor=output_scale_factor,
+            add_downsample=add_downsample,
+            downsample_padding=downsample_padding,
+        )
+        resnets = []
+        temporal_modules = []
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                # [Override] Replace module.
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=None,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    inflation_mode=inflation_mode,
+                    time_receptive_field=time_receptive_field,
+                )
+            )
+            temporal_modules.append(nn.Identity())
+
+        self.resnets = nn.ModuleList(resnets)
+        self.temporal_modules = nn.ModuleList(temporal_modules)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    # [Override] Replace module.
+                    Downsample3D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                        temporal_down=temporal_down,
+                        spatial_down=spatial_down,
+                        inflation_mode=inflation_mode,
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        memory_state: MemoryState = MemoryState.DISABLED,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        for resnet, temporal in zip(self.resnets, self.temporal_modules):
+            hidden_states = resnet(hidden_states, temb=None, memory_state=memory_state)
+            hidden_states = temporal(hidden_states)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, memory_state=memory_state)
+
+        return hidden_states
+
+
+class UpDecoderBlock3D(UpDecoderBlock2D):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",  # default, spatial
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_upsample: bool = True,
+        temb_channels: Optional[int] = None,
+        inflation_mode: _inflation_mode_t = "tail",
+        time_receptive_field: _receptive_field_t = "half",
+        temporal_up: bool = True,
+        spatial_up: bool = True,
+        slicing: bool = False,
+    ):
+        super().__init__(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            dropout=dropout,
+            num_layers=num_layers,
+            resnet_eps=resnet_eps,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_pre_norm=resnet_pre_norm,
+            output_scale_factor=output_scale_factor,
+            add_upsample=add_upsample,
+            temb_channels=temb_channels,
+        )
+        resnets = []
+        temporal_modules = []
+
+        for i in range(num_layers):
+            input_channels = in_channels if i == 0 else out_channels
+
+            resnets.append(
+                # [Override] Replace module.
+                ResnetBlock3D(
+                    in_channels=input_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    inflation_mode=inflation_mode,
+                    time_receptive_field=time_receptive_field,
+                    slicing=slicing,
+                )
+            )
+
+            temporal_modules.append(nn.Identity())
+
+        self.resnets = nn.ModuleList(resnets)
+        self.temporal_modules = nn.ModuleList(temporal_modules)
+
+        if add_upsample:
+            # [Override] Replace module & use learnable upsample
+            self.upsamplers = nn.ModuleList(
+                [
+                    Upsample3D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        temporal_up=temporal_up,
+                        spatial_up=spatial_up,
+                        interpolate=False,
+                        inflation_mode=inflation_mode,
+                        slicing=slicing,
+                    )
+                ]
+            )
+        else:
+            self.upsamplers = None
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        memory_state: MemoryState = MemoryState.DISABLED,
+    ) -> torch.FloatTensor:
+        for resnet, temporal in zip(self.resnets, self.temporal_modules):
+            hidden_states = resnet(hidden_states, temb=None, memory_state=memory_state)
+            hidden_states = temporal(hidden_states)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, memory_state=memory_state)
+
+        return hidden_states
+
+
+class UNetMidBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",  # default, spatial
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        add_attention: bool = True,
+        attention_head_dim: int = 1,
+        output_scale_factor: float = 1.0,
+        inflation_mode: _inflation_mode_t = "tail",
+        time_receptive_field: _receptive_field_t = "half",
+    ):
+        super().__init__()
+        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        self.add_attention = add_attention
+
+        # there is always at least one resnet
+        resnets = [
+            # [Override] Replace module.
+            ResnetBlock3D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+                inflation_mode=inflation_mode,
+                time_receptive_field=time_receptive_field,
+            )
+        ]
+        attentions = []
+
+        if attention_head_dim is None:
+            logger.warn(
+                f"It is not recommend to pass `attention_head_dim=None`. "
+                f"Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
+            )
+            attention_head_dim = in_channels
+
+        for _ in range(num_layers):
+            if self.add_attention:
+                attentions.append(
+                    Attention(
+                        in_channels,
+                        heads=in_channels // attention_head_dim,
+                        dim_head=attention_head_dim,
+                        rescale_output_factor=output_scale_factor,
+                        eps=resnet_eps,
+                        norm_num_groups=(
+                            resnet_groups if resnet_time_scale_shift == "default" else None
+                        ),
+                        spatial_norm_dim=(
+                            temb_channels if resnet_time_scale_shift == "spatial" else None
+                        ),
+                        residual_connection=True,
+                        bias=True,
+                        upcast_softmax=True,
+                        _from_deprecated_attn_block=True,
+                    )
+                )
+            else:
+                attentions.append(None)
+
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    inflation_mode=inflation_mode,
+                    time_receptive_field=time_receptive_field,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def forward(self, hidden_states, temb=None, memory_state: MemoryState = MemoryState.DISABLED):
+        video_length, frame_height, frame_width = hidden_states.size()[-3:]
+        hidden_states = self.resnets[0](hidden_states, temb, memory_state=memory_state)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if attn is not None:
+                hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
+                hidden_states = attn(hidden_states, temb=temb)
+                hidden_states = rearrange(
+                    hidden_states, "(b f) c h w -> b c f h w", f=video_length
+                )
+            hidden_states = resnet(hidden_states, temb, memory_state=memory_state)
+
+        return hidden_states
+
+
+class Encoder3D(nn.Module):
+    r"""
+    [Override] override most logics to support extra condition input and causal conv
+
+    The `Encoder` layer of a variational autoencoder that encodes
+    its input into a latent representation.
+
+    Args:
+        in_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        out_channels (`int`, *optional*, defaults to 3):
+            The number of output channels.
+        down_block_types (`Tuple[str, ...]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+            The types of down blocks to use.
+            See `~diffusers.models.unet_2d_blocks.get_down_block`
+            for available options.
+        block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
+            The number of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2):
+            The number of layers per block.
+        norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups for normalization.
+        act_fn (`str`, *optional*, defaults to `"silu"`):
+            The activation function to use.
+            See `~diffusers.models.activations.get_activation` for available options.
+        double_z (`bool`, *optional*, defaults to `True`):
+            Whether to double the number of output channels for the last block.
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        down_block_types: Tuple[str, ...] = ("DownEncoderBlock3D",),
+        block_out_channels: Tuple[int, ...] = (64,),
+        layers_per_block: int = 2,
+        norm_num_groups: int = 32,
+        act_fn: str = "silu",
+        double_z: bool = True,
+        mid_block_add_attention=True,
+        # [Override] add extra_cond_dim, temporal down num
+        temporal_down_num: int = 2,
+        extra_cond_dim: int = None,
+        gradient_checkpoint: bool = False,
+        inflation_mode: _inflation_mode_t = "tail",
+        time_receptive_field: _receptive_field_t = "half",
+    ):
+        super().__init__()
+        self.layers_per_block = layers_per_block
+        self.temporal_down_num = temporal_down_num
+
+        self.conv_in = init_causal_conv3d(
+            in_channels,
+            block_out_channels[0],
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            inflation_mode=inflation_mode,
+        )
+
+        self.mid_block = None
+        self.down_blocks = nn.ModuleList([])
+        self.extra_cond_dim = extra_cond_dim
+
+        self.conv_extra_cond = nn.ModuleList([])
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+            # [Override] to support temporal down block design
+            is_temporal_down_block = i >= len(block_out_channels) - self.temporal_down_num - 1
+            # Note: take the last ones
+
+            assert down_block_type == "DownEncoderBlock3D"
+
+            down_block = DownEncoderBlock3D(
+                num_layers=self.layers_per_block,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                add_downsample=not is_final_block,
+                resnet_eps=1e-6,
+                downsample_padding=0,
+                # Note: Don't know why set it as 0
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                temporal_down=is_temporal_down_block,
+                spatial_down=True,
+                inflation_mode=inflation_mode,
+                time_receptive_field=time_receptive_field,
+            )
+            self.down_blocks.append(down_block)
+
+            def zero_module(module):
+                # Zero out the parameters of a module and return it.
+                for p in module.parameters():
+                    p.detach().zero_()
+                return module
+
+            self.conv_extra_cond.append(
+                zero_module(
+                    nn.Conv3d(extra_cond_dim, output_channel, kernel_size=1, stride=1, padding=0)
+                )
+                if self.extra_cond_dim is not None and self.extra_cond_dim > 0
+                else None
+            )
+
+        # mid
+        self.mid_block = UNetMidBlock3D(
+            in_channels=block_out_channels[-1],
+            resnet_eps=1e-6,
+            resnet_act_fn=act_fn,
+            output_scale_factor=1,
+            resnet_time_scale_shift="default",
+            attention_head_dim=block_out_channels[-1],
+            resnet_groups=norm_num_groups,
+            temb_channels=None,
+            add_attention=mid_block_add_attention,
+            inflation_mode=inflation_mode,
+            time_receptive_field=time_receptive_field,
+        )
+
+        # out
+        self.conv_norm_out = nn.GroupNorm(
+            num_channels=block_out_channels[-1], num_groups=norm_num_groups, eps=1e-6
+        )
+        self.conv_act = nn.SiLU()
+
+        conv_out_channels = 2 * out_channels if double_z else out_channels
+        self.conv_out = init_causal_conv3d(
+            block_out_channels[-1], conv_out_channels, 3, padding=1, inflation_mode=inflation_mode
+        )
+
+        self.gradient_checkpointing = gradient_checkpoint
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        extra_cond=None,
+        memory_state: MemoryState = MemoryState.DISABLED,
+    ) -> torch.FloatTensor:
+        r"""The forward method of the `Encoder` class."""
+        sample = self.conv_in(sample, memory_state=memory_state)
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            # down
+            # [Override] add extra block and extra cond
+            for down_block, extra_block in zip(self.down_blocks, self.conv_extra_cond):
+                sample = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(down_block), sample, memory_state, use_reentrant=False
+                )
+                if extra_block is not None:
+                    sample = sample + F.interpolate(extra_block(extra_cond), size=sample.shape[2:])
+
+            # middle
+            sample = self.mid_block(sample, memory_state=memory_state)
+
+            # sample = torch.utils.checkpoint.checkpoint(
+            #     create_custom_forward(self.mid_block), sample, use_reentrant=False
+            # )
+
+        else:
+            # down
+            # [Override] add extra block and extra cond
+            for down_block, extra_block in zip(self.down_blocks, self.conv_extra_cond):
+                sample = down_block(sample, memory_state=memory_state)
+                if extra_block is not None:
+                    sample = sample + F.interpolate(extra_block(extra_cond), size=sample.shape[2:])
+
+            # middle
+            sample = self.mid_block(sample, memory_state=memory_state)
+
+        # post-process
+        sample = causal_norm_wrapper(self.conv_norm_out, sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample, memory_state=memory_state)
+
+        return sample
+
+
+class Decoder3D(nn.Module):
+    r"""
+    The `Decoder` layer of a variational autoencoder that
+    decodes its latent representation into an output sample.
+
+    Args:
+        in_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        out_channels (`int`, *optional*, defaults to 3):
+            The number of output channels.
+        up_block_types (`Tuple[str, ...]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+            The types of up blocks to use.
+            See `~diffusers.models.unet_2d_blocks.get_up_block` for available options.
+        block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
+            The number of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2):
+            The number of layers per block.
+        norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups for normalization.
+        act_fn (`str`, *optional*, defaults to `"silu"`):
+            The activation function to use.
+            See `~diffusers.models.activations.get_activation` for available options.
+        norm_type (`str`, *optional*, defaults to `"group"`):
+            The normalization type to use. Can be either `"group"` or `"spatial"`.
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        up_block_types: Tuple[str, ...] = ("UpDecoderBlock3D",),
+        block_out_channels: Tuple[int, ...] = (64,),
+        layers_per_block: int = 2,
+        norm_num_groups: int = 32,
+        act_fn: str = "silu",
+        norm_type: str = "group",  # group, spatial
+        mid_block_add_attention=True,
+        # [Override] add temporal up block
+        inflation_mode: _inflation_mode_t = "tail",
+        time_receptive_field: _receptive_field_t = "half",
+        temporal_up_num: int = 2,
+        slicing_up_num: int = 0,
+        gradient_checkpoint: bool = False,
+    ):
+        super().__init__()
+        self.layers_per_block = layers_per_block
+        self.temporal_up_num = temporal_up_num
+
+        self.conv_in = init_causal_conv3d(
+            in_channels,
+            block_out_channels[-1],
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            inflation_mode=inflation_mode,
+        )
+
+        self.mid_block = None
+        self.up_blocks = nn.ModuleList([])
+
+        temb_channels = in_channels if norm_type == "spatial" else None
+
+        # mid
+        self.mid_block = UNetMidBlock3D(
+            in_channels=block_out_channels[-1],
+            resnet_eps=1e-6,
+            resnet_act_fn=act_fn,
+            output_scale_factor=1,
+            resnet_time_scale_shift="default" if norm_type == "group" else norm_type,
+            attention_head_dim=block_out_channels[-1],
+            resnet_groups=norm_num_groups,
+            temb_channels=temb_channels,
+            add_attention=mid_block_add_attention,
+            inflation_mode=inflation_mode,
+            time_receptive_field=time_receptive_field,
+        )
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        output_channel = reversed_block_out_channels[0]
+        print(f"slicing_up_num: {slicing_up_num}")
+        for i, up_block_type in enumerate(up_block_types):
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+
+            is_final_block = i == len(block_out_channels) - 1
+            is_temporal_up_block = i < self.temporal_up_num
+            is_slicing_up_block = i >= len(block_out_channels) - slicing_up_num
+            # Note: Keep symmetric
+
+            assert up_block_type == "UpDecoderBlock3D"
+            up_block = UpDecoderBlock3D(
+                num_layers=self.layers_per_block + 1,
+                in_channels=prev_output_channel,
+                out_channels=output_channel,
+                add_upsample=not is_final_block,
+                resnet_eps=1e-6,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=norm_type,
+                temb_channels=temb_channels,
+                temporal_up=is_temporal_up_block,
+                slicing=is_slicing_up_block,
+                inflation_mode=inflation_mode,
+                time_receptive_field=time_receptive_field,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if norm_type == "spatial":
+            self.conv_norm_out = SpatialNorm(block_out_channels[0], temb_channels)
+        else:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=1e-6
+            )
+        self.conv_act = nn.SiLU()
+        self.conv_out = init_causal_conv3d(
+            block_out_channels[0], out_channels, 3, padding=1, inflation_mode=inflation_mode
+        )
+
+        self.gradient_checkpointing = gradient_checkpoint
+
+    # Note: Just copy from Decoder.
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        latent_embeds: Optional[torch.FloatTensor] = None,
+        memory_state: MemoryState = MemoryState.DISABLED,
+    ) -> torch.FloatTensor:
+        r"""The forward method of the `Decoder` class."""
+
+        sample = self.conv_in(sample, memory_state=memory_state)
+
+        upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            if is_torch_version(">=", "1.11.0"):
+                sample = self.mid_block(sample, latent_embeds, memory_state=memory_state)
+                sample = sample.to(upscale_dtype)
+
+                # up
+                for up_block in self.up_blocks:
+                    sample = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(up_block),
+                        sample,
+                        latent_embeds,
+                        memory_state,
+                        use_reentrant=False,
+                    )
+            else:
+                # middle
+                sample = self.mid_block(sample, latent_embeds, memory_state=memory_state)
+                sample = sample.to(upscale_dtype)
+
+                # up
+                for up_block in self.up_blocks:
+                    sample = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(up_block), sample, latent_embeds, memory_state
+                    )
+        else:
+            # middle
+            sample = self.mid_block(sample, latent_embeds, memory_state=memory_state)
+            sample = sample.to(upscale_dtype)
+
+            # up
+            for up_block in self.up_blocks:
+                sample = up_block(sample, latent_embeds, memory_state=memory_state)
+
+        # post-process
+        sample = causal_norm_wrapper(self.conv_norm_out, sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample, memory_state=memory_state)
+
+        return sample
+
+
+class AutoencoderKL(diffusers.AutoencoderKL):
+    """
+    We simply inherit the model code from diffusers
+    """
+
+    def __init__(self, attention: bool = True, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # A hacky way to remove attention.
+        if not attention:
+            self.encoder.mid_block.attentions = torch.nn.ModuleList([None])
+            self.decoder.mid_block.attentions = torch.nn.ModuleList([None])
+
+    def load_state_dict(self, state_dict, strict=True):
+        # Newer version of diffusers changed the model keys,
+        # causing incompatibility with old checkpoints.
+        # They provided a method for conversion. We call conversion before loading state_dict.
+        convert_deprecated_attention_blocks = getattr(
+            self, "_convert_deprecated_attention_blocks", None
+        )
+        if callable(convert_deprecated_attention_blocks):
+            convert_deprecated_attention_blocks(state_dict)
+        return super().load_state_dict(state_dict, strict)
+
+
+class VideoAutoencoderKL(diffusers.AutoencoderKL):
+    """
+    We simply inherit the model code from diffusers
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        down_block_types: Tuple[str] = ("DownEncoderBlock3D",),
+        up_block_types: Tuple[str] = ("UpDecoderBlock3D",),
+        block_out_channels: Tuple[int] = (64,),
+        layers_per_block: int = 1,
+        act_fn: str = "silu",
+        latent_channels: int = 4,
+        norm_num_groups: int = 32,
+        sample_size: int = 32,
+        scaling_factor: float = 0.18215,
+        force_upcast: float = True,
+        attention: bool = True,
+        temporal_scale_num: int = 2,
+        slicing_up_num: int = 0,
+        gradient_checkpoint: bool = False,
+        inflation_mode: _inflation_mode_t = "tail",
+        time_receptive_field: _receptive_field_t = "full",
+        slicing_sample_min_size: int = 32,
+        use_quant_conv: bool = True,
+        use_post_quant_conv: bool = True,
+        *args,
+        **kwargs,
+    ):
+        extra_cond_dim = kwargs.pop("extra_cond_dim") if "extra_cond_dim" in kwargs else None
+        self.slicing_sample_min_size = slicing_sample_min_size
+        self.slicing_latent_min_size = slicing_sample_min_size // (2**temporal_scale_num)
+
+        super().__init__(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            # [Override] make sure it can be normally initialized
+            down_block_types=tuple(
+                [down_block_type.replace("3D", "2D") for down_block_type in down_block_types]
+            ),
+            up_block_types=tuple(
+                [up_block_type.replace("3D", "2D") for up_block_type in up_block_types]
+            ),
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            act_fn=act_fn,
+            latent_channels=latent_channels,
+            norm_num_groups=norm_num_groups,
+            sample_size=sample_size,
+            scaling_factor=scaling_factor,
+            force_upcast=force_upcast,
+            *args,
+            **kwargs,
+        )
+
+        # pass init params to Encoder
+        self.encoder = Encoder3D(
+            in_channels=in_channels,
+            out_channels=latent_channels,
+            down_block_types=down_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            act_fn=act_fn,
+            norm_num_groups=norm_num_groups,
+            double_z=True,
+            extra_cond_dim=extra_cond_dim,
+            # [Override] add temporal_down_num parameter
+            temporal_down_num=temporal_scale_num,
+            gradient_checkpoint=gradient_checkpoint,
+            inflation_mode=inflation_mode,
+            time_receptive_field=time_receptive_field,
+        )
+
+        # pass init params to Decoder
+        self.decoder = Decoder3D(
+            in_channels=latent_channels,
+            out_channels=out_channels,
+            up_block_types=up_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            norm_num_groups=norm_num_groups,
+            act_fn=act_fn,
+            # [Override] add temporal_up_num parameter
+            temporal_up_num=temporal_scale_num,
+            slicing_up_num=slicing_up_num,
+            gradient_checkpoint=gradient_checkpoint,
+            inflation_mode=inflation_mode,
+            time_receptive_field=time_receptive_field,
+        )
+
+        self.quant_conv = (
+            init_causal_conv3d(
+                in_channels=2 * latent_channels,
+                out_channels=2 * latent_channels,
+                kernel_size=1,
+                inflation_mode=inflation_mode,
+            )
+            if use_quant_conv
+            else None
+        )
+        self.post_quant_conv = (
+            init_causal_conv3d(
+                in_channels=latent_channels,
+                out_channels=latent_channels,
+                kernel_size=1,
+                inflation_mode=inflation_mode,
+            )
+            if use_post_quant_conv
+            else None
+        )
+
+        # A hacky way to remove attention.
+        if not attention:
+            self.encoder.mid_block.attentions = torch.nn.ModuleList([None])
+            self.decoder.mid_block.attentions = torch.nn.ModuleList([None])
+
+    @apply_forward_hook
+    def encode(self, x: torch.FloatTensor, return_dict: bool = True) -> AutoencoderKLOutput:
+        h = self.slicing_encode(x)
+        posterior = DiagonalGaussianDistribution(h)
+
+        if not return_dict:
+            return (posterior,)
+
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    @apply_forward_hook
+    def decode(
+        self, z: torch.Tensor, return_dict: bool = True
+    ) -> Union[DecoderOutput, torch.Tensor]:
+        decoded = self.slicing_decode(z)
+
+        if not return_dict:
+            return (decoded,)
+
+        return DecoderOutput(sample=decoded)
+
+    def _encode(
+        self, x: torch.Tensor, memory_state: MemoryState = MemoryState.DISABLED
+    ) -> torch.Tensor:
+        _x = x.to(self.device)
+        _x = causal_conv_slice_inputs(_x, self.slicing_sample_min_size, memory_state=memory_state)
+        h = self.encoder(_x, memory_state=memory_state)
+        if self.quant_conv is not None:
+            output = self.quant_conv(h, memory_state=memory_state)
+        else:
+            output = h
+        output = causal_conv_gather_outputs(output)
+        return output.to(x.device)
+
+    def _decode(
+        self, z: torch.Tensor, memory_state: MemoryState = MemoryState.DISABLED
+    ) -> torch.Tensor:
+        _z = z.to(self.device)
+        _z = causal_conv_slice_inputs(_z, self.slicing_latent_min_size, memory_state=memory_state)
+        if self.post_quant_conv is not None:
+            _z = self.post_quant_conv(_z, memory_state=memory_state)
+        output = self.decoder(_z, memory_state=memory_state)
+        output = causal_conv_gather_outputs(output)
+        return output.to(z.device)
+
+    def slicing_encode(self, x: torch.Tensor) -> torch.Tensor:
+        sp_size = get_sequence_parallel_world_size()
+        if self.use_slicing and (x.shape[2] - 1) > self.slicing_sample_min_size * sp_size:
+            x_slices = x[:, :, 1:].split(split_size=self.slicing_sample_min_size * sp_size, dim=2)
+            encoded_slices = [
+                self._encode(
+                    torch.cat((x[:, :, :1], x_slices[0]), dim=2),
+                    memory_state=MemoryState.INITIALIZING,
+                )
+            ]
+            for x_idx in range(1, len(x_slices)):
+                encoded_slices.append(
+                    self._encode(x_slices[x_idx], memory_state=MemoryState.ACTIVE)
+                )
+            return torch.cat(encoded_slices, dim=2)
+        else:
+            return self._encode(x)
+
+    def slicing_decode(self, z: torch.Tensor) -> torch.Tensor:
+        sp_size = get_sequence_parallel_world_size()
+        if self.use_slicing and (z.shape[2] - 1) > self.slicing_latent_min_size * sp_size:
+            z_slices = z[:, :, 1:].split(split_size=self.slicing_latent_min_size * sp_size, dim=2)
+            decoded_slices = [
+                self._decode(
+                    torch.cat((z[:, :, :1], z_slices[0]), dim=2),
+                    memory_state=MemoryState.INITIALIZING,
+                )
+            ]
+            for z_idx in range(1, len(z_slices)):
+                decoded_slices.append(
+                    self._decode(z_slices[z_idx], memory_state=MemoryState.ACTIVE)
+                )
+            return torch.cat(decoded_slices, dim=2)
+        else:
+            return self._decode(z)
+
+    def tiled_encode(self, x: torch.Tensor, **kwargs) -> torch.Tensor:
+        raise NotImplementedError
+
+    def tiled_decode(self, z: torch.Tensor, **kwargs) -> torch.Tensor:
+        raise NotImplementedError
+
+    def forward(
+        self, x: torch.FloatTensor, mode: Literal["encode", "decode", "all"] = "all", **kwargs
+    ):
+        # x: [b c t h w]
+        if mode == "encode":
+            h = self.encode(x)
+            return h.latent_dist
+        elif mode == "decode":
+            h = self.decode(x)
+            return h.sample
+        else:
+            h = self.encode(x)
+            h = self.decode(h.latent_dist.mode())
+            return h.sample
+
+    def load_state_dict(self, state_dict, strict=False):
+        # Newer version of diffusers changed the model keys,
+        # causing incompatibility with old checkpoints.
+        # They provided a method for conversion.
+        # We call conversion before loading state_dict.
+        convert_deprecated_attention_blocks = getattr(
+            self, "_convert_deprecated_attention_blocks", None
+        )
+        if callable(convert_deprecated_attention_blocks):
+            convert_deprecated_attention_blocks(state_dict)
+        return super().load_state_dict(state_dict, strict)
+
+
+class VideoAutoencoderKLWrapper(VideoAutoencoderKL):
+    def __init__(
+        self,
+        *args,
+        spatial_downsample_factor: int,
+        temporal_downsample_factor: int,
+        freeze_encoder: bool,
+        **kwargs,
+    ):
+        self.spatial_downsample_factor = spatial_downsample_factor
+        self.temporal_downsample_factor = temporal_downsample_factor
+        self.freeze_encoder = freeze_encoder
+        super().__init__(*args, **kwargs)
+
+    def forward(self, x: torch.FloatTensor) -> CausalAutoencoderOutput:
+        with torch.no_grad() if self.freeze_encoder else nullcontext():
+            z, p = self.encode(x)
+        x = self.decode(z).sample
+        return CausalAutoencoderOutput(x, z, p)
+
+    def encode(self, x: torch.FloatTensor) -> CausalEncoderOutput:
+        if x.ndim == 4:
+            x = x.unsqueeze(2)
+        p = super().encode(x).latent_dist
+        z = p.sample().squeeze(2)
+        return CausalEncoderOutput(z, p)
+
+    def decode(self, z: torch.FloatTensor) -> CausalDecoderOutput:
+        if z.ndim == 4:
+            z = z.unsqueeze(2)
+        x = super().decode(z).sample.squeeze(2)
+        return CausalDecoderOutput(x)
+
+    def preprocess(self, x: torch.Tensor):
+        # x should in [B, C, T, H, W], [B, C, H, W]
+        assert x.ndim == 4 or x.size(2) % 4 == 1
+        return x
+
+    def postprocess(self, x: torch.Tensor):
+        # x should in [B, C, T, H, W], [B, C, H, W]
+        return x
+
+    def set_causal_slicing(
+        self,
+        *,
+        split_size: Optional[int],
+        memory_device: _memory_device_t,
+    ):
+        assert (
+            split_size is None or memory_device is not None
+        ), "if split_size is set, memory_device must not be None."
+        if split_size is not None:
+            self.enable_slicing()
+            self.slicing_sample_min_size = split_size
+            self.slicing_latent_min_size = split_size // self.temporal_downsample_factor
+        else:
+            self.disable_slicing()
+        for module in self.modules():
+            if isinstance(module, InflatedCausalConv3d):
+                module.set_memory_device(memory_device)
+
+    def set_memory_limit(self, conv_max_mem: Optional[float], norm_max_mem: Optional[float]):
+        set_norm_limit(norm_max_mem)
+        for m in self.modules():
+            if isinstance(m, InflatedCausalConv3d):
+                m.set_memory_limit(conv_max_mem if conv_max_mem is not None else float("inf"))

models/video_vae_v3/modules/causal_inflation_lib.py

@@ -0,0 +1,460 @@
+# // 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
+from contextlib import contextmanager
+from typing import List, Optional, Union
+import torch
+import torch.distributed as dist
+import torch.nn.functional as F
+from diffusers.models.normalization import RMSNorm
+from einops import rearrange
+from torch import Tensor, nn
+from torch.nn import Conv3d
+
+from common.distributed.advanced import (
+    get_next_sequence_parallel_rank,
+    get_prev_sequence_parallel_rank,
+    get_sequence_parallel_group,
+    get_sequence_parallel_rank,
+    get_sequence_parallel_world_size,
+)
+from common.logger import get_logger
+from models.video_vae_v3.modules.context_parallel_lib import cache_send_recv, get_cache_size
+from models.video_vae_v3.modules.global_config import get_norm_limit
+from models.video_vae_v3.modules.types import MemoryState, _inflation_mode_t, _memory_device_t
+
+logger = get_logger(__name__)
+
+
+@contextmanager
+def ignore_padding(model):
+    orig_padding = model.padding
+    model.padding = (0, 0, 0)
+    try:
+        yield
+    finally:
+        model.padding = orig_padding
+
+
+class InflatedCausalConv3d(Conv3d):
+    def __init__(
+        self,
+        *args,
+        inflation_mode: _inflation_mode_t,
+        memory_device: _memory_device_t = "same",
+        **kwargs,
+    ):
+        self.inflation_mode = inflation_mode
+        self.memory = None
+        super().__init__(*args, **kwargs)
+        self.temporal_padding = self.padding[0]
+        self.memory_device = memory_device
+        self.padding = (0, *self.padding[1:])  # Remove temporal pad to keep causal.
+        self.memory_limit = float("inf")
+
+    def set_memory_limit(self, value: float):
+        self.memory_limit = value
+
+    def set_memory_device(self, memory_device: _memory_device_t):
+        self.memory_device = memory_device
+
+    def memory_limit_conv(
+        self,
+        x,
+        *,
+        split_dim=3,
+        padding=(0, 0, 0, 0, 0, 0),
+        prev_cache=None,
+    ):
+        # Compatible with no limit.
+        if math.isinf(self.memory_limit):
+            if prev_cache is not None:
+                x = torch.cat([prev_cache, x], dim=split_dim - 1)
+            return super().forward(x)
+
+        # Compute tensor shape after concat & padding.
+        shape = torch.tensor(x.size())
+        if prev_cache is not None:
+            shape[split_dim - 1] += prev_cache.size(split_dim - 1)
+        shape[-3:] += torch.tensor(padding).view(3, 2).sum(-1).flip(0)
+        memory_occupy = shape.prod() * x.element_size() / 1024**3  # GiB
+        logger.debug(
+            f"x:{(shape, x.dtype)} {memory_occupy:.3f}GiB "
+            f"prev_cache:{prev_cache.shape if prev_cache is not None else None}"
+        )
+        if memory_occupy < self.memory_limit or split_dim == x.ndim:
+            if prev_cache is not None:
+                x = torch.cat([prev_cache, x], dim=split_dim - 1)
+            x = F.pad(x, padding, value=0.0)
+            with ignore_padding(self):
+                return super().forward(x)
+
+        logger.debug(
+            f"Exceed memory limit {memory_occupy} > {self.memory_limit}, split dim {split_dim}"
+        )
+
+        # Split input (& prev_cache).
+        num_splits = math.ceil(memory_occupy / self.memory_limit)
+        size_per_split = x.size(split_dim) // num_splits
+        split_sizes = [size_per_split] * (num_splits - 1)
+        split_sizes += [x.size(split_dim) - sum(split_sizes)]
+
+        x = list(x.split(split_sizes, dim=split_dim))
+        logger.debug(f"Conv inputs: {[inp.size() for inp in x]} {x[0].dtype}")
+        if prev_cache is not None:
+            prev_cache = list(prev_cache.split(split_sizes, dim=split_dim))
+
+        # Loop Fwd.
+        cache = None
+        for idx in range(len(x)):
+            # Concat prev cache from last dim
+            if prev_cache is not None:
+                x[idx] = torch.cat([prev_cache[idx], x[idx]], dim=split_dim - 1)
+
+            # Get padding pattern.
+            lpad_dim = (x[idx].ndim - split_dim - 1) * 2
+            rpad_dim = lpad_dim + 1
+            padding = list(padding)
+            padding[lpad_dim] = self.padding[split_dim - 2] if idx == 0 else 0
+            padding[rpad_dim] = self.padding[split_dim - 2] if idx == len(x) - 1 else 0
+            pad_len = padding[lpad_dim] + padding[rpad_dim]
+            padding = tuple(padding)
+
+            # Prepare cache for next slice (this dim).
+            next_cache = None
+            cache_len = cache.size(split_dim) if cache is not None else 0
+            next_catch_size = get_cache_size(
+                conv_module=self,
+                input_len=x[idx].size(split_dim) + cache_len,
+                pad_len=pad_len,
+                dim=split_dim - 2,
+            )
+            if next_catch_size != 0:
+                assert next_catch_size <= x[idx].size(split_dim)
+                next_cache = (
+                    x[idx].transpose(0, split_dim)[-next_catch_size:].transpose(0, split_dim)
+                )
+
+            # Recursive.
+            x[idx] = self.memory_limit_conv(
+                x[idx],
+                split_dim=split_dim + 1,
+                padding=padding,
+                prev_cache=cache,
+            )
+
+            # Update cache.
+            cache = next_cache
+
+        logger.debug(f"Conv outputs, concat(dim={split_dim}): {[d.size() for d in x]}")
+        return torch.cat(x, split_dim)
+
+    def forward(
+        self,
+        input: Union[Tensor, List[Tensor]],
+        memory_state: MemoryState = MemoryState.UNSET,
+    ) -> Tensor:
+        assert memory_state != MemoryState.UNSET
+        if memory_state != MemoryState.ACTIVE:
+            self.memory = None
+        if (
+            math.isinf(self.memory_limit)
+            and torch.is_tensor(input)
+            and get_sequence_parallel_group() is None
+        ):
+            return self.basic_forward(input, memory_state)
+        return self.slicing_forward(input, memory_state)
+
+    def basic_forward(self, input: Tensor, memory_state: MemoryState = MemoryState.UNSET):
+        mem_size = self.stride[0] - self.kernel_size[0]
+        if (self.memory is not None) and (memory_state == MemoryState.ACTIVE):
+            input = extend_head(input, memory=self.memory, times=-1)
+        else:
+            input = extend_head(input, times=self.temporal_padding * 2)
+        memory = (
+            input[:, :, mem_size:].detach()
+            if (mem_size != 0 and memory_state != MemoryState.DISABLED)
+            else None
+        )
+        if (
+            memory_state != MemoryState.DISABLED
+            and not self.training
+            and (self.memory_device is not None)
+        ):
+            self.memory = memory
+            if self.memory_device == "cpu" and self.memory is not None:
+                self.memory = self.memory.to("cpu")
+        return super().forward(input)
+
+    def slicing_forward(
+        self,
+        input: Union[Tensor, List[Tensor]],
+        memory_state: MemoryState = MemoryState.UNSET,
+    ) -> Tensor:
+        squeeze_out = False
+        if torch.is_tensor(input):
+            input = [input]
+            squeeze_out = True
+
+        cache_size = self.kernel_size[0] - self.stride[0]
+        cache = cache_send_recv(
+            input, cache_size=cache_size, memory=self.memory, times=self.temporal_padding * 2
+        )
+
+        # For slice=4 and sp=2, and 17 frames in total
+        #                  sp0                  sp1
+        # slice 0: [`0 0` 0 1 2 {3 4}]   [{3 4} 5 6 (7 8)]    extend=`0 0` cache={3 4} memory=(7 8)
+        # slice 1: [(7 8) 9 10 {11 12}]  [{11 12} 13 14 15 16]
+        sp_rank = get_sequence_parallel_rank()
+        sp_size = get_sequence_parallel_world_size()
+        sp_group = get_sequence_parallel_group()
+        send_dst = get_next_sequence_parallel_rank()
+        recv_src = get_prev_sequence_parallel_rank()
+        if (
+            memory_state in [MemoryState.INITIALIZING, MemoryState.ACTIVE]  # use_slicing
+            and not self.training
+            and (self.memory_device is not None)
+            and sp_rank in [0, sp_size - 1]
+            and cache_size != 0
+        ):
+            if cache_size > input[-1].size(2) and cache is not None and len(input) == 1:
+                input[0] = torch.cat([cache, input[0]], dim=2)
+                cache = None
+            assert cache_size <= input[-1].size(2)
+            if sp_size == 1:
+                self.memory = input[-1][:, :, -cache_size:].detach().contiguous()
+            else:
+                if sp_rank == sp_size - 1:
+                    dist.send(
+                        input[-1][:, :, -cache_size:].detach().contiguous(),
+                        send_dst,
+                        group=sp_group,
+                    )
+                if sp_rank == 0:
+                    shape = list(input[0].size())
+                    shape[2] = cache_size
+                    self.memory = torch.empty(
+                        *shape, device=input[0].device, dtype=input[0].dtype
+                    ).contiguous()
+                    dist.recv(self.memory, recv_src, group=sp_group)
+            if self.memory_device == "cpu" and self.memory is not None:
+                self.memory = self.memory.to("cpu")
+
+        padding = tuple(x for x in reversed(self.padding) for _ in range(2))
+        for i in range(len(input)):
+            # Prepare cache for next input slice.
+            next_cache = None
+            cache_size = 0
+            if i < len(input) - 1:
+                cache_len = cache.size(2) if cache is not None else 0
+                cache_size = get_cache_size(self, input[i].size(2) + cache_len, pad_len=0)
+            if cache_size != 0:
+                if cache_size > input[i].size(2) and cache is not None:
+                    input[i] = torch.cat([cache, input[i]], dim=2)
+                    cache = None
+                assert cache_size <= input[i].size(2), f"{cache_size} > {input[i].size(2)}"
+                next_cache = input[i][:, :, -cache_size:]
+
+            # Conv forward for this input slice.
+            input[i] = self.memory_limit_conv(
+                input[i],
+                padding=padding,
+                prev_cache=cache,
+            )
+
+            # Update cache.
+            cache = next_cache
+
+        return input[0] if squeeze_out else input
+
+    def tflops(self, args, kwargs, output) -> float:
+        if torch.is_tensor(output):
+            output_numel = output.numel()
+        elif isinstance(output, list):
+            output_numel = sum(o.numel() for o in output)
+        else:
+            raise NotImplementedError
+        return (2 * math.prod(self.kernel_size) * self.in_channels * (output_numel / 1e6)) / 1e6
+
+    def _load_from_state_dict(
+        self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
+    ):
+        if self.inflation_mode != "none":
+            state_dict = modify_state_dict(
+                self,
+                state_dict,
+                prefix,
+                inflate_weight_fn=inflate_weight,
+                inflate_bias_fn=inflate_bias,
+            )
+        super()._load_from_state_dict(
+            state_dict,
+            prefix,
+            local_metadata,
+            (strict and self.inflation_mode == "none"),
+            missing_keys,
+            unexpected_keys,
+            error_msgs,
+        )
+
+
+def init_causal_conv3d(
+    *args,
+    inflation_mode: _inflation_mode_t,
+    **kwargs,
+):
+    """
+    Initialize a Causal-3D convolution layer.
+    Parameters:
+        inflation_mode: Listed as below. It's compatible with all the 3D-VAE checkpoints we have.
+            - none: No inflation will be conducted.
+                    The loading logic of state dict will fall back to default.
+            - tail / replicate: Refer to the definition of `InflatedCausalConv3d`.
+    """
+    return InflatedCausalConv3d(*args, inflation_mode=inflation_mode, **kwargs)
+
+
+def causal_norm_wrapper(norm_layer: nn.Module, x: torch.Tensor) -> torch.Tensor:
+    input_dtype = x.dtype
+    if isinstance(norm_layer, (nn.LayerNorm, RMSNorm)):
+        if x.ndim == 4:
+            x = rearrange(x, "b c h w -> b h w c")
+            x = norm_layer(x)
+            x = rearrange(x, "b h w c -> b c h w")
+            return x.to(input_dtype)
+        if x.ndim == 5:
+            x = rearrange(x, "b c t h w -> b t h w c")
+            x = norm_layer(x)
+            x = rearrange(x, "b t h w c -> b c t h w")
+            return x.to(input_dtype)
+    if isinstance(norm_layer, (nn.GroupNorm, nn.BatchNorm2d, nn.SyncBatchNorm)):
+        if x.ndim <= 4:
+            return norm_layer(x).to(input_dtype)
+        if x.ndim == 5:
+            t = x.size(2)
+            x = rearrange(x, "b c t h w -> (b t) c h w")
+            memory_occupy = x.numel() * x.element_size() / 1024**3
+            if isinstance(norm_layer, nn.GroupNorm) and memory_occupy > get_norm_limit():
+                num_chunks = min(4 if x.element_size() == 2 else 2, norm_layer.num_groups)
+                logger.debug(f"large tensor {x.shape}, norm in {num_chunks} chunks")
+                assert norm_layer.num_groups % num_chunks == 0
+                num_groups_per_chunk = norm_layer.num_groups // num_chunks
+
+                x = list(x.chunk(num_chunks, dim=1))
+                weights = norm_layer.weight.chunk(num_chunks, dim=0)
+                biases = norm_layer.bias.chunk(num_chunks, dim=0)
+                for i, (w, b) in enumerate(zip(weights, biases)):
+                    x[i] = F.group_norm(x[i], num_groups_per_chunk, w, b, norm_layer.eps)
+                    x[i] = x[i].to(input_dtype)
+                x = torch.cat(x, dim=1)
+            else:
+                x = norm_layer(x)
+            x = rearrange(x, "(b t) c h w -> b c t h w", t=t)
+            return x.to(input_dtype)
+    raise NotImplementedError
+
+
+def remove_head(tensor: Tensor, times: int = 1) -> Tensor:
+    """
+    Remove duplicated first frame features in the up-sampling process.
+    """
+    sp_rank = get_sequence_parallel_rank()
+    if times == 0 or sp_rank > 0:
+        return tensor
+    return torch.cat(tensors=(tensor[:, :, :1], tensor[:, :, times + 1 :]), dim=2)
+
+
+def extend_head(tensor: Tensor, times: int = 2, memory: Optional[Tensor] = None) -> Tensor:
+    """
+    When memory is None:
+        - Duplicate first frame features in the down-sampling process.
+    When memory is not None:
+        - Concatenate memory features with the input features to keep temporal consistency.
+    """
+    if memory is not None:
+        return torch.cat((memory.to(tensor), tensor), dim=2)
+    assert times >= 0, "Invalid input for function 'extend_head'!"
+    if times == 0:
+        return tensor
+    else:
+        tile_repeat = [1] * tensor.ndim
+        tile_repeat[2] = times
+        return torch.cat(tensors=(torch.tile(tensor[:, :, :1], tile_repeat), tensor), dim=2)
+
+
+def inflate_weight(weight_2d: torch.Tensor, weight_3d: torch.Tensor, inflation_mode: str):
+    """
+    Inflate a 2D convolution weight matrix to a 3D one.
+    Parameters:
+        weight_2d:      The weight matrix of 2D conv to be inflated.
+        weight_3d:      The weight matrix of 3D conv to be initialized.
+        inflation_mode: the mode of inflation
+    """
+    assert inflation_mode in ["tail", "replicate"]
+    assert weight_3d.shape[:2] == weight_2d.shape[:2]
+    with torch.no_grad():
+        if inflation_mode == "replicate":
+            depth = weight_3d.size(2)
+            weight_3d.copy_(weight_2d.unsqueeze(2).repeat(1, 1, depth, 1, 1) / depth)
+        else:
+            weight_3d.fill_(0.0)
+            weight_3d[:, :, -1].copy_(weight_2d)
+    return weight_3d
+
+
+def inflate_bias(bias_2d: torch.Tensor, bias_3d: torch.Tensor, inflation_mode: str):
+    """
+    Inflate a 2D convolution bias tensor to a 3D one
+    Parameters:
+        bias_2d:        The bias tensor of 2D conv to be inflated.
+        bias_3d:        The bias tensor of 3D conv to be initialized.
+        inflation_mode: Placeholder to align `inflate_weight`.
+    """
+    assert bias_3d.shape == bias_2d.shape
+    with torch.no_grad():
+        bias_3d.copy_(bias_2d)
+    return bias_3d
+
+
+def modify_state_dict(layer, state_dict, prefix, inflate_weight_fn, inflate_bias_fn):
+    """
+    the main function to inflated 2D parameters to 3D.
+    """
+    weight_name = prefix + "weight"
+    bias_name = prefix + "bias"
+    if weight_name in state_dict:
+        weight_2d = state_dict[weight_name]
+        if weight_2d.dim() == 4:
+            # Assuming the 2D weights are 4D tensors (out_channels, in_channels, h, w)
+            weight_3d = inflate_weight_fn(
+                weight_2d=weight_2d,
+                weight_3d=layer.weight,
+                inflation_mode=layer.inflation_mode,
+            )
+            state_dict[weight_name] = weight_3d
+        else:
+            return state_dict
+            # It's a 3d state dict, should not do inflation on both bias and weight.
+    if bias_name in state_dict:
+        bias_2d = state_dict[bias_name]
+        if bias_2d.dim() == 1:
+            # Assuming the 2D biases are 1D tensors (out_channels,)
+            bias_3d = inflate_bias_fn(
+                bias_2d=bias_2d,
+                bias_3d=layer.bias,
+                inflation_mode=layer.inflation_mode,
+            )
+            state_dict[bias_name] = bias_3d
+    return state_dict

models/video_vae_v3/modules/context_parallel_lib.py

@@ -0,0 +1,164 @@
+# // 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 List
+import torch
+import torch.distributed as dist
+import torch.nn.functional as F
+from torch import Tensor
+
+from common.distributed import get_device
+from common.distributed.advanced import (
+    get_next_sequence_parallel_rank,
+    get_prev_sequence_parallel_rank,
+    get_sequence_parallel_group,
+    get_sequence_parallel_rank,
+    get_sequence_parallel_world_size,
+)
+from common.distributed.ops import Gather
+from common.logger import get_logger
+from models.video_vae_v3.modules.types import MemoryState
+
+logger = get_logger(__name__)
+
+
+def causal_conv_slice_inputs(x, split_size, memory_state):
+    sp_size = get_sequence_parallel_world_size()
+    sp_group = get_sequence_parallel_group()
+    sp_rank = get_sequence_parallel_rank()
+    if sp_group is None:
+        return x
+
+    assert memory_state != MemoryState.UNSET
+    leave_out = 1 if memory_state != MemoryState.ACTIVE else 0
+
+    # Should have at least sp_size slices.
+    num_slices = (x.size(2) - leave_out) // split_size
+    assert num_slices >= sp_size, f"{num_slices} < {sp_size}"
+
+    split_sizes = [split_size + leave_out] + [split_size] * (num_slices - 1)
+    split_sizes += [x.size(2) - sum(split_sizes)]
+    assert sum(split_sizes) == x.size(2)
+
+    split_sizes = torch.tensor(split_sizes)
+    slices_per_rank = len(split_sizes) // sp_size
+    split_sizes = split_sizes.split(
+        [slices_per_rank] * (sp_size - 1) + [len(split_sizes) - slices_per_rank * (sp_size - 1)]
+    )
+    split_sizes = list(map(lambda s: s.sum().item(), split_sizes))
+    logger.debug(f"split_sizes: {split_sizes}")
+    return x.split(split_sizes, dim=2)[sp_rank]
+
+
+def causal_conv_gather_outputs(x):
+    sp_group = get_sequence_parallel_group()
+    sp_size = get_sequence_parallel_world_size()
+    if sp_group is None:
+        return x
+
+    # Communicate shapes.
+    unpad_lens = torch.empty((sp_size,), device=get_device(), dtype=torch.long)
+    local_unpad_len = torch.tensor([x.size(2)], device=get_device(), dtype=torch.long)
+    torch.distributed.all_gather_into_tensor(unpad_lens, local_unpad_len, group=sp_group)
+
+    # Padding to max_len for gather.
+    max_len = unpad_lens.max()
+    x_pad = F.pad(x, (0, 0, 0, 0, 0, max_len - x.size(2))).contiguous()
+
+    # Gather outputs.
+    x_pad = Gather.apply(sp_group, x_pad, 2, True)
+
+    # Remove padding.
+    x_pad_lists = list(x_pad.chunk(sp_size, dim=2))
+    for i, (x_pad, unpad_len) in enumerate(zip(x_pad_lists, unpad_lens)):
+        x_pad_lists[i] = x_pad[:, :, :unpad_len]
+
+    return torch.cat(x_pad_lists, dim=2)
+
+
+def get_output_len(conv_module, input_len, pad_len, dim=0):
+    dilated_kernerl_size = conv_module.dilation[dim] * (conv_module.kernel_size[dim] - 1) + 1
+    output_len = (input_len + pad_len - dilated_kernerl_size) // conv_module.stride[dim] + 1
+    return output_len
+
+
+def get_cache_size(conv_module, input_len, pad_len, dim=0):
+    dilated_kernerl_size = conv_module.dilation[dim] * (conv_module.kernel_size[dim] - 1) + 1
+    output_len = (input_len + pad_len - dilated_kernerl_size) // conv_module.stride[dim] + 1
+    remain_len = (
+        input_len + pad_len - ((output_len - 1) * conv_module.stride[dim] + dilated_kernerl_size)
+    )
+    overlap_len = dilated_kernerl_size - conv_module.stride[dim]
+    cache_len = overlap_len + remain_len  # >= 0
+    logger.debug(
+        f"I:{input_len}, "
+        f"P:{pad_len}, "
+        f"K:{conv_module.kernel_size[dim]}, "
+        f"S:{conv_module.stride[dim]}, "
+        f"O:{output_len}, "
+        f"Cache:{cache_len}"
+    )
+    assert output_len > 0
+    return cache_len
+
+
+def cache_send_recv(tensor: List[Tensor], cache_size, times, memory=None):
+    sp_group = get_sequence_parallel_group()
+    sp_rank = get_sequence_parallel_rank()
+    sp_size = get_sequence_parallel_world_size()
+    send_dst = get_next_sequence_parallel_rank()
+    recv_src = get_prev_sequence_parallel_rank()
+    recv_buffer = None
+    recv_req = None
+
+    logger.debug(
+        f"[sp{sp_rank}] cur_tensors:{[(t.size(), t.dtype) for t in tensor]}, times: {times}"
+    )
+    if sp_rank == 0 or sp_group is None:
+        if memory is not None:
+            recv_buffer = memory.to(tensor[0])
+        elif times > 0:
+            tile_repeat = [1] * tensor[0].ndim
+            tile_repeat[2] = times
+            recv_buffer = torch.tile(tensor[0][:, :, :1], tile_repeat)
+
+    if cache_size != 0 and sp_group is not None:
+        if sp_rank > 0:
+            shape = list(tensor[0].size())
+            shape[2] = cache_size
+            recv_buffer = torch.empty(
+                *shape, device=tensor[0].device, dtype=tensor[0].dtype
+            ).contiguous()
+            recv_req = dist.irecv(recv_buffer, recv_src, group=sp_group)
+        if sp_rank < sp_size - 1:
+            if cache_size > tensor[-1].size(2) and len(tensor) == 1:
+                logger.debug(f"[sp{sp_rank}] force concat before send {tensor[-1].size()}")
+                if recv_req is not None:
+                    recv_req.wait()
+                tensor[0] = torch.cat([recv_buffer, tensor[0]], dim=2)
+                recv_buffer = None
+            assert cache_size <= tensor[-1].size(
+                2
+            ), f"Not enough value to cache, got {tensor[-1].size()}, cache_size={cache_size}"
+            dist.isend(
+                tensor[-1][:, :, -cache_size:].detach().contiguous(), send_dst, group=sp_group
+            )
+        if recv_req is not None:
+            recv_req.wait()
+
+    logger.debug(
+        f"[sp{sp_rank}] recv_src:{recv_src}, "
+        f"recv_buffer:{recv_buffer.size() if recv_buffer is not None else None}"
+    )
+    return recv_buffer