Duplicate from THUDM/CogVideo

Co-authored-by: Ahsen Khaliq <akhaliq@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,27 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zstandard filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,133 @@
+gradio_queue.db*
+pretrained/*
+icetk_models/*
+!*/.gitkeep
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "CogVideo"]
+	path = CogVideo
+	url = https://github.com/THUDM/CogVideo

.pre-commit-config.yaml

@@ -0,0 +1,46 @@
+exclude: ^patch
+repos:
+- repo: https://github.com/pre-commit/pre-commit-hooks
+  rev: v4.2.0
+  hooks:
+  - id: check-executables-have-shebangs
+  - id: check-json
+  - id: check-merge-conflict
+  - id: check-shebang-scripts-are-executable
+  - id: check-toml
+  - id: check-yaml
+  - id: double-quote-string-fixer
+  - id: end-of-file-fixer
+  - id: mixed-line-ending
+    args: ['--fix=lf']
+  - id: requirements-txt-fixer
+  - id: trailing-whitespace
+- repo: https://github.com/myint/docformatter
+  rev: v1.4
+  hooks:
+  - id: docformatter
+    args: ['--in-place']
+- repo: https://github.com/pycqa/isort
+  rev: 5.10.1
+  hooks:
+    - id: isort
+- repo: https://github.com/pre-commit/mirrors-mypy
+  rev: v0.812
+  hooks:
+    - id: mypy
+      args: ['--ignore-missing-imports']
+- repo: https://github.com/google/yapf
+  rev: v0.32.0
+  hooks:
+  - id: yapf
+    args: ['--parallel', '--in-place']
+- repo: https://github.com/kynan/nbstripout
+  rev: 0.5.0
+  hooks:
+    - id: nbstripout
+      args: ['--extra-keys', 'metadata.interpreter metadata.kernelspec cell.metadata.pycharm']
+- repo: https://github.com/nbQA-dev/nbQA
+  rev: 1.3.1
+  hooks:
+    - id: nbqa-isort
+    - id: nbqa-yapf

.style.yapf

@@ -0,0 +1,5 @@
+[style]
+based_on_style = pep8
+blank_line_before_nested_class_or_def = false
+spaces_before_comment = 2
+split_before_logical_operator = true

CogVideo

@@ -0,0 +1 @@
+Subproject commit ff423aa169978fb2f636f761e348631fa3178b03

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 hysts
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

LICENSE.CogVideo

@@ -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 [yyyy] [name of copyright owner]
+
+   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

@@ -0,0 +1,14 @@
+---
+title: CogVideo
+emoji: 🌍
+colorFrom: indigo
+colorTo: yellow
+sdk: gradio
+sdk_version: 3.1.6
+python_version: 3.9.13
+app_file: app.py
+pinned: false
+duplicated_from: THUDM/CogVideo
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,80 @@
+#!/usr/bin/env python
+
+from __future__ import annotations
+
+import gradio as gr
+
+from model import AppModel
+
+DESCRIPTION = '''# <a href="https://github.com/THUDM/CogVideo">CogVideo</a>
+
+Currently, this Space only supports the first stage of the CogVideo pipeline due to hardware limitations.
+
+The model accepts only Chinese as input.
+By checking the "Translate to Chinese" checkbox, the results of English to Chinese translation with [this Space](https://huggingface.co/spaces/chinhon/translation_eng2ch) will be used as input.
+Since the translation model may mistranslate, you may want to use the translation results from other translation services.
+'''
+NOTES = 'This app is adapted from <a href="https://github.com/hysts/CogVideo_demo">https://github.com/hysts/CogVideo_demo</a>. It would be recommended to use the repo if you want to run the app yourself.'
+FOOTER = '<img id="visitor-badge" alt="visitor badge" src="https://visitor-badge.glitch.me/badge?page_id=THUDM.CogVideo" />'
+
+
+def main():
+    only_first_stage = True
+    model = AppModel(only_first_stage)
+
+    with gr.Blocks(css='style.css') as demo:
+        gr.Markdown(DESCRIPTION)
+
+        with gr.Row():
+            with gr.Column():
+                with gr.Group():
+                    text = gr.Textbox(label='Input Text')
+                    translate = gr.Checkbox(label='Translate to Chinese',
+                                            value=False)
+                    seed = gr.Slider(0,
+                                     100000,
+                                     step=1,
+                                     value=1234,
+                                     label='Seed')
+                    only_first_stage = gr.Checkbox(
+                        label='Only First Stage',
+                        value=only_first_stage,
+                        visible=not only_first_stage)
+                    image_prompt = gr.Image(type="filepath",
+                                            label="Image Prompt",
+                                            value=None)
+                    run_button = gr.Button('Run')
+
+            with gr.Column():
+                with gr.Group():
+                    translated_text = gr.Textbox(label='Translated Text')
+                    with gr.Tabs():
+                        with gr.TabItem('Output (Video)'):
+                            result_video = gr.Video(show_label=False)
+
+        examples = gr.Examples(
+            examples=[['骑滑板的皮卡丘', False, 1234, True,None],
+                      ['a cat playing chess', True, 1253, True,None]],
+            fn=model.run_with_translation,
+            inputs=[text, translate, seed, only_first_stage,image_prompt],
+            outputs=[translated_text, result_video],
+            cache_examples=True)
+
+        gr.Markdown(NOTES)
+        gr.Markdown(FOOTER)
+        print(gr.__version__)
+        run_button.click(fn=model.run_with_translation,
+                         inputs=[
+                             text,
+                             translate,
+                             seed,
+                             only_first_stage,
+                             image_prompt
+                         ],
+                         outputs=[translated_text, result_video])
+        print(gr.__version__)
+    demo.launch()
+
+
+if __name__ == '__main__':
+    main()

model.py

@@ -0,0 +1,1243 @@
+# This code is adapted from https://github.com/THUDM/CogVideo/blob/ff423aa169978fb2f636f761e348631fa3178b03/cogvideo_pipeline.py
+
+from __future__ import annotations
+
+import argparse
+import logging
+import os
+import pathlib
+import shutil
+import subprocess
+import sys
+import tempfile
+import time
+import zipfile
+from typing import Any
+
+if os.getenv('SYSTEM') == 'spaces':
+    subprocess.run('pip install icetk==0.0.4'.split())
+    subprocess.run('pip install SwissArmyTransformer==0.2.9'.split())
+    subprocess.run(
+        'pip install git+https://github.com/Sleepychord/Image-Local-Attention@43fee31'
+        .split())
+    #subprocess.run('git clone https://github.com/NVIDIA/apex'.split())
+    #subprocess.run('git checkout 1403c21'.split(), cwd='apex')
+    #with open('patch.apex') as f:
+    #    subprocess.run('patch -p1'.split(), cwd='apex', stdin=f)
+    #subprocess.run(
+    #    'pip install -v --disable-pip-version-check --no-cache-dir --global-option --cpp_ext --global-option --cuda_ext ./'
+    #    .split(),
+    #    cwd='apex')
+    #subprocess.run('rm -rf apex'.split())
+    with open('patch') as f:
+        subprocess.run('patch -p1'.split(), cwd='CogVideo', stdin=f)
+
+    from huggingface_hub import hf_hub_download
+
+    def download_and_extract_icetk_models() -> None:
+        icetk_model_dir = pathlib.Path('/home/user/.icetk_models')
+        icetk_model_dir.mkdir()
+        path = hf_hub_download('THUDM/icetk',
+                               'models.zip',
+                               use_auth_token=os.getenv('HF_TOKEN'))
+        with zipfile.ZipFile(path) as f:
+            f.extractall(path=icetk_model_dir.as_posix())
+
+    def download_and_extract_cogvideo_models(name: str) -> None:
+        path = hf_hub_download('THUDM/CogVideo',
+                               name,
+                               use_auth_token=os.getenv('HF_TOKEN'))
+        with zipfile.ZipFile(path) as f:
+            f.extractall('pretrained')
+        os.remove(path)
+
+    def download_and_extract_cogview2_models(name: str) -> None:
+        path = hf_hub_download('THUDM/CogView2', name)
+        with zipfile.ZipFile(path) as f:
+            f.extractall()
+        shutil.move('/home/user/app/sharefs/cogview-new/cogview2-dsr',
+                    'pretrained')
+        shutil.rmtree('/home/user/app/sharefs/')
+        os.remove(path)
+
+    download_and_extract_icetk_models()
+    download_and_extract_cogvideo_models('cogvideo-stage1.zip')
+    #download_and_extract_cogvideo_models('cogvideo-stage2.zip')
+    #download_and_extract_cogview2_models('cogview2-dsr.zip')
+
+    os.environ['SAT_HOME'] = '/home/user/app/pretrained'
+
+import gradio as gr
+import imageio.v2 as iio
+import numpy as np
+import torch
+from icetk import IceTokenizer
+from SwissArmyTransformer import get_args
+from SwissArmyTransformer.arguments import set_random_seed
+from SwissArmyTransformer.generation.sampling_strategies import BaseStrategy
+from SwissArmyTransformer.resources import auto_create
+
+app_dir = pathlib.Path(__file__).parent
+submodule_dir = app_dir / 'CogVideo'
+sys.path.insert(0, submodule_dir.as_posix())
+
+from coglm_strategy import CoglmStrategy
+from models.cogvideo_cache_model import CogVideoCacheModel
+from sr_pipeline import DirectSuperResolution
+
+formatter = logging.Formatter(
+    '[%(asctime)s] %(name)s %(levelname)s: %(message)s',
+    datefmt='%Y-%m-%d %H:%M:%S')
+stream_handler = logging.StreamHandler(stream=sys.stdout)
+stream_handler.setLevel(logging.INFO)
+stream_handler.setFormatter(formatter)
+logger = logging.getLogger(__name__)
+logger.setLevel(logging.INFO)
+logger.propagate = False
+logger.addHandler(stream_handler)
+
+ICETK_MODEL_DIR = app_dir / 'icetk_models'
+
+
+def get_masks_and_position_ids_stage1(data, textlen, framelen):
+    # Extract batch size and sequence length.
+    tokens = data
+    seq_length = len(data[0])
+    # Attention mask (lower triangular).
+    attention_mask = torch.ones((1, textlen + framelen, textlen + framelen),
+                                device=data.device)
+    attention_mask[:, :textlen, textlen:] = 0
+    attention_mask[:, textlen:, textlen:].tril_()
+    attention_mask.unsqueeze_(1)
+    # Unaligned version
+    position_ids = torch.zeros(seq_length,
+                               dtype=torch.long,
+                               device=data.device)
+    torch.arange(textlen,
+                 out=position_ids[:textlen],
+                 dtype=torch.long,
+                 device=data.device)
+    torch.arange(512,
+                 512 + seq_length - textlen,
+                 out=position_ids[textlen:],
+                 dtype=torch.long,
+                 device=data.device)
+    position_ids = position_ids.unsqueeze(0)
+
+    return tokens, attention_mask, position_ids
+
+
+def get_masks_and_position_ids_stage2(data, textlen, framelen):
+    # Extract batch size and sequence length.
+    tokens = data
+    seq_length = len(data[0])
+
+    # Attention mask (lower triangular).
+    attention_mask = torch.ones((1, textlen + framelen, textlen + framelen),
+                                device=data.device)
+    attention_mask[:, :textlen, textlen:] = 0
+    attention_mask[:, textlen:, textlen:].tril_()
+    attention_mask.unsqueeze_(1)
+
+    # Unaligned version
+    position_ids = torch.zeros(seq_length,
+                               dtype=torch.long,
+                               device=data.device)
+    torch.arange(textlen,
+                 out=position_ids[:textlen],
+                 dtype=torch.long,
+                 device=data.device)
+    frame_num = (seq_length - textlen) // framelen
+    assert frame_num == 5
+    torch.arange(512,
+                 512 + framelen,
+                 out=position_ids[textlen:textlen + framelen],
+                 dtype=torch.long,
+                 device=data.device)
+    torch.arange(512 + framelen * 2,
+                 512 + framelen * 3,
+                 out=position_ids[textlen + framelen:textlen + framelen * 2],
+                 dtype=torch.long,
+                 device=data.device)
+    torch.arange(512 + framelen * (frame_num - 1),
+                 512 + framelen * frame_num,
+                 out=position_ids[textlen + framelen * 2:textlen +
+                                  framelen * 3],
+                 dtype=torch.long,
+                 device=data.device)
+    torch.arange(512 + framelen * 1,
+                 512 + framelen * 2,
+                 out=position_ids[textlen + framelen * 3:textlen +
+                                  framelen * 4],
+                 dtype=torch.long,
+                 device=data.device)
+    torch.arange(512 + framelen * 3,
+                 512 + framelen * 4,
+                 out=position_ids[textlen + framelen * 4:textlen +
+                                  framelen * 5],
+                 dtype=torch.long,
+                 device=data.device)
+
+    position_ids = position_ids.unsqueeze(0)
+
+    return tokens, attention_mask, position_ids
+
+
+def my_update_mems(hiddens, mems_buffers, mems_indexs,
+                   limited_spatial_channel_mem, text_len, frame_len):
+    if hiddens is None:
+        return None, mems_indexs
+    mem_num = len(hiddens)
+    ret_mem = []
+    with torch.no_grad():
+        for id in range(mem_num):
+            if hiddens[id][0] is None:
+                ret_mem.append(None)
+            else:
+                if id == 0 and limited_spatial_channel_mem and mems_indexs[
+                        id] + hiddens[0][0].shape[1] >= text_len + frame_len:
+                    if mems_indexs[id] == 0:
+                        for layer, hidden in enumerate(hiddens[id]):
+                            mems_buffers[id][
+                                layer, :, :text_len] = hidden.expand(
+                                    mems_buffers[id].shape[1], -1,
+                                    -1)[:, :text_len]
+                    new_mem_len_part2 = (mems_indexs[id] +
+                                         hiddens[0][0].shape[1] -
+                                         text_len) % frame_len
+                    if new_mem_len_part2 > 0:
+                        for layer, hidden in enumerate(hiddens[id]):
+                            mems_buffers[id][
+                                layer, :, text_len:text_len +
+                                new_mem_len_part2] = hidden.expand(
+                                    mems_buffers[id].shape[1], -1,
+                                    -1)[:, -new_mem_len_part2:]
+                    mems_indexs[id] = text_len + new_mem_len_part2
+                else:
+                    for layer, hidden in enumerate(hiddens[id]):
+                        mems_buffers[id][layer, :,
+                                         mems_indexs[id]:mems_indexs[id] +
+                                         hidden.shape[1]] = hidden.expand(
+                                             mems_buffers[id].shape[1], -1, -1)
+                    mems_indexs[id] += hidden.shape[1]
+                ret_mem.append(mems_buffers[id][:, :, :mems_indexs[id]])
+    return ret_mem, mems_indexs
+
+
+def calc_next_tokens_frame_begin_id(text_len, frame_len, total_len):
+    # The fisrt token's position id of the frame that the next token belongs to;
+    if total_len < text_len:
+        return None
+    return (total_len - text_len) // frame_len * frame_len + text_len
+
+
+def my_filling_sequence(
+        model,
+        tokenizer,
+        args,
+        seq,
+        batch_size,
+        get_masks_and_position_ids,
+        text_len,
+        frame_len,
+        strategy=BaseStrategy(),
+        strategy2=BaseStrategy(),
+        mems=None,
+        log_text_attention_weights=0,  # default to 0: no artificial change
+        mode_stage1=True,
+        enforce_no_swin=False,
+        guider_seq=None,
+        guider_text_len=0,
+        guidance_alpha=1,
+        limited_spatial_channel_mem=False,  # 空间通道的存储限制在本帧内
+        **kw_args):
+    '''
+        seq: [2, 3, 5, ..., -1(to be generated), -1, ...]
+        mems: [num_layers, batch_size, len_mems(index), mem_hidden_size]
+            cache, should be first mems.shape[1] parts of context_tokens.
+            mems are the first-level citizens here, but we don't assume what is memorized.
+            input mems are used when multi-phase generation.
+    '''
+    if guider_seq is not None:
+        logger.debug('Using Guidance In Inference')
+    if limited_spatial_channel_mem:
+        logger.debug("Limit spatial-channel's mem to current frame")
+    assert len(seq.shape) == 2
+
+    # building the initial tokens, attention_mask, and position_ids
+    actual_context_length = 0
+
+    while seq[-1][
+            actual_context_length] >= 0:  # the last seq has least given tokens
+        actual_context_length += 1  # [0, context_length-1] are given
+    assert actual_context_length > 0
+    current_frame_num = (actual_context_length - text_len) // frame_len
+    assert current_frame_num >= 0
+    context_length = text_len + current_frame_num * frame_len
+
+    tokens, attention_mask, position_ids = get_masks_and_position_ids(
+        seq, text_len, frame_len)
+    tokens = tokens[..., :context_length]
+    input_tokens = tokens.clone()
+
+    if guider_seq is not None:
+        guider_index_delta = text_len - guider_text_len
+        guider_tokens, guider_attention_mask, guider_position_ids = get_masks_and_position_ids(
+            guider_seq, guider_text_len, frame_len)
+        guider_tokens = guider_tokens[..., :context_length -
+                                      guider_index_delta]
+        guider_input_tokens = guider_tokens.clone()
+
+    for fid in range(current_frame_num):
+        input_tokens[:, text_len + 400 * fid] = tokenizer['<start_of_image>']
+        if guider_seq is not None:
+            guider_input_tokens[:, guider_text_len +
+                                400 * fid] = tokenizer['<start_of_image>']
+
+    attention_mask = attention_mask.type_as(next(
+        model.parameters()))  # if fp16
+    # initialize generation
+    counter = context_length - 1  # Last fixed index is ``counter''
+    index = 0  # Next forward starting index, also the length of cache.
+    mems_buffers_on_GPU = False
+    mems_indexs = [0, 0]
+    mems_len = [(400 + 74) if limited_spatial_channel_mem else 5 * 400 + 74,
+                5 * 400 + 74]
+    mems_buffers = [
+        torch.zeros(args.num_layers,
+                    batch_size,
+                    mem_len,
+                    args.hidden_size * 2,
+                    dtype=next(model.parameters()).dtype)
+        for mem_len in mems_len
+    ]
+
+    if guider_seq is not None:
+        guider_attention_mask = guider_attention_mask.type_as(
+            next(model.parameters()))  # if fp16
+        guider_mems_buffers = [
+            torch.zeros(args.num_layers,
+                        batch_size,
+                        mem_len,
+                        args.hidden_size * 2,
+                        dtype=next(model.parameters()).dtype)
+            for mem_len in mems_len
+        ]
+        guider_mems_indexs = [0, 0]
+        guider_mems = None
+
+    torch.cuda.empty_cache()
+    # step-by-step generation
+    while counter < len(seq[0]) - 1:
+        # we have generated counter+1 tokens
+        # Now, we want to generate seq[counter + 1],
+        # token[:, index: counter+1] needs forwarding.
+        if index == 0:
+            group_size = 2 if (input_tokens.shape[0] == batch_size
+                               and not mode_stage1) else batch_size
+
+            logits_all = None
+            for batch_idx in range(0, input_tokens.shape[0], group_size):
+                logits, *output_per_layers = model(
+                    input_tokens[batch_idx:batch_idx + group_size, index:],
+                    position_ids[..., index:counter + 1],
+                    attention_mask,  # TODO memlen
+                    mems=mems,
+                    text_len=text_len,
+                    frame_len=frame_len,
+                    counter=counter,
+                    log_text_attention_weights=log_text_attention_weights,
+                    enforce_no_swin=enforce_no_swin,
+                    **kw_args)
+                logits_all = torch.cat(
+                    (logits_all,
+                     logits), dim=0) if logits_all is not None else logits
+                mem_kv01 = [[o['mem_kv'][0] for o in output_per_layers],
+                            [o['mem_kv'][1] for o in output_per_layers]]
+                next_tokens_frame_begin_id = calc_next_tokens_frame_begin_id(
+                    text_len, frame_len, mem_kv01[0][0].shape[1])
+                for id, mem_kv in enumerate(mem_kv01):
+                    for layer, mem_kv_perlayer in enumerate(mem_kv):
+                        if limited_spatial_channel_mem and id == 0:
+                            mems_buffers[id][
+                                layer, batch_idx:batch_idx + group_size, :
+                                text_len] = mem_kv_perlayer.expand(
+                                    min(group_size,
+                                        input_tokens.shape[0] - batch_idx), -1,
+                                    -1)[:, :text_len]
+                            mems_buffers[id][layer, batch_idx:batch_idx+group_size, text_len:text_len+mem_kv_perlayer.shape[1]-next_tokens_frame_begin_id] =\
+                                mem_kv_perlayer.expand(min(group_size, input_tokens.shape[0]-batch_idx), -1, -1)[:, next_tokens_frame_begin_id:]
+                        else:
+                            mems_buffers[id][
+                                layer, batch_idx:batch_idx +
+                                group_size, :mem_kv_perlayer.
+                                shape[1]] = mem_kv_perlayer.expand(
+                                    min(group_size,
+                                        input_tokens.shape[0] - batch_idx), -1,
+                                    -1)
+                mems_indexs[0], mems_indexs[1] = mem_kv01[0][0].shape[
+                    1], mem_kv01[1][0].shape[1]
+                if limited_spatial_channel_mem:
+                    mems_indexs[0] -= (next_tokens_frame_begin_id - text_len)
+
+            mems = [
+                mems_buffers[id][:, :, :mems_indexs[id]] for id in range(2)
+            ]
+            logits = logits_all
+
+            # Guider
+            if guider_seq is not None:
+                guider_logits_all = None
+                for batch_idx in range(0, guider_input_tokens.shape[0],
+                                       group_size):
+                    guider_logits, *guider_output_per_layers = model(
+                        guider_input_tokens[batch_idx:batch_idx + group_size,
+                                            max(index -
+                                                guider_index_delta, 0):],
+                        guider_position_ids[
+                            ...,
+                            max(index - guider_index_delta, 0):counter + 1 -
+                            guider_index_delta],
+                        guider_attention_mask,
+                        mems=guider_mems,
+                        text_len=guider_text_len,
+                        frame_len=frame_len,
+                        counter=counter - guider_index_delta,
+                        log_text_attention_weights=log_text_attention_weights,
+                        enforce_no_swin=enforce_no_swin,
+                        **kw_args)
+                    guider_logits_all = torch.cat(
+                        (guider_logits_all, guider_logits), dim=0
+                    ) if guider_logits_all is not None else guider_logits
+                    guider_mem_kv01 = [[
+                        o['mem_kv'][0] for o in guider_output_per_layers
+                    ], [o['mem_kv'][1] for o in guider_output_per_layers]]
+                    for id, guider_mem_kv in enumerate(guider_mem_kv01):
+                        for layer, guider_mem_kv_perlayer in enumerate(
+                                guider_mem_kv):
+                            if limited_spatial_channel_mem and id == 0:
+                                guider_mems_buffers[id][
+                                    layer, batch_idx:batch_idx + group_size, :
+                                    guider_text_len] = guider_mem_kv_perlayer.expand(
+                                        min(group_size,
+                                            input_tokens.shape[0] - batch_idx),
+                                        -1, -1)[:, :guider_text_len]
+                                guider_next_tokens_frame_begin_id = calc_next_tokens_frame_begin_id(
+                                    guider_text_len, frame_len,
+                                    guider_mem_kv_perlayer.shape[1])
+                                guider_mems_buffers[id][layer, batch_idx:batch_idx+group_size, guider_text_len:guider_text_len+guider_mem_kv_perlayer.shape[1]-guider_next_tokens_frame_begin_id] =\
+                                    guider_mem_kv_perlayer.expand(min(group_size, input_tokens.shape[0]-batch_idx), -1, -1)[:, guider_next_tokens_frame_begin_id:]
+                            else:
+                                guider_mems_buffers[id][
+                                    layer, batch_idx:batch_idx +
+                                    group_size, :guider_mem_kv_perlayer.
+                                    shape[1]] = guider_mem_kv_perlayer.expand(
+                                        min(group_size,
+                                            input_tokens.shape[0] - batch_idx),
+                                        -1, -1)
+                    guider_mems_indexs[0], guider_mems_indexs[
+                        1] = guider_mem_kv01[0][0].shape[1], guider_mem_kv01[
+                            1][0].shape[1]
+                    if limited_spatial_channel_mem:
+                        guider_mems_indexs[0] -= (
+                            guider_next_tokens_frame_begin_id -
+                            guider_text_len)
+                guider_mems = [
+                    guider_mems_buffers[id][:, :, :guider_mems_indexs[id]]
+                    for id in range(2)
+                ]
+                guider_logits = guider_logits_all
+        else:
+            if not mems_buffers_on_GPU:
+                if not mode_stage1:
+                    torch.cuda.empty_cache()
+                    for idx, mem in enumerate(mems):
+                        mems[idx] = mem.to(next(model.parameters()).device)
+                    if guider_seq is not None:
+                        for idx, mem in enumerate(guider_mems):
+                            guider_mems[idx] = mem.to(
+                                next(model.parameters()).device)
+                else:
+                    torch.cuda.empty_cache()
+                    for idx, mem_buffer in enumerate(mems_buffers):
+                        mems_buffers[idx] = mem_buffer.to(
+                            next(model.parameters()).device)
+                    mems = [
+                        mems_buffers[id][:, :, :mems_indexs[id]]
+                        for id in range(2)
+                    ]
+                    if guider_seq is not None:
+                        for idx, guider_mem_buffer in enumerate(
+                                guider_mems_buffers):
+                            guider_mems_buffers[idx] = guider_mem_buffer.to(
+                                next(model.parameters()).device)
+                        guider_mems = [
+                            guider_mems_buffers[id]
+                            [:, :, :guider_mems_indexs[id]] for id in range(2)
+                        ]
+                    mems_buffers_on_GPU = True
+
+            logits, *output_per_layers = model(
+                input_tokens[:, index:],
+                position_ids[..., index:counter + 1],
+                attention_mask,  # TODO memlen
+                mems=mems,
+                text_len=text_len,
+                frame_len=frame_len,
+                counter=counter,
+                log_text_attention_weights=log_text_attention_weights,
+                enforce_no_swin=enforce_no_swin,
+                limited_spatial_channel_mem=limited_spatial_channel_mem,
+                **kw_args)
+            mem_kv0, mem_kv1 = [o['mem_kv'][0] for o in output_per_layers
+                                ], [o['mem_kv'][1] for o in output_per_layers]
+
+            if guider_seq is not None:
+                guider_logits, *guider_output_per_layers = model(
+                    guider_input_tokens[:,
+                                        max(index - guider_index_delta, 0):],
+                    guider_position_ids[...,
+                                        max(index -
+                                            guider_index_delta, 0):counter +
+                                        1 - guider_index_delta],
+                    guider_attention_mask,
+                    mems=guider_mems,
+                    text_len=guider_text_len,
+                    frame_len=frame_len,
+                    counter=counter - guider_index_delta,
+                    log_text_attention_weights=0,
+                    enforce_no_swin=enforce_no_swin,
+                    limited_spatial_channel_mem=limited_spatial_channel_mem,
+                    **kw_args)
+                guider_mem_kv0, guider_mem_kv1 = [
+                    o['mem_kv'][0] for o in guider_output_per_layers
+                ], [o['mem_kv'][1] for o in guider_output_per_layers]
+
+            if not mems_buffers_on_GPU:
+                torch.cuda.empty_cache()
+                for idx, mem_buffer in enumerate(mems_buffers):
+                    mems_buffers[idx] = mem_buffer.to(
+                        next(model.parameters()).device)
+                if guider_seq is not None:
+                    for idx, guider_mem_buffer in enumerate(
+                            guider_mems_buffers):
+                        guider_mems_buffers[idx] = guider_mem_buffer.to(
+                            next(model.parameters()).device)
+                mems_buffers_on_GPU = True
+
+            mems, mems_indexs = my_update_mems([mem_kv0, mem_kv1],
+                                               mems_buffers, mems_indexs,
+                                               limited_spatial_channel_mem,
+                                               text_len, frame_len)
+            if guider_seq is not None:
+                guider_mems, guider_mems_indexs = my_update_mems(
+                    [guider_mem_kv0, guider_mem_kv1], guider_mems_buffers,
+                    guider_mems_indexs, limited_spatial_channel_mem,
+                    guider_text_len, frame_len)
+
+        counter += 1
+        index = counter
+
+        logits = logits[:, -1].expand(batch_size,
+                                      -1)  # [batch size, vocab size]
+        tokens = tokens.expand(batch_size, -1)
+        if guider_seq is not None:
+            guider_logits = guider_logits[:, -1].expand(batch_size, -1)
+            guider_tokens = guider_tokens.expand(batch_size, -1)
+
+        if seq[-1][counter].item() < 0:
+            # sampling
+            guided_logits = guider_logits + (
+                logits - guider_logits
+            ) * guidance_alpha if guider_seq is not None else logits
+            if mode_stage1 and counter < text_len + 400:
+                tokens, mems = strategy.forward(guided_logits, tokens, mems)
+            else:
+                tokens, mems = strategy2.forward(guided_logits, tokens, mems)
+            if guider_seq is not None:
+                guider_tokens = torch.cat((guider_tokens, tokens[:, -1:]),
+                                          dim=1)
+
+            if seq[0][counter].item() >= 0:
+                for si in range(seq.shape[0]):
+                    if seq[si][counter].item() >= 0:
+                        tokens[si, -1] = seq[si, counter]
+                        if guider_seq is not None:
+                            guider_tokens[si,
+                                          -1] = guider_seq[si, counter -
+                                                           guider_index_delta]
+
+        else:
+            tokens = torch.cat(
+                (tokens, seq[:, counter:counter + 1].clone().expand(
+                    tokens.shape[0], 1).to(device=tokens.device,
+                                           dtype=tokens.dtype)),
+                dim=1)
+            if guider_seq is not None:
+                guider_tokens = torch.cat(
+                    (guider_tokens,
+                     guider_seq[:, counter - guider_index_delta:counter + 1 -
+                                guider_index_delta].clone().expand(
+                                    guider_tokens.shape[0], 1).to(
+                                        device=guider_tokens.device,
+                                        dtype=guider_tokens.dtype)),
+                    dim=1)
+
+        input_tokens = tokens.clone()
+        if guider_seq is not None:
+            guider_input_tokens = guider_tokens.clone()
+        if (index - text_len - 1) // 400 < (input_tokens.shape[-1] - text_len -
+                                            1) // 400:
+            boi_idx = ((index - text_len - 1) // 400 + 1) * 400 + text_len
+            while boi_idx < input_tokens.shape[-1]:
+                input_tokens[:, boi_idx] = tokenizer['<start_of_image>']
+                if guider_seq is not None:
+                    guider_input_tokens[:, boi_idx -
+                                        guider_index_delta] = tokenizer[
+                                            '<start_of_image>']
+                boi_idx += 400
+
+        if strategy.is_done:
+            break
+    return strategy.finalize(tokens, mems)
+
+
+class InferenceModel_Sequential(CogVideoCacheModel):
+    def __init__(self, args, transformer=None, parallel_output=True):
+        super().__init__(args,
+                         transformer=transformer,
+                         parallel_output=parallel_output,
+                         window_size=-1,
+                         cogvideo_stage=1)
+
+    # TODO: check it
+
+    def final_forward(self, logits, **kwargs):
+        logits_parallel = logits
+        logits_parallel = torch.nn.functional.linear(
+            logits_parallel.float(),
+            self.transformer.word_embeddings.weight[:20000].float())
+        return logits_parallel
+
+
+class InferenceModel_Interpolate(CogVideoCacheModel):
+    def __init__(self, args, transformer=None, parallel_output=True):
+        super().__init__(args,
+                         transformer=transformer,
+                         parallel_output=parallel_output,
+                         window_size=10,
+                         cogvideo_stage=2)
+
+    # TODO: check it
+
+    def final_forward(self, logits, **kwargs):
+        logits_parallel = logits
+        logits_parallel = torch.nn.functional.linear(
+            logits_parallel.float(),
+            self.transformer.word_embeddings.weight[:20000].float())
+        return logits_parallel
+
+
+def get_default_args() -> argparse.Namespace:
+    known = argparse.Namespace(generate_frame_num=5,
+                               coglm_temperature2=0.89,
+                               use_guidance_stage1=True,
+                               use_guidance_stage2=False,
+                               guidance_alpha=3.0,
+                               stage_1=True,
+                               stage_2=False,
+                               both_stages=False,
+                               parallel_size=1,
+                               stage1_max_inference_batch_size=-1,
+                               multi_gpu=False,
+                               layout='64, 464, 2064',
+                               window_size=10,
+                               additional_seqlen=2000,
+                               cogvideo_stage=1)
+
+    args_list = [
+        '--tokenizer-type',
+        'fake',
+        '--mode',
+        'inference',
+        '--distributed-backend',
+        'nccl',
+        '--fp16',
+        '--model-parallel-size',
+        '1',
+        '--temperature',
+        '1.05',
+        '--top_k',
+        '12',
+        '--sandwich-ln',
+        '--seed',
+        '1234',
+        '--num-workers',
+        '0',
+        '--batch-size',
+        '1',
+        '--max-inference-batch-size',
+        '8',
+    ]
+    args = get_args(args_list)
+    args = argparse.Namespace(**vars(args), **vars(known))
+    args.layout = [int(x) for x in args.layout.split(',')]
+    args.do_train = False
+    return args
+
+
+class Model:
+    def __init__(self, only_first_stage: bool = False):
+        self.args = get_default_args()
+        if only_first_stage:
+            self.args.stage_1 = True
+            self.args.both_stages = False
+        else:
+            self.args.stage_1 = False
+            self.args.both_stages = True
+
+        self.tokenizer = self.load_tokenizer()
+
+        self.model_stage1, self.args = self.load_model_stage1()
+        self.model_stage2, self.args = self.load_model_stage2()
+
+        self.strategy_cogview2, self.strategy_cogvideo = self.load_strategies()
+        self.dsr = self.load_dsr()
+
+        self.device = torch.device(self.args.device)
+
+    def load_tokenizer(self) -> IceTokenizer:
+        logger.info('--- load_tokenizer ---')
+        start = time.perf_counter()
+
+        tokenizer = IceTokenizer(ICETK_MODEL_DIR.as_posix())
+        tokenizer.add_special_tokens(
+            ['<start_of_image>', '<start_of_english>', '<start_of_chinese>'])
+
+        elapsed = time.perf_counter() - start
+        logger.info(f'--- done ({elapsed=:.3f}) ---')
+        return tokenizer
+
+    def load_model_stage1(
+            self) -> tuple[CogVideoCacheModel, argparse.Namespace]:
+        logger.info('--- load_model_stage1 ---')
+        start = time.perf_counter()
+
+        args = self.args
+        model_stage1, args = InferenceModel_Sequential.from_pretrained(
+            args, 'cogvideo-stage1')
+        model_stage1.eval()
+        if args.both_stages:
+            model_stage1 = model_stage1.cpu()
+
+        elapsed = time.perf_counter() - start
+        logger.info(f'--- done ({elapsed=:.3f}) ---')
+        return model_stage1, args
+
+    def load_model_stage2(
+            self) -> tuple[CogVideoCacheModel | None, argparse.Namespace]:
+        logger.info('--- load_model_stage2 ---')
+        start = time.perf_counter()
+
+        args = self.args
+        if args.both_stages:
+            model_stage2, args = InferenceModel_Interpolate.from_pretrained(
+                args, 'cogvideo-stage2')
+            model_stage2.eval()
+            if args.both_stages:
+                model_stage2 = model_stage2.cpu()
+        else:
+            model_stage2 = None
+
+        elapsed = time.perf_counter() - start
+        logger.info(f'--- done ({elapsed=:.3f}) ---')
+        return model_stage2, args
+
+    def load_strategies(self) -> tuple[CoglmStrategy, CoglmStrategy]:
+        logger.info('--- load_strategies ---')
+        start = time.perf_counter()
+
+        invalid_slices = [slice(self.tokenizer.num_image_tokens, None)]
+        strategy_cogview2 = CoglmStrategy(invalid_slices,
+                                          temperature=1.0,
+                                          top_k=16)
+        strategy_cogvideo = CoglmStrategy(
+            invalid_slices,
+            temperature=self.args.temperature,
+            top_k=self.args.top_k,
+            temperature2=self.args.coglm_temperature2)
+
+        elapsed = time.perf_counter() - start
+        logger.info(f'--- done ({elapsed=:.3f}) ---')
+        return strategy_cogview2, strategy_cogvideo
+
+    def load_dsr(self) -> DirectSuperResolution | None:
+        logger.info('--- load_dsr ---')
+        start = time.perf_counter()
+
+        if self.args.both_stages:
+            path = auto_create('cogview2-dsr', path=None)
+            dsr = DirectSuperResolution(self.args,
+                                        path,
+                                        max_bz=12,
+                                        onCUDA=False)
+        else:
+            dsr = None
+
+        elapsed = time.perf_counter() - start
+        logger.info(f'--- done ({elapsed=:.3f}) ---')
+        return dsr
+
+    @torch.inference_mode()
+    def process_stage1(self,
+                       model,
+                       seq_text,
+                       duration,
+                       video_raw_text=None,
+                       video_guidance_text='视频',
+                       image_text_suffix='',
+                       batch_size=1,
+                       image_prompt=None):
+        process_start_time = time.perf_counter()
+
+        generate_frame_num = self.args.generate_frame_num
+        tokenizer = self.tokenizer
+        use_guide = self.args.use_guidance_stage1
+
+        if next(model.parameters()).device != self.device:
+            move_start_time = time.perf_counter()
+            logger.debug('moving stage 1 model to cuda')
+
+            model = model.to(self.device)
+
+            elapsed = time.perf_counter() - move_start_time
+            logger.debug(f'moving in model1 takes time: {elapsed:.2f}')
+
+        if video_raw_text is None:
+            video_raw_text = seq_text
+        mbz = self.args.stage1_max_inference_batch_size if self.args.stage1_max_inference_batch_size > 0 else self.args.max_inference_batch_size
+        assert batch_size < mbz or batch_size % mbz == 0
+        frame_len = 400
+
+        # generate the first frame:
+        enc_text = tokenizer.encode(seq_text + image_text_suffix)
+        seq_1st = enc_text + [tokenizer['<start_of_image>']] + [-1] * 400
+        logger.info(
+            f'[Generating First Frame with CogView2] Raw text: {tokenizer.decode(enc_text):s}'
+        )
+        text_len_1st = len(seq_1st) - frame_len * 1 - 1
+
+        seq_1st = torch.tensor(seq_1st, dtype=torch.long,
+                               device=self.device).unsqueeze(0)
+        if image_prompt is None:
+            output_list_1st = []
+            for tim in range(max(batch_size // mbz, 1)):
+                start_time = time.perf_counter()
+                output_list_1st.append(
+                    my_filling_sequence(
+                        model,
+                        tokenizer,
+                        self.args,
+                        seq_1st.clone(),
+                        batch_size=min(batch_size, mbz),
+                        get_masks_and_position_ids=
+                        get_masks_and_position_ids_stage1,
+                        text_len=text_len_1st,
+                        frame_len=frame_len,
+                        strategy=self.strategy_cogview2,
+                        strategy2=self.strategy_cogvideo,
+                        log_text_attention_weights=1.4,
+                        enforce_no_swin=True,
+                        mode_stage1=True,
+                    )[0])
+                elapsed = time.perf_counter() - start_time
+                logger.info(f'[First Frame] Elapsed: {elapsed:.2f}')
+            output_tokens_1st = torch.cat(output_list_1st, dim=0)
+            given_tokens = output_tokens_1st[:, text_len_1st + 1:text_len_1st +
+                                            401].unsqueeze(
+                                                1
+                                            )  # given_tokens.shape: [bs, frame_num, 400]
+        else:
+            given_tokens = tokenizer.encode(image_path=image_prompt, image_size=160).repeat(batch_size, 1).unsqueeze(1)
+
+        # generate subsequent frames:
+        total_frames = generate_frame_num
+        enc_duration = tokenizer.encode(f'{float(duration)}秒')
+        if use_guide:
+            video_raw_text = video_raw_text + ' 视频'
+        enc_text_video = tokenizer.encode(video_raw_text)
+        seq = enc_duration + [tokenizer['<n>']] + enc_text_video + [
+            tokenizer['<start_of_image>']
+        ] + [-1] * 400 * generate_frame_num
+        guider_seq = enc_duration + [tokenizer['<n>']] + tokenizer.encode(
+            video_guidance_text) + [tokenizer['<start_of_image>']
+                                    ] + [-1] * 400 * generate_frame_num
+        logger.info(
+            f'[Stage1: Generating Subsequent Frames, Frame Rate {4/duration:.1f}] raw text: {tokenizer.decode(enc_text_video):s}'
+        )
+
+        text_len = len(seq) - frame_len * generate_frame_num - 1
+        guider_text_len = len(guider_seq) - frame_len * generate_frame_num - 1
+        seq = torch.tensor(seq, dtype=torch.long,
+                           device=self.device).unsqueeze(0).repeat(
+                               batch_size, 1)
+        guider_seq = torch.tensor(guider_seq,
+                                  dtype=torch.long,
+                                  device=self.device).unsqueeze(0).repeat(
+                                      batch_size, 1)
+
+        for given_frame_id in range(given_tokens.shape[1]):
+            seq[:, text_len + 1 + given_frame_id * 400:text_len + 1 +
+                (given_frame_id + 1) * 400] = given_tokens[:, given_frame_id]
+            guider_seq[:, guider_text_len + 1 +
+                       given_frame_id * 400:guider_text_len + 1 +
+                       (given_frame_id + 1) *
+                       400] = given_tokens[:, given_frame_id]
+        output_list = []
+
+        if use_guide:
+            video_log_text_attention_weights = 0
+        else:
+            guider_seq = None
+            video_log_text_attention_weights = 1.4
+
+        for tim in range(max(batch_size // mbz, 1)):
+            input_seq = seq[:min(batch_size, mbz)].clone(
+            ) if tim == 0 else seq[mbz * tim:mbz * (tim + 1)].clone()
+            guider_seq2 = (guider_seq[:min(batch_size, mbz)].clone()
+                           if tim == 0 else guider_seq[mbz * tim:mbz *
+                                                       (tim + 1)].clone()
+                           ) if guider_seq is not None else None
+            output_list.append(
+                my_filling_sequence(
+                    model,
+                    tokenizer,
+                    self.args,
+                    input_seq,
+                    batch_size=min(batch_size, mbz),
+                    get_masks_and_position_ids=
+                    get_masks_and_position_ids_stage1,
+                    text_len=text_len,
+                    frame_len=frame_len,
+                    strategy=self.strategy_cogview2,
+                    strategy2=self.strategy_cogvideo,
+                    log_text_attention_weights=video_log_text_attention_weights,
+                    guider_seq=guider_seq2,
+                    guider_text_len=guider_text_len,
+                    guidance_alpha=self.args.guidance_alpha,
+                    limited_spatial_channel_mem=True,
+                    mode_stage1=True,
+                )[0])
+
+        output_tokens = torch.cat(output_list, dim=0)[:, 1 + text_len:]
+
+        if self.args.both_stages:
+            move_start_time = time.perf_counter()
+            logger.debug('moving stage 1 model to cpu')
+            model = model.cpu()
+            torch.cuda.empty_cache()
+            elapsed = time.perf_counter() - move_start_time
+            logger.debug(f'moving in model1 takes time: {elapsed:.2f}')
+
+        # decoding
+        res = []
+        for seq in output_tokens:
+            decoded_imgs = [
+                self.postprocess(
+                    torch.nn.functional.interpolate(tokenizer.decode(
+                        image_ids=seq.tolist()[i * 400:(i + 1) * 400]),
+                                                    size=(480, 480))[0])
+                for i in range(total_frames)
+            ]
+            res.append(decoded_imgs)  # only the last image (target)
+
+        assert len(res) == batch_size
+        tokens = output_tokens[:, :+total_frames * 400].reshape(
+            -1, total_frames, 400).cpu()
+
+        elapsed = time.perf_counter() - process_start_time
+        logger.info(f'--- done ({elapsed=:.3f}) ---')
+        return tokens, res[0]
+
+    @torch.inference_mode()
+    def process_stage2(self,
+                       model,
+                       seq_text,
+                       duration,
+                       parent_given_tokens,
+                       video_raw_text=None,
+                       video_guidance_text='视频',
+                       gpu_rank=0,
+                       gpu_parallel_size=1):
+        process_start_time = time.perf_counter()
+
+        generate_frame_num = self.args.generate_frame_num
+        tokenizer = self.tokenizer
+        use_guidance = self.args.use_guidance_stage2
+
+        stage2_start_time = time.perf_counter()
+
+        if next(model.parameters()).device != self.device:
+            move_start_time = time.perf_counter()
+            logger.debug('moving stage-2 model to cuda')
+
+            model = model.to(self.device)
+
+            elapsed = time.perf_counter() - move_start_time
+            logger.debug(f'moving in stage-2 model takes time: {elapsed:.2f}')
+
+        try:
+            sample_num_allgpu = parent_given_tokens.shape[0]
+            sample_num = sample_num_allgpu // gpu_parallel_size
+            assert sample_num * gpu_parallel_size == sample_num_allgpu
+            parent_given_tokens = parent_given_tokens[gpu_rank *
+                                                      sample_num:(gpu_rank +
+                                                                  1) *
+                                                      sample_num]
+        except:
+            logger.critical('No frame_tokens found in interpolation, skip')
+            return False, []
+
+        # CogVideo Stage2 Generation
+        while duration >= 0.5:  # TODO: You can change the boundary to change the frame rate
+            parent_given_tokens_num = parent_given_tokens.shape[1]
+            generate_batchsize_persample = (parent_given_tokens_num - 1) // 2
+            generate_batchsize_total = generate_batchsize_persample * sample_num
+            total_frames = generate_frame_num
+            frame_len = 400
+            enc_text = tokenizer.encode(seq_text)
+            enc_duration = tokenizer.encode(str(float(duration)) + '秒')
+            seq = enc_duration + [tokenizer['<n>']] + enc_text + [
+                tokenizer['<start_of_image>']
+            ] + [-1] * 400 * generate_frame_num
+            text_len = len(seq) - frame_len * generate_frame_num - 1
+
+            logger.info(
+                f'[Stage2: Generating Frames, Frame Rate {int(4/duration):d}] raw text: {tokenizer.decode(enc_text):s}'
+            )
+
+            # generation
+            seq = torch.tensor(seq, dtype=torch.long,
+                               device=self.device).unsqueeze(0).repeat(
+                                   generate_batchsize_total, 1)
+            for sample_i in range(sample_num):
+                for i in range(generate_batchsize_persample):
+                    seq[sample_i * generate_batchsize_persample +
+                        i][text_len + 1:text_len + 1 +
+                           400] = parent_given_tokens[sample_i][2 * i]
+                    seq[sample_i * generate_batchsize_persample +
+                        i][text_len + 1 + 400:text_len + 1 +
+                           800] = parent_given_tokens[sample_i][2 * i + 1]
+                    seq[sample_i * generate_batchsize_persample +
+                        i][text_len + 1 + 800:text_len + 1 +
+                           1200] = parent_given_tokens[sample_i][2 * i + 2]
+
+            if use_guidance:
+                guider_seq = enc_duration + [
+                    tokenizer['<n>']
+                ] + tokenizer.encode(video_guidance_text) + [
+                    tokenizer['<start_of_image>']
+                ] + [-1] * 400 * generate_frame_num
+                guider_text_len = len(
+                    guider_seq) - frame_len * generate_frame_num - 1
+                guider_seq = torch.tensor(
+                    guider_seq, dtype=torch.long,
+                    device=self.device).unsqueeze(0).repeat(
+                        generate_batchsize_total, 1)
+                for sample_i in range(sample_num):
+                    for i in range(generate_batchsize_persample):
+                        guider_seq[sample_i * generate_batchsize_persample +
+                                   i][text_len + 1:text_len + 1 +
+                                      400] = parent_given_tokens[sample_i][2 *
+                                                                           i]
+                        guider_seq[sample_i * generate_batchsize_persample +
+                                   i][text_len + 1 + 400:text_len + 1 +
+                                      800] = parent_given_tokens[sample_i][2 *
+                                                                           i +
+                                                                           1]
+                        guider_seq[sample_i * generate_batchsize_persample +
+                                   i][text_len + 1 + 800:text_len + 1 +
+                                      1200] = parent_given_tokens[sample_i][2 *
+                                                                            i +
+                                                                            2]
+                video_log_text_attention_weights = 0
+            else:
+                guider_seq = None
+                guider_text_len = 0
+                video_log_text_attention_weights = 1.4
+
+            mbz = self.args.max_inference_batch_size
+
+            assert generate_batchsize_total < mbz or generate_batchsize_total % mbz == 0
+            output_list = []
+            start_time = time.perf_counter()
+            for tim in range(max(generate_batchsize_total // mbz, 1)):
+                input_seq = seq[:min(generate_batchsize_total, mbz)].clone(
+                ) if tim == 0 else seq[mbz * tim:mbz * (tim + 1)].clone()
+                guider_seq2 = (
+                    guider_seq[:min(generate_batchsize_total, mbz)].clone()
+                    if tim == 0 else guider_seq[mbz * tim:mbz *
+                                                (tim + 1)].clone()
+                ) if guider_seq is not None else None
+                output_list.append(
+                    my_filling_sequence(
+                        model,
+                        tokenizer,
+                        self.args,
+                        input_seq,
+                        batch_size=min(generate_batchsize_total, mbz),
+                        get_masks_and_position_ids=
+                        get_masks_and_position_ids_stage2,
+                        text_len=text_len,
+                        frame_len=frame_len,
+                        strategy=self.strategy_cogview2,
+                        strategy2=self.strategy_cogvideo,
+                        log_text_attention_weights=
+                        video_log_text_attention_weights,
+                        mode_stage1=False,
+                        guider_seq=guider_seq2,
+                        guider_text_len=guider_text_len,
+                        guidance_alpha=self.args.guidance_alpha,
+                        limited_spatial_channel_mem=True,
+                    )[0])
+            elapsed = time.perf_counter() - start_time
+            logger.info(f'Duration {duration:.2f}, Elapsed: {elapsed:.2f}\n')
+
+            output_tokens = torch.cat(output_list, dim=0)
+            output_tokens = output_tokens[:, text_len + 1:text_len + 1 +
+                                          (total_frames) * 400].reshape(
+                                              sample_num, -1,
+                                              400 * total_frames)
+            output_tokens_merge = torch.cat(
+                (output_tokens[:, :, :1 * 400], output_tokens[:, :,
+                                                              400 * 3:4 * 400],
+                 output_tokens[:, :, 400 * 1:2 * 400],
+                 output_tokens[:, :, 400 * 4:(total_frames) * 400]),
+                dim=2).reshape(sample_num, -1, 400)
+
+            output_tokens_merge = torch.cat(
+                (output_tokens_merge, output_tokens[:, -1:, 400 * 2:3 * 400]),
+                dim=1)
+            duration /= 2
+            parent_given_tokens = output_tokens_merge
+
+        if self.args.both_stages:
+            move_start_time = time.perf_counter()
+            logger.debug('moving stage 2 model to cpu')
+            model = model.cpu()
+            torch.cuda.empty_cache()
+            elapsed = time.perf_counter() - move_start_time
+            logger.debug(f'moving out model2 takes time: {elapsed:.2f}')
+
+        elapsed = time.perf_counter() - stage2_start_time
+        logger.info(f'CogVideo Stage2 completed. Elapsed: {elapsed:.2f}\n')
+
+        # direct super-resolution by CogView2
+        logger.info('[Direct super-resolution]')
+        dsr_start_time = time.perf_counter()
+
+        enc_text = tokenizer.encode(seq_text)
+        frame_num_per_sample = parent_given_tokens.shape[1]
+        parent_given_tokens_2d = parent_given_tokens.reshape(-1, 400)
+        text_seq = torch.tensor(enc_text, dtype=torch.long,
+                                device=self.device).unsqueeze(0).repeat(
+                                    parent_given_tokens_2d.shape[0], 1)
+        sred_tokens = self.dsr(text_seq, parent_given_tokens_2d)
+
+        decoded_sr_videos = []
+        for sample_i in range(sample_num):
+            decoded_sr_imgs = []
+            for frame_i in range(frame_num_per_sample):
+                decoded_sr_img = tokenizer.decode(
+                    image_ids=sred_tokens[frame_i + sample_i *
+                                          frame_num_per_sample][-3600:])
+                decoded_sr_imgs.append(
+                    self.postprocess(
+                        torch.nn.functional.interpolate(decoded_sr_img,
+                                                        size=(480, 480))[0]))
+            decoded_sr_videos.append(decoded_sr_imgs)
+
+        elapsed = time.perf_counter() - dsr_start_time
+        logger.info(
+            f'Direct super-resolution completed. Elapsed: {elapsed:.2f}')
+
+        elapsed = time.perf_counter() - process_start_time
+        logger.info(f'--- done ({elapsed=:.3f}) ---')
+        return True, decoded_sr_videos[0]
+
+    @staticmethod
+    def postprocess(tensor: torch.Tensor) -> np.ndarray:
+        return tensor.cpu().mul(255).add_(0.5).clamp_(0, 255).permute(
+            1, 2, 0).to(torch.uint8).numpy()
+
+    def run(self, text: str, seed: int,
+            only_first_stage: bool,image_prompt: None) -> list[np.ndarray]:
+        logger.info('==================== run ====================')
+        start = time.perf_counter()
+
+        set_random_seed(seed)
+        self.args.seed = seed
+
+        if only_first_stage:
+            self.args.stage_1 = True
+            self.args.both_stages = False
+        else:
+            self.args.stage_1 = False
+            self.args.both_stages = True
+
+        parent_given_tokens, res = self.process_stage1(
+            self.model_stage1,
+            text,
+            duration=4.0,
+            video_raw_text=text,
+            video_guidance_text='视频',
+            image_text_suffix=' 高清摄影',
+            batch_size=self.args.batch_size,
+            image_prompt=image_prompt)
+        if not only_first_stage:
+            _, res = self.process_stage2(
+                self.model_stage2,
+                text,
+                duration=2.0,
+                parent_given_tokens=parent_given_tokens,
+                video_raw_text=text + ' 视频',
+                video_guidance_text='视频',
+                gpu_rank=0,
+                gpu_parallel_size=1)  # TODO: 修改
+
+        elapsed = time.perf_counter() - start
+        logger.info(f'Elapsed: {elapsed:.3f}')
+        logger.info('==================== done ====================')
+        return res
+
+
+class AppModel(Model):
+    def __init__(self, only_first_stage: bool):
+        super().__init__(only_first_stage)
+        self.translator = gr.Interface.load(
+            'spaces/chinhon/translation_eng2ch')
+
+    def to_video(self, frames: list[np.ndarray]) -> str:
+        out_file = tempfile.NamedTemporaryFile(suffix='.mp4', delete=False)
+        if self.args.stage_1:
+            fps = 4
+        else:
+            fps = 8
+        writer = iio.get_writer(out_file.name, fps=fps)
+        for frame in frames:
+            writer.append_data(frame)
+        writer.close()
+        return out_file.name
+
+    def run_with_translation(
+            self, text: str, translate: bool, seed: int,
+            only_first_stage: bool,image_prompt: None) -> tuple[str | None, str | None]:
+            
+        logger.info(f'{text=}, {translate=}, {seed=}, {only_first_stage=},{image_prompt=}')
+        if translate:
+            text = translated_text = self.translator(text)
+        else:
+            translated_text = None
+        frames = self.run(text, seed, only_first_stage,image_prompt)
+        video_path = self.to_video(frames)
+        return translated_text, video_path

patch

@@ -0,0 +1,51 @@
+diff --git a/coglm_strategy.py b/coglm_strategy.py
+index d485715..a9eab3b 100644
+--- a/coglm_strategy.py
++++ b/coglm_strategy.py
+@@ -8,6 +8,7 @@
+ 
+ # here put the import lib
+ import os
++import pathlib
+ import sys
+ import math
+ import random
+@@ -58,7 +59,8 @@ class CoglmStrategy:
+         self._is_done = False
+         self.outlier_count_down = torch.zeros(16) 
+         self.vis_list = [[]for i in range(16)]
+-        self.cluster_labels = torch.tensor(np.load('cluster_label2.npy'), device='cuda', dtype=torch.long)
++        cluster_label_path = pathlib.Path(__file__).parent / 'cluster_label2.npy'
++        self.cluster_labels = torch.tensor(np.load(cluster_label_path), device='cuda', dtype=torch.long)
+         self.start_pos = -1
+         self.white_cluster = []
+         # self.fout = open('tmp.txt', 'w')
+@@ -98,4 +100,4 @@ class CoglmStrategy:
+ 
+     def finalize(self, tokens, mems):
+         self._is_done = False
+-        return tokens, mems
+
++        return tokens, mems
+diff --git a/sr_pipeline/dsr_sampling.py b/sr_pipeline/dsr_sampling.py
+index 5b8dded..07e97fd 100644
+--- a/sr_pipeline/dsr_sampling.py
++++ b/sr_pipeline/dsr_sampling.py
+@@ -8,6 +8,7 @@
+ 
+ # here put the import lib
+ import os
++import pathlib
+ import sys
+ import math
+ import random
+@@ -28,7 +29,8 @@ class IterativeEntfilterStrategy:
+         self.invalid_slices = invalid_slices
+         self.temperature = temperature
+         self.topk = topk        
+-        self.cluster_labels = torch.tensor(np.load('cluster_label2.npy'), device='cuda', dtype=torch.long)
++        cluster_label_path = pathlib.Path(__file__).parents[1] / 'cluster_label2.npy'
++        self.cluster_labels = torch.tensor(np.load(cluster_label_path), device='cuda', dtype=torch.long)
+ 
+ 
+     def forward(self, logits_, tokens, temperature=None, entfilter=None, filter_topk=5, temperature2=None):

requirements.txt

@@ -0,0 +1,7 @@
+--extra-index-url https://download.pytorch.org/whl/cu113
+imageio==2.19.5
+imageio-ffmpeg==0.4.7
+numpy==1.22.4
+opencv-python-headless==4.6.0.66
+torch==1.12.0+cu113
+torchvision==0.13.0+cu113

samples.txt

@@ -0,0 +1,2 @@
+骑滑板的皮卡丘
+a cat playing chess

style.css

@@ -0,0 +1,7 @@
+h1 {
+  text-align: center;
+}
+img#visitor-badge {
+  display: block;
+  margin: auto;
+}