diff --git a/.gitattributes b/.gitattributes index f6f5207665f8b2d614da6ac4d8ff8a1040a0083f..221951040d7db45fd81e2690558e8fe45b33f76e 100644 --- a/.gitattributes +++ b/.gitattributes @@ -67,3 +67,5 @@ venv/bin/python3.10 filter=lfs diff=lfs merge=lfs -text venv/bin/uv filter=lfs diff=lfs merge=lfs -text venv/lib/python3.10/site-packages/chardet/__pycache__/johabfreq.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text venv/lib/python3.10/site-packages/nacl/_sodium.abi3.so filter=lfs diff=lfs merge=lfs -text +venv/lib/python3.10/site-packages/pygments/lexers/__pycache__/lisp.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text +venv/lib/python3.10/site-packages/wrapt/_wrappers.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text diff --git a/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/AUTHORS.rst b/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/AUTHORS.rst new file mode 100644 index 0000000000000000000000000000000000000000..88e2b6ad752425c0c372874e5b3b7695408e7ee3 --- /dev/null +++ b/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/AUTHORS.rst @@ -0,0 +1,7 @@ +Authors +======= + +``pyjwt`` is currently written and maintained by `Jose Padilla `_. +Originally written and maintained by `Jeff Lindsay `_. + +A full list of contributors can be found on GitHub’s `overview `_. diff --git a/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/INSTALLER b/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/INSTALLER new file mode 100644 index 0000000000000000000000000000000000000000..a1b589e38a32041e49332e5e81c2d363dc418d68 --- /dev/null +++ b/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/INSTALLER @@ -0,0 +1 @@ +pip diff --git a/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/LICENSE b/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..fd0ecbc889c067a562b3a745c9d2d7934f87d64b --- /dev/null +++ b/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/LICENSE @@ -0,0 +1,21 @@ +The MIT License (MIT) + +Copyright (c) 2015-2022 José Padilla + +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. diff --git a/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/METADATA b/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/METADATA new file mode 100644 index 0000000000000000000000000000000000000000..f31b700e64c9dd3c2cf4ed16a262787fe61e6f8d --- /dev/null +++ b/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/METADATA @@ -0,0 +1,106 @@ +Metadata-Version: 2.1 +Name: PyJWT +Version: 2.10.1 +Summary: JSON Web Token implementation in Python +Author-email: Jose Padilla +License: MIT +Project-URL: Homepage, https://github.com/jpadilla/pyjwt +Keywords: json,jwt,security,signing,token,web +Classifier: Development Status :: 5 - Production/Stable +Classifier: Intended Audience :: Developers +Classifier: License :: OSI Approved :: MIT License +Classifier: Natural Language :: English +Classifier: Programming Language :: Python +Classifier: Programming Language :: Python :: 3 +Classifier: Programming Language :: Python :: 3 :: Only +Classifier: Programming Language :: Python :: 3.9 +Classifier: Programming Language :: Python :: 3.10 +Classifier: Programming Language :: Python :: 3.11 +Classifier: Programming Language :: Python :: 3.12 +Classifier: Programming Language :: Python :: 3.13 +Classifier: Topic :: Utilities +Requires-Python: >=3.9 +Description-Content-Type: text/x-rst +License-File: LICENSE +License-File: AUTHORS.rst +Provides-Extra: crypto +Requires-Dist: cryptography>=3.4.0; extra == "crypto" +Provides-Extra: dev +Requires-Dist: coverage[toml]==5.0.4; extra == "dev" +Requires-Dist: cryptography>=3.4.0; extra == "dev" +Requires-Dist: pre-commit; extra == "dev" +Requires-Dist: pytest<7.0.0,>=6.0.0; extra == "dev" +Requires-Dist: sphinx; extra == "dev" +Requires-Dist: sphinx-rtd-theme; extra == "dev" +Requires-Dist: zope.interface; extra == "dev" +Provides-Extra: docs +Requires-Dist: sphinx; extra == "docs" +Requires-Dist: sphinx-rtd-theme; extra == "docs" +Requires-Dist: zope.interface; extra == "docs" +Provides-Extra: tests +Requires-Dist: coverage[toml]==5.0.4; extra == "tests" +Requires-Dist: pytest<7.0.0,>=6.0.0; extra == "tests" + +PyJWT +===== + +.. image:: https://github.com/jpadilla/pyjwt/workflows/CI/badge.svg + :target: https://github.com/jpadilla/pyjwt/actions?query=workflow%3ACI + +.. image:: https://img.shields.io/pypi/v/pyjwt.svg + :target: https://pypi.python.org/pypi/pyjwt + +.. image:: https://codecov.io/gh/jpadilla/pyjwt/branch/master/graph/badge.svg + :target: https://codecov.io/gh/jpadilla/pyjwt + +.. image:: https://readthedocs.org/projects/pyjwt/badge/?version=stable + :target: https://pyjwt.readthedocs.io/en/stable/ + +A Python implementation of `RFC 7519 `_. Original implementation was written by `@progrium `_. + +Sponsor +------- + +.. |auth0-logo| image:: https://github.com/user-attachments/assets/ee98379e-ee76-4bcb-943a-e25c4ea6d174 + :width: 160px + ++--------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ +| |auth0-logo| | If you want to quickly add secure token-based authentication to Python projects, feel free to check Auth0's Python SDK and free plan at `auth0.com/signup `_. | ++--------------+-----------------------------------------------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ + +Installing +---------- + +Install with **pip**: + +.. code-block:: console + + $ pip install PyJWT + + +Usage +----- + +.. code-block:: pycon + + >>> import jwt + >>> encoded = jwt.encode({"some": "payload"}, "secret", algorithm="HS256") + >>> print(encoded) + eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJzb21lIjoicGF5bG9hZCJ9.4twFt5NiznN84AWoo1d7KO1T_yoc0Z6XOpOVswacPZg + >>> jwt.decode(encoded, "secret", algorithms=["HS256"]) + {'some': 'payload'} + +Documentation +------------- + +View the full docs online at https://pyjwt.readthedocs.io/en/stable/ + + +Tests +----- + +You can run tests from the project root after cloning with: + +.. code-block:: console + + $ tox diff --git a/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/RECORD b/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/RECORD new file mode 100644 index 0000000000000000000000000000000000000000..bcc25fd4c9f246b69129bb84143ccbb6743cec19 --- /dev/null +++ b/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/RECORD @@ -0,0 +1,32 @@ +PyJWT-2.10.1.dist-info/AUTHORS.rst,sha256=klzkNGECnu2_VY7At89_xLBF3vUSDruXk3xwgUBxzwc,322 +PyJWT-2.10.1.dist-info/INSTALLER,sha256=zuuue4knoyJ-UwPPXg8fezS7VCrXJQrAP7zeNuwvFQg,4 +PyJWT-2.10.1.dist-info/LICENSE,sha256=eXp6ICMdTEM-nxkR2xcx0GtYKLmPSZgZoDT3wPVvXOU,1085 +PyJWT-2.10.1.dist-info/METADATA,sha256=EkewF6D6KU8SGaaQzVYfxUUU1P_gs_dp1pYTkoYvAx8,3990 +PyJWT-2.10.1.dist-info/RECORD,, +PyJWT-2.10.1.dist-info/WHEEL,sha256=PZUExdf71Ui_so67QXpySuHtCi3-J3wvF4ORK6k_S8U,91 +PyJWT-2.10.1.dist-info/top_level.txt,sha256=RP5DHNyJbMq2ka0FmfTgoSaQzh7e3r5XuCWCO8a00k8,4 +jwt/__init__.py,sha256=VB2vFKuboTjcDGeZ8r-UqK_dz3NsQSQEqySSICby8Xg,1711 +jwt/__pycache__/__init__.cpython-310.pyc,, +jwt/__pycache__/algorithms.cpython-310.pyc,, +jwt/__pycache__/api_jwk.cpython-310.pyc,, +jwt/__pycache__/api_jws.cpython-310.pyc,, +jwt/__pycache__/api_jwt.cpython-310.pyc,, +jwt/__pycache__/exceptions.cpython-310.pyc,, +jwt/__pycache__/help.cpython-310.pyc,, +jwt/__pycache__/jwk_set_cache.cpython-310.pyc,, +jwt/__pycache__/jwks_client.cpython-310.pyc,, +jwt/__pycache__/types.cpython-310.pyc,, +jwt/__pycache__/utils.cpython-310.pyc,, +jwt/__pycache__/warnings.cpython-310.pyc,, +jwt/algorithms.py,sha256=cKr-XEioe0mBtqJMCaHEswqVOA1Z8Purt5Sb3Bi-5BE,30409 +jwt/api_jwk.py,sha256=6F1r7rmm8V5qEnBKA_xMjS9R7VoANe1_BL1oD2FrAjE,4451 +jwt/api_jws.py,sha256=aM8vzqQf6mRrAw7bRy-Moj_pjWsKSVQyYK896AfMjJU,11762 +jwt/api_jwt.py,sha256=OGT4hok1l5A6FH_KdcrU5g6u6EQ8B7em0r9kGM9SYgA,14512 +jwt/exceptions.py,sha256=bUIOJ-v9tjopTLS-FYOTc3kFx5WP5IZt7ksN_HE1G9Q,1211 +jwt/help.py,sha256=vFdNzjQoAch04XCMYpCkyB2blaqHAGAqQrtf9nSPkdk,1808 +jwt/jwk_set_cache.py,sha256=hBKmN-giU7-G37L_XKgc_OZu2ah4wdbj1ZNG_GkoSE8,959 +jwt/jwks_client.py,sha256=p9b-IbQqo2tEge9Zit3oSPBFNePqwho96VLbnUrHUWs,4259 +jwt/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0 +jwt/types.py,sha256=VnhGv_VFu5a7_mrPoSCB7HaNLrJdhM8Sq1sSfEg0gLU,99 +jwt/utils.py,sha256=hxOjvDBheBYhz-RIPiEz7Q88dSUSTMzEdKE_Ww2VdJw,3640 +jwt/warnings.py,sha256=50XWOnyNsIaqzUJTk6XHNiIDykiL763GYA92MjTKmok,59 diff --git a/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/WHEEL b/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/WHEEL new file mode 100644 index 0000000000000000000000000000000000000000..ae527e7d64811439e61b93aa375defb30e06edfe --- /dev/null +++ b/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/WHEEL @@ -0,0 +1,5 @@ +Wheel-Version: 1.0 +Generator: setuptools (75.6.0) +Root-Is-Purelib: true +Tag: py3-none-any + diff --git a/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/top_level.txt b/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/top_level.txt new file mode 100644 index 0000000000000000000000000000000000000000..27ccc9bc3a9ca1cd3dab3b2530e67b3309fde50a --- /dev/null +++ b/venv/lib/python3.10/site-packages/PyJWT-2.10.1.dist-info/top_level.txt @@ -0,0 +1 @@ +jwt diff --git a/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/INSTALLER b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/INSTALLER new file mode 100644 index 0000000000000000000000000000000000000000..a1b589e38a32041e49332e5e81c2d363dc418d68 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/INSTALLER @@ -0,0 +1 @@ +pip diff --git a/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/LICENSE b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..261eeb9e9f8b2b4b0d119366dda99c6fd7d35c64 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/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 [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. diff --git a/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/METADATA b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/METADATA new file mode 100644 index 0000000000000000000000000000000000000000..5abe77c7a776232b22cbd4343c1864141e38a49f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/METADATA @@ -0,0 +1,372 @@ +Metadata-Version: 2.2 +Name: diffusers +Version: 0.34.0 +Summary: State-of-the-art diffusion in PyTorch and JAX. +Home-page: https://github.com/huggingface/diffusers +Author: The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/diffusers/graphs/contributors) +Author-email: diffusers@huggingface.co +License: Apache 2.0 License +Keywords: deep learning diffusion jax pytorch stable diffusion audioldm +Classifier: Development Status :: 5 - Production/Stable +Classifier: Intended Audience :: Developers +Classifier: Intended Audience :: Education +Classifier: Intended Audience :: Science/Research +Classifier: License :: OSI Approved :: Apache Software License +Classifier: Operating System :: OS Independent +Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence +Classifier: Programming Language :: Python :: 3 +Classifier: Programming Language :: Python :: 3.8 +Classifier: Programming Language :: Python :: 3.9 +Classifier: Programming Language :: Python :: 3.10 +Requires-Python: >=3.8.0 +Description-Content-Type: text/markdown +License-File: LICENSE +Requires-Dist: importlib_metadata +Requires-Dist: filelock +Requires-Dist: huggingface-hub>=0.27.0 +Requires-Dist: numpy +Requires-Dist: regex!=2019.12.17 +Requires-Dist: requests +Requires-Dist: safetensors>=0.3.1 +Requires-Dist: Pillow +Provides-Extra: quality +Requires-Dist: urllib3<=2.0.0; extra == "quality" +Requires-Dist: isort>=5.5.4; extra == "quality" +Requires-Dist: ruff==0.9.10; extra == "quality" +Requires-Dist: hf-doc-builder>=0.3.0; extra == "quality" +Provides-Extra: docs +Requires-Dist: hf-doc-builder>=0.3.0; extra == "docs" +Provides-Extra: training +Requires-Dist: accelerate>=0.31.0; extra == "training" +Requires-Dist: datasets; extra == "training" +Requires-Dist: protobuf<4,>=3.20.3; extra == "training" +Requires-Dist: tensorboard; extra == "training" +Requires-Dist: Jinja2; extra == "training" +Requires-Dist: peft>=0.15.0; extra == "training" +Provides-Extra: test +Requires-Dist: compel==0.1.8; extra == "test" +Requires-Dist: GitPython<3.1.19; extra == "test" +Requires-Dist: datasets; extra == "test" +Requires-Dist: Jinja2; extra == "test" +Requires-Dist: invisible-watermark>=0.2.0; extra == "test" +Requires-Dist: k-diffusion>=0.0.12; extra == "test" +Requires-Dist: librosa; extra == "test" +Requires-Dist: parameterized; extra == "test" +Requires-Dist: pytest; extra == "test" +Requires-Dist: pytest-timeout; extra == "test" +Requires-Dist: pytest-xdist; extra == "test" +Requires-Dist: requests-mock==1.10.0; extra == "test" +Requires-Dist: safetensors>=0.3.1; extra == "test" +Requires-Dist: sentencepiece!=0.1.92,>=0.1.91; extra == "test" +Requires-Dist: scipy; extra == "test" +Requires-Dist: tiktoken>=0.7.0; extra == "test" +Requires-Dist: torchvision; extra == "test" +Requires-Dist: transformers>=4.41.2; extra == "test" +Requires-Dist: phonemizer; extra == "test" +Provides-Extra: torch +Requires-Dist: torch>=1.4; extra == "torch" +Requires-Dist: accelerate>=0.31.0; extra == "torch" +Provides-Extra: bitsandbytes +Requires-Dist: bitsandbytes>=0.43.3; extra == "bitsandbytes" +Requires-Dist: accelerate>=0.31.0; extra == "bitsandbytes" +Provides-Extra: gguf +Requires-Dist: gguf>=0.10.0; extra == "gguf" +Requires-Dist: accelerate>=0.31.0; extra == "gguf" +Provides-Extra: optimum-quanto +Requires-Dist: optimum_quanto>=0.2.6; extra == "optimum-quanto" +Requires-Dist: accelerate>=0.31.0; extra == "optimum-quanto" +Provides-Extra: torchao +Requires-Dist: torchao>=0.7.0; extra == "torchao" +Requires-Dist: accelerate>=0.31.0; extra == "torchao" +Provides-Extra: flax +Requires-Dist: jax>=0.4.1; extra == "flax" +Requires-Dist: jaxlib>=0.4.1; extra == "flax" +Requires-Dist: flax>=0.4.1; extra == "flax" +Provides-Extra: dev +Requires-Dist: urllib3<=2.0.0; extra == "dev" +Requires-Dist: isort>=5.5.4; extra == "dev" +Requires-Dist: ruff==0.9.10; extra == "dev" +Requires-Dist: hf-doc-builder>=0.3.0; extra == "dev" +Requires-Dist: compel==0.1.8; extra == "dev" +Requires-Dist: GitPython<3.1.19; extra == "dev" +Requires-Dist: datasets; extra == "dev" +Requires-Dist: Jinja2; extra == "dev" +Requires-Dist: invisible-watermark>=0.2.0; extra == "dev" +Requires-Dist: k-diffusion>=0.0.12; extra == "dev" +Requires-Dist: librosa; extra == "dev" +Requires-Dist: parameterized; extra == "dev" +Requires-Dist: pytest; extra == "dev" +Requires-Dist: pytest-timeout; extra == "dev" +Requires-Dist: pytest-xdist; extra == "dev" +Requires-Dist: requests-mock==1.10.0; extra == "dev" +Requires-Dist: safetensors>=0.3.1; extra == "dev" +Requires-Dist: sentencepiece!=0.1.92,>=0.1.91; extra == "dev" +Requires-Dist: scipy; extra == "dev" +Requires-Dist: tiktoken>=0.7.0; extra == "dev" +Requires-Dist: torchvision; extra == "dev" +Requires-Dist: transformers>=4.41.2; extra == "dev" +Requires-Dist: phonemizer; extra == "dev" +Requires-Dist: accelerate>=0.31.0; extra == "dev" +Requires-Dist: datasets; extra == "dev" +Requires-Dist: protobuf<4,>=3.20.3; extra == "dev" +Requires-Dist: tensorboard; extra == "dev" +Requires-Dist: Jinja2; extra == "dev" +Requires-Dist: peft>=0.15.0; extra == "dev" +Requires-Dist: hf-doc-builder>=0.3.0; extra == "dev" +Requires-Dist: torch>=1.4; extra == "dev" +Requires-Dist: accelerate>=0.31.0; extra == "dev" +Requires-Dist: jax>=0.4.1; extra == "dev" +Requires-Dist: jaxlib>=0.4.1; extra == "dev" +Requires-Dist: flax>=0.4.1; extra == "dev" +Dynamic: author +Dynamic: author-email +Dynamic: classifier +Dynamic: description +Dynamic: description-content-type +Dynamic: home-page +Dynamic: keywords +Dynamic: license +Dynamic: provides-extra +Dynamic: requires-dist +Dynamic: requires-python +Dynamic: summary + + + +

+
+ +
+

+

+ GitHub + GitHub release + GitHub release + Contributor Covenant + X account +

+ +🤗 Diffusers is the go-to library for state-of-the-art pretrained diffusion models for generating images, audio, and even 3D structures of molecules. Whether you're looking for a simple inference solution or training your own diffusion models, 🤗 Diffusers is a modular toolbox that supports both. Our library is designed with a focus on [usability over performance](https://huggingface.co/docs/diffusers/conceptual/philosophy#usability-over-performance), [simple over easy](https://huggingface.co/docs/diffusers/conceptual/philosophy#simple-over-easy), and [customizability over abstractions](https://huggingface.co/docs/diffusers/conceptual/philosophy#tweakable-contributorfriendly-over-abstraction). + +🤗 Diffusers offers three core components: + +- State-of-the-art [diffusion pipelines](https://huggingface.co/docs/diffusers/api/pipelines/overview) that can be run in inference with just a few lines of code. +- Interchangeable noise [schedulers](https://huggingface.co/docs/diffusers/api/schedulers/overview) for different diffusion speeds and output quality. +- Pretrained [models](https://huggingface.co/docs/diffusers/api/models/overview) that can be used as building blocks, and combined with schedulers, for creating your own end-to-end diffusion systems. + +## Installation + +We recommend installing 🤗 Diffusers in a virtual environment from PyPI or Conda. For more details about installing [PyTorch](https://pytorch.org/get-started/locally/) and [Flax](https://flax.readthedocs.io/en/latest/#installation), please refer to their official documentation. + +### PyTorch + +With `pip` (official package): + +```bash +pip install --upgrade diffusers[torch] +``` + +With `conda` (maintained by the community): + +```sh +conda install -c conda-forge diffusers +``` + +### Flax + +With `pip` (official package): + +```bash +pip install --upgrade diffusers[flax] +``` + +### Apple Silicon (M1/M2) support + +Please refer to the [How to use Stable Diffusion in Apple Silicon](https://huggingface.co/docs/diffusers/optimization/mps) guide. + +## Quickstart + +Generating outputs is super easy with 🤗 Diffusers. To generate an image from text, use the `from_pretrained` method to load any pretrained diffusion model (browse the [Hub](https://huggingface.co/models?library=diffusers&sort=downloads) for 30,000+ checkpoints): + +```python +from diffusers import DiffusionPipeline +import torch + +pipeline = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16) +pipeline.to("cuda") +pipeline("An image of a squirrel in Picasso style").images[0] +``` + +You can also dig into the models and schedulers toolbox to build your own diffusion system: + +```python +from diffusers import DDPMScheduler, UNet2DModel +from PIL import Image +import torch + +scheduler = DDPMScheduler.from_pretrained("google/ddpm-cat-256") +model = UNet2DModel.from_pretrained("google/ddpm-cat-256").to("cuda") +scheduler.set_timesteps(50) + +sample_size = model.config.sample_size +noise = torch.randn((1, 3, sample_size, sample_size), device="cuda") +input = noise + +for t in scheduler.timesteps: + with torch.no_grad(): + noisy_residual = model(input, t).sample + prev_noisy_sample = scheduler.step(noisy_residual, t, input).prev_sample + input = prev_noisy_sample + +image = (input / 2 + 0.5).clamp(0, 1) +image = image.cpu().permute(0, 2, 3, 1).numpy()[0] +image = Image.fromarray((image * 255).round().astype("uint8")) +image +``` + +Check out the [Quickstart](https://huggingface.co/docs/diffusers/quicktour) to launch your diffusion journey today! + +## How to navigate the documentation + +| **Documentation** | **What can I learn?** | +|---------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| +| [Tutorial](https://huggingface.co/docs/diffusers/tutorials/tutorial_overview) | A basic crash course for learning how to use the library's most important features like using models and schedulers to build your own diffusion system, and training your own diffusion model. | +| [Loading](https://huggingface.co/docs/diffusers/using-diffusers/loading) | Guides for how to load and configure all the components (pipelines, models, and schedulers) of the library, as well as how to use different schedulers. | +| [Pipelines for inference](https://huggingface.co/docs/diffusers/using-diffusers/overview_techniques) | Guides for how to use pipelines for different inference tasks, batched generation, controlling generated outputs and randomness, and how to contribute a pipeline to the library. | +| [Optimization](https://huggingface.co/docs/diffusers/optimization/fp16) | Guides for how to optimize your diffusion model to run faster and consume less memory. | +| [Training](https://huggingface.co/docs/diffusers/training/overview) | Guides for how to train a diffusion model for different tasks with different training techniques. | +## Contribution + +We ❤️ contributions from the open-source community! +If you want to contribute to this library, please check out our [Contribution guide](https://github.com/huggingface/diffusers/blob/main/CONTRIBUTING.md). +You can look out for [issues](https://github.com/huggingface/diffusers/issues) you'd like to tackle to contribute to the library. +- See [Good first issues](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22good+first+issue%22) for general opportunities to contribute +- See [New model/pipeline](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22New+pipeline%2Fmodel%22) to contribute exciting new diffusion models / diffusion pipelines +- See [New scheduler](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22New+scheduler%22) + +Also, say 👋 in our public Discord channel Join us on Discord. We discuss the hottest trends about diffusion models, help each other with contributions, personal projects or just hang out ☕. + + +## Popular Tasks & Pipelines + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
TaskPipeline🤗 Hub
Unconditional Image Generation DDPM google/ddpm-ema-church-256
Text-to-ImageStable Diffusion Text-to-Image stable-diffusion-v1-5/stable-diffusion-v1-5
Text-to-ImageunCLIP kakaobrain/karlo-v1-alpha
Text-to-ImageDeepFloyd IF DeepFloyd/IF-I-XL-v1.0
Text-to-ImageKandinsky kandinsky-community/kandinsky-2-2-decoder
Text-guided Image-to-ImageControlNet lllyasviel/sd-controlnet-canny
Text-guided Image-to-ImageInstructPix2Pix timbrooks/instruct-pix2pix
Text-guided Image-to-ImageStable Diffusion Image-to-Image stable-diffusion-v1-5/stable-diffusion-v1-5
Text-guided Image InpaintingStable Diffusion Inpainting runwayml/stable-diffusion-inpainting
Image VariationStable Diffusion Image Variation lambdalabs/sd-image-variations-diffusers
Super ResolutionStable Diffusion Upscale stabilityai/stable-diffusion-x4-upscaler
Super ResolutionStable Diffusion Latent Upscale stabilityai/sd-x2-latent-upscaler
+ +## Popular libraries using 🧨 Diffusers + +- https://github.com/microsoft/TaskMatrix +- https://github.com/invoke-ai/InvokeAI +- https://github.com/InstantID/InstantID +- https://github.com/apple/ml-stable-diffusion +- https://github.com/Sanster/lama-cleaner +- https://github.com/IDEA-Research/Grounded-Segment-Anything +- https://github.com/ashawkey/stable-dreamfusion +- https://github.com/deep-floyd/IF +- https://github.com/bentoml/BentoML +- https://github.com/bmaltais/kohya_ss +- +14,000 other amazing GitHub repositories 💪 + +Thank you for using us ❤️. + +## Credits + +This library concretizes previous work by many different authors and would not have been possible without their great research and implementations. We'd like to thank, in particular, the following implementations which have helped us in our development and without which the API could not have been as polished today: + +- @CompVis' latent diffusion models library, available [here](https://github.com/CompVis/latent-diffusion) +- @hojonathanho original DDPM implementation, available [here](https://github.com/hojonathanho/diffusion) as well as the extremely useful translation into PyTorch by @pesser, available [here](https://github.com/pesser/pytorch_diffusion) +- @ermongroup's DDIM implementation, available [here](https://github.com/ermongroup/ddim) +- @yang-song's Score-VE and Score-VP implementations, available [here](https://github.com/yang-song/score_sde_pytorch) + +We also want to thank @heejkoo for the very helpful overview of papers, code and resources on diffusion models, available [here](https://github.com/heejkoo/Awesome-Diffusion-Models) as well as @crowsonkb and @rromb for useful discussions and insights. + +## Citation + +```bibtex +@misc{von-platen-etal-2022-diffusers, + author = {Patrick von Platen and Suraj Patil and Anton Lozhkov and Pedro Cuenca and Nathan Lambert and Kashif Rasul and Mishig Davaadorj and Dhruv Nair and Sayak Paul and William Berman and Yiyi Xu and Steven Liu and Thomas Wolf}, + title = {Diffusers: State-of-the-art diffusion models}, + year = {2022}, + publisher = {GitHub}, + journal = {GitHub repository}, + howpublished = {\url{https://github.com/huggingface/diffusers}} +} +``` diff --git a/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/RECORD b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/RECORD new file mode 100644 index 0000000000000000000000000000000000000000..2065e508ceb59e540ff8fa8be1c0b042147a0c67 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/RECORD @@ -0,0 +1,1273 @@ +../../../bin/diffusers-cli,sha256=F0-KZOMEcKsoGXC_Y_BQwRcKF0b8z3YQHhxLuFhDhwE,249 +diffusers-0.34.0.dist-info/INSTALLER,sha256=zuuue4knoyJ-UwPPXg8fezS7VCrXJQrAP7zeNuwvFQg,4 +diffusers-0.34.0.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357 +diffusers-0.34.0.dist-info/METADATA,sha256=9cVvp1z0cfI46dpVFR5F4UK3F2rL9hWyNLol9CH1r_w,20120 +diffusers-0.34.0.dist-info/RECORD,, +diffusers-0.34.0.dist-info/REQUESTED,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0 +diffusers-0.34.0.dist-info/WHEEL,sha256=In9FTNxeP60KnTkGw7wk6mJPYd_dQSjEZmXdBdMCI-8,91 +diffusers-0.34.0.dist-info/entry_points.txt,sha256=_1bvshKV_6_b63_FAkcUs9W6tUKGeIoQ3SHEZsovEWs,72 +diffusers-0.34.0.dist-info/top_level.txt,sha256=axJl2884vMSvhzrFrSoht36QXA_6gZN9cKtg4xOO72o,10 +diffusers/__init__.py,sha256=pDjLfKQ_7dmCkQT0crSvex7rBfURdqh0ZnjZWiLOhBE,46619 +diffusers/__pycache__/__init__.cpython-310.pyc,, +diffusers/__pycache__/callbacks.cpython-310.pyc,, +diffusers/__pycache__/configuration_utils.cpython-310.pyc,, +diffusers/__pycache__/dependency_versions_check.cpython-310.pyc,, +diffusers/__pycache__/dependency_versions_table.cpython-310.pyc,, +diffusers/__pycache__/image_processor.cpython-310.pyc,, +diffusers/__pycache__/optimization.cpython-310.pyc,, +diffusers/__pycache__/training_utils.cpython-310.pyc,, +diffusers/__pycache__/video_processor.cpython-310.pyc,, +diffusers/callbacks.py,sha256=wmLFSUrnEZ9gz6gay2pRWfha5cErTSAbB9PK482_KGY,8749 +diffusers/commands/__init__.py,sha256=KQXlWjcjH6qmwD5B6be8kpniSOPbyWZPTghdXrNquXU,920 +diffusers/commands/__pycache__/__init__.cpython-310.pyc,, +diffusers/commands/__pycache__/diffusers_cli.cpython-310.pyc,, +diffusers/commands/__pycache__/env.cpython-310.pyc,, +diffusers/commands/__pycache__/fp16_safetensors.cpython-310.pyc,, +diffusers/commands/diffusers_cli.py,sha256=bCLOzLKq2s8wj9AifPD6yEbT0gkpmgDM5Kvoiw9DPvQ,1317 +diffusers/commands/env.py,sha256=IZ2e2id8ZjNQ-BNEG3qt58TyEsM6Qs337kL5KdZmqZ0,6224 +diffusers/commands/fp16_safetensors.py,sha256=w6r3FnfiLHKQMpiVcs3vWIZIVENh_Z8um1nPN1HRqus,5423 +diffusers/configuration_utils.py,sha256=bPpagE2tmsaQjoksYORdOaSiMaSgaxoCx7i4iDwlA-g,34480 +diffusers/dependency_versions_check.py,sha256=PcT_deWuvIKrNkjkCnQKi0ZTWCl77tHC02lhttbqQHM,1271 +diffusers/dependency_versions_table.py,sha256=BKcEswVLknflk4QNnPDd9Q7HYAQd7qgj7h6dAtVuJDA,1766 +diffusers/experimental/__init__.py,sha256=0C9ExG0XYiGZuzFJkZuJ53K6Ix5ylF2kWe4PGASchtY,38 +diffusers/experimental/__pycache__/__init__.cpython-310.pyc,, +diffusers/experimental/rl/__init__.py,sha256=Gcoznw9rYjfMvswH0seXekKYDAAN1YXXxZ-RWMdzvrE,57 +diffusers/experimental/rl/__pycache__/__init__.cpython-310.pyc,, +diffusers/experimental/rl/__pycache__/value_guided_sampling.cpython-310.pyc,, +diffusers/experimental/rl/value_guided_sampling.py,sha256=a1F7YC--u1OgoGC8WGgi2pDUhYBI6XYnF2-RUzPD3dI,6033 +diffusers/hooks/__init__.py,sha256=sImnLjLQ2WnTbhCNu_NcawZJk4MUamWMTduczLqCJ7U,439 +diffusers/hooks/__pycache__/__init__.cpython-310.pyc,, +diffusers/hooks/__pycache__/faster_cache.cpython-310.pyc,, +diffusers/hooks/__pycache__/group_offloading.cpython-310.pyc,, +diffusers/hooks/__pycache__/hooks.cpython-310.pyc,, +diffusers/hooks/__pycache__/layerwise_casting.cpython-310.pyc,, +diffusers/hooks/__pycache__/pyramid_attention_broadcast.cpython-310.pyc,, +diffusers/hooks/faster_cache.py,sha256=Xe4m3fUqJItffff5F3g73uS04cd-scRDYiO4GXAp3fo,34950 +diffusers/hooks/group_offloading.py,sha256=T7R7eUSvjh0vp6NIp3-XFBVdNye23ACr5Hv-rPRDdV8,43294 +diffusers/hooks/hooks.py,sha256=rmbl-KsdsYkkoOCY7ySJMXNRvIUtyPmVGAs94h3Vrnc,8795 +diffusers/hooks/layerwise_casting.py,sha256=7vafpuCnKWKWN8hM0BehMMLNIBI8yIkdUge08cuLWrA,10589 +diffusers/hooks/pyramid_attention_broadcast.py,sha256=jGj67TTlk5H0WhtKep8ZkF9EckomPld6P7dQlyBL5kc,15760 +diffusers/image_processor.py,sha256=6qop7filTHdq3L-ZGVWmJMn0-0s4Bg7W-AaAFBNdO0Q,52868 +diffusers/loaders/__init__.py,sha256=6JRHLAuQPsNjEI38uTklH-xy4BvBzUqx6t-mf7ZXLcc,5280 +diffusers/loaders/__pycache__/__init__.cpython-310.pyc,, +diffusers/loaders/__pycache__/ip_adapter.cpython-310.pyc,, +diffusers/loaders/__pycache__/lora_base.cpython-310.pyc,, +diffusers/loaders/__pycache__/lora_conversion_utils.cpython-310.pyc,, +diffusers/loaders/__pycache__/lora_pipeline.cpython-310.pyc,, +diffusers/loaders/__pycache__/peft.cpython-310.pyc,, +diffusers/loaders/__pycache__/single_file.cpython-310.pyc,, +diffusers/loaders/__pycache__/single_file_model.cpython-310.pyc,, +diffusers/loaders/__pycache__/single_file_utils.cpython-310.pyc,, +diffusers/loaders/__pycache__/textual_inversion.cpython-310.pyc,, +diffusers/loaders/__pycache__/transformer_flux.cpython-310.pyc,, +diffusers/loaders/__pycache__/transformer_sd3.cpython-310.pyc,, +diffusers/loaders/__pycache__/unet.cpython-310.pyc,, +diffusers/loaders/__pycache__/unet_loader_utils.cpython-310.pyc,, +diffusers/loaders/__pycache__/utils.cpython-310.pyc,, +diffusers/loaders/ip_adapter.py,sha256=TB1HLgy0u6rKezWbPhXNQttmUNpq7nMNhWnW-ljlTIs,45068 +diffusers/loaders/lora_base.py,sha256=-DCuq6UAGlw0-fmP5vwYZzHNDhZN-NmyuB7VTDaUjEE,43666 +diffusers/loaders/lora_conversion_utils.py,sha256=h2nb6eL8JrRkTNDemOceW2jFPda-P0I95_yPvBk7gzs,90021 +diffusers/loaders/lora_pipeline.py,sha256=02HBNwDA9xMIQkVKeogbM8U7a_lfZvkpWON-iRXQoZ4,289014 +diffusers/loaders/peft.py,sha256=rFBo4wYogICBEsJC6y8euB3TSauI4sNC0QNe828pZws,38003 +diffusers/loaders/single_file.py,sha256=NyDO7tx2ASAENP5rXsWKxpOD320L24wl4QqpPixYYLE,25003 +diffusers/loaders/single_file_model.py,sha256=ZysRXed49QWEqZM2Uvz2jaNZNWIgxwn0BiAcD90Bm9Q,20083 +diffusers/loaders/single_file_utils.py,sha256=pas9vPmcPM0rH51RvrrUCQVffdIGcYoGobEM_tPolD0,156096 +diffusers/loaders/textual_inversion.py,sha256=Yh4Mw-3EAUWaQpDXptIMuNbyAYhNRir-irqoHSPvOx8,26925 +diffusers/loaders/transformer_flux.py,sha256=PKutmTlDU_T7DlssyGgSExeL8f6gw9lRz3JlnneWvSY,7915 +diffusers/loaders/transformer_sd3.py,sha256=LdPxicRUw_YelTZgJ_E9R4C2AlXpije_fgEG08kuWyQ,8482 +diffusers/loaders/unet.py,sha256=TX3a1PHxs0-IktVTU8mvoQZS_ZLQwYd-dbWyiSNecP0,45145 +diffusers/loaders/unet_loader_utils.py,sha256=tVuE5Xd6PiC_GIl_3SHfa_jDxfgn9rgqEZnYUjTCwBI,6220 +diffusers/loaders/utils.py,sha256=tknQHwXnDlc7piArCSDZjFFdgJI3DY6qfffsZlIxLnY,2423 +diffusers/models/__init__.py,sha256=dJpGFKbTJcRRcc7LeYsADK9sv7JGVqgFuY3LJxHYcJU,10601 +diffusers/models/__pycache__/__init__.cpython-310.pyc,, +diffusers/models/__pycache__/activations.cpython-310.pyc,, +diffusers/models/__pycache__/adapter.cpython-310.pyc,, +diffusers/models/__pycache__/attention.cpython-310.pyc,, +diffusers/models/__pycache__/attention_flax.cpython-310.pyc,, +diffusers/models/__pycache__/attention_processor.cpython-310.pyc,, +diffusers/models/__pycache__/auto_model.cpython-310.pyc,, +diffusers/models/__pycache__/cache_utils.cpython-310.pyc,, +diffusers/models/__pycache__/controlnet.cpython-310.pyc,, +diffusers/models/__pycache__/controlnet_flux.cpython-310.pyc,, +diffusers/models/__pycache__/controlnet_sd3.cpython-310.pyc,, +diffusers/models/__pycache__/controlnet_sparsectrl.cpython-310.pyc,, +diffusers/models/__pycache__/downsampling.cpython-310.pyc,, +diffusers/models/__pycache__/embeddings.cpython-310.pyc,, +diffusers/models/__pycache__/embeddings_flax.cpython-310.pyc,, +diffusers/models/__pycache__/lora.cpython-310.pyc,, +diffusers/models/__pycache__/model_loading_utils.cpython-310.pyc,, +diffusers/models/__pycache__/modeling_flax_pytorch_utils.cpython-310.pyc,, +diffusers/models/__pycache__/modeling_flax_utils.cpython-310.pyc,, +diffusers/models/__pycache__/modeling_outputs.cpython-310.pyc,, +diffusers/models/__pycache__/modeling_pytorch_flax_utils.cpython-310.pyc,, +diffusers/models/__pycache__/modeling_utils.cpython-310.pyc,, +diffusers/models/__pycache__/normalization.cpython-310.pyc,, +diffusers/models/__pycache__/resnet.cpython-310.pyc,, +diffusers/models/__pycache__/resnet_flax.cpython-310.pyc,, +diffusers/models/__pycache__/upsampling.cpython-310.pyc,, +diffusers/models/__pycache__/vae_flax.cpython-310.pyc,, +diffusers/models/__pycache__/vq_model.cpython-310.pyc,, +diffusers/models/activations.py,sha256=qxdn6OROfUvxyxgpm6M2VDKeJxH6mDsUI_xP4S3iw6s,6511 +diffusers/models/adapter.py,sha256=NDnqBqD53fg2fWWHt-LUHDyhuH6J-4R7PoStr2ggp-4,24507 +diffusers/models/attention.py,sha256=k3lQlDPoEUtuj-HBkxcMeQslATqCEKRQc_v8X6LkId8,53548 +diffusers/models/attention_flax.py,sha256=NJTCmsALDnRScOy2EG7r6fZGXaNrJGBVTHQxAluWZEs,20392 +diffusers/models/attention_processor.py,sha256=8B2iWzgEYfmgYcDyf-_07N0W_Dm7VsFQrjt_9EqUNcQ,263926 +diffusers/models/auto_model.py,sha256=7bLn0pgTf5TG2txeRLFgUgtYGmrylZB9eZZI6jC401s,11155 +diffusers/models/autoencoders/__init__.py,sha256=uEgP7tgXHRQdw8BSbJ8pEgqk7gnzyHsEgqKGt01txlk,854 +diffusers/models/autoencoders/__pycache__/__init__.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/autoencoder_asym_kl.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/autoencoder_dc.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/autoencoder_kl.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/autoencoder_kl_allegro.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/autoencoder_kl_cogvideox.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/autoencoder_kl_cosmos.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/autoencoder_kl_hunyuan_video.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/autoencoder_kl_ltx.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/autoencoder_kl_magvit.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/autoencoder_kl_mochi.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/autoencoder_kl_temporal_decoder.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/autoencoder_kl_wan.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/autoencoder_oobleck.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/autoencoder_tiny.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/consistency_decoder_vae.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/vae.cpython-310.pyc,, +diffusers/models/autoencoders/__pycache__/vq_model.cpython-310.pyc,, +diffusers/models/autoencoders/autoencoder_asym_kl.py,sha256=_67adzZxCMsV8N7s0JzOCSOz4_Kewb49c0CuA2KZS-I,7788 +diffusers/models/autoencoders/autoencoder_dc.py,sha256=Fvi2jKJQqmhyWgH9XZ58nJzq9WKVlYttL-l8P97I_aU,30503 +diffusers/models/autoencoders/autoencoder_kl.py,sha256=lHKSeG3m5tpfkDdcS0fg1UMqaZ4NVITulBjUJZ602_o,25079 +diffusers/models/autoencoders/autoencoder_kl_allegro.py,sha256=N1tLNJjqtCOGPQREU647RA5LdUsl4fgK2Uiii-mYx2c,45152 +diffusers/models/autoencoders/autoencoder_kl_cogvideox.py,sha256=YDCPc0NugDpHkw4JKBkrNHd0sOgxBbM0QXgovnAToTU,60406 +diffusers/models/autoencoders/autoencoder_kl_cosmos.py,sha256=nHasHRysFvmAbWteH1IscQ7z3rQGSoA15udZPWmSGLs,53706 +diffusers/models/autoencoders/autoencoder_kl_hunyuan_video.py,sha256=T4ijgoUTrJou5DMnUBkaCnPA7pMZuxutTSHxE31Y3zk,45481 +diffusers/models/autoencoders/autoencoder_kl_ltx.py,sha256=GQEIy5muXUndwYRH4jlCFUQanbgiUCHLNaEnVcL286g,64433 +diffusers/models/autoencoders/autoencoder_kl_magvit.py,sha256=6jrWs8thq0DERB2zOsGBIQHPtFFyjYuuekFYM-fhE9I,45098 +diffusers/models/autoencoders/autoencoder_kl_mochi.py,sha256=EJBSXMnIE4clpodfDKy4tffZbQkC15iYGmlZRowDQHA,46811 +diffusers/models/autoencoders/autoencoder_kl_temporal_decoder.py,sha256=qJ52TwMtQQz7QtczEsFLKXVceS7RLDRSvMMZI-RsjIY,14778 +diffusers/models/autoencoders/autoencoder_kl_wan.py,sha256=1xtFoijZaHBb1040yY9JO1aTGBOGce8yXmjLF5xcOrE,41721 +diffusers/models/autoencoders/autoencoder_oobleck.py,sha256=HV30_00e21-mts-JdT9K9Q-s8Uerk_cMD9_OX9uaNtU,17085 +diffusers/models/autoencoders/autoencoder_tiny.py,sha256=6kLUGAR0fDx7fwy5BiRLaUzkqJg9VQFKUJyisloxF4M,15853 +diffusers/models/autoencoders/consistency_decoder_vae.py,sha256=3uvrzIR2vEgj0etcRvUiBw8YmCQ5RfQ6RdhX4LcCoeM,19761 +diffusers/models/autoencoders/vae.py,sha256=XXm10FfQ9wLhozWomRO2ldfcb6qR85ePF-ShUxYq-6Y,32733 +diffusers/models/autoencoders/vq_model.py,sha256=wCvrranGbiCR8s4Qhq5qNSy_1BP9PGWvWyxvWJEHA8A,7912 +diffusers/models/cache_utils.py,sha256=A88pJJikRVSXZgiyyx4q36xdItvXvmA1DqDT6F1eBXo,4232 +diffusers/models/controlnet.py,sha256=RG8RSUtkCk77Dxw6e1nc2o6vdikTVKtskxgALI7ijsg,5774 +diffusers/models/controlnet_flux.py,sha256=bVej9TQJhszOMJOtHqzh2kwK1PLseuwiGSYdbT6b7f4,3442 +diffusers/models/controlnet_sd3.py,sha256=5cPugxXJrWroX4bjIG_DQGMlOWiFlwl2sMZz6pTSLTc,3377 +diffusers/models/controlnet_sparsectrl.py,sha256=ewfJCGXU-bvthUH2OP93RtMI3zZFNfnioYs8YSa-tmo,5904 +diffusers/models/controlnets/__init__.py,sha256=oKACrY23NZqmM0Hbl-S_nUuDFyQ4NesJMoKBYPzeK0Y,1058 +diffusers/models/controlnets/__pycache__/__init__.cpython-310.pyc,, +diffusers/models/controlnets/__pycache__/controlnet.cpython-310.pyc,, +diffusers/models/controlnets/__pycache__/controlnet_flax.cpython-310.pyc,, +diffusers/models/controlnets/__pycache__/controlnet_flux.cpython-310.pyc,, +diffusers/models/controlnets/__pycache__/controlnet_hunyuan.cpython-310.pyc,, +diffusers/models/controlnets/__pycache__/controlnet_sana.cpython-310.pyc,, +diffusers/models/controlnets/__pycache__/controlnet_sd3.cpython-310.pyc,, +diffusers/models/controlnets/__pycache__/controlnet_sparsectrl.cpython-310.pyc,, +diffusers/models/controlnets/__pycache__/controlnet_union.cpython-310.pyc,, +diffusers/models/controlnets/__pycache__/controlnet_xs.cpython-310.pyc,, +diffusers/models/controlnets/__pycache__/multicontrolnet.cpython-310.pyc,, +diffusers/models/controlnets/__pycache__/multicontrolnet_union.cpython-310.pyc,, +diffusers/models/controlnets/controlnet.py,sha256=dzIASI0h_bNeSjRi3BQTW2K4Jbo8-eQfJ7F5PeOzpsA,43149 +diffusers/models/controlnets/controlnet_flax.py,sha256=DekqQPnow0gSWDCUlIwcP6ezT16ppiYUGFVn1Ixi0zM,16715 +diffusers/models/controlnets/controlnet_flux.py,sha256=u_kd28VbTsro-FMc1j6rrPoTXvzQOoXkwPz6VzxPqx0,22889 +diffusers/models/controlnets/controlnet_hunyuan.py,sha256=UCyQL-GtHIQEeEu6bP04eIm-DDhsqG92I3hAJJE_m2g,16928 +diffusers/models/controlnets/controlnet_sana.py,sha256=2rYVE1EJAE-BavlAWtkkX4zcVk3nrP_BJkG8ewfOYXc,12519 +diffusers/models/controlnets/controlnet_sd3.py,sha256=E-r-_y2M3sN3sT4rQBm3CYKcTNrYTEvnIVUDqpjzj7c,23227 +diffusers/models/controlnets/controlnet_sparsectrl.py,sha256=ui450jzQEfI-OUU90nrmb1Oud4lcTmw7aiVMLtF3uTE,38357 +diffusers/models/controlnets/controlnet_union.py,sha256=mOK2L2xYcGivPqixZbqsj7Ev9JeTUY5BAG4Uv5f39SA,41604 +diffusers/models/controlnets/controlnet_xs.py,sha256=aF13vgEAeCf1UygSzuiktKzA_OrGXPiycRg3YzHKOHg,85416 +diffusers/models/controlnets/multicontrolnet.py,sha256=VoccQ64iwREXC04DseYLZjub2bdph9y8tBdga4br1XM,9332 +diffusers/models/controlnets/multicontrolnet_union.py,sha256=o_5BMCsvXUehw7Jni0Fn6e_qNm_66XnAyN4Ull4txrE,10164 +diffusers/models/downsampling.py,sha256=ss0au04TEj_eMz1vCgi20a1ZGoUEOLgC1zxeQGTN85s,15829 +diffusers/models/embeddings.py,sha256=8BXCFaefAd7LDMqX438v1_sk7i3Qezz8cnlp1hDZzPg,103970 +diffusers/models/embeddings_flax.py,sha256=FMuPD4wNU0wUXOaih81VcFjkKgOg4lyWi7hYIlmnYT4,4361 +diffusers/models/lora.py,sha256=kSZzH-taCGJCSH9PHYEcS1M1LXUGUpIti-taQxRRnS4,18851 +diffusers/models/model_loading_utils.py,sha256=khiNv_O8xz1VG6BvL_tnEzK0FfIpNn_6vQf9nyNrzKQ,21712 +diffusers/models/modeling_flax_pytorch_utils.py,sha256=f-j9Y-AhcrRp9UvLldjZOHPYBluMC5BXwxmAi6gS1rA,5332 +diffusers/models/modeling_flax_utils.py,sha256=j7BYsBBzYlUO7FUlWXs3Y3bWCTo8gtBL0kU0atuvyN4,26954 +diffusers/models/modeling_outputs.py,sha256=XH3sJO34MRW6UuWqqKo05mVqxGSBFRazpap_-YLwO2I,1042 +diffusers/models/modeling_pytorch_flax_utils.py,sha256=Hz5IoBV0vygRekVw8OZc9Jji22gyuEZ84uAWzXpYLvs,6973 +diffusers/models/modeling_utils.py,sha256=p5_UPBrCXIFQGt_Btarq4VN0E_7EO7z5Ja0yFQHsaFI,85234 +diffusers/models/normalization.py,sha256=Hke3BXA1GYAVBKrkJi7X8A0DLT6_NqP2dv2Q7lZFTRU,24616 +diffusers/models/resnet.py,sha256=DIFQHFSh8F0NcrAGjvxaca1fK8ZoArlpTtkTTLm50VU,32254 +diffusers/models/resnet_flax.py,sha256=aXKT-L0yMwFZz-xhSsbLsB5n7M7bNOE2se6RJhPI9Tk,4021 +diffusers/models/transformers/__init__.py,sha256=eSBzRwfhMlec09pYQcMMoH70-ZFn18h-6ZWSoumt9KA,2077 +diffusers/models/transformers/__pycache__/__init__.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/auraflow_transformer_2d.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/cogvideox_transformer_3d.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/consisid_transformer_3d.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/dit_transformer_2d.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/dual_transformer_2d.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/hunyuan_transformer_2d.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/latte_transformer_3d.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/lumina_nextdit2d.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/pixart_transformer_2d.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/prior_transformer.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/sana_transformer.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/stable_audio_transformer.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/t5_film_transformer.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_2d.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_allegro.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_chroma.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_cogview3plus.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_cogview4.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_cosmos.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_easyanimate.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_flux.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_hidream_image.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_hunyuan_video.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_hunyuan_video_framepack.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_ltx.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_lumina2.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_mochi.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_omnigen.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_sd3.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_temporal.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_wan.cpython-310.pyc,, +diffusers/models/transformers/__pycache__/transformer_wan_vace.cpython-310.pyc,, +diffusers/models/transformers/auraflow_transformer_2d.py,sha256=Lnk2i4c0Ay3SZZXiRHIsTvL-IuTyxIhZvyYMK2X0unU,23439 +diffusers/models/transformers/cogvideox_transformer_3d.py,sha256=gCZBGA7R2kFhf0m-nLxg3eISapfABKxVmzWN09J5rCA,22827 +diffusers/models/transformers/consisid_transformer_3d.py,sha256=ZT4plc4SnCJZczuKVhFTjws_3gPU6kCxm0-lIwAQA14,36023 +diffusers/models/transformers/dit_transformer_2d.py,sha256=4c-KTXuQagfdIr0m6VoysBYaNbAOTGlEt_AAoAfhoI0,10544 +diffusers/models/transformers/dual_transformer_2d.py,sha256=4iU7Ljh9_HQa9uiNwwW2xa7FqZOiEdw2TrTBuWdmR4M,7711 +diffusers/models/transformers/hunyuan_transformer_2d.py,sha256=OKYOjAse6jQn0eeimtPgQ5vhnV7720rznzeZWdVvQoA,24311 +diffusers/models/transformers/latte_transformer_3d.py,sha256=30LmMFjvNQyUKB3DHoVQfFph9wQP5i6BWIcS1zWdVhg,15686 +diffusers/models/transformers/lumina_nextdit2d.py,sha256=THVMjGaMNqvZDy9RXusVMH4lTR-e_q3ns6Vmjq5JZf4,14486 +diffusers/models/transformers/pixart_transformer_2d.py,sha256=K-qrAK5OqbgkZzoPqFZz8KiqKTMSyHCjoFOiCbfJ4hU,20940 +diffusers/models/transformers/prior_transformer.py,sha256=PHWVtGDOIo29k7xyb8BukUFbs-yT9s6Vr4Y8hcgYjo0,17467 +diffusers/models/transformers/sana_transformer.py,sha256=pqXzq-npEsee5ateLL4PS2oHsx8Rz_x36VcSyKWNcYg,25256 +diffusers/models/transformers/stable_audio_transformer.py,sha256=ikFDgj4O5rEk2Nxp3sL9RH76MLM5vzQnfMrR0Donlro,18653 +diffusers/models/transformers/t5_film_transformer.py,sha256=YicaHEHbkOkh7HLjzj8jzIjLA9MobcolI_3ZyCKETCE,16040 +diffusers/models/transformers/transformer_2d.py,sha256=9yiEGNq4poMiHaWIh3OfsfB2dsRKlcIqgTNDW_mV3vY,28305 +diffusers/models/transformers/transformer_allegro.py,sha256=b8fdklklJlmepbUgEwpn83L8AvmhycZao17_-sK2Ehg,17186 +diffusers/models/transformers/transformer_chroma.py,sha256=5t8b4LuAt6DTlxM6FFVi8SFzOON2_ueT67Uo-hR2oIo,31322 +diffusers/models/transformers/transformer_cogview3plus.py,sha256=shbvxhjREErZRCl2T16JZC2gnm39PIG6zgyAgMTDNLo,15811 +diffusers/models/transformers/transformer_cogview4.py,sha256=sK5RUMTzQsxMayJ8ZDxnfqTBN7TeYcFaQjO-n1-BTzk,33638 +diffusers/models/transformers/transformer_cosmos.py,sha256=o-iHkWwVOjudOAF0uLZfcCDpKElj34T9uXfycnPm4FA,24311 +diffusers/models/transformers/transformer_easyanimate.py,sha256=Z5innEWRbaOVoC16o-h-uobbYI6tPqzFG-S2H_Dohgo,22001 +diffusers/models/transformers/transformer_flux.py,sha256=LN86DeAqnhX5QxdxYuyml1EvYk9-gpoNowLoMERbKIc,23502 +diffusers/models/transformers/transformer_hidream_image.py,sha256=ciuyPRprws2K77apFzuLHmDZADOJ3AzAwsVze3FmR_E,39322 +diffusers/models/transformers/transformer_hunyuan_video.py,sha256=HN2zgdACpgw29JvZqnQS2X6xMcEX_LAIKRzSs0F6X28,46897 +diffusers/models/transformers/transformer_hunyuan_video_framepack.py,sha256=p9RgBqN5SWWqQU_wvEmaPZ2W5sCsYZoiGblqFKg6GpM,18698 +diffusers/models/transformers/transformer_ltx.py,sha256=iJQ1tR6Ihof0Rzk0-UsO57QnQNJ05xq-ICkVykbJr_I,18854 +diffusers/models/transformers/transformer_lumina2.py,sha256=ACu9X7vatGMoiKSZKkEOcbEQjIvZwmqZ6nmHaSkD9Wo,22075 +diffusers/models/transformers/transformer_mochi.py,sha256=FbNpuQR3MrH7I7CU8tJ_8Nf00Q2VB5hNUpqwuYXS50Y,18521 +diffusers/models/transformers/transformer_omnigen.py,sha256=nxmuNqRXRG51Mw2G8BNu-DZrRKWMDc3G99DFgPS9yZA,20029 +diffusers/models/transformers/transformer_sd3.py,sha256=2Aw1Di240iPngdlg_TZewlgi4WT2fGAOpB5XFa3RmO0,19232 +diffusers/models/transformers/transformer_temporal.py,sha256=GMn5WUbWWX7ZvyqVhO12g6bID7dnrMndYrDN-UZEI0Q,16812 +diffusers/models/transformers/transformer_wan.py,sha256=nUSN30izYESv4ETeL1LD6SkI0G9zN0qK859hd717mf8,20144 +diffusers/models/transformers/transformer_wan_vace.py,sha256=K_n6VSeAxVgl6p0HufXt4JTeYCcc_WvN4eX5MO41gr4,16866 +diffusers/models/unets/__init__.py,sha256=srYFA7zEcDY7LxyUB2jz3TdRgsLz8elrWCpT6Y4YXuU,695 +diffusers/models/unets/__pycache__/__init__.cpython-310.pyc,, +diffusers/models/unets/__pycache__/unet_1d.cpython-310.pyc,, +diffusers/models/unets/__pycache__/unet_1d_blocks.cpython-310.pyc,, +diffusers/models/unets/__pycache__/unet_2d.cpython-310.pyc,, +diffusers/models/unets/__pycache__/unet_2d_blocks.cpython-310.pyc,, +diffusers/models/unets/__pycache__/unet_2d_blocks_flax.cpython-310.pyc,, +diffusers/models/unets/__pycache__/unet_2d_condition.cpython-310.pyc,, +diffusers/models/unets/__pycache__/unet_2d_condition_flax.cpython-310.pyc,, +diffusers/models/unets/__pycache__/unet_3d_blocks.cpython-310.pyc,, +diffusers/models/unets/__pycache__/unet_3d_condition.cpython-310.pyc,, +diffusers/models/unets/__pycache__/unet_i2vgen_xl.cpython-310.pyc,, +diffusers/models/unets/__pycache__/unet_kandinsky3.cpython-310.pyc,, +diffusers/models/unets/__pycache__/unet_motion_model.cpython-310.pyc,, +diffusers/models/unets/__pycache__/unet_spatio_temporal_condition.cpython-310.pyc,, +diffusers/models/unets/__pycache__/unet_stable_cascade.cpython-310.pyc,, +diffusers/models/unets/__pycache__/uvit_2d.cpython-310.pyc,, +diffusers/models/unets/unet_1d.py,sha256=tmSBsH3cPzNj9xCQt1zrV2lApufLEB2xiFTcXulE1Wo,10853 +diffusers/models/unets/unet_1d_blocks.py,sha256=LIuM8MwkcJ0n8wYwS6FBGYSMfl9wYv0-TbQ6FHO6A7k,26829 +diffusers/models/unets/unet_2d.py,sha256=ue2Xbgmra55kSf6D7sq4b_ZCDr4ZmarrZiGWT4qNNXk,16905 +diffusers/models/unets/unet_2d_blocks.py,sha256=r7e-UDzbBRdbx4Iusu-fNGmSm70Vf2YkaDRihltn12w,141758 +diffusers/models/unets/unet_2d_blocks_flax.py,sha256=RWQ-c2gOHFD99nbtWN2TTGvO2khRWNJ6BoQYDg-hqjs,15624 +diffusers/models/unets/unet_2d_condition.py,sha256=zzRzpVkxmP73brYE_QI4n4JK4K6X7ADhkBv_mSiTIMk,67030 +diffusers/models/unets/unet_2d_condition_flax.py,sha256=62kO3PaPNfcPvWbOIL4nws7m_SZ2j4_GMAA1lXVPR4I,22289 +diffusers/models/unets/unet_3d_blocks.py,sha256=OyEIov5NMa_T9kIWzcMoMfmKZVcgDRbe6vH1Wb4ar9o,51676 +diffusers/models/unets/unet_3d_condition.py,sha256=_OscxwREczYttMFghevnaGu-LnIZLlV63LiS8HszdkU,34379 +diffusers/models/unets/unet_i2vgen_xl.py,sha256=BcsJO2K_GDZV3aPt9-VDsJN9hD07xAkFCVf5iE8PA2A,32505 +diffusers/models/unets/unet_kandinsky3.py,sha256=B9kql5-5U7rG7sbZg6Hy49P-4MusRQDwZT78fc2v6FY,20584 +diffusers/models/unets/unet_motion_model.py,sha256=5ct4afFiyEQzq-GhDHbO1HcV0rj-YvaM3gr7jfMpXGo,99370 +diffusers/models/unets/unet_spatio_temporal_condition.py,sha256=LsBj9J6nltqs2vrkIt-OT4weL1GUxJMmcNx3wTMM-vY,23259 +diffusers/models/unets/unet_stable_cascade.py,sha256=6fgr1ksDUhAe54xd_4uRHEqbeGzlVnd7R80GyKAG9_4,27229 +diffusers/models/unets/uvit_2d.py,sha256=0cV3yDVyC1KA9bAMnhcHGGVbT7n9bFl09gX61bcQLeo,17227 +diffusers/models/upsampling.py,sha256=7oUwSN_b8Wsl3nTuxWgZyXiIid5FOBnTpdeGwS64Xfk,19573 +diffusers/models/vae_flax.py,sha256=862gk7zaRCWBHNY1ZDIRJl7rnoHLWJBJyt6wN4anZ4w,31950 +diffusers/models/vq_model.py,sha256=J6dVRtVAQNUtHbl5LY5KjHYYsab55AbuoATHQttDGX4,1524 +diffusers/optimization.py,sha256=CAFI9pabb3C2KD3y_LCr3o_dBISPU_dfDKODtzbkdvs,14741 +diffusers/pipelines/__init__.py,sha256=5tYzTmjhNadauLrwT7LYyNcESP4MQAtFKVPsqpUzUck,33270 +diffusers/pipelines/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/__pycache__/auto_pipeline.cpython-310.pyc,, +diffusers/pipelines/__pycache__/free_init_utils.cpython-310.pyc,, +diffusers/pipelines/__pycache__/free_noise_utils.cpython-310.pyc,, +diffusers/pipelines/__pycache__/onnx_utils.cpython-310.pyc,, +diffusers/pipelines/__pycache__/pipeline_flax_utils.cpython-310.pyc,, +diffusers/pipelines/__pycache__/pipeline_loading_utils.cpython-310.pyc,, +diffusers/pipelines/__pycache__/pipeline_utils.cpython-310.pyc,, +diffusers/pipelines/__pycache__/transformers_loading_utils.cpython-310.pyc,, +diffusers/pipelines/allegro/__init__.py,sha256=T1MLZgDf8Fhh6YunF8a4Ta6NNIqneWsJIvmBhiy1ABM,1290 +diffusers/pipelines/allegro/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/allegro/__pycache__/pipeline_allegro.cpython-310.pyc,, +diffusers/pipelines/allegro/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/allegro/pipeline_allegro.py,sha256=6h-lqIRh_nktZAZeorkOdeMwuaS5UCXskIAVrxhQ9nw,44067 +diffusers/pipelines/allegro/pipeline_output.py,sha256=Q2W_16pT5o5xbzSrCKO8AB3IweqAyvv6BGTlPwAUNhE,722 +diffusers/pipelines/amused/__init__.py,sha256=pzqLeLosNQ29prMLhTxvPpmoIDPB3OFMQMlErOIRkmI,1793 +diffusers/pipelines/amused/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/amused/__pycache__/pipeline_amused.cpython-310.pyc,, +diffusers/pipelines/amused/__pycache__/pipeline_amused_img2img.cpython-310.pyc,, +diffusers/pipelines/amused/__pycache__/pipeline_amused_inpaint.cpython-310.pyc,, +diffusers/pipelines/amused/pipeline_amused.py,sha256=EJbwGd3ePgst-GmA9CSfIOncKaxRozr4zAi1QzGRKks,16018 +diffusers/pipelines/amused/pipeline_amused_img2img.py,sha256=PvInrMAHCOqES6HjltqGbkAmLWa0T_WvZub4IAAAoIE,17495 +diffusers/pipelines/amused/pipeline_amused_inpaint.py,sha256=zmpdSChNiXl8v4C7GKnlEmyGtuPfhYKXVo3OPpFW0uQ,19181 +diffusers/pipelines/animatediff/__init__.py,sha256=8e7xkGr1MrQerNXsfFBaDT8f7ELe5aoPX9v2qRN1hvg,2324 +diffusers/pipelines/animatediff/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/animatediff/__pycache__/pipeline_animatediff.cpython-310.pyc,, +diffusers/pipelines/animatediff/__pycache__/pipeline_animatediff_controlnet.cpython-310.pyc,, +diffusers/pipelines/animatediff/__pycache__/pipeline_animatediff_sdxl.cpython-310.pyc,, +diffusers/pipelines/animatediff/__pycache__/pipeline_animatediff_sparsectrl.cpython-310.pyc,, +diffusers/pipelines/animatediff/__pycache__/pipeline_animatediff_video2video.cpython-310.pyc,, +diffusers/pipelines/animatediff/__pycache__/pipeline_animatediff_video2video_controlnet.cpython-310.pyc,, +diffusers/pipelines/animatediff/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/animatediff/pipeline_animatediff.py,sha256=LZWsnSJ0PaWdLcn4hipA3TKSuZxllqYUIAn0M3Uk5oo,42434 +diffusers/pipelines/animatediff/pipeline_animatediff_controlnet.py,sha256=wyKBvLx6UMF_vqgmu42HNkj0wMUBjLggXsBrqyi61Jk,55940 +diffusers/pipelines/animatediff/pipeline_animatediff_sdxl.py,sha256=O1vn-vdJSyLOxk_o9VO2gPzsKV4WZIS18lcgtpkmeP0,66566 +diffusers/pipelines/animatediff/pipeline_animatediff_sparsectrl.py,sha256=hZJ4Bqy857KN0tbzyjVpRQlUCeJE3z9d0byZcNbYCm8,51514 +diffusers/pipelines/animatediff/pipeline_animatediff_video2video.py,sha256=v-KmxoDFQJnWPcToTTrEScO3UIK2gqhLjOJxuKzA71Q,52258 +diffusers/pipelines/animatediff/pipeline_animatediff_video2video_controlnet.py,sha256=6NNBLmnPnr9VoB0Uw6plQR_8HEan3a8yJXo8Mx2pTjI,67610 +diffusers/pipelines/animatediff/pipeline_output.py,sha256=Ggp2OfMwdOPjHh4wIEN5aHJHDiSU0ORyzWzfdysrtcA,729 +diffusers/pipelines/audioldm/__init__.py,sha256=HMUjKqEf7OAtgIeV2CQoGIoDE6oY7b26N55yn4qCIpU,1419 +diffusers/pipelines/audioldm/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/audioldm/__pycache__/pipeline_audioldm.cpython-310.pyc,, +diffusers/pipelines/audioldm/pipeline_audioldm.py,sha256=R4WZVEilSXA339bzZWMFzcpVDnf8cn1L1Vd-CGlNNqQ,26375 +diffusers/pipelines/audioldm2/__init__.py,sha256=gR7gTyh-YGI4uxTCPnz_LnCGbErpFGtNMEzM_CQdqgE,1605 +diffusers/pipelines/audioldm2/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/audioldm2/__pycache__/modeling_audioldm2.cpython-310.pyc,, +diffusers/pipelines/audioldm2/__pycache__/pipeline_audioldm2.cpython-310.pyc,, +diffusers/pipelines/audioldm2/modeling_audioldm2.py,sha256=SpKP4UVjYRznsInmq4gvobBnX2uTlA4wILzyWRHWdY0,70483 +diffusers/pipelines/audioldm2/pipeline_audioldm2.py,sha256=2r_aHtDrCijivNAPbGIyxzp_zz8NppHMpzTNQOo9ayg,54856 +diffusers/pipelines/aura_flow/__init__.py,sha256=TOGRbwqwr7j1XIVGAxIBwAp4lM2zt21C_hYm5dFb76o,1296 +diffusers/pipelines/aura_flow/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/aura_flow/__pycache__/pipeline_aura_flow.cpython-310.pyc,, +diffusers/pipelines/aura_flow/pipeline_aura_flow.py,sha256=c2d06A6gdkM09Hciw0An8Hfu5y44bDmggyDptFTe_wU,33217 +diffusers/pipelines/auto_pipeline.py,sha256=7DobcWehSghE8ASytRf292ISW1XrhB7DjjBkWYSBRvc,58301 +diffusers/pipelines/blip_diffusion/__init__.py,sha256=v_PoaUspuKZG54FdKtITSccYo6eIhMnO0d6n7Pf3JJU,697 +diffusers/pipelines/blip_diffusion/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/blip_diffusion/__pycache__/blip_image_processing.cpython-310.pyc,, +diffusers/pipelines/blip_diffusion/__pycache__/modeling_blip2.cpython-310.pyc,, +diffusers/pipelines/blip_diffusion/__pycache__/modeling_ctx_clip.cpython-310.pyc,, +diffusers/pipelines/blip_diffusion/__pycache__/pipeline_blip_diffusion.cpython-310.pyc,, +diffusers/pipelines/blip_diffusion/blip_image_processing.py,sha256=xRyx14zBBKrERosHMm_Pj9HjSfESxFAh6nPmhj42dgE,16742 +diffusers/pipelines/blip_diffusion/modeling_blip2.py,sha256=eBrI0_E7mBN0uctDOHVdcv_tYhHHWpaXGuk7DZoxAQQ,27139 +diffusers/pipelines/blip_diffusion/modeling_ctx_clip.py,sha256=53ggZ2e0gs0Oh7UwpKvSXdrcr70IlYkAUNB-hnkpMYk,9002 +diffusers/pipelines/blip_diffusion/pipeline_blip_diffusion.py,sha256=mxCrWForS6kYGJ8wkg9Im_Nic0gVwQb2KpfNUdQfM9A,15322 +diffusers/pipelines/chroma/__init__.py,sha256=iRH8V4N01bnqzev2gozrhjVu1cRhLAn61RNFCBxFzgg,1611 +diffusers/pipelines/chroma/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/chroma/__pycache__/pipeline_chroma.cpython-310.pyc,, +diffusers/pipelines/chroma/__pycache__/pipeline_chroma_img2img.cpython-310.pyc,, +diffusers/pipelines/chroma/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/chroma/pipeline_chroma.py,sha256=oqZxPLKh_shVr52SiAd74Ra0v9WTRnup09WyJC_J_WY,44783 +diffusers/pipelines/chroma/pipeline_chroma_img2img.py,sha256=fWvioL2M6BSwU5ewishYX8fr3fNMa7W7waK4mAnEifA,49244 +diffusers/pipelines/chroma/pipeline_output.py,sha256=_LYDjLRSV0vIaNBSwddFVpWwbNGPQ6E5GkNAccpaOQo,600 +diffusers/pipelines/cogvideo/__init__.py,sha256=84bmJbrCvjUtEXFgyCKvX5N4HNtAWorMjQTNvWPh8ZU,1816 +diffusers/pipelines/cogvideo/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/cogvideo/__pycache__/pipeline_cogvideox.cpython-310.pyc,, +diffusers/pipelines/cogvideo/__pycache__/pipeline_cogvideox_fun_control.cpython-310.pyc,, +diffusers/pipelines/cogvideo/__pycache__/pipeline_cogvideox_image2video.cpython-310.pyc,, +diffusers/pipelines/cogvideo/__pycache__/pipeline_cogvideox_video2video.cpython-310.pyc,, +diffusers/pipelines/cogvideo/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/cogvideo/pipeline_cogvideox.py,sha256=v4ECnmnyJDebRQvMN-tbNhLBy6sFmav5aaM2NUEhd4U,37686 +diffusers/pipelines/cogvideo/pipeline_cogvideox_fun_control.py,sha256=bephSzLcsNUQVmFw2AI484mAl1BL0RL_D-im0Bs9d3s,40247 +diffusers/pipelines/cogvideo/pipeline_cogvideox_image2video.py,sha256=7uvUgL33C0enGCqBxj2yZhSEtzplTwhAiZwi0m3kU74,42690 +diffusers/pipelines/cogvideo/pipeline_cogvideox_video2video.py,sha256=u83mFoVRGOusErNc2dpLMZJYKuRGvOUCLCZ_cGZvRmQ,41301 +diffusers/pipelines/cogvideo/pipeline_output.py,sha256=QOyumhJJERjm7moyxnYzU_X27hvN9p99MIkjT_Vf1x0,616 +diffusers/pipelines/cogview3/__init__.py,sha256=ophRMlB8W7AocUEWUJLbmK1o4yJpHEfKvwyDceyMu00,1497 +diffusers/pipelines/cogview3/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/cogview3/__pycache__/pipeline_cogview3plus.cpython-310.pyc,, +diffusers/pipelines/cogview3/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/cogview3/pipeline_cogview3plus.py,sha256=oCX9IvHfGfY7UbeYWa8jc1-q7xiVhUDjuYNPfzZUV7E,33881 +diffusers/pipelines/cogview3/pipeline_output.py,sha256=qU187W2KZY8KloD6E5EOpkbgJOYrsQFSGHyNHKHqwxs,594 +diffusers/pipelines/cogview4/__init__.py,sha256=DSW0f5XIu2bcirGYm6FE9lhyxADbV5nbiayq1x2ttJg,1633 +diffusers/pipelines/cogview4/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/cogview4/__pycache__/pipeline_cogview4.cpython-310.pyc,, +diffusers/pipelines/cogview4/__pycache__/pipeline_cogview4_control.cpython-310.pyc,, +diffusers/pipelines/cogview4/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/cogview4/pipeline_cogview4.py,sha256=d5ZI1KVrd9b_0Fr6_chIeYWEB7I8jl-k1A8Iz-wMoKE,33662 +diffusers/pipelines/cogview4/pipeline_cogview4_control.py,sha256=oqZJ5XbE3RblxqH5VzIbMpT1AnnQJV-krevmRc4CXyY,35345 +diffusers/pipelines/cogview4/pipeline_output.py,sha256=l3XZTqGTse6epEQn_VZsGyRAyQF9TpfYRaQrqO5O_Dw,594 +diffusers/pipelines/consisid/__init__.py,sha256=wi4mmbsztby5LLgmrtDhz857JWT_4Jbc2sRzRyL0EpY,1367 +diffusers/pipelines/consisid/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/consisid/__pycache__/consisid_utils.cpython-310.pyc,, +diffusers/pipelines/consisid/__pycache__/pipeline_consisid.cpython-310.pyc,, +diffusers/pipelines/consisid/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/consisid/consisid_utils.py,sha256=ZoAiyDYhNV_nMJHr2pkhtHgXVFzdcsIhWvAWWzgpoek,14522 +diffusers/pipelines/consisid/pipeline_consisid.py,sha256=Di5fV28wCWfQrRX8fzqhQj9BnwTTBHat1XLpS9emNwA,46715 +diffusers/pipelines/consisid/pipeline_output.py,sha256=bCTmHqOyRXggYJLwx_nBxNg0-yci9Dhge8Div0k_D1U,615 +diffusers/pipelines/consistency_models/__init__.py,sha256=q_nrLK9DH0_kLcLmRIvgvLP-vDVwloC3lBus776596c,484 +diffusers/pipelines/consistency_models/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/consistency_models/__pycache__/pipeline_consistency_models.cpython-310.pyc,, +diffusers/pipelines/consistency_models/pipeline_consistency_models.py,sha256=-QlWGbE6nflcWLYLY5s7F5drUOUxaR7IXPd35bC3rB8,12596 +diffusers/pipelines/controlnet/__init__.py,sha256=pqndp8HbyQ2D45STcpMp37nO5M4SagpfwADCCOC_2CU,4057 +diffusers/pipelines/controlnet/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/controlnet/__pycache__/multicontrolnet.cpython-310.pyc,, +diffusers/pipelines/controlnet/__pycache__/pipeline_controlnet.cpython-310.pyc,, +diffusers/pipelines/controlnet/__pycache__/pipeline_controlnet_blip_diffusion.cpython-310.pyc,, +diffusers/pipelines/controlnet/__pycache__/pipeline_controlnet_img2img.cpython-310.pyc,, +diffusers/pipelines/controlnet/__pycache__/pipeline_controlnet_inpaint.cpython-310.pyc,, +diffusers/pipelines/controlnet/__pycache__/pipeline_controlnet_inpaint_sd_xl.cpython-310.pyc,, +diffusers/pipelines/controlnet/__pycache__/pipeline_controlnet_sd_xl.cpython-310.pyc,, +diffusers/pipelines/controlnet/__pycache__/pipeline_controlnet_sd_xl_img2img.cpython-310.pyc,, +diffusers/pipelines/controlnet/__pycache__/pipeline_controlnet_union_inpaint_sd_xl.cpython-310.pyc,, +diffusers/pipelines/controlnet/__pycache__/pipeline_controlnet_union_sd_xl.cpython-310.pyc,, +diffusers/pipelines/controlnet/__pycache__/pipeline_controlnet_union_sd_xl_img2img.cpython-310.pyc,, +diffusers/pipelines/controlnet/__pycache__/pipeline_flax_controlnet.cpython-310.pyc,, +diffusers/pipelines/controlnet/multicontrolnet.py,sha256=-sluVPEM3oDV_dscvtnklZf3S_KsbRK98-V1fwGKUtg,684 +diffusers/pipelines/controlnet/pipeline_controlnet.py,sha256=ctdh4Nt3TzEHSi8Hsa5xUz4FDvLK7AIFEw_lxUk4fgk,69320 +diffusers/pipelines/controlnet/pipeline_controlnet_blip_diffusion.py,sha256=8ajTqSf96k6OgFyeCCHohAANKk-rxnL7GDHUF1I5nTY,17596 +diffusers/pipelines/controlnet/pipeline_controlnet_img2img.py,sha256=vr7XO-3CfGRVeLlS4QUAtqyekPZYrWVb0pyTXfcHwn8,67610 +diffusers/pipelines/controlnet/pipeline_controlnet_inpaint.py,sha256=JxW2wNkNUUY4HiI3IhivKwvG7qEDSgJ3kKk_ILS5YBQ,76660 +diffusers/pipelines/controlnet/pipeline_controlnet_inpaint_sd_xl.py,sha256=DsTeSK8OqT9racfudAHSSyPttoXAewKF2hL3hJcgDlA,95313 +diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py,sha256=5yUeJMSfgbQ-po-5ljpfPPeCQwK_m_UEc5Gv1ZiJOs4,82765 +diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py,sha256=lBQ72LPNhz7JB6zLRhkDTsMGOH_OwdCN24iFaoysqGA,87279 +diffusers/pipelines/controlnet/pipeline_controlnet_union_inpaint_sd_xl.py,sha256=26INHgrfnw7QCWEifp6t6aPbefVmXVpOHoy01jgdMJY,90688 +diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl.py,sha256=twIFMu935FUjrtYNkMDs_B_LBSHQSwEjMcmHFMAdol4,83286 +diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl_img2img.py,sha256=I9PxHFLz81vf7BrP6oVoA72Jfsu81Khsr0TEOdgiQkw,83525 +diffusers/pipelines/controlnet/pipeline_flax_controlnet.py,sha256=gJ54OTBKXNYMk1wPzaaiE7a60F_mp9S4me7DAElgIsQ,22771 +diffusers/pipelines/controlnet_hunyuandit/__init__.py,sha256=LvB-TNhPTnUIdinVZfxzUX40RFWvNWxrjAzsDDiLBfM,1344 +diffusers/pipelines/controlnet_hunyuandit/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/controlnet_hunyuandit/__pycache__/pipeline_hunyuandit_controlnet.cpython-310.pyc,, +diffusers/pipelines/controlnet_hunyuandit/pipeline_hunyuandit_controlnet.py,sha256=JmU_qrB_oPBVnu0hnrc5HEgtRCSfWicIZQwSwUFEdRg,51005 +diffusers/pipelines/controlnet_sd3/__init__.py,sha256=_-t5_Jac1hvUKbjACSwVDVWx1lFIBQDCeCE9CVMbsW0,1903 +diffusers/pipelines/controlnet_sd3/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/controlnet_sd3/__pycache__/pipeline_stable_diffusion_3_controlnet.cpython-310.pyc,, +diffusers/pipelines/controlnet_sd3/__pycache__/pipeline_stable_diffusion_3_controlnet_inpainting.cpython-310.pyc,, +diffusers/pipelines/controlnet_sd3/pipeline_stable_diffusion_3_controlnet.py,sha256=-2xtwBvccIne5Bau5P7_wOw7SdkOMq5dC9AmEAzGVLI,62633 +diffusers/pipelines/controlnet_sd3/pipeline_stable_diffusion_3_controlnet_inpainting.py,sha256=MOxSarLvTPdiFqFwaVWmAGCqWzOINPSda09taRqPD68,63597 +diffusers/pipelines/controlnet_xs/__init__.py,sha256=TuIgTKgY4MVB6zaoNTduQAEVRsNptBZQZhnxxQ3hpyg,2403 +diffusers/pipelines/controlnet_xs/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/controlnet_xs/__pycache__/pipeline_controlnet_xs.cpython-310.pyc,, +diffusers/pipelines/controlnet_xs/__pycache__/pipeline_controlnet_xs_sd_xl.cpython-310.pyc,, +diffusers/pipelines/controlnet_xs/pipeline_controlnet_xs.py,sha256=I_wHrWyaxKh8DccRvvOn7SBfjn2AoYTJxEFvOQndyiQ,46056 +diffusers/pipelines/controlnet_xs/pipeline_controlnet_xs_sd_xl.py,sha256=xnp5qtiDjVZeq4c6DgleZE8a7qJ7Y1VFv-j54ygvAfE,57342 +diffusers/pipelines/cosmos/__init__.py,sha256=pX1QX9zkM2e4Vo_0A3tvOVE0X-vHxYuFOedVtHUPmfs,1820 +diffusers/pipelines/cosmos/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/cosmos/__pycache__/pipeline_cosmos2_text2image.cpython-310.pyc,, +diffusers/pipelines/cosmos/__pycache__/pipeline_cosmos2_video2world.cpython-310.pyc,, +diffusers/pipelines/cosmos/__pycache__/pipeline_cosmos_text2world.cpython-310.pyc,, +diffusers/pipelines/cosmos/__pycache__/pipeline_cosmos_video2world.cpython-310.pyc,, +diffusers/pipelines/cosmos/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/cosmos/pipeline_cosmos2_text2image.py,sha256=sO54yib9payqZjZpBhvodLdPDkIEtlIonJAgCMJLI5Q,32962 +diffusers/pipelines/cosmos/pipeline_cosmos2_video2world.py,sha256=u94ClHRWm87hYfcoDSbN_Hd4WLX9cx0H0mbRyqdfva4,39284 +diffusers/pipelines/cosmos/pipeline_cosmos_text2world.py,sha256=JxExXGuCdoKaahFC1iux5BwLJnWFCGqA0T9JC9undP8,31639 +diffusers/pipelines/cosmos/pipeline_cosmos_video2world.py,sha256=8NWyFa78NXtHxgwSFB0ecA-Qz29EoqU6CrGh8mOX15o,39983 +diffusers/pipelines/cosmos/pipeline_output.py,sha256=iJIxvS8s2YCd790FLEd3Bl3WNm1TS67MGJlgOF6rqjA,1214 +diffusers/pipelines/dance_diffusion/__init__.py,sha256=SOwr8mpuw34oKEUuy4uVLlhjfHuLRCP0kpMjoSPXADU,453 +diffusers/pipelines/dance_diffusion/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/dance_diffusion/__pycache__/pipeline_dance_diffusion.cpython-310.pyc,, +diffusers/pipelines/dance_diffusion/pipeline_dance_diffusion.py,sha256=VDSk8B05ddFRCTRD3_cjyVH4qCM0MvqdxSIrg_MPDms,6732 +diffusers/pipelines/ddim/__init__.py,sha256=-zCVlqBSKWZdwY5HSsoiRT4nUEuT6dckiD_KIFen3bs,411 +diffusers/pipelines/ddim/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/ddim/__pycache__/pipeline_ddim.cpython-310.pyc,, +diffusers/pipelines/ddim/pipeline_ddim.py,sha256=YkknZ2EFsC7Knf-0KJB-ckEU5-xCHjctxUa7karayAk,6910 +diffusers/pipelines/ddpm/__init__.py,sha256=DAj0i0-iba7KACShx0bzGa9gqAV7yxGgf9sy_Hf095Q,425 +diffusers/pipelines/ddpm/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/ddpm/__pycache__/pipeline_ddpm.cpython-310.pyc,, +diffusers/pipelines/ddpm/pipeline_ddpm.py,sha256=M0hWjZbUbWL2VCiqsJ5oJ2VNDZSjS9jscFUXsF4_bKs,5344 +diffusers/pipelines/deepfloyd_if/__init__.py,sha256=gh1fQ5u6q0d-o3XGExCGD0jPaUK-gWCturfHU-TYIi8,2975 +diffusers/pipelines/deepfloyd_if/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/deepfloyd_if/__pycache__/pipeline_if.cpython-310.pyc,, +diffusers/pipelines/deepfloyd_if/__pycache__/pipeline_if_img2img.cpython-310.pyc,, +diffusers/pipelines/deepfloyd_if/__pycache__/pipeline_if_img2img_superresolution.cpython-310.pyc,, +diffusers/pipelines/deepfloyd_if/__pycache__/pipeline_if_inpainting.cpython-310.pyc,, +diffusers/pipelines/deepfloyd_if/__pycache__/pipeline_if_inpainting_superresolution.cpython-310.pyc,, +diffusers/pipelines/deepfloyd_if/__pycache__/pipeline_if_superresolution.cpython-310.pyc,, +diffusers/pipelines/deepfloyd_if/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/deepfloyd_if/__pycache__/safety_checker.cpython-310.pyc,, +diffusers/pipelines/deepfloyd_if/__pycache__/timesteps.cpython-310.pyc,, +diffusers/pipelines/deepfloyd_if/__pycache__/watermark.cpython-310.pyc,, +diffusers/pipelines/deepfloyd_if/pipeline_if.py,sha256=r7908WHOwr0MiDK0IBCQYfDUVncIGZ1nvSkhL84wBXY,35624 +diffusers/pipelines/deepfloyd_if/pipeline_if_img2img.py,sha256=K6d5O6eDU8OJ8GcCa5H3ck_PGX9ttRqpgJN_b0IHApw,40033 +diffusers/pipelines/deepfloyd_if/pipeline_if_img2img_superresolution.py,sha256=G6pRC9JmRlWQn5h7jf9-D2URPzWSJdz4ABJ0PrBvBBo,45073 +diffusers/pipelines/deepfloyd_if/pipeline_if_inpainting.py,sha256=2rT9uoczvSbqvIppIkzZRnJt0KL14gG7X0r9iIJvg-0,45231 +diffusers/pipelines/deepfloyd_if/pipeline_if_inpainting_superresolution.py,sha256=2Osy4qE8PfrKPRUC6yDf536tVVnuAePrvvQF_Hwy8ZY,50070 +diffusers/pipelines/deepfloyd_if/pipeline_if_superresolution.py,sha256=eSFTCAH0GE_Pn6hrMAk_CfQrpZdqHdHw7pm8HDMB-k8,39915 +diffusers/pipelines/deepfloyd_if/pipeline_output.py,sha256=RBxF2VMyXgIiZzJs9ZnmRRPom6GdFzb1DrOnuYmd6uQ,1145 +diffusers/pipelines/deepfloyd_if/safety_checker.py,sha256=zqN0z4Mvf7AtrxlUb6qAoiw_QuxGdDk-6js5YuarxTo,2117 +diffusers/pipelines/deepfloyd_if/timesteps.py,sha256=JO8b-8zlcvk_Tb6s6GGY7MgRPRADs35y0KBcSkqmNDM,5164 +diffusers/pipelines/deepfloyd_if/watermark.py,sha256=d-43jrlsjyJt1NJrXrl7U1LgCPlFD5C1gzJ83GVoijc,1601 +diffusers/pipelines/deprecated/__init__.py,sha256=mXBnea22TkkUdiGxUpZDXTSb1RlURczuRcGeIzn9DcQ,5470 +diffusers/pipelines/deprecated/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/deprecated/alt_diffusion/__init__.py,sha256=1SiGoNJytgnMwGmR48q8erVnU9JP5uz5E6XgHvlFDTc,1783 +diffusers/pipelines/deprecated/alt_diffusion/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/deprecated/alt_diffusion/__pycache__/modeling_roberta_series.cpython-310.pyc,, +diffusers/pipelines/deprecated/alt_diffusion/__pycache__/pipeline_alt_diffusion.cpython-310.pyc,, +diffusers/pipelines/deprecated/alt_diffusion/__pycache__/pipeline_alt_diffusion_img2img.cpython-310.pyc,, +diffusers/pipelines/deprecated/alt_diffusion/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/deprecated/alt_diffusion/modeling_roberta_series.py,sha256=CmrX8Y2bvoTr1_kZYQ-13nXL9ttMsX6gYX_0yPx1F3g,5530 +diffusers/pipelines/deprecated/alt_diffusion/pipeline_alt_diffusion.py,sha256=AgwVAwzfqBp0xXay6gTADpLygznVEpCUyPuxVGAd-zU,50256 +diffusers/pipelines/deprecated/alt_diffusion/pipeline_alt_diffusion_img2img.py,sha256=cGf3Fi3sjqJTRpgjAub3IGU52lPkrq0fc_WOLJ1vwgk,52976 +diffusers/pipelines/deprecated/alt_diffusion/pipeline_output.py,sha256=wtKrIaa_f-rfw5_bbEGi5mdNnQ6qmsaTGcDVBNBnvJ8,928 +diffusers/pipelines/deprecated/audio_diffusion/__init__.py,sha256=SiFqPmeNbqOYTwuTx2WUaMIpMzgSnJ2SZ_97tIDryOE,507 +diffusers/pipelines/deprecated/audio_diffusion/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/deprecated/audio_diffusion/__pycache__/mel.cpython-310.pyc,, +diffusers/pipelines/deprecated/audio_diffusion/__pycache__/pipeline_audio_diffusion.cpython-310.pyc,, +diffusers/pipelines/deprecated/audio_diffusion/mel.py,sha256=BjN1eYahKBBWn3FTSlHpn5VqbC1pym253k7u8WfucXw,5764 +diffusers/pipelines/deprecated/audio_diffusion/pipeline_audio_diffusion.py,sha256=E-e-P_zb5rQQ61QYFP_A_ONMkFLaCSB1FNGrt-C06GQ,13239 +diffusers/pipelines/deprecated/latent_diffusion_uncond/__init__.py,sha256=ZWWt671s-zbWawgtJNoIstZsvOE5ucP2M_vp7OMUMeM,448 +diffusers/pipelines/deprecated/latent_diffusion_uncond/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/deprecated/latent_diffusion_uncond/__pycache__/pipeline_latent_diffusion_uncond.cpython-310.pyc,, +diffusers/pipelines/deprecated/latent_diffusion_uncond/pipeline_latent_diffusion_uncond.py,sha256=pl4GRcCIjzva81en87KPzAiX19z3ARqUwrgapzPbKV8,5381 +diffusers/pipelines/deprecated/pndm/__init__.py,sha256=R8RavcZ5QXU-fR4o4HT_xvypifWUcqRKF3bduCgieEI,412 +diffusers/pipelines/deprecated/pndm/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/deprecated/pndm/__pycache__/pipeline_pndm.cpython-310.pyc,, +diffusers/pipelines/deprecated/pndm/pipeline_pndm.py,sha256=H9NvlZdnQ1MfSrsaicaCdvjpVPJdba6yh1a1i9Sa2UY,4662 +diffusers/pipelines/deprecated/repaint/__init__.py,sha256=mlHI_qG20VS7yuags8W0HXpbHkZgObu-jUBuYnOfffo,425 +diffusers/pipelines/deprecated/repaint/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/deprecated/repaint/__pycache__/pipeline_repaint.cpython-310.pyc,, +diffusers/pipelines/deprecated/repaint/pipeline_repaint.py,sha256=nkJmBYTDiF8Znogu769JFGFQDfX1qNY-MjMC201fLl8,10094 +diffusers/pipelines/deprecated/score_sde_ve/__init__.py,sha256=7CLXxU1JqmMFbdm0bLwCHxGUjGJFvS64xueOQdD2X7s,441 +diffusers/pipelines/deprecated/score_sde_ve/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/deprecated/score_sde_ve/__pycache__/pipeline_score_sde_ve.cpython-310.pyc,, +diffusers/pipelines/deprecated/score_sde_ve/pipeline_score_sde_ve.py,sha256=_1vGqiY3cZY-1h_74x7XGUmfT7IZicTl3K6XYlgg49g,4390 +diffusers/pipelines/deprecated/spectrogram_diffusion/__init__.py,sha256=lOJEU-CHJhv0N2BCEM9-dzKmm1Y-HPt1FuF9lGBgIpg,2588 +diffusers/pipelines/deprecated/spectrogram_diffusion/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/deprecated/spectrogram_diffusion/__pycache__/continuous_encoder.cpython-310.pyc,, +diffusers/pipelines/deprecated/spectrogram_diffusion/__pycache__/midi_utils.cpython-310.pyc,, +diffusers/pipelines/deprecated/spectrogram_diffusion/__pycache__/notes_encoder.cpython-310.pyc,, +diffusers/pipelines/deprecated/spectrogram_diffusion/__pycache__/pipeline_spectrogram_diffusion.cpython-310.pyc,, +diffusers/pipelines/deprecated/spectrogram_diffusion/continuous_encoder.py,sha256=7hqiAFb5gYob4MPvNMDaDJPX1IlJNlHzHRfnC52yhGQ,3100 +diffusers/pipelines/deprecated/spectrogram_diffusion/midi_utils.py,sha256=UY5j5rCm2rzkvX1aGrjqzl-dkvh9nHGaV67PiIDXAqk,25096 +diffusers/pipelines/deprecated/spectrogram_diffusion/notes_encoder.py,sha256=1dphTvTIpodYsic9Rn9DpX9MOchZW_PjyS626MIWw1A,2923 +diffusers/pipelines/deprecated/spectrogram_diffusion/pipeline_spectrogram_diffusion.py,sha256=DaKLh1UdQ0jlwo92eU--Js5lL9A027akOx7UaDT5vXs,11528 +diffusers/pipelines/deprecated/stable_diffusion_variants/__init__.py,sha256=mnIQupN59oc3JmKGaQZia7MO92E08wswJrP9QITzWQs,2111 +diffusers/pipelines/deprecated/stable_diffusion_variants/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_cycle_diffusion.cpython-310.pyc,, +diffusers/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_onnx_stable_diffusion_inpaint_legacy.cpython-310.pyc,, +diffusers/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_inpaint_legacy.cpython-310.pyc,, +diffusers/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_model_editing.cpython-310.pyc,, +diffusers/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_paradigms.cpython-310.pyc,, +diffusers/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_pix2pix_zero.cpython-310.pyc,, +diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_cycle_diffusion.py,sha256=4iCc0Z2B6RCWsfQ3s03Xj7kOC2RX43RHM5paJwDqztM,48043 +diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_onnx_stable_diffusion_inpaint_legacy.py,sha256=d2dTEqsup3V20i9dI8EM1ROGJlBMSrbVLecptK1tBnY,27841 +diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_inpaint_legacy.py,sha256=cr9owvhcPbRtB85QQbJ8CK4hVqdJ__TeacljBe6BunI,42556 +diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_model_editing.py,sha256=CDlbEGKY8M5ZOEwSexMxyq0Re_0PfVwzaFmV8OoFjkQ,41530 +diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_paradigms.py,sha256=DzGD_KpDwTModZFDiRxfrWBrgi-goFEglS3A_w6XJZ8,41216 +diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_pix2pix_zero.py,sha256=xa7GE5lLiAmb_c_5A4y0BqaLV9yBMTIzjdjEOQx6vPc,63563 +diffusers/pipelines/deprecated/stochastic_karras_ve/__init__.py,sha256=WOKqWaBgVgNkDUUf4ZL1--TauXKeaPqtGf3P2fTFYMw,453 +diffusers/pipelines/deprecated/stochastic_karras_ve/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/deprecated/stochastic_karras_ve/__pycache__/pipeline_stochastic_karras_ve.cpython-310.pyc,, +diffusers/pipelines/deprecated/stochastic_karras_ve/pipeline_stochastic_karras_ve.py,sha256=jRhJFQxTQ18_Kr05blYzF415U1xGMuIVPEycF5UGxX8,5277 +diffusers/pipelines/deprecated/versatile_diffusion/__init__.py,sha256=_CRp2PIJD6loFlES3hMcPigZNOUMf2OgTaRFgoit7hc,2838 +diffusers/pipelines/deprecated/versatile_diffusion/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/deprecated/versatile_diffusion/__pycache__/modeling_text_unet.cpython-310.pyc,, +diffusers/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion.cpython-310.pyc,, +diffusers/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion_dual_guided.cpython-310.pyc,, +diffusers/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion_image_variation.cpython-310.pyc,, +diffusers/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion_text_to_image.cpython-310.pyc,, +diffusers/pipelines/deprecated/versatile_diffusion/modeling_text_unet.py,sha256=s4_K2WJXXXVXXJfrw1pBU14gme12kfgs1Eakjgjm2oc,113024 +diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion.py,sha256=lBCiSYM8hQrILIEhtxs2bGDQqwsd1RHA2rWHl_xOY2c,21912 +diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_dual_guided.py,sha256=Or5-U70TQ716NM2PH5UdEQErlY5trDPj1TsYbJnl5BU,27228 +diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_image_variation.py,sha256=PqrXJWL-ZvPhaF3UMwPgbsYcxYMfJ5liVqkNS1mfwyQ,19742 +diffusers/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_text_to_image.py,sha256=Ut-bWQmxYf6g6zw7YOf_TRnWR8Om0V3Bx9B7l0fMt5k,22946 +diffusers/pipelines/deprecated/vq_diffusion/__init__.py,sha256=CD0X20a3_61pBaOzDxgU_33PLjxN1W8V46TCAwykUgE,1650 +diffusers/pipelines/deprecated/vq_diffusion/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/deprecated/vq_diffusion/__pycache__/pipeline_vq_diffusion.cpython-310.pyc,, +diffusers/pipelines/deprecated/vq_diffusion/pipeline_vq_diffusion.py,sha256=6El_KY-vwVUvYF3LGiwk6K5q5UjwhgT9J9cBw_V5Jv0,15444 +diffusers/pipelines/dit/__init__.py,sha256=w6yUFMbGzaUGPKpLfEfvHlYmrKD0UErczwsHDaDtLuQ,408 +diffusers/pipelines/dit/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/dit/__pycache__/pipeline_dit.cpython-310.pyc,, +diffusers/pipelines/dit/pipeline_dit.py,sha256=D4TI2xAQnUcsKEBput0EDV5p4y2uKsiolRxd-5Us_uQ,10275 +diffusers/pipelines/easyanimate/__init__.py,sha256=CeJHlQus6mhlN2rmFk1LA44ygm4jYJVEyMYWMckcxZI,1634 +diffusers/pipelines/easyanimate/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/easyanimate/__pycache__/pipeline_easyanimate.cpython-310.pyc,, +diffusers/pipelines/easyanimate/__pycache__/pipeline_easyanimate_control.cpython-310.pyc,, +diffusers/pipelines/easyanimate/__pycache__/pipeline_easyanimate_inpaint.cpython-310.pyc,, +diffusers/pipelines/easyanimate/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/easyanimate/pipeline_easyanimate.py,sha256=iNg09pgsMUdZzkvgqxbq5olN0fZBt87jK6Yhiws1xII,35903 +diffusers/pipelines/easyanimate/pipeline_easyanimate_control.py,sha256=evNzX0HHy9HO83ob-yTALTAlDm_ZUyx62-Z92wzKiH8,46131 +diffusers/pipelines/easyanimate/pipeline_easyanimate_inpaint.py,sha256=9OZA8OnhHqaRtxyZFX7SUbrlA6VRMFwPpa7pRYggXdY,58526 +diffusers/pipelines/easyanimate/pipeline_output.py,sha256=cpEvM-GStcMexQOQnPunDqrCOHP_Fk4aNxMvUHYeuxQ,621 +diffusers/pipelines/flux/__init__.py,sha256=wVmWJNnXfm5LIsV0eB1ugtnz_lbRoEDC7jdJZxwWxlE,3143 +diffusers/pipelines/flux/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/flux/__pycache__/modeling_flux.cpython-310.pyc,, +diffusers/pipelines/flux/__pycache__/pipeline_flux.cpython-310.pyc,, +diffusers/pipelines/flux/__pycache__/pipeline_flux_control.cpython-310.pyc,, +diffusers/pipelines/flux/__pycache__/pipeline_flux_control_img2img.cpython-310.pyc,, +diffusers/pipelines/flux/__pycache__/pipeline_flux_control_inpaint.cpython-310.pyc,, +diffusers/pipelines/flux/__pycache__/pipeline_flux_controlnet.cpython-310.pyc,, +diffusers/pipelines/flux/__pycache__/pipeline_flux_controlnet_image_to_image.cpython-310.pyc,, +diffusers/pipelines/flux/__pycache__/pipeline_flux_controlnet_inpainting.cpython-310.pyc,, +diffusers/pipelines/flux/__pycache__/pipeline_flux_fill.cpython-310.pyc,, +diffusers/pipelines/flux/__pycache__/pipeline_flux_img2img.cpython-310.pyc,, +diffusers/pipelines/flux/__pycache__/pipeline_flux_inpaint.cpython-310.pyc,, +diffusers/pipelines/flux/__pycache__/pipeline_flux_prior_redux.cpython-310.pyc,, +diffusers/pipelines/flux/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/flux/modeling_flux.py,sha256=46xvV8iCF77ifCJFPQ9tjtvNxBorVnBrQZw_eHRBujQ,1549 +diffusers/pipelines/flux/pipeline_flux.py,sha256=G9_TO9pX2aEwMPCiMpCgOe6PIq-RApxQhxLHtmOOPaY,47045 +diffusers/pipelines/flux/pipeline_flux_control.py,sha256=rQodOu9pEtTJR-dpoD9RfaoJQ2RORTPxDwBr041747Q,40671 +diffusers/pipelines/flux/pipeline_flux_control_img2img.py,sha256=OBl-7AMaDpApUD09U8CI1AOFsmuxJXNtxN4FBniQ2Rg,44561 +diffusers/pipelines/flux/pipeline_flux_control_inpaint.py,sha256=rx4GqLHlJkZ6P5qIhgyINUYi3U2RdfEp5ca6odp12_8,53655 +diffusers/pipelines/flux/pipeline_flux_controlnet.py,sha256=0qKCt57GyPw0DWTUHuSCxthRZWBH28Ug0aiZXlVp9b0,57915 +diffusers/pipelines/flux/pipeline_flux_controlnet_image_to_image.py,sha256=k3MlFbme1HKJ_2e07E1XL90A0zCRcRjwaAPlD_8TBgE,46497 +diffusers/pipelines/flux/pipeline_flux_controlnet_inpainting.py,sha256=-1R7oFr-cXRE5uPYBKypIrWG1bw4DDdNlIXMUwBADAg,55525 +diffusers/pipelines/flux/pipeline_flux_fill.py,sha256=L3cEwHTltNl7NCZeEy_VYkA9V6ejxHhzdkvejtC3oDQ,49397 +diffusers/pipelines/flux/pipeline_flux_img2img.py,sha256=9LY5esG__sgSS7c8hAjps37E2Xhsp6jTtj12Dx6d5Gc,52256 +diffusers/pipelines/flux/pipeline_flux_inpaint.py,sha256=rFTDZoCChYfGIflrdhL6trvS5e7rOnQgIGAs1O9bVRs,59241 +diffusers/pipelines/flux/pipeline_flux_prior_redux.py,sha256=ruDOkehjql-hiImJXBiArJ-HxwjFdx_ep4NFEYyvVvI,21808 +diffusers/pipelines/flux/pipeline_output.py,sha256=0w0lxW102Yf5upqYyGO2JLFYygcJNwg9ql92yKZZlTI,1102 +diffusers/pipelines/free_init_utils.py,sha256=SHrGV68cii9sYCKZLbIdBEOB5tANOVbN9vv7KznOlII,7699 +diffusers/pipelines/free_noise_utils.py,sha256=SlcvpUInyDDOOq7CkzXsjbcsC3Z7nY_JBzy6MJorHc4,29691 +diffusers/pipelines/hidream_image/__init__.py,sha256=SeMI0Ae_K8guNBe8doZIX_vJndXz6jWFWhCqtH1s5I0,1499 +diffusers/pipelines/hidream_image/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/hidream_image/__pycache__/pipeline_hidream_image.cpython-310.pyc,, +diffusers/pipelines/hidream_image/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/hidream_image/pipeline_hidream_image.py,sha256=eTkd0terZVFjMRYg43e46iCRuqydOElLXhzXJ_IX4Tk,51391 +diffusers/pipelines/hidream_image/pipeline_output.py,sha256=nfVKQcyZW-RIHpoVd5XUBzP_ekbWvk6xVikTuPbASgU,1229 +diffusers/pipelines/hunyuan_video/__init__.py,sha256=2IhKqUmvwUZpkBWZI1HBbsq15U-_YEC3aSwiNDMogDE,1878 +diffusers/pipelines/hunyuan_video/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/hunyuan_video/__pycache__/pipeline_hunyuan_skyreels_image2video.cpython-310.pyc,, +diffusers/pipelines/hunyuan_video/__pycache__/pipeline_hunyuan_video.cpython-310.pyc,, +diffusers/pipelines/hunyuan_video/__pycache__/pipeline_hunyuan_video_framepack.cpython-310.pyc,, +diffusers/pipelines/hunyuan_video/__pycache__/pipeline_hunyuan_video_image2video.cpython-310.pyc,, +diffusers/pipelines/hunyuan_video/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/hunyuan_video/pipeline_hunyuan_skyreels_image2video.py,sha256=8Z7MjSkajuroRhVI4Vlno578R_eOuMRkxatgk2lR-VU,38826 +diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video.py,sha256=froIVMQ8BiZxPvH1o69r-Tind4qZKZK_Q6wmyTPNr6s,35591 +diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video_framepack.py,sha256=MKjgjLDy9OEDN32vKWyfqZ7WlHObgpi3s_bpVVSLeG0,53966 +diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video_image2video.py,sha256=yzvQhYFGcHXa0sfP9sKxRCwN0HfsOjIgouK2qJWOF4w,46321 +diffusers/pipelines/hunyuan_video/pipeline_output.py,sha256=xSreG3mREfUvF_dfDzWDLF2pSdqNiUp47whS-tVl7xI,1417 +diffusers/pipelines/hunyuandit/__init__.py,sha256=Zby0yEsLNAoa4cf6W92QXIzyGoijI54xXRVhmrHGHsc,1302 +diffusers/pipelines/hunyuandit/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/hunyuandit/__pycache__/pipeline_hunyuandit.cpython-310.pyc,, +diffusers/pipelines/hunyuandit/pipeline_hunyuandit.py,sha256=Rag59kh4-CYhoUhO3eB38c1ZLTKPaQqt-xdQpguQpCM,43442 +diffusers/pipelines/i2vgen_xl/__init__.py,sha256=5Stj50A-AIJ1pPhilpDRx1PARMs_n8OKTDl64cq0LAY,1307 +diffusers/pipelines/i2vgen_xl/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/i2vgen_xl/__pycache__/pipeline_i2vgen_xl.cpython-310.pyc,, +diffusers/pipelines/i2vgen_xl/pipeline_i2vgen_xl.py,sha256=Uhe3P_65v4M0CaoM5cRqO5PcHrYRH2c-uNNYWEHEQcY,37427 +diffusers/pipelines/kandinsky/__init__.py,sha256=wrxuhSw_CunNhm7TdzA_fm__092mibGxp5_ep1boZmQ,2312 +diffusers/pipelines/kandinsky/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/kandinsky/__pycache__/pipeline_kandinsky.cpython-310.pyc,, +diffusers/pipelines/kandinsky/__pycache__/pipeline_kandinsky_combined.cpython-310.pyc,, +diffusers/pipelines/kandinsky/__pycache__/pipeline_kandinsky_img2img.cpython-310.pyc,, +diffusers/pipelines/kandinsky/__pycache__/pipeline_kandinsky_inpaint.cpython-310.pyc,, +diffusers/pipelines/kandinsky/__pycache__/pipeline_kandinsky_prior.cpython-310.pyc,, +diffusers/pipelines/kandinsky/__pycache__/text_encoder.cpython-310.pyc,, +diffusers/pipelines/kandinsky/pipeline_kandinsky.py,sha256=9bjqUCjOFgsuFX38fzNTOHvLSQVXWLSUH8-2aVHPf1c,17913 +diffusers/pipelines/kandinsky/pipeline_kandinsky_combined.py,sha256=vLWCvPd6Qpv3XNlvpm4t5bFCXaz1Pcz8CA9-hP_GvdA,39610 +diffusers/pipelines/kandinsky/pipeline_kandinsky_img2img.py,sha256=Gbl8MhTC6cfifzDqS_sO0winbWbi3MfLx9oMOyr7r6Y,21920 +diffusers/pipelines/kandinsky/pipeline_kandinsky_inpaint.py,sha256=7nepe82E_RqUAKaYY76hb8TONq70SKcEomsmKTNEYOE,28740 +diffusers/pipelines/kandinsky/pipeline_kandinsky_prior.py,sha256=29gDoi2jh95jyubYbbY_7TDpYiQ3XzmSsrbLonOQhkE,24044 +diffusers/pipelines/kandinsky/text_encoder.py,sha256=zDi1K-p-rPii0ZugI-83D75DR6AW36pkl8SvGBO77bA,1022 +diffusers/pipelines/kandinsky2_2/__init__.py,sha256=WeV8KWoCLj6KTvJ-f3Do87IoX_dR_AZNylBz7_Iu87s,2796 +diffusers/pipelines/kandinsky2_2/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2.cpython-310.pyc,, +diffusers/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_combined.cpython-310.pyc,, +diffusers/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_controlnet.cpython-310.pyc,, +diffusers/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_controlnet_img2img.cpython-310.pyc,, +diffusers/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_img2img.cpython-310.pyc,, +diffusers/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_inpainting.cpython-310.pyc,, +diffusers/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_prior.cpython-310.pyc,, +diffusers/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_prior_emb2emb.cpython-310.pyc,, +diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2.py,sha256=Mk61oIckbX_h2Rz6jn93qR46vNPuNAvpKFdUFqz7zmg,14377 +diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py,sha256=7BW9hMirrEn7QaYjukHgM3C1fKW1lMFOhXIW2URCoX0,44243 +diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet.py,sha256=TZzib5OqmKlFuWemhCkMxJZubFXsONofoxXNlixqe3M,14342 +diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet_img2img.py,sha256=nbhz22EZYB-S9Qt_srCwwywYSfLskGQOplcAaR8wQR0,16849 +diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_img2img.py,sha256=9blo8FSUgj4SwVMLJK4YtNNpbphlNeJsNFjZcK85r6o,17167 +diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_inpainting.py,sha256=ljn17ynhnbomb6nJFNJCWX81zhwq7fTkSgNA07D92Xc,25019 +diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior.py,sha256=2mIAe7glGkd9aGMfTKjS5t4LV4StkcXYbwZ4uyD-BuA,25648 +diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior_emb2emb.py,sha256=f3dL1Bqr8GZP9MV7Fvb8bWlMsfMo7W_clyeZ3agrSzk,25200 +diffusers/pipelines/kandinsky3/__init__.py,sha256=7Mv8Ov-XstHMLmRQU7psdheFn_e_qXJWWTYV7z7uj4U,1461 +diffusers/pipelines/kandinsky3/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/kandinsky3/__pycache__/convert_kandinsky3_unet.cpython-310.pyc,, +diffusers/pipelines/kandinsky3/__pycache__/pipeline_kandinsky3.cpython-310.pyc,, +diffusers/pipelines/kandinsky3/__pycache__/pipeline_kandinsky3_img2img.cpython-310.pyc,, +diffusers/pipelines/kandinsky3/convert_kandinsky3_unet.py,sha256=FJ8psagvZtQHJupm0hgMUI2mto3IHEXjaoLDXip1LMA,3273 +diffusers/pipelines/kandinsky3/pipeline_kandinsky3.py,sha256=l3_pd1OpfKe34n4HQrHKVqFNcBL3YN01PeW03-3yOqM,27789 +diffusers/pipelines/kandinsky3/pipeline_kandinsky3_img2img.py,sha256=cqQnHdUl_nEr_HRNTFUZ3P4_tHA2i-9a6PF3TUW29_Q,30506 +diffusers/pipelines/kolors/__init__.py,sha256=6Xp5M_K6PfByqqnK1HuMD9RKLkOZYekeNNqrGk4HToM,1791 +diffusers/pipelines/kolors/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/kolors/__pycache__/pipeline_kolors.cpython-310.pyc,, +diffusers/pipelines/kolors/__pycache__/pipeline_kolors_img2img.cpython-310.pyc,, +diffusers/pipelines/kolors/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/kolors/__pycache__/text_encoder.cpython-310.pyc,, +diffusers/pipelines/kolors/__pycache__/tokenizer.cpython-310.pyc,, +diffusers/pipelines/kolors/pipeline_kolors.py,sha256=im6XM1e4I13kqVFT_KOZrKLYAQ4GGgDMo2H63XstZSo,56001 +diffusers/pipelines/kolors/pipeline_kolors_img2img.py,sha256=CsA_xfEmCZbzgxB6hl1hOImhs0WQt5y9unkwJxXHrMM,65879 +diffusers/pipelines/kolors/pipeline_output.py,sha256=1POR3PAcQ-ZtBgf8GOwjVI46TCjdtpXPsv1sdslBFA0,590 +diffusers/pipelines/kolors/text_encoder.py,sha256=xpm-7Qv1gLUMTCdh5oQegPq7n1Yh56hbi4kmjlPpXD4,35059 +diffusers/pipelines/kolors/tokenizer.py,sha256=TzqdOwu1gyi8KJ_tb8lAc7_5M6AdvhkvJW2LKOCcDrI,13428 +diffusers/pipelines/latent_consistency_models/__init__.py,sha256=SfUylLTTBCs_wlGOPpW899lgE1E0GOLGu4GhDPFx-Ls,1560 +diffusers/pipelines/latent_consistency_models/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/latent_consistency_models/__pycache__/pipeline_latent_consistency_img2img.cpython-310.pyc,, +diffusers/pipelines/latent_consistency_models/__pycache__/pipeline_latent_consistency_text2img.cpython-310.pyc,, +diffusers/pipelines/latent_consistency_models/pipeline_latent_consistency_img2img.py,sha256=lDaK8Pg1pokNh97zhbPYavsF3HTsErdGntq-IW6xZO0,49548 +diffusers/pipelines/latent_consistency_models/pipeline_latent_consistency_text2img.py,sha256=LzLz1Fkyz65EzxTRE3kFvT2-gXLoEG6Tb6gonb0S3tA,46056 +diffusers/pipelines/latent_diffusion/__init__.py,sha256=iUkMRZY-pteRsvsROOz2Pacm7t02Q6QvbsgQedJt6-E,1542 +diffusers/pipelines/latent_diffusion/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/latent_diffusion/__pycache__/pipeline_latent_diffusion.cpython-310.pyc,, +diffusers/pipelines/latent_diffusion/__pycache__/pipeline_latent_diffusion_superresolution.cpython-310.pyc,, +diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion.py,sha256=wDR9k8KbxUoKlvbzQ0UQ6bw_gF8An2YXtP0_-OGvq2w,32607 +diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion_superresolution.py,sha256=Z50tjLw7uJkngAZ-GHFd92iDT8gnyZBhDFzqTDbRnEg,8294 +diffusers/pipelines/latte/__init__.py,sha256=1XMhkoAvpw2akbDmMTsKJbTU4PsR9H6boq4FEhCGbwo,1282 +diffusers/pipelines/latte/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/latte/__pycache__/pipeline_latte.cpython-310.pyc,, +diffusers/pipelines/latte/pipeline_latte.py,sha256=71-fHckM7jLr84nVlFxZKAnIGN-xILcGCLNsrq-A4ho,42687 +diffusers/pipelines/ledits_pp/__init__.py,sha256=3VaqGS1d39iC5flUifb4vAD_bDJ-sIUFaLIYhBuHbwE,1783 +diffusers/pipelines/ledits_pp/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/ledits_pp/__pycache__/pipeline_leditspp_stable_diffusion.cpython-310.pyc,, +diffusers/pipelines/ledits_pp/__pycache__/pipeline_leditspp_stable_diffusion_xl.cpython-310.pyc,, +diffusers/pipelines/ledits_pp/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion.py,sha256=vK7RFmkCw5tdrYbgW2JI3q3WB5hj6Pc2qUFj79Q8y1Y,77462 +diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py,sha256=TgulOyUe1gjg_8jCutH1018qMqqTsnJAl9rNxlT_qiw,89461 +diffusers/pipelines/ledits_pp/pipeline_output.py,sha256=xiAplyxGWB6uCHIdryai6UP7ghtFuXhES52ZYpO3k8A,1579 +diffusers/pipelines/ltx/__init__.py,sha256=6STasTKI0b2ALlgCUi9GMmb_aScZzLWcfKKZxdHWOqo,1878 +diffusers/pipelines/ltx/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/ltx/__pycache__/modeling_latent_upsampler.cpython-310.pyc,, +diffusers/pipelines/ltx/__pycache__/pipeline_ltx.cpython-310.pyc,, +diffusers/pipelines/ltx/__pycache__/pipeline_ltx_condition.cpython-310.pyc,, +diffusers/pipelines/ltx/__pycache__/pipeline_ltx_image2video.cpython-310.pyc,, +diffusers/pipelines/ltx/__pycache__/pipeline_ltx_latent_upsample.cpython-310.pyc,, +diffusers/pipelines/ltx/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/ltx/modeling_latent_upsampler.py,sha256=2Fj74GY82S8By8l2mHKq1DIGaYUZQzUSI3Nf_1s00rA,7482 +diffusers/pipelines/ltx/pipeline_ltx.py,sha256=Pxm7U9QGYCuH7KtrpdpJoYIkW_nLOJtRoSsKxp5m_Yg,40849 +diffusers/pipelines/ltx/pipeline_ltx_condition.py,sha256=zMQCxCBCZNtwzQBRXO0ZqoSGU9RFeJrMXk2YgYMC8y4,62329 +diffusers/pipelines/ltx/pipeline_ltx_image2video.py,sha256=wfnWLsiaJER2uGFVNnW9eYZCOJLLdmhrHSNlgvOnIA8,45691 +diffusers/pipelines/ltx/pipeline_ltx_latent_upsample.py,sha256=6g5DXLlkMm4U3V4aWjHsGvBpHeO96Y9FSMIrWYlQNJk,11813 +diffusers/pipelines/ltx/pipeline_output.py,sha256=amPUfWP2XV5Qgb0BpfVIbC6L3vIHCMpTWdqgr5nfrvw,605 +diffusers/pipelines/lumina/__init__.py,sha256=AzWsnxikODkQnCxliBN7eDi83TxcSihxfehfLYxRPD4,1336 +diffusers/pipelines/lumina/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/lumina/__pycache__/pipeline_lumina.cpython-310.pyc,, +diffusers/pipelines/lumina/pipeline_lumina.py,sha256=5LTTbsltMVJg83iMMEjwhstpw64b3QbJfcUjYj8RKWE,44964 +diffusers/pipelines/lumina2/__init__.py,sha256=ZnlJglaTqtwptFJ0uelQ4MKg_p-Lxwzu7eQCd1CFxtc,1342 +diffusers/pipelines/lumina2/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/lumina2/__pycache__/pipeline_lumina2.cpython-310.pyc,, +diffusers/pipelines/lumina2/pipeline_lumina2.py,sha256=iiuTXnVG1AsOl--0R2F73VX2EKUh82VDUW9NVwHFs7c,38521 +diffusers/pipelines/marigold/__init__.py,sha256=kAs3DZB4oxiYHqLOq9kgAvBCYw3ptpiQKGr01hM2BDQ,1926 +diffusers/pipelines/marigold/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/marigold/__pycache__/marigold_image_processing.cpython-310.pyc,, +diffusers/pipelines/marigold/__pycache__/pipeline_marigold_depth.cpython-310.pyc,, +diffusers/pipelines/marigold/__pycache__/pipeline_marigold_intrinsics.cpython-310.pyc,, +diffusers/pipelines/marigold/__pycache__/pipeline_marigold_normals.cpython-310.pyc,, +diffusers/pipelines/marigold/marigold_image_processing.py,sha256=veWoIY4QDBKp74-MWHjbrQwgPtD5dOECVPAiTUs2DDE,31409 +diffusers/pipelines/marigold/pipeline_marigold_depth.py,sha256=IQqLiRvuSqFmwr37JlPEMaXZe0DVH8YeCm12MqvewO8,41201 +diffusers/pipelines/marigold/pipeline_marigold_intrinsics.py,sha256=MADfO3Yzh30E5usYZSsf7EurYG3tIar01cy39chDsRY,35861 +diffusers/pipelines/marigold/pipeline_marigold_normals.py,sha256=osc4gL7CVCH5IH530Vej6Q06EGSbDXTeaaAIJBh_oao,34791 +diffusers/pipelines/mochi/__init__.py,sha256=8yDkp3YgOvbC4VhO4Tfin2myNxRlWiX1Mi8rY_UvAh4,1282 +diffusers/pipelines/mochi/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/mochi/__pycache__/pipeline_mochi.cpython-310.pyc,, +diffusers/pipelines/mochi/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/mochi/pipeline_mochi.py,sha256=ONMqBpSi-bzFbWssfEaB1CBcATniiUGoo7YtRY2lYk4,35591 +diffusers/pipelines/mochi/pipeline_output.py,sha256=RyFrgJRzyCNzbHurysrFsN4wtZLdcax8wTarxhUq-50,609 +diffusers/pipelines/musicldm/__init__.py,sha256=l1I5QzvTwMOOltJkcwpTb6nNcr93bWiP_ErHbDdwz6Y,1411 +diffusers/pipelines/musicldm/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/musicldm/__pycache__/pipeline_musicldm.cpython-310.pyc,, +diffusers/pipelines/musicldm/pipeline_musicldm.py,sha256=h5Li1cDVFDi-lOwL5i_fPXy-BfuqfC9kkYXCMEiYK08,30670 +diffusers/pipelines/omnigen/__init__.py,sha256=9596QBScCQfCbrybjkWJ7p0N4CnYB7W3hQJsNlGn3dU,1292 +diffusers/pipelines/omnigen/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/omnigen/__pycache__/pipeline_omnigen.cpython-310.pyc,, +diffusers/pipelines/omnigen/__pycache__/processor_omnigen.cpython-310.pyc,, +diffusers/pipelines/omnigen/pipeline_omnigen.py,sha256=h2-QMBcfcgYXhimDRIwwQ4Vt_dAGSXIVz4SBxxrLkaQ,23752 +diffusers/pipelines/omnigen/processor_omnigen.py,sha256=iGzPTREszduXazqwAHBmXKfoB2bKPBLITwLsnOJ0Wek,14191 +diffusers/pipelines/onnx_utils.py,sha256=oTRc_iLHEKpf_IGFw_ka1bloAI-XUa6_ASMLW2LAH4w,8810 +diffusers/pipelines/pag/__init__.py,sha256=pyv70bIvWZpMgXJgf8I8JH4OWzvdqJqfvxlFax9SzTg,3986 +diffusers/pipelines/pag/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pag_utils.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_controlnet_sd.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_controlnet_sd_inpaint.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_controlnet_sd_xl.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_controlnet_sd_xl_img2img.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_hunyuandit.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_kolors.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_pixart_sigma.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_sana.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_sd.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_sd_3.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_sd_3_img2img.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_sd_animatediff.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_sd_img2img.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_sd_inpaint.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_sd_xl.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_sd_xl_img2img.cpython-310.pyc,, +diffusers/pipelines/pag/__pycache__/pipeline_pag_sd_xl_inpaint.cpython-310.pyc,, +diffusers/pipelines/pag/pag_utils.py,sha256=FGN5B2UocvZW7Y_EKwm_0B4BToIzfo0qemoHLW9jPvo,10221 +diffusers/pipelines/pag/pipeline_pag_controlnet_sd.py,sha256=DIkyn-6Q0F0cp9DHx84zTZ9bzkMWYATd-Rh_m3LEc30,68337 +diffusers/pipelines/pag/pipeline_pag_controlnet_sd_inpaint.py,sha256=aFeuyLzqCZgT9EoVTEyfsHRQxT88Yab_25Hzi2k3024,79030 +diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl.py,sha256=hs2zkLmgb5eOHM4O6wUXiDFSSLFWteXcyYsY9CxtfLI,83631 +diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl_img2img.py,sha256=jxbc1AoLlYVSZHLp_bvyG68TYV185w0a5TvTpeRLbfU,88473 +diffusers/pipelines/pag/pipeline_pag_hunyuandit.py,sha256=Sszu4cI7T9pUM2CXv35k0-sFs5ecDJX8qnZBwKqjsbA,46649 +diffusers/pipelines/pag/pipeline_pag_kolors.py,sha256=G4ezp_hF7AF5LScEswyNAgt7ziU9XhwRQmfetD541ek,59694 +diffusers/pipelines/pag/pipeline_pag_pixart_sigma.py,sha256=2DneQFcaXGRkurJvgmQ2jkvtI0zwzTcbPN4xj2d7Tv0,42103 +diffusers/pipelines/pag/pipeline_pag_sana.py,sha256=aJzap_XDStpd0kSHLJF136lV7gnqxd3_BBtwPCIhyp0,45728 +diffusers/pipelines/pag/pipeline_pag_sd.py,sha256=g2Fac2_uG1qGr0QzdGGLGPoYVR3rOX0J5BrnxT69OQI,55556 +diffusers/pipelines/pag/pipeline_pag_sd_3.py,sha256=08LmpiuzuMZRzVWyIGiIHvq5Axpwa7UXw0sBq1hRTu4,49479 +diffusers/pipelines/pag/pipeline_pag_sd_3_img2img.py,sha256=IFND8xNQbWG2jRxadoIkYdkAqPJHMPhC2z41ZgV-0Fc,53952 +diffusers/pipelines/pag/pipeline_pag_sd_animatediff.py,sha256=GcDAKtLel_OAZ5eoXbSq_Q3iWTRk9SN05BFPFcTwvTQ,43567 +diffusers/pipelines/pag/pipeline_pag_sd_img2img.py,sha256=7v9waEdxFItMC7iJOy3Ac1djLXe0lAgRE_SJ8TEM1hQ,57891 +diffusers/pipelines/pag/pipeline_pag_sd_inpaint.py,sha256=XUjGF65RTE7bylt-mRKjzEng3oIbxC3xF20RBgf5pyo,69978 +diffusers/pipelines/pag/pipeline_pag_sd_xl.py,sha256=gfgXXZw81JPMTgIvMPpw4Vwsf6IstEgXHchJHt-gyRA,70531 +diffusers/pipelines/pag/pipeline_pag_sd_xl_img2img.py,sha256=Pps9YyXn-4QboB2nIqoXl9a--lrF5IN8zWt5IcZrGUo,81967 +diffusers/pipelines/pag/pipeline_pag_sd_xl_inpaint.py,sha256=001uhtBD-1gykVLFeckJP2Sy1t8DxYcKemzlLut54cw,93720 +diffusers/pipelines/paint_by_example/__init__.py,sha256=EL3EGhjCG7CMzwloJRauSDHc6oArjVsETUCj8mOauRs,1566 +diffusers/pipelines/paint_by_example/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/paint_by_example/__pycache__/image_encoder.cpython-310.pyc,, +diffusers/pipelines/paint_by_example/__pycache__/pipeline_paint_by_example.cpython-310.pyc,, +diffusers/pipelines/paint_by_example/image_encoder.py,sha256=NT3RadNaIVztZWGP628BDHpvaMRPZ-w3J3j096y3L9s,2484 +diffusers/pipelines/paint_by_example/pipeline_paint_by_example.py,sha256=3PrO9c5yVuNajZ-fDmAJBxcMcCvQZB49YQLoQl3fJ8E,31344 +diffusers/pipelines/pia/__init__.py,sha256=md5F8G279iZg4WGSmLP7N8apWkuHkfssjLQFzv6c2zI,1299 +diffusers/pipelines/pia/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/pia/__pycache__/pipeline_pia.cpython-310.pyc,, +diffusers/pipelines/pia/pipeline_pia.py,sha256=TxnQBCeUWggseAABRyspi13LDs2zdyn1GbEOy7SJKlA,46525 +diffusers/pipelines/pipeline_flax_utils.py,sha256=q4emCFj0dF2xQTqeaC4JmwX0eR4JfXxoahY5OJjgk1Y,27034 +diffusers/pipelines/pipeline_loading_utils.py,sha256=VPIua1FpHmlF8bjm8q6VolMYgRTDxoRDbeE8Zdm0GpU,47268 +diffusers/pipelines/pipeline_utils.py,sha256=Bn3oM0fR3cqZL2t4MLeIn8pnaMZ18H1ULZfiB7BnSjg,105342 +diffusers/pipelines/pixart_alpha/__init__.py,sha256=QxcTJF9ryOIejEHQVw3bZAYHn2dah-WPT5pZudE8XxU,1595 +diffusers/pipelines/pixart_alpha/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/pixart_alpha/__pycache__/pipeline_pixart_alpha.cpython-310.pyc,, +diffusers/pipelines/pixart_alpha/__pycache__/pipeline_pixart_sigma.cpython-310.pyc,, +diffusers/pipelines/pixart_alpha/pipeline_pixart_alpha.py,sha256=9qear7s6XP5Zt8g1fgpmQC4DxhVCMcYg26ZHdu7zvNg,44823 +diffusers/pipelines/pixart_alpha/pipeline_pixart_sigma.py,sha256=FdbUbM8SKn3StH3pS4mHqvMy-3hRjaXBT_v7Ik77-m8,41217 +diffusers/pipelines/sana/__init__.py,sha256=qkgbJxOAEH4gmyQ4FX4USnOd-PPEDkZGjZ3QO0ID0pA,1719 +diffusers/pipelines/sana/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/sana/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/sana/__pycache__/pipeline_sana.cpython-310.pyc,, +diffusers/pipelines/sana/__pycache__/pipeline_sana_controlnet.cpython-310.pyc,, +diffusers/pipelines/sana/__pycache__/pipeline_sana_sprint.cpython-310.pyc,, +diffusers/pipelines/sana/__pycache__/pipeline_sana_sprint_img2img.cpython-310.pyc,, +diffusers/pipelines/sana/pipeline_output.py,sha256=ErM82CTECCPbaLIpJsEzwl0r_hGNi29IbxKcsJ1mEMM,586 +diffusers/pipelines/sana/pipeline_sana.py,sha256=Y7W4MXlppwU04EWCeE6tkRI31tAystB8mnGiYRdWZ1U,47262 +diffusers/pipelines/sana/pipeline_sana_controlnet.py,sha256=SkMUSXUPVX-Md9mBqtwacqoPWB92ibART5CUE2IPAFo,51742 +diffusers/pipelines/sana/pipeline_sana_sprint.py,sha256=Q4ckz9JXmOngptRQh0up3zKCO8sf5T1FAnIaNB8UpI0,41342 +diffusers/pipelines/sana/pipeline_sana_sprint_img2img.py,sha256=X_PURGwytjQh5S_XRLQEqKE1zERVtHAFNITa775_qE8,45583 +diffusers/pipelines/semantic_stable_diffusion/__init__.py,sha256=4jDvmgpXRVXGeSAcfGN90iQoJJBBRgE7NXzBE_8AYxM,1443 +diffusers/pipelines/semantic_stable_diffusion/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/semantic_stable_diffusion/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/semantic_stable_diffusion/__pycache__/pipeline_semantic_stable_diffusion.cpython-310.pyc,, +diffusers/pipelines/semantic_stable_diffusion/pipeline_output.py,sha256=YBxNQ2JiY3jYW-GB44nzNZxeADAswMQBJfnr2tBX0eY,822 +diffusers/pipelines/semantic_stable_diffusion/pipeline_semantic_stable_diffusion.py,sha256=w3qDw26PbhhhPgankdSOnHIACBrrb-9_X5hM8mrmj1s,39024 +diffusers/pipelines/shap_e/__init__.py,sha256=LGToZwsVeVBEsE5eveY0Hc2GgI6UgDz6H_6cB_Snn0Y,2093 +diffusers/pipelines/shap_e/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/shap_e/__pycache__/camera.cpython-310.pyc,, +diffusers/pipelines/shap_e/__pycache__/pipeline_shap_e.cpython-310.pyc,, +diffusers/pipelines/shap_e/__pycache__/pipeline_shap_e_img2img.cpython-310.pyc,, +diffusers/pipelines/shap_e/__pycache__/renderer.cpython-310.pyc,, +diffusers/pipelines/shap_e/camera.py,sha256=oIKN8kcD6gqpeQWABRDfkdm3wW0MXFCZJSW_qdmUMLo,4942 +diffusers/pipelines/shap_e/pipeline_shap_e.py,sha256=YG8TDX2t_2TVseeyS7Rotxorh2prpJsp8ol4ucUxM48,13398 +diffusers/pipelines/shap_e/pipeline_shap_e_img2img.py,sha256=uCfAsOKkrI85NPR3X1lhfK_10D2IhfOVkDrUJBmxpps,13205 +diffusers/pipelines/shap_e/renderer.py,sha256=Wx9GqWWLUmZjOIXdMgPKGyotodqyrdNzbKIkBrmY9ck,39153 +diffusers/pipelines/stable_audio/__init__.py,sha256=R8Tuxx2LsaWWR0lncRJ0faKOmAdaQ0ilvftdBC_07Eo,1561 +diffusers/pipelines/stable_audio/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/stable_audio/__pycache__/modeling_stable_audio.cpython-310.pyc,, +diffusers/pipelines/stable_audio/__pycache__/pipeline_stable_audio.cpython-310.pyc,, +diffusers/pipelines/stable_audio/modeling_stable_audio.py,sha256=vakcN7KjC2PjVWoYbpZSiqW7Fnd63LpvgmYtmoXfSyE,6127 +diffusers/pipelines/stable_audio/pipeline_stable_audio.py,sha256=aph9UJQsqVQs2EJzl2iu6jSY65Nk7TTen7sSd6FWij0,35613 +diffusers/pipelines/stable_cascade/__init__.py,sha256=buKExLbA-qdePd19JSEF29AhOCIaDgqFfLajEmo-Kg4,1672 +diffusers/pipelines/stable_cascade/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/stable_cascade/__pycache__/pipeline_stable_cascade.cpython-310.pyc,, +diffusers/pipelines/stable_cascade/__pycache__/pipeline_stable_cascade_combined.cpython-310.pyc,, +diffusers/pipelines/stable_cascade/__pycache__/pipeline_stable_cascade_prior.cpython-310.pyc,, +diffusers/pipelines/stable_cascade/pipeline_stable_cascade.py,sha256=xXRM34zQrVgayRBVKJrGA11hjU_2u-V-h5slnVcdh0c,26124 +diffusers/pipelines/stable_cascade/pipeline_stable_cascade_combined.py,sha256=CcFI-qOl1VCXQPgbB196gqB12CAcMu23I7m5RjIRnm4,18130 +diffusers/pipelines/stable_cascade/pipeline_stable_cascade_prior.py,sha256=fBpaAE1aPZhvq4EuEg-7Uq6khJLFTP3fYI-hPeIZOFY,31444 +diffusers/pipelines/stable_diffusion/__init__.py,sha256=4DkYPvosNLHumc2wQcV10KkWsGRCRhcyeUogvraHSEM,8555 +diffusers/pipelines/stable_diffusion/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/clip_image_project_model.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/convert_from_ckpt.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_flax_stable_diffusion.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_flax_stable_diffusion_img2img.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_flax_stable_diffusion_inpaint.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_onnx_stable_diffusion.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_onnx_stable_diffusion_img2img.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_onnx_stable_diffusion_inpaint.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_onnx_stable_diffusion_upscale.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_depth2img.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_image_variation.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_img2img.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_inpaint.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_instruct_pix2pix.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_latent_upscale.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_upscale.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_stable_unclip.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/pipeline_stable_unclip_img2img.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/safety_checker.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/safety_checker_flax.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/__pycache__/stable_unclip_image_normalizer.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion/clip_image_project_model.py,sha256=9LPlAopNG-vpWhmh8i8pz7CHttqZ203-RnGlvy74xxY,1094 +diffusers/pipelines/stable_diffusion/convert_from_ckpt.py,sha256=7AkKESPubz4W7bYqzGOK87gxsP_HZg5nX5VhRmoCPgg,81729 +diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion.py,sha256=3SsX3YOJJbZ5El0cLf1jf-9M83ZMenIf7NXiaQA86ho,20756 +diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_img2img.py,sha256=G6V7Okambalqs8-Agpa0iREwSyc_9GZH5-iRhYP_No8,22554 +diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_inpaint.py,sha256=7F-z40HNgb0a6a_dV_T46nj6ZqSmBDSlsGMPClEPd7U,26130 +diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion.py,sha256=tVEekoMVfQ1CyF0VSYAU5UnzNvDbv7R44IbvluSU-EQ,24341 +diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_img2img.py,sha256=J-FbiSOueXBLk153qyUUgAcaAzVLHdHTzXwJIi7BPbc,28552 +diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py,sha256=heCESPus4qrmryvhHK-Rsq6JzpHf-88gJKgSD9gZDRk,29170 +diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_upscale.py,sha256=UFZ8HlQuuHF8mnT7Fq0xO0FeOaW5SD7X29SMd6QJBh0,27950 +diffusers/pipelines/stable_diffusion/pipeline_output.py,sha256=Io-12AumvYjBOKP4Qq1BJ2rak5pKdtMO-ACcorY6hFE,1496 +diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py,sha256=E9x8m3TsPZnmCguwNBL6-k5MPglX4m4THlXCQwKouF0,55652 +diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py,sha256=sO7DpfNobnuAal0kDM3mD3z2fe5GRlyTHG7u2x_rZIY,44509 +diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py,sha256=XNLerrOA1__emc2DHlPt7-16pxJR_plP3usFH4_DVGQ,22720 +diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py,sha256=Th3gTxeKNQnAHjnPF2DgAB7dKHzLsWNZ6yoGw1XO094,59513 +diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py,sha256=hHJegbumSJD5-Tf213dLcNFy0YY6UTKorL5Y8DuMqbQ,70158 +diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py,sha256=_vsAtsPHV40l1LFYPTD-uJRzaaj0_ACMu2Jb4fT-bFA,45826 +diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py,sha256=jFgn7dOCcGRa_W8ALhRM2fTSq6umDxqTD5fX6LYRY6U,30985 +diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py,sha256=91cjNp53ErZtTQqltnHz6LDHtiOapmCXE_UIXc_6LQY,39757 +diffusers/pipelines/stable_diffusion/pipeline_stable_unclip.py,sha256=-ga6Cxugyl9N1OlJiL5W2-BRAcapI0Us4G2XccZcEcU,45544 +diffusers/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py,sha256=55kewO9zpxnTmcXrmqwoGDheiRXjSKEdZ5TpedeX3b4,40310 +diffusers/pipelines/stable_diffusion/safety_checker.py,sha256=h3WwFCcuJpWxEebxwgiRs0x7t4-4u4ma0GtLOeERaLA,5759 +diffusers/pipelines/stable_diffusion/safety_checker_flax.py,sha256=RhHxBbrcBrEosONc4uVlA9LgftN3WJv7wn_nV-nSDCk,4476 +diffusers/pipelines/stable_diffusion/stable_unclip_image_normalizer.py,sha256=wgtRMD_dzwanDdbqHTqlc_Ro_1q51SnotP8uqyC7Yr8,1890 +diffusers/pipelines/stable_diffusion_3/__init__.py,sha256=4JrcTgfij4mGbSSnCaHSqRRNhCUry8-HH3zQaUIq3DE,1922 +diffusers/pipelines/stable_diffusion_3/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_stable_diffusion_3.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_stable_diffusion_3_img2img.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_stable_diffusion_3_inpaint.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_3/pipeline_output.py,sha256=empNHoFAmdz6__yOCX2kuJqZtVdtoGAvVmH5mW42-3s,610 +diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3.py,sha256=3Wdp8VpfehM0pwTGH3jmi_DUqZgPuEJLKuWkMe1BhhY,56846 +diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_img2img.py,sha256=HDuLdQDpuWnxvSBk_ATg5qcKjbyo7j2RbgkTmD3c-Bc,57765 +diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_inpaint.py,sha256=KvElHFo4l47wHmXx4GpK9EH-J-6PhNGqwMylnZB07FQ,69805 +diffusers/pipelines/stable_diffusion_attend_and_excite/__init__.py,sha256=VpZ5FPx9ACTOT4qiEqun2QYeUtx9Rp0YVDwqhYe28QM,1390 +diffusers/pipelines/stable_diffusion_attend_and_excite/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_attend_and_excite/__pycache__/pipeline_stable_diffusion_attend_and_excite.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_attend_and_excite/pipeline_stable_diffusion_attend_and_excite.py,sha256=RckQAPyM_I7oqlwmA10pXjGG8Jg2vUTSRCAh5SxbRdg,51576 +diffusers/pipelines/stable_diffusion_diffedit/__init__.py,sha256=JlcUNahRBm0uaPzappogqfjyLDsNW6IeyOfuLs4af5M,1358 +diffusers/pipelines/stable_diffusion_diffedit/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_diffedit/__pycache__/pipeline_stable_diffusion_diffedit.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_diffedit/pipeline_stable_diffusion_diffedit.py,sha256=Eq_Wm1HEgtPR5wbLU7mKecM0qlta-nHY8jg_Q0zuTLk,78549 +diffusers/pipelines/stable_diffusion_gligen/__init__.py,sha256=b4dZB5bUuZmEAcg7MmCyWZpyxNmMrlrByEQW_xwGGgI,1568 +diffusers/pipelines/stable_diffusion_gligen/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_gligen/__pycache__/pipeline_stable_diffusion_gligen.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_gligen/__pycache__/pipeline_stable_diffusion_gligen_text_image.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen.py,sha256=8Z_67jeBdl2ZCrukTkxYkRW73NazIYUWWGJK5WpWZOY,43505 +diffusers/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen_text_image.py,sha256=QR2zoPiBxX_6lsfeYyBuiZNoXwrtXgkW0rTfuCOR8I0,52092 +diffusers/pipelines/stable_diffusion_k_diffusion/__init__.py,sha256=EBpyQedEN-jfJ0qeLCFg9t28cFPNbNaniKIGM4ZMF14,1924 +diffusers/pipelines/stable_diffusion_k_diffusion/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_k_diffusion/__pycache__/pipeline_stable_diffusion_k_diffusion.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_k_diffusion/__pycache__/pipeline_stable_diffusion_xl_k_diffusion.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_k_diffusion.py,sha256=b7pXB2GE2n7W-_he4RoUxiQyvU-JtuUdudhA_tBuv1k,34096 +diffusers/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_xl_k_diffusion.py,sha256=KYwMUrnxseg_O9fLme-wVBOoNYSekS8YYr6K2lIf2aU,45496 +diffusers/pipelines/stable_diffusion_ldm3d/__init__.py,sha256=8p2npGKPPJbPaTa4swOWRMd24x36E563Bhc_mM29va0,1346 +diffusers/pipelines/stable_diffusion_ldm3d/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_ldm3d/__pycache__/pipeline_stable_diffusion_ldm3d.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_ldm3d/pipeline_stable_diffusion_ldm3d.py,sha256=vrIqC9I1bJvfv79G8SzuU2slJo3CXXmHzvSX9XMv-tE,51801 +diffusers/pipelines/stable_diffusion_panorama/__init__.py,sha256=af52eZSYshuw1d6kqKwx0C5Teopkx8UpO9ph_A4WI0Q,1358 +diffusers/pipelines/stable_diffusion_panorama/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_panorama/__pycache__/pipeline_stable_diffusion_panorama.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_panorama/pipeline_stable_diffusion_panorama.py,sha256=l1SmYHbJE-Nz1Iq82WckJEAmeg5xXbwBjxztcsk4X2k,60262 +diffusers/pipelines/stable_diffusion_safe/__init__.py,sha256=rRKtzOjuaHLDqSLSavcy2W8sEljso9MLhmEwrNiJFJ0,2751 +diffusers/pipelines/stable_diffusion_safe/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_safe/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_safe/__pycache__/pipeline_stable_diffusion_safe.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_safe/__pycache__/safety_checker.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_safe/pipeline_output.py,sha256=WGQS6-k9dPH0hYBj_dZMlHFkOvUUti9fjVv0Sf8LCjQ,1459 +diffusers/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py,sha256=zTC3C_MaqPnEgJjAqoCqpPU14JkKBtGG95AD5PaDstU,39634 +diffusers/pipelines/stable_diffusion_safe/safety_checker.py,sha256=Ewt9xIW9sKMGL0sbdjBgu4OhuzmjN_xqI87E6CwLBg0,5039 +diffusers/pipelines/stable_diffusion_sag/__init__.py,sha256=06vnWbASiG3o4sQ7CDlDrqEm6dSCerKdLODz1FS-EFE,1338 +diffusers/pipelines/stable_diffusion_sag/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_sag/__pycache__/pipeline_stable_diffusion_sag.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_sag/pipeline_stable_diffusion_sag.py,sha256=66XjcrcYb2AmD2jNsoqJ1fColkmBWBvSQeojNbGlUeg,48087 +diffusers/pipelines/stable_diffusion_xl/__init__.py,sha256=6lTMI458kVDLzQDeZxEBacdFxpj4xAY9CSZ6Xr_FWoY,3022 +diffusers/pipelines/stable_diffusion_xl/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_flax_stable_diffusion_xl.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_img2img.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_inpaint.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_instruct_pix2pix.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_xl/__pycache__/watermark.cpython-310.pyc,, +diffusers/pipelines/stable_diffusion_xl/pipeline_flax_stable_diffusion_xl.py,sha256=dGnKmKC_XzmMre6JTtdeTbXEYlenI-7CBmnyULK2BME,11280 +diffusers/pipelines/stable_diffusion_xl/pipeline_output.py,sha256=Isy1wE8hgoScXXHWVel5jRAzgPTelP-aZieugTOTgUc,1037 +diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py,sha256=QgIJOG9PV7jVKgsneQRDGdbdFq0JZM_s6UPT36GgJ5I,67847 +diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py,sha256=B6z9lXb432zNiY3aiwLvln2u8mXFULjQHlP5EuaYMaM,78850 +diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py,sha256=XqCUCWpzy2tEA3rpKjRK2TXjhCL87cO23-hqiTTPRKw,90574 +diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py,sha256=NkET5qoCbXLJYVvOELEacOWYjzBk53JfvVepIrFhPtk,52910 +diffusers/pipelines/stable_diffusion_xl/watermark.py,sha256=LDItvRnZKokIUchP0oIrO2Ew9AARhAP4MMrQY8maQ6Q,1458 +diffusers/pipelines/stable_video_diffusion/__init__.py,sha256=QtcDxzfLJ7loCDspiulKyKU6kd-l3twJyWBDPraD_94,1551 +diffusers/pipelines/stable_video_diffusion/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/stable_video_diffusion/__pycache__/pipeline_stable_video_diffusion.cpython-310.pyc,, +diffusers/pipelines/stable_video_diffusion/pipeline_stable_video_diffusion.py,sha256=x6jEr8eURAGh55q6Iey1pIIhxbWEzxD0NQmzvY_Vd-k,32628 +diffusers/pipelines/t2i_adapter/__init__.py,sha256=PgIg_SzwFAqWOML5BLHvuCTmu4p06MPT66xBpDShx8c,1556 +diffusers/pipelines/t2i_adapter/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/t2i_adapter/__pycache__/pipeline_stable_diffusion_adapter.cpython-310.pyc,, +diffusers/pipelines/t2i_adapter/__pycache__/pipeline_stable_diffusion_xl_adapter.cpython-310.pyc,, +diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_adapter.py,sha256=izK-JFnaVGq-hbOcPwJVaymM-MQVBJEJHRo-QcT1PiA,47764 +diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py,sha256=qxnmqebGNnOutOPnQowpa8rmGUS-iJmQND4euuO4z9w,69409 +diffusers/pipelines/text_to_video_synthesis/__init__.py,sha256=7-NplGtgnp5GUu4XN_STE9fqAtFCAc6FF3lphjbDBhs,1979 +diffusers/pipelines/text_to_video_synthesis/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_synth.cpython-310.pyc,, +diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_synth_img2img.cpython-310.pyc,, +diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_zero.cpython-310.pyc,, +diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_zero_sdxl.cpython-310.pyc,, +diffusers/pipelines/text_to_video_synthesis/pipeline_output.py,sha256=12i4JmK2TgksR46kwOSw02McNrV7qksA4MFAw6KB6_Q,735 +diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth.py,sha256=GEH-FNE5tn-dTVB3yt91f_suYhtBzDnrfwEUHnxlvr4,31922 +diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py,sha256=EcVA37GTRP5XWFFmKxoGQzW0BH7KLRkM_4uwB0r-EsM,35440 +diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero.py,sha256=Y6oKIsoXcLxPOcmeCtZ-YNa7Jg7XX3DU1zocXDWGoIE,45882 +diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero_sdxl.py,sha256=Z6hH40KmN_ZD8lMG7xqjW4MeGf1XVF-kyvzNnt7CaV4,64869 +diffusers/pipelines/transformers_loading_utils.py,sha256=98wKUHN89Q1nmmat046hgQxLDlnZNj9Ww4TLB5W52pQ,5281 +diffusers/pipelines/unclip/__init__.py,sha256=jBYZIN7NhTKM_Oq7ipJ4JaMXO-GtdchmFWe07gDerfA,1752 +diffusers/pipelines/unclip/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/unclip/__pycache__/pipeline_unclip.cpython-310.pyc,, +diffusers/pipelines/unclip/__pycache__/pipeline_unclip_image_variation.cpython-310.pyc,, +diffusers/pipelines/unclip/__pycache__/text_proj.cpython-310.pyc,, +diffusers/pipelines/unclip/pipeline_unclip.py,sha256=yR6XocwNUDPzOknyGqq-OepEGMxTJEu9Y99fVN3attc,22445 +diffusers/pipelines/unclip/pipeline_unclip_image_variation.py,sha256=08S5ysw2k0bw0gM3JKU_PEqWWvXcUUBraN02ejB1k7E,19347 +diffusers/pipelines/unclip/text_proj.py,sha256=UkICP2P1vL8tVJr92A6gPFuqrjl03fAs-HiFFkZDvfU,4286 +diffusers/pipelines/unidiffuser/__init__.py,sha256=GvGtf-AToJXNHxv3RAo5_I_9zPQjDFbMTAHICCt-4xY,1814 +diffusers/pipelines/unidiffuser/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/unidiffuser/__pycache__/modeling_text_decoder.cpython-310.pyc,, +diffusers/pipelines/unidiffuser/__pycache__/modeling_uvit.cpython-310.pyc,, +diffusers/pipelines/unidiffuser/__pycache__/pipeline_unidiffuser.cpython-310.pyc,, +diffusers/pipelines/unidiffuser/modeling_text_decoder.py,sha256=pG_QeaB4gBm4sVBgR_LfH_o6Jz91zRqZHrO5ac6xn6o,14111 +diffusers/pipelines/unidiffuser/modeling_uvit.py,sha256=To1JmzOZ1nIKTkHI0Df1qS3Z2ddVuEMECVEjqBhhT8M,54286 +diffusers/pipelines/unidiffuser/pipeline_unidiffuser.py,sha256=7kc_9YFKE4TMoQDQYCZFrgbJc0kUJvxZfG5lAhn8dVo,68891 +diffusers/pipelines/visualcloze/__init__.py,sha256=Cgc6UqhelXaUEYJbmEf_kGBY90hS3wx7eIiO1C5A0g4,1502 +diffusers/pipelines/visualcloze/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/visualcloze/__pycache__/pipeline_visualcloze_combined.cpython-310.pyc,, +diffusers/pipelines/visualcloze/__pycache__/pipeline_visualcloze_generation.cpython-310.pyc,, +diffusers/pipelines/visualcloze/__pycache__/visualcloze_utils.cpython-310.pyc,, +diffusers/pipelines/visualcloze/pipeline_visualcloze_combined.py,sha256=bTEZei9dyD81tk0NBsBFV5RyEv-658XQpVvMTZWONs8,23999 +diffusers/pipelines/visualcloze/pipeline_visualcloze_generation.py,sha256=16iiLTqGAtUEYeLYQRoHpQyLxiq2dfsp4zUAiCNP5fA,45329 +diffusers/pipelines/visualcloze/visualcloze_utils.py,sha256=4DPWGxtCkC51lPoQTMe3Xlv5Lx8RKo5Uoo6tnDAwr-Y,10864 +diffusers/pipelines/wan/__init__.py,sha256=SOHLX5iUuab80gLF9ZYo54pzFWj0hiSoZU4hgWesuuQ,1677 +diffusers/pipelines/wan/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/wan/__pycache__/pipeline_output.cpython-310.pyc,, +diffusers/pipelines/wan/__pycache__/pipeline_wan.cpython-310.pyc,, +diffusers/pipelines/wan/__pycache__/pipeline_wan_i2v.cpython-310.pyc,, +diffusers/pipelines/wan/__pycache__/pipeline_wan_vace.cpython-310.pyc,, +diffusers/pipelines/wan/__pycache__/pipeline_wan_video2video.cpython-310.pyc,, +diffusers/pipelines/wan/pipeline_output.py,sha256=EjM_BX0gQD1hQ78lbUM40OtvwPqTtncuQdvP5YMBjvE,605 +diffusers/pipelines/wan/pipeline_wan.py,sha256=b2HYq72rpwhlF0E2JfP0w2xzqtDsl1xr7C_GMuO8PnI,27276 +diffusers/pipelines/wan/pipeline_wan_i2v.py,sha256=WTgNIgEC468rpMJFk5hFbxbe_ONDBfdzVjaCIsBFr0Y,34851 +diffusers/pipelines/wan/pipeline_wan_vace.py,sha256=iucIWvDyyiZqluK5nNcRFQ_XmsekaI1pcSlmVgMgBbs,48324 +diffusers/pipelines/wan/pipeline_wan_video2video.py,sha256=tHloW45GFTDBzE1fggRUZDPKqv2ZznT55-5yQ5jdKAE,33331 +diffusers/pipelines/wuerstchen/__init__.py,sha256=JSCoPCwV_rBJiCy4jbILRoAgQSITS4-j77qOPmzy284,2100 +diffusers/pipelines/wuerstchen/__pycache__/__init__.cpython-310.pyc,, +diffusers/pipelines/wuerstchen/__pycache__/modeling_paella_vq_model.cpython-310.pyc,, +diffusers/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_common.cpython-310.pyc,, +diffusers/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_diffnext.cpython-310.pyc,, +diffusers/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_prior.cpython-310.pyc,, +diffusers/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen.cpython-310.pyc,, +diffusers/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen_combined.cpython-310.pyc,, +diffusers/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen_prior.cpython-310.pyc,, +diffusers/pipelines/wuerstchen/modeling_paella_vq_model.py,sha256=lmTIqLgXK7PXwZw3cGBHpEYn0oCZMtFSGsrGrXHkLTI,6925 +diffusers/pipelines/wuerstchen/modeling_wuerstchen_common.py,sha256=mx0bj5b87g590UQhoFWY_L0ht_RTIynaPQa9DLk9MTU,2713 +diffusers/pipelines/wuerstchen/modeling_wuerstchen_diffnext.py,sha256=-vKLdD7RkbDlUPpX8sgjKUTpbq7bwBdHP6y8L8kPwS0,10423 +diffusers/pipelines/wuerstchen/modeling_wuerstchen_prior.py,sha256=VhlC4cLgRwe0r7Ng5vrJad9NwK_3BOFzcklNpvXzSrI,7335 +diffusers/pipelines/wuerstchen/pipeline_wuerstchen.py,sha256=yujVa8eLz7_NHdbrqhPCg02gR0cz_soxAZQXCA0wHeM,20832 +diffusers/pipelines/wuerstchen/pipeline_wuerstchen_combined.py,sha256=zP-zrPDO0_DMT0BRN_kjkiuw2O9eO0f3uOYfvnvGtyo,16686 +diffusers/pipelines/wuerstchen/pipeline_wuerstchen_prior.py,sha256=d4iA2SSyMsLm6RhHCZliks58a2D55kLXuvB5nMAKPcg,24127 +diffusers/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0 +diffusers/quantizers/__init__.py,sha256=CtSmZkDrIAraPPxpKO-M5Wzeb0d7JPc-IRZnEgNUkg4,8867 +diffusers/quantizers/__pycache__/__init__.cpython-310.pyc,, +diffusers/quantizers/__pycache__/auto.cpython-310.pyc,, +diffusers/quantizers/__pycache__/base.cpython-310.pyc,, +diffusers/quantizers/__pycache__/quantization_config.cpython-310.pyc,, +diffusers/quantizers/auto.py,sha256=_rlfK4_5XQuP3Af0ArlPPWgq3E0_gwxv6z-nSQZTls8,5789 +diffusers/quantizers/base.py,sha256=lx6FKnuNn_v88zcDf4E_IDJd0PDjCD4tlhxe7xlZLi0,9682 +diffusers/quantizers/bitsandbytes/__init__.py,sha256=ILCM6ZopnzrhM_fW1oh4J_YCNsaEQAptcoTuSVgXab8,170 +diffusers/quantizers/bitsandbytes/__pycache__/__init__.cpython-310.pyc,, +diffusers/quantizers/bitsandbytes/__pycache__/bnb_quantizer.cpython-310.pyc,, +diffusers/quantizers/bitsandbytes/__pycache__/utils.cpython-310.pyc,, +diffusers/quantizers/bitsandbytes/bnb_quantizer.py,sha256=hPg52A4NwlYqW1u5M0lhccXeW-sUGAHH0ipzdI7Etj0,26268 +diffusers/quantizers/bitsandbytes/utils.py,sha256=48clXsL2q2NQ41n4dcCAhXsRNxyrkGgXYoJUUfHmHlY,13690 +diffusers/quantizers/gguf/__init__.py,sha256=2bxvfZbFr4xqm953cZaGJMgSCRiGJAWwbwKxNRQIEs4,42 +diffusers/quantizers/gguf/__pycache__/__init__.cpython-310.pyc,, +diffusers/quantizers/gguf/__pycache__/gguf_quantizer.cpython-310.pyc,, +diffusers/quantizers/gguf/__pycache__/utils.cpython-310.pyc,, +diffusers/quantizers/gguf/gguf_quantizer.py,sha256=AP6kUgk9bDh-gBuvLqWhUfaMgrpTF7O_VWEI6QDAfUM,6023 +diffusers/quantizers/gguf/utils.py,sha256=K9yweZftcOz8-4KaSr5gAo8RmXGH7rP8I775Yw9n5ww,16198 +diffusers/quantizers/quantization_config.py,sha256=vkZTAdtuV0vG0L_HOHURI-Qm91Deb__SEdmFPtvP2EE,32541 +diffusers/quantizers/quanto/__init__.py,sha256=ynS7j_VTG-QtimbyxHAaihUmi6eVqEDxA5dnKGjeS5M,46 +diffusers/quantizers/quanto/__pycache__/__init__.cpython-310.pyc,, +diffusers/quantizers/quanto/__pycache__/quanto_quantizer.cpython-310.pyc,, +diffusers/quantizers/quanto/__pycache__/utils.cpython-310.pyc,, +diffusers/quantizers/quanto/quanto_quantizer.py,sha256=nB7c2FHGzZUseLaiwTBKLSUftmLO7kZSOJGvk5I4hGw,6336 +diffusers/quantizers/quanto/utils.py,sha256=6-EaqWTbhb0dkuJ0C8XRDIpMmowPVjJ_C4rPaNNHMkc,2448 +diffusers/quantizers/torchao/__init__.py,sha256=tJimVpSGQsz3owo3yzh2SuCg7NQfiMnrHkAHyYHkmGA,662 +diffusers/quantizers/torchao/__pycache__/__init__.cpython-310.pyc,, +diffusers/quantizers/torchao/__pycache__/torchao_quantizer.cpython-310.pyc,, +diffusers/quantizers/torchao/torchao_quantizer.py,sha256=QaI_oVApbNqh9h81Eb8jlRSJo6MuhR1sO1sc7c-k4nM,14177 +diffusers/schedulers/__init__.py,sha256=HYxAaUCP3JSJXKm4FVJBGxqqwyqqgdDooE0YupPh7V4,11182 +diffusers/schedulers/__pycache__/__init__.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_amused.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_consistency_decoder.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_consistency_models.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_cosine_dpmsolver_multistep.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_ddim.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_ddim_cogvideox.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_ddim_flax.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_ddim_inverse.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_ddim_parallel.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_ddpm.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_ddpm_flax.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_ddpm_parallel.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_ddpm_wuerstchen.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_deis_multistep.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_dpm_cogvideox.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_dpmsolver_multistep.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_dpmsolver_multistep_flax.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_dpmsolver_multistep_inverse.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_dpmsolver_sde.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_dpmsolver_singlestep.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_edm_dpmsolver_multistep.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_edm_euler.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_euler_ancestral_discrete.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_euler_discrete.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_euler_discrete_flax.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_flow_match_euler_discrete.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_flow_match_heun_discrete.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_flow_match_lcm.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_heun_discrete.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_ipndm.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_k_dpm_2_ancestral_discrete.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_k_dpm_2_discrete.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_karras_ve_flax.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_lcm.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_lms_discrete.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_lms_discrete_flax.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_pndm.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_pndm_flax.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_repaint.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_sasolver.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_scm.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_sde_ve.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_sde_ve_flax.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_tcd.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_unclip.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_unipc_multistep.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_utils.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_utils_flax.cpython-310.pyc,, +diffusers/schedulers/__pycache__/scheduling_vq_diffusion.cpython-310.pyc,, +diffusers/schedulers/deprecated/__init__.py,sha256=3QlQ4gSBFu4zUkY3S5KLxd9sukbxLv8Aj4eO0Rymaq0,1349 +diffusers/schedulers/deprecated/__pycache__/__init__.cpython-310.pyc,, +diffusers/schedulers/deprecated/__pycache__/scheduling_karras_ve.cpython-310.pyc,, +diffusers/schedulers/deprecated/__pycache__/scheduling_sde_vp.cpython-310.pyc,, +diffusers/schedulers/deprecated/scheduling_karras_ve.py,sha256=CrJZjLD0QQ1mY_cE80aIDDMLzh9U6sn1iS2eEuCZFCc,9724 +diffusers/schedulers/deprecated/scheduling_sde_vp.py,sha256=4TK-8FOzNWnvlSrFIaOjhPVdNk57zeTDQYjtBdDgr3k,4294 +diffusers/schedulers/scheduling_amused.py,sha256=pioDeoYfXWP2CnAiI3-lczDdfSbkfXRC9m9REN_kmvI,6590 +diffusers/schedulers/scheduling_consistency_decoder.py,sha256=IKNAkeVZIxTPk_hS6QBvFbkjv_h8yWbtU3FmIUSXKVI,6817 +diffusers/schedulers/scheduling_consistency_models.py,sha256=wIgDb_44DIa8phTMn7WPyzAtoWVXRKSkyWLYLF1hI0Q,18699 +diffusers/schedulers/scheduling_cosine_dpmsolver_multistep.py,sha256=JZgxQvw4BJHG2RCUe9wA9dwmjKZbJa9fuR0CJVZVu-w,24873 +diffusers/schedulers/scheduling_ddim.py,sha256=a7PIMKeTepas2b_Du6ttdUBM7aAYnEI4mdA8nTTahZk,24919 +diffusers/schedulers/scheduling_ddim_cogvideox.py,sha256=Fa4lj0jnQDms_cpr26bITHnnylU4HLrWpDQkmgqLDsE,21351 +diffusers/schedulers/scheduling_ddim_flax.py,sha256=4w8-PFsXdvtfx7cE8HKwt50xBF5mR30am7BYIaYOIlY,13146 +diffusers/schedulers/scheduling_ddim_inverse.py,sha256=PPEfEduk6yAHGtok30IfG2uAn_3kvHx0H3BvgpcYzr0,17792 +diffusers/schedulers/scheduling_ddim_parallel.py,sha256=vPmnAxJvA_9Fq66rnzgkxr9hdzc1hDF41iryM51BS-E,31598 +diffusers/schedulers/scheduling_ddpm.py,sha256=g9KnAlLfVxUN0uZRcUsLmwSe9MmjSuY_xUIzmr9im6g,26019 +diffusers/schedulers/scheduling_ddpm_flax.py,sha256=IxYvg6SH17czD-N8zHdvuhGxpCI7JePypvJfOTQsLjk,12574 +diffusers/schedulers/scheduling_ddpm_parallel.py,sha256=2yoo64zf6I0RIAJ9aPDpfnbk7qv9t5GEYuHlgXy_Sog,31041 +diffusers/schedulers/scheduling_ddpm_wuerstchen.py,sha256=w7feuL53AY_363GyU1PWVGXfDHcyzJSbWOOtcXaJrAE,8938 +diffusers/schedulers/scheduling_deis_multistep.py,sha256=9ZFbpqO01jlhaKa3REl5X4gcMMChYABxujmhr_q3AG8,39716 +diffusers/schedulers/scheduling_dpm_cogvideox.py,sha256=QQ_3P5LLDrQvltlSuXYVkFl7AHSPXvlwBy_GnrJj2CA,23347 +diffusers/schedulers/scheduling_dpmsolver_multistep.py,sha256=wSYtAGNtxcWmWxp8rIj6yOK6faOi2KY3RZ2J_fRkGBY,55336 +diffusers/schedulers/scheduling_dpmsolver_multistep_flax.py,sha256=UEiHAqI4HWvItbfTqEy90pmDtrig51Owx3I-m66jwc8,28849 +diffusers/schedulers/scheduling_dpmsolver_multistep_inverse.py,sha256=fSTExThSvMHA9nbNCq096KkXSGUkJwKeVMDsybvn860,48871 +diffusers/schedulers/scheduling_dpmsolver_sde.py,sha256=MV6XfmTZhUyJed1uJv4S857Wxlz5gZrV1-R4MEdQDig,29491 +diffusers/schedulers/scheduling_dpmsolver_singlestep.py,sha256=ulU5srNvJFpVZWqxpSwXEPmd5CUOTcET9NffFgaWzLk,54259 +diffusers/schedulers/scheduling_edm_dpmsolver_multistep.py,sha256=0b8CYA_EQ0C2CL1AonsiyN6AC5hTBmUi0Venz9vFe90,31915 +diffusers/schedulers/scheduling_edm_euler.py,sha256=lU2uJnAOgIyi2IEr5NI396G049J2Fejm30OpltCEru0,19500 +diffusers/schedulers/scheduling_euler_ancestral_discrete.py,sha256=Z5Azbf97q3XIriLTXtPq92coqTUHHZ3gTST-2aRr5Vg,21098 +diffusers/schedulers/scheduling_euler_discrete.py,sha256=gKhj6feeZ48ORwArAwwaWWmnU5As73MP6BAhkf7zKzY,34944 +diffusers/schedulers/scheduling_euler_discrete_flax.py,sha256=UHWWpKBY-wC6wgXs8f0qMxp1LPH_VXNiTLRrK3vEXcU,10809 +diffusers/schedulers/scheduling_flow_match_euler_discrete.py,sha256=CQ-0P0WfqFHl_w84bsyDThiCbZavW_3MVYKRjwP5Lac,23837 +diffusers/schedulers/scheduling_flow_match_heun_discrete.py,sha256=Yv0bU0HKMhthpzEYULiZs-O8H1Hg934r7T17bgAh8-Y,12154 +diffusers/schedulers/scheduling_flow_match_lcm.py,sha256=gxRWyVmy49gC8Gp7kfWhsnDkTQq6oHs3-4xYO5mpPJ0,23951 +diffusers/schedulers/scheduling_heun_discrete.py,sha256=JkVoEF-vdtBweqjUs57o8JUA3K9uvGW_gHGY5f2tr_8,27708 +diffusers/schedulers/scheduling_ipndm.py,sha256=Ev9x-4E5QnubgRjwLQQimMLXhpEV7NiffKK7RGYWF-c,8768 +diffusers/schedulers/scheduling_k_dpm_2_ancestral_discrete.py,sha256=Zk9Q4-pIu0w0jSkpqPKuDA7UWQmCwN0uA5KYbCeSDr8,27593 +diffusers/schedulers/scheduling_k_dpm_2_discrete.py,sha256=fE24rs9yyxtAR6LAPbtYtgEwYjmv2wG0l5EGWM55vOA,26140 +diffusers/schedulers/scheduling_karras_ve_flax.py,sha256=HFO9Xjg4SBTRFe3e0UgtpDE5dAEHUWOxVA1R3avuXhQ,9630 +diffusers/schedulers/scheduling_lcm.py,sha256=Wcul9ObZdS3JcT2DlW5P9x25AFZLbJYy6Pt4eYdIHx0,32010 +diffusers/schedulers/scheduling_lms_discrete.py,sha256=YbnEAXPU4Rho1-BcMbaNJ0lk4m2dfZ5We1Ih0Cyyq_Y,24372 +diffusers/schedulers/scheduling_lms_discrete_flax.py,sha256=JSeHrj700b-rrDqWpPOUREktzrKwb4S9sHYxUwtlBxc,11077 +diffusers/schedulers/scheduling_pndm.py,sha256=n8P3-s6I6LKA6pU0yZfs6CoopFvArDerMdcSz12zA88,21731 +diffusers/schedulers/scheduling_pndm_flax.py,sha256=o2EiaUcOeeoJKsj4uABXrs3QUH_MwrgQrXCihEXay7U,21555 +diffusers/schedulers/scheduling_repaint.py,sha256=M1dddhIGRnQgnRPMxbOR1m2jgrsBkZo-fPpiSvKbT70,15762 +diffusers/schedulers/scheduling_sasolver.py,sha256=7dYuwAbKLsbr0gjGLwwusqzEhM1eBLa770sC9mNFQMI,55116 +diffusers/schedulers/scheduling_scm.py,sha256=3UgiM11atSskN-LYf9TloBC3fztxmRHu9umhBOtZ698,11402 +diffusers/schedulers/scheduling_sde_ve.py,sha256=_y-9IBk2yxbz1iwTKngBJipWSkiFrdk8ky9LDkLvYo4,13321 +diffusers/schedulers/scheduling_sde_ve_flax.py,sha256=6flCFfSIiGpAwEZK_UhjKmWjzsIXOD5UAORrpkIAbgM,12142 +diffusers/schedulers/scheduling_tcd.py,sha256=tiAvTdewr6CtMe2wn4z7_-iUDcVCvWdK27aiMxq5kJM,34749 +diffusers/schedulers/scheduling_unclip.py,sha256=T7HJXANwbhVCo5FaFkvmEuUl6geBDuz-DxVm5s81Z7o,15056 +diffusers/schedulers/scheduling_unipc_multistep.py,sha256=GNlYBYL4Pk_74A_dv7aHvWfOImW2JHqxZUK-V_0uxUY,46681 +diffusers/schedulers/scheduling_utils.py,sha256=6ArJlV6zTYUwbvcfKH0WuFFr_jBK8Qs7G0e5OZCv-qI,8708 +diffusers/schedulers/scheduling_utils_flax.py,sha256=1vxGnNAjeLqxx0VMOEkvfeDEopQqfAmr2pMMJtaKXNg,12153 +diffusers/schedulers/scheduling_vq_diffusion.py,sha256=xkYjAqC_dF5kfazvKnct6V0Rg19GdBe9Qq2AHlNYi2Y,22954 +diffusers/training_utils.py,sha256=6YUrhwteivj69J0_dqKmvsqaWZt-j-BbgSkBz38pE8E,26769 +diffusers/utils/__init__.py,sha256=9Ilx8KvWskzj3WQiJrUBckDWCGSF0-naGmjZMnwN98U,4587 +diffusers/utils/__pycache__/__init__.cpython-310.pyc,, +diffusers/utils/__pycache__/accelerate_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/constants.cpython-310.pyc,, +diffusers/utils/__pycache__/deprecation_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/doc_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_bitsandbytes_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_flax_and_transformers_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_flax_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_gguf_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_note_seq_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_onnx_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_optimum_quanto_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_pt_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_torch_and_librosa_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_torch_and_scipy_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_torch_and_torchsde_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_torch_and_transformers_and_k_diffusion_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_torch_and_transformers_and_onnx_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_torch_and_transformers_and_opencv_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_torch_and_transformers_and_sentencepiece_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_torch_and_transformers_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_torchao_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dummy_transformers_and_torch_and_note_seq_objects.cpython-310.pyc,, +diffusers/utils/__pycache__/dynamic_modules_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/export_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/hub_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/import_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/loading_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/logging.cpython-310.pyc,, +diffusers/utils/__pycache__/outputs.cpython-310.pyc,, +diffusers/utils/__pycache__/peft_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/pil_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/remote_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/source_code_parsing_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/state_dict_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/testing_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/torch_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/typing_utils.cpython-310.pyc,, +diffusers/utils/__pycache__/versions.cpython-310.pyc,, +diffusers/utils/accelerate_utils.py,sha256=ZkopOK29_6QrrdkIPidHnO078B8BtzzFSts3ul2cP8Q,1839 +diffusers/utils/constants.py,sha256=FIBvtiRBdXih2T03vmKSVYX8X0KAfoMUhZ0Sf-hjGfw,3307 +diffusers/utils/deprecation_utils.py,sha256=WWegSa1ZBX2DNfFp-L2wQuKAyGqXsc275Eua0vw7Z8o,2103 +diffusers/utils/doc_utils.py,sha256=3x-UhJrm7eNcJiQAYnYefoH7Q-V2qPmc04TzbEr25-o,1348 +diffusers/utils/dummy_bitsandbytes_objects.py,sha256=7uoVirIvcuylaDyUas94Wc9AAKJMJS3GEv-in3bEqJA,527 +diffusers/utils/dummy_flax_and_transformers_objects.py,sha256=XyiqnjacRb86sS9F_VwniBrLLEmff2cgJM2X4T_RAg4,2358 +diffusers/utils/dummy_flax_objects.py,sha256=EIyO7jYPH4yjuBIxysZWE0rka3qPLEl1TmMBt5SwXNA,5316 +diffusers/utils/dummy_gguf_objects.py,sha256=H0SYZuOON9cFlkyYSTcUJJk4skYjgjIu-wDausCm0sU,499 +diffusers/utils/dummy_note_seq_objects.py,sha256=DffX40mDzWTMCyYhKudgIeBhtqTSpiSkVzcAMRue8dY,506 +diffusers/utils/dummy_onnx_objects.py,sha256=4Z61m3P9NUwbebsK58wAKs6y32Id6UaiSRyeHXo3ecA,493 +diffusers/utils/dummy_optimum_quanto_objects.py,sha256=_k-3g7WAYcJO0-38rJrHX7aIAzkkICtj2ISiggVFiz8,529 +diffusers/utils/dummy_pt_objects.py,sha256=d6DDSSV-h2-TYWXrzj8t06vdWLhQDOPQ9hYgvcBM1o8,52845 +diffusers/utils/dummy_torch_and_librosa_objects.py,sha256=JUfqU2n3tSKHyWbjSXrpdW_jr-YbMxAvAhLlPa2_Rxs,948 +diffusers/utils/dummy_torch_and_scipy_objects.py,sha256=zOLdmqbtma5nakkdYgoErISV28yaztmBLI3wrC2Z_bU,537 +diffusers/utils/dummy_torch_and_torchsde_objects.py,sha256=EJiExfXva8tnRJEn-VaCkcII31WnPr2HqdTh3PBQ-jk,985 +diffusers/utils/dummy_torch_and_transformers_and_k_diffusion_objects.py,sha256=IMw6Qs9tTdRrMUXyM_Bc_BuJBvw0OVVHNZMOk3suF7g,1151 +diffusers/utils/dummy_torch_and_transformers_and_onnx_objects.py,sha256=SiKni7YZ-pmZrurHU3-lhbDGKOGCCVxSK3GJbrARqgU,3023 +diffusers/utils/dummy_torch_and_transformers_and_opencv_objects.py,sha256=Hvskt_HEoCkRikDyiYWQ95CsSv0fX7jYqxDxjJuNIZc,601 +diffusers/utils/dummy_torch_and_transformers_and_sentencepiece_objects.py,sha256=rauUQkG4sLSyBVeEbfp5fhJFJUGqw273oXbN_KC8NIM,1637 +diffusers/utils/dummy_torch_and_transformers_objects.py,sha256=gXYEKtwJUG1-XZ_6S9Fx_orowRGHZrFNO2CI1xZgcgA,89839 +diffusers/utils/dummy_torchao_objects.py,sha256=XiJAoV11rr_7aSgkb0vgYkdZOGjM1ouTuQZ4YrXzJ4g,502 +diffusers/utils/dummy_transformers_and_torch_and_note_seq_objects.py,sha256=z-JrPgPo2dWv-buMytUqBd6QqEx8Uha6M1cKa6gR4Dc,621 +diffusers/utils/dynamic_modules_utils.py,sha256=u3VLfAM7JpDHjW-MCBePAE-xelL3IEJrhOgaOMnq3Dc,20680 +diffusers/utils/export_utils.py,sha256=YIleDkRb3-m7a2NryRlnvSXLBbjjpkEfR7U5z-cUdO4,7749 +diffusers/utils/hub_utils.py,sha256=kSbSCJ3MHsJOytCuwiqDT-RS09WBhsraRtGmboSxW7g,23904 +diffusers/utils/import_utils.py,sha256=rX4gMytXrMFTBKimJKtoe4Ep1_dxuqIKISA7ppiLy8g,31444 +diffusers/utils/loading_utils.py,sha256=6Qgm5vRGICI7daRewJOKGSBob4tGaoBOApAHhF5waPY,5300 +diffusers/utils/logging.py,sha256=cFr8Bo_UydpL0NEeIwx7CHMUwyTWaFrZztY__nPfP3s,9470 +diffusers/utils/model_card_template.md,sha256=ZhGhzjnMT2oPLbHnC0UYRNEpVC-okH-MLKjvkYsh-Ds,550 +diffusers/utils/outputs.py,sha256=-bBcKV7uRT5nwB8QnqEN83yk4HhoJ9ToA729sFN3A9E,5030 +diffusers/utils/peft_utils.py,sha256=kedTtDw2WtpijASU1_pXepfR06BUmWpzaIvmBJ7H_L8,14131 +diffusers/utils/pil_utils.py,sha256=mNv1FfHvtdW-lpOemxwe-dNoSfSF_sptgpYELP-bA20,1979 +diffusers/utils/remote_utils.py,sha256=DHtS6_2ggX45YltaE9i_o5mb5gJd-VnGjr7UJ18wg2Q,16117 +diffusers/utils/source_code_parsing_utils.py,sha256=Mk4KHfymwXHYmV2zZdVsDSn-VJ97HvUV6fGfYSuYn8g,1863 +diffusers/utils/state_dict_utils.py,sha256=ygt17dCeHowH5QxwQldJHZpc0sS_E-aCuzpjKmAZ3HA,14470 +diffusers/utils/testing_utils.py,sha256=3Vqk3Yvv2gwt0O_5wyJV28nrO4RwmXnk8_3a1BHX7Kc,50706 +diffusers/utils/torch_utils.py,sha256=2q-WsALki6ONS-V1gYnouAK_dAA1J242tjTnv6WyZUw,7250 +diffusers/utils/typing_utils.py,sha256=yeuCJmb1t5n5rG1JRPJo33KO7tg_m9ZwSXQcPKiKyFA,3400 +diffusers/utils/versions.py,sha256=-e7XW1TzZ-tsRo9PMQHp-hNGYHuVDFzLtwg3uAJzqdI,4333 +diffusers/video_processor.py,sha256=jKcIhqCs4EUcEbKEF23u0TJs1nfZrPwdXKr-YzrZ3O0,5398 diff --git a/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/REQUESTED b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/REQUESTED new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/WHEEL b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/WHEEL new file mode 100644 index 0000000000000000000000000000000000000000..505164bc02d63fe6b0b3299f849a77c5f1beeb41 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/WHEEL @@ -0,0 +1,5 @@ +Wheel-Version: 1.0 +Generator: setuptools (75.8.0) +Root-Is-Purelib: true +Tag: py3-none-any + diff --git a/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/entry_points.txt b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/entry_points.txt new file mode 100644 index 0000000000000000000000000000000000000000..4d8cdefa4968bbed7a122fa7a34940bf9a15360b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/entry_points.txt @@ -0,0 +1,2 @@ +[console_scripts] +diffusers-cli = diffusers.commands.diffusers_cli:main diff --git a/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/top_level.txt b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/top_level.txt new file mode 100644 index 0000000000000000000000000000000000000000..6033efb6dbaaff9bc81792fd75a6b39d9f195aeb --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers-0.34.0.dist-info/top_level.txt @@ -0,0 +1 @@ +diffusers diff --git a/venv/lib/python3.10/site-packages/diffusers/__init__.py b/venv/lib/python3.10/site-packages/diffusers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..4c880d794cb96051c42a9201ff78dc4b79ee3ef8 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/__init__.py @@ -0,0 +1,1244 @@ +__version__ = "0.34.0" + +from typing import TYPE_CHECKING + +from .utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + is_accelerate_available, + is_bitsandbytes_available, + is_flax_available, + is_gguf_available, + is_k_diffusion_available, + is_librosa_available, + is_note_seq_available, + is_onnx_available, + is_opencv_available, + is_optimum_quanto_available, + is_scipy_available, + is_sentencepiece_available, + is_torch_available, + is_torchao_available, + is_torchsde_available, + is_transformers_available, +) + + +# Lazy Import based on +# https://github.com/huggingface/transformers/blob/main/src/transformers/__init__.py + +# When adding a new object to this init, please add it to `_import_structure`. The `_import_structure` is a dictionary submodule to list of object names, +# and is used to defer the actual importing for when the objects are requested. +# This way `import diffusers` provides the names in the namespace without actually importing anything (and especially none of the backends). + +_import_structure = { + "configuration_utils": ["ConfigMixin"], + "hooks": [], + "loaders": ["FromOriginalModelMixin"], + "models": [], + "pipelines": [], + "quantizers.quantization_config": [], + "schedulers": [], + "utils": [ + "OptionalDependencyNotAvailable", + "is_flax_available", + "is_inflect_available", + "is_invisible_watermark_available", + "is_k_diffusion_available", + "is_k_diffusion_version", + "is_librosa_available", + "is_note_seq_available", + "is_onnx_available", + "is_scipy_available", + "is_torch_available", + "is_torchsde_available", + "is_transformers_available", + "is_transformers_version", + "is_unidecode_available", + "logging", + ], +} + +try: + if not is_torch_available() and not is_accelerate_available() and not is_bitsandbytes_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_bitsandbytes_objects + + _import_structure["utils.dummy_bitsandbytes_objects"] = [ + name for name in dir(dummy_bitsandbytes_objects) if not name.startswith("_") + ] +else: + _import_structure["quantizers.quantization_config"].append("BitsAndBytesConfig") + +try: + if not is_torch_available() and not is_accelerate_available() and not is_gguf_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_gguf_objects + + _import_structure["utils.dummy_gguf_objects"] = [ + name for name in dir(dummy_gguf_objects) if not name.startswith("_") + ] +else: + _import_structure["quantizers.quantization_config"].append("GGUFQuantizationConfig") + +try: + if not is_torch_available() and not is_accelerate_available() and not is_torchao_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_torchao_objects + + _import_structure["utils.dummy_torchao_objects"] = [ + name for name in dir(dummy_torchao_objects) if not name.startswith("_") + ] +else: + _import_structure["quantizers.quantization_config"].append("TorchAoConfig") + +try: + if not is_torch_available() and not is_accelerate_available() and not is_optimum_quanto_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_optimum_quanto_objects + + _import_structure["utils.dummy_optimum_quanto_objects"] = [ + name for name in dir(dummy_optimum_quanto_objects) if not name.startswith("_") + ] +else: + _import_structure["quantizers.quantization_config"].append("QuantoConfig") + +try: + if not is_onnx_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_onnx_objects # noqa F403 + + _import_structure["utils.dummy_onnx_objects"] = [ + name for name in dir(dummy_onnx_objects) if not name.startswith("_") + ] + +else: + _import_structure["pipelines"].extend(["OnnxRuntimeModel"]) + +try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_pt_objects # noqa F403 + + _import_structure["utils.dummy_pt_objects"] = [name for name in dir(dummy_pt_objects) if not name.startswith("_")] + +else: + _import_structure["hooks"].extend( + [ + "FasterCacheConfig", + "HookRegistry", + "PyramidAttentionBroadcastConfig", + "apply_faster_cache", + "apply_pyramid_attention_broadcast", + ] + ) + _import_structure["models"].extend( + [ + "AllegroTransformer3DModel", + "AsymmetricAutoencoderKL", + "AuraFlowTransformer2DModel", + "AutoencoderDC", + "AutoencoderKL", + "AutoencoderKLAllegro", + "AutoencoderKLCogVideoX", + "AutoencoderKLCosmos", + "AutoencoderKLHunyuanVideo", + "AutoencoderKLLTXVideo", + "AutoencoderKLMagvit", + "AutoencoderKLMochi", + "AutoencoderKLTemporalDecoder", + "AutoencoderKLWan", + "AutoencoderOobleck", + "AutoencoderTiny", + "AutoModel", + "CacheMixin", + "ChromaTransformer2DModel", + "CogVideoXTransformer3DModel", + "CogView3PlusTransformer2DModel", + "CogView4Transformer2DModel", + "ConsisIDTransformer3DModel", + "ConsistencyDecoderVAE", + "ControlNetModel", + "ControlNetUnionModel", + "ControlNetXSAdapter", + "CosmosTransformer3DModel", + "DiTTransformer2DModel", + "EasyAnimateTransformer3DModel", + "FluxControlNetModel", + "FluxMultiControlNetModel", + "FluxTransformer2DModel", + "HiDreamImageTransformer2DModel", + "HunyuanDiT2DControlNetModel", + "HunyuanDiT2DModel", + "HunyuanDiT2DMultiControlNetModel", + "HunyuanVideoFramepackTransformer3DModel", + "HunyuanVideoTransformer3DModel", + "I2VGenXLUNet", + "Kandinsky3UNet", + "LatteTransformer3DModel", + "LTXVideoTransformer3DModel", + "Lumina2Transformer2DModel", + "LuminaNextDiT2DModel", + "MochiTransformer3DModel", + "ModelMixin", + "MotionAdapter", + "MultiAdapter", + "MultiControlNetModel", + "OmniGenTransformer2DModel", + "PixArtTransformer2DModel", + "PriorTransformer", + "SanaControlNetModel", + "SanaTransformer2DModel", + "SD3ControlNetModel", + "SD3MultiControlNetModel", + "SD3Transformer2DModel", + "SparseControlNetModel", + "StableAudioDiTModel", + "StableCascadeUNet", + "T2IAdapter", + "T5FilmDecoder", + "Transformer2DModel", + "TransformerTemporalModel", + "UNet1DModel", + "UNet2DConditionModel", + "UNet2DModel", + "UNet3DConditionModel", + "UNetControlNetXSModel", + "UNetMotionModel", + "UNetSpatioTemporalConditionModel", + "UVit2DModel", + "VQModel", + "WanTransformer3DModel", + "WanVACETransformer3DModel", + ] + ) + _import_structure["optimization"] = [ + "get_constant_schedule", + "get_constant_schedule_with_warmup", + "get_cosine_schedule_with_warmup", + "get_cosine_with_hard_restarts_schedule_with_warmup", + "get_linear_schedule_with_warmup", + "get_polynomial_decay_schedule_with_warmup", + "get_scheduler", + ] + _import_structure["pipelines"].extend( + [ + "AudioPipelineOutput", + "AutoPipelineForImage2Image", + "AutoPipelineForInpainting", + "AutoPipelineForText2Image", + "ConsistencyModelPipeline", + "DanceDiffusionPipeline", + "DDIMPipeline", + "DDPMPipeline", + "DiffusionPipeline", + "DiTPipeline", + "ImagePipelineOutput", + "KarrasVePipeline", + "LDMPipeline", + "LDMSuperResolutionPipeline", + "PNDMPipeline", + "RePaintPipeline", + "ScoreSdeVePipeline", + "StableDiffusionMixin", + ] + ) + _import_structure["quantizers"] = ["DiffusersQuantizer"] + _import_structure["schedulers"].extend( + [ + "AmusedScheduler", + "CMStochasticIterativeScheduler", + "CogVideoXDDIMScheduler", + "CogVideoXDPMScheduler", + "DDIMInverseScheduler", + "DDIMParallelScheduler", + "DDIMScheduler", + "DDPMParallelScheduler", + "DDPMScheduler", + "DDPMWuerstchenScheduler", + "DEISMultistepScheduler", + "DPMSolverMultistepInverseScheduler", + "DPMSolverMultistepScheduler", + "DPMSolverSinglestepScheduler", + "EDMDPMSolverMultistepScheduler", + "EDMEulerScheduler", + "EulerAncestralDiscreteScheduler", + "EulerDiscreteScheduler", + "FlowMatchEulerDiscreteScheduler", + "FlowMatchHeunDiscreteScheduler", + "FlowMatchLCMScheduler", + "HeunDiscreteScheduler", + "IPNDMScheduler", + "KarrasVeScheduler", + "KDPM2AncestralDiscreteScheduler", + "KDPM2DiscreteScheduler", + "LCMScheduler", + "PNDMScheduler", + "RePaintScheduler", + "SASolverScheduler", + "SchedulerMixin", + "SCMScheduler", + "ScoreSdeVeScheduler", + "TCDScheduler", + "UnCLIPScheduler", + "UniPCMultistepScheduler", + "VQDiffusionScheduler", + ] + ) + _import_structure["training_utils"] = ["EMAModel"] + +try: + if not (is_torch_available() and is_scipy_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_torch_and_scipy_objects # noqa F403 + + _import_structure["utils.dummy_torch_and_scipy_objects"] = [ + name for name in dir(dummy_torch_and_scipy_objects) if not name.startswith("_") + ] + +else: + _import_structure["schedulers"].extend(["LMSDiscreteScheduler"]) + +try: + if not (is_torch_available() and is_torchsde_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_torch_and_torchsde_objects # noqa F403 + + _import_structure["utils.dummy_torch_and_torchsde_objects"] = [ + name for name in dir(dummy_torch_and_torchsde_objects) if not name.startswith("_") + ] + +else: + _import_structure["schedulers"].extend(["CosineDPMSolverMultistepScheduler", "DPMSolverSDEScheduler"]) + +try: + if not (is_torch_available() and is_transformers_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_torch_and_transformers_objects # noqa F403 + + _import_structure["utils.dummy_torch_and_transformers_objects"] = [ + name for name in dir(dummy_torch_and_transformers_objects) if not name.startswith("_") + ] + +else: + _import_structure["pipelines"].extend( + [ + "AllegroPipeline", + "AltDiffusionImg2ImgPipeline", + "AltDiffusionPipeline", + "AmusedImg2ImgPipeline", + "AmusedInpaintPipeline", + "AmusedPipeline", + "AnimateDiffControlNetPipeline", + "AnimateDiffPAGPipeline", + "AnimateDiffPipeline", + "AnimateDiffSDXLPipeline", + "AnimateDiffSparseControlNetPipeline", + "AnimateDiffVideoToVideoControlNetPipeline", + "AnimateDiffVideoToVideoPipeline", + "AudioLDM2Pipeline", + "AudioLDM2ProjectionModel", + "AudioLDM2UNet2DConditionModel", + "AudioLDMPipeline", + "AuraFlowPipeline", + "BlipDiffusionControlNetPipeline", + "BlipDiffusionPipeline", + "ChromaImg2ImgPipeline", + "ChromaPipeline", + "CLIPImageProjection", + "CogVideoXFunControlPipeline", + "CogVideoXImageToVideoPipeline", + "CogVideoXPipeline", + "CogVideoXVideoToVideoPipeline", + "CogView3PlusPipeline", + "CogView4ControlPipeline", + "CogView4Pipeline", + "ConsisIDPipeline", + "Cosmos2TextToImagePipeline", + "Cosmos2VideoToWorldPipeline", + "CosmosTextToWorldPipeline", + "CosmosVideoToWorldPipeline", + "CycleDiffusionPipeline", + "EasyAnimateControlPipeline", + "EasyAnimateInpaintPipeline", + "EasyAnimatePipeline", + "FluxControlImg2ImgPipeline", + "FluxControlInpaintPipeline", + "FluxControlNetImg2ImgPipeline", + "FluxControlNetInpaintPipeline", + "FluxControlNetPipeline", + "FluxControlPipeline", + "FluxFillPipeline", + "FluxImg2ImgPipeline", + "FluxInpaintPipeline", + "FluxPipeline", + "FluxPriorReduxPipeline", + "HiDreamImagePipeline", + "HunyuanDiTControlNetPipeline", + "HunyuanDiTPAGPipeline", + "HunyuanDiTPipeline", + "HunyuanSkyreelsImageToVideoPipeline", + "HunyuanVideoFramepackPipeline", + "HunyuanVideoImageToVideoPipeline", + "HunyuanVideoPipeline", + "I2VGenXLPipeline", + "IFImg2ImgPipeline", + "IFImg2ImgSuperResolutionPipeline", + "IFInpaintingPipeline", + "IFInpaintingSuperResolutionPipeline", + "IFPipeline", + "IFSuperResolutionPipeline", + "ImageTextPipelineOutput", + "Kandinsky3Img2ImgPipeline", + "Kandinsky3Pipeline", + "KandinskyCombinedPipeline", + "KandinskyImg2ImgCombinedPipeline", + "KandinskyImg2ImgPipeline", + "KandinskyInpaintCombinedPipeline", + "KandinskyInpaintPipeline", + "KandinskyPipeline", + "KandinskyPriorPipeline", + "KandinskyV22CombinedPipeline", + "KandinskyV22ControlnetImg2ImgPipeline", + "KandinskyV22ControlnetPipeline", + "KandinskyV22Img2ImgCombinedPipeline", + "KandinskyV22Img2ImgPipeline", + "KandinskyV22InpaintCombinedPipeline", + "KandinskyV22InpaintPipeline", + "KandinskyV22Pipeline", + "KandinskyV22PriorEmb2EmbPipeline", + "KandinskyV22PriorPipeline", + "LatentConsistencyModelImg2ImgPipeline", + "LatentConsistencyModelPipeline", + "LattePipeline", + "LDMTextToImagePipeline", + "LEditsPPPipelineStableDiffusion", + "LEditsPPPipelineStableDiffusionXL", + "LTXConditionPipeline", + "LTXImageToVideoPipeline", + "LTXLatentUpsamplePipeline", + "LTXPipeline", + "Lumina2Pipeline", + "Lumina2Text2ImgPipeline", + "LuminaPipeline", + "LuminaText2ImgPipeline", + "MarigoldDepthPipeline", + "MarigoldIntrinsicsPipeline", + "MarigoldNormalsPipeline", + "MochiPipeline", + "MusicLDMPipeline", + "OmniGenPipeline", + "PaintByExamplePipeline", + "PIAPipeline", + "PixArtAlphaPipeline", + "PixArtSigmaPAGPipeline", + "PixArtSigmaPipeline", + "ReduxImageEncoder", + "SanaControlNetPipeline", + "SanaPAGPipeline", + "SanaPipeline", + "SanaSprintImg2ImgPipeline", + "SanaSprintPipeline", + "SemanticStableDiffusionPipeline", + "ShapEImg2ImgPipeline", + "ShapEPipeline", + "StableAudioPipeline", + "StableAudioProjectionModel", + "StableCascadeCombinedPipeline", + "StableCascadeDecoderPipeline", + "StableCascadePriorPipeline", + "StableDiffusion3ControlNetInpaintingPipeline", + "StableDiffusion3ControlNetPipeline", + "StableDiffusion3Img2ImgPipeline", + "StableDiffusion3InpaintPipeline", + "StableDiffusion3PAGImg2ImgPipeline", + "StableDiffusion3PAGImg2ImgPipeline", + "StableDiffusion3PAGPipeline", + "StableDiffusion3Pipeline", + "StableDiffusionAdapterPipeline", + "StableDiffusionAttendAndExcitePipeline", + "StableDiffusionControlNetImg2ImgPipeline", + "StableDiffusionControlNetInpaintPipeline", + "StableDiffusionControlNetPAGInpaintPipeline", + "StableDiffusionControlNetPAGPipeline", + "StableDiffusionControlNetPipeline", + "StableDiffusionControlNetXSPipeline", + "StableDiffusionDepth2ImgPipeline", + "StableDiffusionDiffEditPipeline", + "StableDiffusionGLIGENPipeline", + "StableDiffusionGLIGENTextImagePipeline", + "StableDiffusionImageVariationPipeline", + "StableDiffusionImg2ImgPipeline", + "StableDiffusionInpaintPipeline", + "StableDiffusionInpaintPipelineLegacy", + "StableDiffusionInstructPix2PixPipeline", + "StableDiffusionLatentUpscalePipeline", + "StableDiffusionLDM3DPipeline", + "StableDiffusionModelEditingPipeline", + "StableDiffusionPAGImg2ImgPipeline", + "StableDiffusionPAGInpaintPipeline", + "StableDiffusionPAGPipeline", + "StableDiffusionPanoramaPipeline", + "StableDiffusionParadigmsPipeline", + "StableDiffusionPipeline", + "StableDiffusionPipelineSafe", + "StableDiffusionPix2PixZeroPipeline", + "StableDiffusionSAGPipeline", + "StableDiffusionUpscalePipeline", + "StableDiffusionXLAdapterPipeline", + "StableDiffusionXLControlNetImg2ImgPipeline", + "StableDiffusionXLControlNetInpaintPipeline", + "StableDiffusionXLControlNetPAGImg2ImgPipeline", + "StableDiffusionXLControlNetPAGPipeline", + "StableDiffusionXLControlNetPipeline", + "StableDiffusionXLControlNetUnionImg2ImgPipeline", + "StableDiffusionXLControlNetUnionInpaintPipeline", + "StableDiffusionXLControlNetUnionPipeline", + "StableDiffusionXLControlNetXSPipeline", + "StableDiffusionXLImg2ImgPipeline", + "StableDiffusionXLInpaintPipeline", + "StableDiffusionXLInstructPix2PixPipeline", + "StableDiffusionXLPAGImg2ImgPipeline", + "StableDiffusionXLPAGInpaintPipeline", + "StableDiffusionXLPAGPipeline", + "StableDiffusionXLPipeline", + "StableUnCLIPImg2ImgPipeline", + "StableUnCLIPPipeline", + "StableVideoDiffusionPipeline", + "TextToVideoSDPipeline", + "TextToVideoZeroPipeline", + "TextToVideoZeroSDXLPipeline", + "UnCLIPImageVariationPipeline", + "UnCLIPPipeline", + "UniDiffuserModel", + "UniDiffuserPipeline", + "UniDiffuserTextDecoder", + "VersatileDiffusionDualGuidedPipeline", + "VersatileDiffusionImageVariationPipeline", + "VersatileDiffusionPipeline", + "VersatileDiffusionTextToImagePipeline", + "VideoToVideoSDPipeline", + "VisualClozeGenerationPipeline", + "VisualClozePipeline", + "VQDiffusionPipeline", + "WanImageToVideoPipeline", + "WanPipeline", + "WanVACEPipeline", + "WanVideoToVideoPipeline", + "WuerstchenCombinedPipeline", + "WuerstchenDecoderPipeline", + "WuerstchenPriorPipeline", + ] + ) + +try: + if not (is_torch_available() and is_transformers_available() and is_opencv_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_torch_and_transformers_and_opencv_objects # noqa F403 + + _import_structure["utils.dummy_torch_and_transformers_and_opencv_objects"] = [ + name for name in dir(dummy_torch_and_transformers_and_opencv_objects) if not name.startswith("_") + ] + +else: + _import_structure["pipelines"].extend(["ConsisIDPipeline"]) + +try: + if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_torch_and_transformers_and_k_diffusion_objects # noqa F403 + + _import_structure["utils.dummy_torch_and_transformers_and_k_diffusion_objects"] = [ + name for name in dir(dummy_torch_and_transformers_and_k_diffusion_objects) if not name.startswith("_") + ] + +else: + _import_structure["pipelines"].extend(["StableDiffusionKDiffusionPipeline", "StableDiffusionXLKDiffusionPipeline"]) + +try: + if not (is_torch_available() and is_transformers_available() and is_sentencepiece_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_torch_and_transformers_and_sentencepiece_objects # noqa F403 + + _import_structure["utils.dummy_torch_and_transformers_and_sentencepiece_objects"] = [ + name for name in dir(dummy_torch_and_transformers_and_sentencepiece_objects) if not name.startswith("_") + ] + +else: + _import_structure["pipelines"].extend(["KolorsImg2ImgPipeline", "KolorsPAGPipeline", "KolorsPipeline"]) + +try: + if not (is_torch_available() and is_transformers_available() and is_onnx_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_torch_and_transformers_and_onnx_objects # noqa F403 + + _import_structure["utils.dummy_torch_and_transformers_and_onnx_objects"] = [ + name for name in dir(dummy_torch_and_transformers_and_onnx_objects) if not name.startswith("_") + ] + +else: + _import_structure["pipelines"].extend( + [ + "OnnxStableDiffusionImg2ImgPipeline", + "OnnxStableDiffusionInpaintPipeline", + "OnnxStableDiffusionInpaintPipelineLegacy", + "OnnxStableDiffusionPipeline", + "OnnxStableDiffusionUpscalePipeline", + "StableDiffusionOnnxPipeline", + ] + ) + +try: + if not (is_torch_available() and is_librosa_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_torch_and_librosa_objects # noqa F403 + + _import_structure["utils.dummy_torch_and_librosa_objects"] = [ + name for name in dir(dummy_torch_and_librosa_objects) if not name.startswith("_") + ] + +else: + _import_structure["pipelines"].extend(["AudioDiffusionPipeline", "Mel"]) + +try: + if not (is_transformers_available() and is_torch_available() and is_note_seq_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_transformers_and_torch_and_note_seq_objects # noqa F403 + + _import_structure["utils.dummy_transformers_and_torch_and_note_seq_objects"] = [ + name for name in dir(dummy_transformers_and_torch_and_note_seq_objects) if not name.startswith("_") + ] + + +else: + _import_structure["pipelines"].extend(["SpectrogramDiffusionPipeline"]) + +try: + if not is_flax_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_flax_objects # noqa F403 + + _import_structure["utils.dummy_flax_objects"] = [ + name for name in dir(dummy_flax_objects) if not name.startswith("_") + ] + + +else: + _import_structure["models.controlnets.controlnet_flax"] = ["FlaxControlNetModel"] + _import_structure["models.modeling_flax_utils"] = ["FlaxModelMixin"] + _import_structure["models.unets.unet_2d_condition_flax"] = ["FlaxUNet2DConditionModel"] + _import_structure["models.vae_flax"] = ["FlaxAutoencoderKL"] + _import_structure["pipelines"].extend(["FlaxDiffusionPipeline"]) + _import_structure["schedulers"].extend( + [ + "FlaxDDIMScheduler", + "FlaxDDPMScheduler", + "FlaxDPMSolverMultistepScheduler", + "FlaxEulerDiscreteScheduler", + "FlaxKarrasVeScheduler", + "FlaxLMSDiscreteScheduler", + "FlaxPNDMScheduler", + "FlaxSchedulerMixin", + "FlaxScoreSdeVeScheduler", + ] + ) + + +try: + if not (is_flax_available() and is_transformers_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_flax_and_transformers_objects # noqa F403 + + _import_structure["utils.dummy_flax_and_transformers_objects"] = [ + name for name in dir(dummy_flax_and_transformers_objects) if not name.startswith("_") + ] + + +else: + _import_structure["pipelines"].extend( + [ + "FlaxStableDiffusionControlNetPipeline", + "FlaxStableDiffusionImg2ImgPipeline", + "FlaxStableDiffusionInpaintPipeline", + "FlaxStableDiffusionPipeline", + "FlaxStableDiffusionXLPipeline", + ] + ) + +try: + if not (is_note_seq_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from .utils import dummy_note_seq_objects # noqa F403 + + _import_structure["utils.dummy_note_seq_objects"] = [ + name for name in dir(dummy_note_seq_objects) if not name.startswith("_") + ] + + +else: + _import_structure["pipelines"].extend(["MidiProcessor"]) + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + from .configuration_utils import ConfigMixin + from .quantizers import PipelineQuantizationConfig + + try: + if not is_bitsandbytes_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_bitsandbytes_objects import * + else: + from .quantizers.quantization_config import BitsAndBytesConfig + + try: + if not is_gguf_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_gguf_objects import * + else: + from .quantizers.quantization_config import GGUFQuantizationConfig + + try: + if not is_torchao_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_torchao_objects import * + else: + from .quantizers.quantization_config import TorchAoConfig + + try: + if not is_optimum_quanto_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_optimum_quanto_objects import * + else: + from .quantizers.quantization_config import QuantoConfig + + try: + if not is_onnx_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_onnx_objects import * # noqa F403 + else: + from .pipelines import OnnxRuntimeModel + + try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_pt_objects import * # noqa F403 + else: + from .hooks import ( + FasterCacheConfig, + HookRegistry, + PyramidAttentionBroadcastConfig, + apply_faster_cache, + apply_pyramid_attention_broadcast, + ) + from .models import ( + AllegroTransformer3DModel, + AsymmetricAutoencoderKL, + AuraFlowTransformer2DModel, + AutoencoderDC, + AutoencoderKL, + AutoencoderKLAllegro, + AutoencoderKLCogVideoX, + AutoencoderKLCosmos, + AutoencoderKLHunyuanVideo, + AutoencoderKLLTXVideo, + AutoencoderKLMagvit, + AutoencoderKLMochi, + AutoencoderKLTemporalDecoder, + AutoencoderKLWan, + AutoencoderOobleck, + AutoencoderTiny, + AutoModel, + CacheMixin, + ChromaTransformer2DModel, + CogVideoXTransformer3DModel, + CogView3PlusTransformer2DModel, + CogView4Transformer2DModel, + ConsisIDTransformer3DModel, + ConsistencyDecoderVAE, + ControlNetModel, + ControlNetUnionModel, + ControlNetXSAdapter, + CosmosTransformer3DModel, + DiTTransformer2DModel, + EasyAnimateTransformer3DModel, + FluxControlNetModel, + FluxMultiControlNetModel, + FluxTransformer2DModel, + HiDreamImageTransformer2DModel, + HunyuanDiT2DControlNetModel, + HunyuanDiT2DModel, + HunyuanDiT2DMultiControlNetModel, + HunyuanVideoFramepackTransformer3DModel, + HunyuanVideoTransformer3DModel, + I2VGenXLUNet, + Kandinsky3UNet, + LatteTransformer3DModel, + LTXVideoTransformer3DModel, + Lumina2Transformer2DModel, + LuminaNextDiT2DModel, + MochiTransformer3DModel, + ModelMixin, + MotionAdapter, + MultiAdapter, + MultiControlNetModel, + OmniGenTransformer2DModel, + PixArtTransformer2DModel, + PriorTransformer, + SanaControlNetModel, + SanaTransformer2DModel, + SD3ControlNetModel, + SD3MultiControlNetModel, + SD3Transformer2DModel, + SparseControlNetModel, + StableAudioDiTModel, + T2IAdapter, + T5FilmDecoder, + Transformer2DModel, + TransformerTemporalModel, + UNet1DModel, + UNet2DConditionModel, + UNet2DModel, + UNet3DConditionModel, + UNetControlNetXSModel, + UNetMotionModel, + UNetSpatioTemporalConditionModel, + UVit2DModel, + VQModel, + WanTransformer3DModel, + WanVACETransformer3DModel, + ) + from .optimization import ( + get_constant_schedule, + get_constant_schedule_with_warmup, + get_cosine_schedule_with_warmup, + get_cosine_with_hard_restarts_schedule_with_warmup, + get_linear_schedule_with_warmup, + get_polynomial_decay_schedule_with_warmup, + get_scheduler, + ) + from .pipelines import ( + AudioPipelineOutput, + AutoPipelineForImage2Image, + AutoPipelineForInpainting, + AutoPipelineForText2Image, + BlipDiffusionControlNetPipeline, + BlipDiffusionPipeline, + CLIPImageProjection, + ConsistencyModelPipeline, + DanceDiffusionPipeline, + DDIMPipeline, + DDPMPipeline, + DiffusionPipeline, + DiTPipeline, + ImagePipelineOutput, + KarrasVePipeline, + LDMPipeline, + LDMSuperResolutionPipeline, + PNDMPipeline, + RePaintPipeline, + ScoreSdeVePipeline, + StableDiffusionMixin, + ) + from .quantizers import DiffusersQuantizer + from .schedulers import ( + AmusedScheduler, + CMStochasticIterativeScheduler, + CogVideoXDDIMScheduler, + CogVideoXDPMScheduler, + DDIMInverseScheduler, + DDIMParallelScheduler, + DDIMScheduler, + DDPMParallelScheduler, + DDPMScheduler, + DDPMWuerstchenScheduler, + DEISMultistepScheduler, + DPMSolverMultistepInverseScheduler, + DPMSolverMultistepScheduler, + DPMSolverSinglestepScheduler, + EDMDPMSolverMultistepScheduler, + EDMEulerScheduler, + EulerAncestralDiscreteScheduler, + EulerDiscreteScheduler, + FlowMatchEulerDiscreteScheduler, + FlowMatchHeunDiscreteScheduler, + FlowMatchLCMScheduler, + HeunDiscreteScheduler, + IPNDMScheduler, + KarrasVeScheduler, + KDPM2AncestralDiscreteScheduler, + KDPM2DiscreteScheduler, + LCMScheduler, + PNDMScheduler, + RePaintScheduler, + SASolverScheduler, + SchedulerMixin, + SCMScheduler, + ScoreSdeVeScheduler, + TCDScheduler, + UnCLIPScheduler, + UniPCMultistepScheduler, + VQDiffusionScheduler, + ) + from .training_utils import EMAModel + + try: + if not (is_torch_available() and is_scipy_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_torch_and_scipy_objects import * # noqa F403 + else: + from .schedulers import LMSDiscreteScheduler + + try: + if not (is_torch_available() and is_torchsde_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_torch_and_torchsde_objects import * # noqa F403 + else: + from .schedulers import CosineDPMSolverMultistepScheduler, DPMSolverSDEScheduler + + try: + if not (is_torch_available() and is_transformers_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipelines import ( + AllegroPipeline, + AltDiffusionImg2ImgPipeline, + AltDiffusionPipeline, + AmusedImg2ImgPipeline, + AmusedInpaintPipeline, + AmusedPipeline, + AnimateDiffControlNetPipeline, + AnimateDiffPAGPipeline, + AnimateDiffPipeline, + AnimateDiffSDXLPipeline, + AnimateDiffSparseControlNetPipeline, + AnimateDiffVideoToVideoControlNetPipeline, + AnimateDiffVideoToVideoPipeline, + AudioLDM2Pipeline, + AudioLDM2ProjectionModel, + AudioLDM2UNet2DConditionModel, + AudioLDMPipeline, + AuraFlowPipeline, + ChromaImg2ImgPipeline, + ChromaPipeline, + CLIPImageProjection, + CogVideoXFunControlPipeline, + CogVideoXImageToVideoPipeline, + CogVideoXPipeline, + CogVideoXVideoToVideoPipeline, + CogView3PlusPipeline, + CogView4ControlPipeline, + CogView4Pipeline, + ConsisIDPipeline, + Cosmos2TextToImagePipeline, + Cosmos2VideoToWorldPipeline, + CosmosTextToWorldPipeline, + CosmosVideoToWorldPipeline, + CycleDiffusionPipeline, + EasyAnimateControlPipeline, + EasyAnimateInpaintPipeline, + EasyAnimatePipeline, + FluxControlImg2ImgPipeline, + FluxControlInpaintPipeline, + FluxControlNetImg2ImgPipeline, + FluxControlNetInpaintPipeline, + FluxControlNetPipeline, + FluxControlPipeline, + FluxFillPipeline, + FluxImg2ImgPipeline, + FluxInpaintPipeline, + FluxPipeline, + FluxPriorReduxPipeline, + HiDreamImagePipeline, + HunyuanDiTControlNetPipeline, + HunyuanDiTPAGPipeline, + HunyuanDiTPipeline, + HunyuanSkyreelsImageToVideoPipeline, + HunyuanVideoFramepackPipeline, + HunyuanVideoImageToVideoPipeline, + HunyuanVideoPipeline, + I2VGenXLPipeline, + IFImg2ImgPipeline, + IFImg2ImgSuperResolutionPipeline, + IFInpaintingPipeline, + IFInpaintingSuperResolutionPipeline, + IFPipeline, + IFSuperResolutionPipeline, + ImageTextPipelineOutput, + Kandinsky3Img2ImgPipeline, + Kandinsky3Pipeline, + KandinskyCombinedPipeline, + KandinskyImg2ImgCombinedPipeline, + KandinskyImg2ImgPipeline, + KandinskyInpaintCombinedPipeline, + KandinskyInpaintPipeline, + KandinskyPipeline, + KandinskyPriorPipeline, + KandinskyV22CombinedPipeline, + KandinskyV22ControlnetImg2ImgPipeline, + KandinskyV22ControlnetPipeline, + KandinskyV22Img2ImgCombinedPipeline, + KandinskyV22Img2ImgPipeline, + KandinskyV22InpaintCombinedPipeline, + KandinskyV22InpaintPipeline, + KandinskyV22Pipeline, + KandinskyV22PriorEmb2EmbPipeline, + KandinskyV22PriorPipeline, + LatentConsistencyModelImg2ImgPipeline, + LatentConsistencyModelPipeline, + LattePipeline, + LDMTextToImagePipeline, + LEditsPPPipelineStableDiffusion, + LEditsPPPipelineStableDiffusionXL, + LTXConditionPipeline, + LTXImageToVideoPipeline, + LTXLatentUpsamplePipeline, + LTXPipeline, + Lumina2Pipeline, + Lumina2Text2ImgPipeline, + LuminaPipeline, + LuminaText2ImgPipeline, + MarigoldDepthPipeline, + MarigoldIntrinsicsPipeline, + MarigoldNormalsPipeline, + MochiPipeline, + MusicLDMPipeline, + OmniGenPipeline, + PaintByExamplePipeline, + PIAPipeline, + PixArtAlphaPipeline, + PixArtSigmaPAGPipeline, + PixArtSigmaPipeline, + ReduxImageEncoder, + SanaControlNetPipeline, + SanaPAGPipeline, + SanaPipeline, + SanaSprintImg2ImgPipeline, + SanaSprintPipeline, + SemanticStableDiffusionPipeline, + ShapEImg2ImgPipeline, + ShapEPipeline, + StableAudioPipeline, + StableAudioProjectionModel, + StableCascadeCombinedPipeline, + StableCascadeDecoderPipeline, + StableCascadePriorPipeline, + StableDiffusion3ControlNetInpaintingPipeline, + StableDiffusion3ControlNetPipeline, + StableDiffusion3Img2ImgPipeline, + StableDiffusion3InpaintPipeline, + StableDiffusion3PAGImg2ImgPipeline, + StableDiffusion3PAGPipeline, + StableDiffusion3Pipeline, + StableDiffusionAdapterPipeline, + StableDiffusionAttendAndExcitePipeline, + StableDiffusionControlNetImg2ImgPipeline, + StableDiffusionControlNetInpaintPipeline, + StableDiffusionControlNetPAGInpaintPipeline, + StableDiffusionControlNetPAGPipeline, + StableDiffusionControlNetPipeline, + StableDiffusionControlNetXSPipeline, + StableDiffusionDepth2ImgPipeline, + StableDiffusionDiffEditPipeline, + StableDiffusionGLIGENPipeline, + StableDiffusionGLIGENTextImagePipeline, + StableDiffusionImageVariationPipeline, + StableDiffusionImg2ImgPipeline, + StableDiffusionInpaintPipeline, + StableDiffusionInpaintPipelineLegacy, + StableDiffusionInstructPix2PixPipeline, + StableDiffusionLatentUpscalePipeline, + StableDiffusionLDM3DPipeline, + StableDiffusionModelEditingPipeline, + StableDiffusionPAGImg2ImgPipeline, + StableDiffusionPAGInpaintPipeline, + StableDiffusionPAGPipeline, + StableDiffusionPanoramaPipeline, + StableDiffusionParadigmsPipeline, + StableDiffusionPipeline, + StableDiffusionPipelineSafe, + StableDiffusionPix2PixZeroPipeline, + StableDiffusionSAGPipeline, + StableDiffusionUpscalePipeline, + StableDiffusionXLAdapterPipeline, + StableDiffusionXLControlNetImg2ImgPipeline, + StableDiffusionXLControlNetInpaintPipeline, + StableDiffusionXLControlNetPAGImg2ImgPipeline, + StableDiffusionXLControlNetPAGPipeline, + StableDiffusionXLControlNetPipeline, + StableDiffusionXLControlNetUnionImg2ImgPipeline, + StableDiffusionXLControlNetUnionInpaintPipeline, + StableDiffusionXLControlNetUnionPipeline, + StableDiffusionXLControlNetXSPipeline, + StableDiffusionXLImg2ImgPipeline, + StableDiffusionXLInpaintPipeline, + StableDiffusionXLInstructPix2PixPipeline, + StableDiffusionXLPAGImg2ImgPipeline, + StableDiffusionXLPAGInpaintPipeline, + StableDiffusionXLPAGPipeline, + StableDiffusionXLPipeline, + StableUnCLIPImg2ImgPipeline, + StableUnCLIPPipeline, + StableVideoDiffusionPipeline, + TextToVideoSDPipeline, + TextToVideoZeroPipeline, + TextToVideoZeroSDXLPipeline, + UnCLIPImageVariationPipeline, + UnCLIPPipeline, + UniDiffuserModel, + UniDiffuserPipeline, + UniDiffuserTextDecoder, + VersatileDiffusionDualGuidedPipeline, + VersatileDiffusionImageVariationPipeline, + VersatileDiffusionPipeline, + VersatileDiffusionTextToImagePipeline, + VideoToVideoSDPipeline, + VisualClozeGenerationPipeline, + VisualClozePipeline, + VQDiffusionPipeline, + WanImageToVideoPipeline, + WanPipeline, + WanVACEPipeline, + WanVideoToVideoPipeline, + WuerstchenCombinedPipeline, + WuerstchenDecoderPipeline, + WuerstchenPriorPipeline, + ) + + try: + if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_torch_and_transformers_and_k_diffusion_objects import * # noqa F403 + else: + from .pipelines import StableDiffusionKDiffusionPipeline, StableDiffusionXLKDiffusionPipeline + + try: + if not (is_torch_available() and is_transformers_available() and is_sentencepiece_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_torch_and_transformers_and_sentencepiece_objects import * # noqa F403 + else: + from .pipelines import KolorsImg2ImgPipeline, KolorsPAGPipeline, KolorsPipeline + + try: + if not (is_torch_available() and is_transformers_available() and is_opencv_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_torch_and_transformers_and_opencv_objects import * # noqa F403 + else: + from .pipelines import ConsisIDPipeline + + try: + if not (is_torch_available() and is_transformers_available() and is_onnx_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_torch_and_transformers_and_onnx_objects import * # noqa F403 + else: + from .pipelines import ( + OnnxStableDiffusionImg2ImgPipeline, + OnnxStableDiffusionInpaintPipeline, + OnnxStableDiffusionInpaintPipelineLegacy, + OnnxStableDiffusionPipeline, + OnnxStableDiffusionUpscalePipeline, + StableDiffusionOnnxPipeline, + ) + + try: + if not (is_torch_available() and is_librosa_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_torch_and_librosa_objects import * # noqa F403 + else: + from .pipelines import AudioDiffusionPipeline, Mel + + try: + if not (is_transformers_available() and is_torch_available() and is_note_seq_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_transformers_and_torch_and_note_seq_objects import * # noqa F403 + else: + from .pipelines import SpectrogramDiffusionPipeline + + try: + if not is_flax_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_flax_objects import * # noqa F403 + else: + from .models.controlnets.controlnet_flax import FlaxControlNetModel + from .models.modeling_flax_utils import FlaxModelMixin + from .models.unets.unet_2d_condition_flax import FlaxUNet2DConditionModel + from .models.vae_flax import FlaxAutoencoderKL + from .pipelines import FlaxDiffusionPipeline + from .schedulers import ( + FlaxDDIMScheduler, + FlaxDDPMScheduler, + FlaxDPMSolverMultistepScheduler, + FlaxEulerDiscreteScheduler, + FlaxKarrasVeScheduler, + FlaxLMSDiscreteScheduler, + FlaxPNDMScheduler, + FlaxSchedulerMixin, + FlaxScoreSdeVeScheduler, + ) + + try: + if not (is_flax_available() and is_transformers_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_flax_and_transformers_objects import * # noqa F403 + else: + from .pipelines import ( + FlaxStableDiffusionControlNetPipeline, + FlaxStableDiffusionImg2ImgPipeline, + FlaxStableDiffusionInpaintPipeline, + FlaxStableDiffusionPipeline, + FlaxStableDiffusionXLPipeline, + ) + + try: + if not (is_note_seq_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from .utils.dummy_note_seq_objects import * # noqa F403 + else: + from .pipelines import MidiProcessor + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + extra_objects={"__version__": __version__}, + ) diff --git a/venv/lib/python3.10/site-packages/diffusers/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ea346bbfa4a58f8d3fd9d1a071b60b740540beaf Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/__pycache__/callbacks.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/__pycache__/callbacks.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0aa3062a5b223d249fbb02bb5440a326f8164a72 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/__pycache__/callbacks.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/__pycache__/configuration_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/__pycache__/configuration_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..10e3f1153ca530af55cdd3da0ccdacbe31279190 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/__pycache__/configuration_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/__pycache__/dependency_versions_check.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/__pycache__/dependency_versions_check.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..191540351e9e59cae17903ef0964cde6a1b35547 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/__pycache__/dependency_versions_check.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/__pycache__/dependency_versions_table.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/__pycache__/dependency_versions_table.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b566cddf728f542bae1d5374869958c8d25e3279 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/__pycache__/dependency_versions_table.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/__pycache__/image_processor.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/__pycache__/image_processor.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..92c1a04094591c86ede889256a68f40e6e712b64 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/__pycache__/image_processor.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/__pycache__/optimization.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/__pycache__/optimization.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8e53f972c13d66e1b987ee996101ef95ef209664 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/__pycache__/optimization.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/__pycache__/training_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/__pycache__/training_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6a3d45a55142c6e3eadee9ce6da712ee28841ba8 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/__pycache__/training_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/__pycache__/video_processor.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/__pycache__/video_processor.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9c0801bc9c621ddc517c814920f042f6b151d765 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/__pycache__/video_processor.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/callbacks.py b/venv/lib/python3.10/site-packages/diffusers/callbacks.py new file mode 100644 index 0000000000000000000000000000000000000000..4b8b15368c4736ce54282e92231a752558663cb6 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/callbacks.py @@ -0,0 +1,209 @@ +from typing import Any, Dict, List + +from .configuration_utils import ConfigMixin, register_to_config +from .utils import CONFIG_NAME + + +class PipelineCallback(ConfigMixin): + """ + Base class for all the official callbacks used in a pipeline. This class provides a structure for implementing + custom callbacks and ensures that all callbacks have a consistent interface. + + Please implement the following: + `tensor_inputs`: This should return a list of tensor inputs specific to your callback. You will only be able to + include + variables listed in the `._callback_tensor_inputs` attribute of your pipeline class. + `callback_fn`: This method defines the core functionality of your callback. + """ + + config_name = CONFIG_NAME + + @register_to_config + def __init__(self, cutoff_step_ratio=1.0, cutoff_step_index=None): + super().__init__() + + if (cutoff_step_ratio is None and cutoff_step_index is None) or ( + cutoff_step_ratio is not None and cutoff_step_index is not None + ): + raise ValueError("Either cutoff_step_ratio or cutoff_step_index should be provided, not both or none.") + + if cutoff_step_ratio is not None and ( + not isinstance(cutoff_step_ratio, float) or not (0.0 <= cutoff_step_ratio <= 1.0) + ): + raise ValueError("cutoff_step_ratio must be a float between 0.0 and 1.0.") + + @property + def tensor_inputs(self) -> List[str]: + raise NotImplementedError(f"You need to set the attribute `tensor_inputs` for {self.__class__}") + + def callback_fn(self, pipeline, step_index, timesteps, callback_kwargs) -> Dict[str, Any]: + raise NotImplementedError(f"You need to implement the method `callback_fn` for {self.__class__}") + + def __call__(self, pipeline, step_index, timestep, callback_kwargs) -> Dict[str, Any]: + return self.callback_fn(pipeline, step_index, timestep, callback_kwargs) + + +class MultiPipelineCallbacks: + """ + This class is designed to handle multiple pipeline callbacks. It accepts a list of PipelineCallback objects and + provides a unified interface for calling all of them. + """ + + def __init__(self, callbacks: List[PipelineCallback]): + self.callbacks = callbacks + + @property + def tensor_inputs(self) -> List[str]: + return [input for callback in self.callbacks for input in callback.tensor_inputs] + + def __call__(self, pipeline, step_index, timestep, callback_kwargs) -> Dict[str, Any]: + """ + Calls all the callbacks in order with the given arguments and returns the final callback_kwargs. + """ + for callback in self.callbacks: + callback_kwargs = callback(pipeline, step_index, timestep, callback_kwargs) + + return callback_kwargs + + +class SDCFGCutoffCallback(PipelineCallback): + """ + Callback function for Stable Diffusion Pipelines. After certain number of steps (set by `cutoff_step_ratio` or + `cutoff_step_index`), this callback will disable the CFG. + + Note: This callback mutates the pipeline by changing the `_guidance_scale` attribute to 0.0 after the cutoff step. + """ + + tensor_inputs = ["prompt_embeds"] + + def callback_fn(self, pipeline, step_index, timestep, callback_kwargs) -> Dict[str, Any]: + cutoff_step_ratio = self.config.cutoff_step_ratio + cutoff_step_index = self.config.cutoff_step_index + + # Use cutoff_step_index if it's not None, otherwise use cutoff_step_ratio + cutoff_step = ( + cutoff_step_index if cutoff_step_index is not None else int(pipeline.num_timesteps * cutoff_step_ratio) + ) + + if step_index == cutoff_step: + prompt_embeds = callback_kwargs[self.tensor_inputs[0]] + prompt_embeds = prompt_embeds[-1:] # "-1" denotes the embeddings for conditional text tokens. + + pipeline._guidance_scale = 0.0 + + callback_kwargs[self.tensor_inputs[0]] = prompt_embeds + return callback_kwargs + + +class SDXLCFGCutoffCallback(PipelineCallback): + """ + Callback function for the base Stable Diffusion XL Pipelines. After certain number of steps (set by + `cutoff_step_ratio` or `cutoff_step_index`), this callback will disable the CFG. + + Note: This callback mutates the pipeline by changing the `_guidance_scale` attribute to 0.0 after the cutoff step. + """ + + tensor_inputs = [ + "prompt_embeds", + "add_text_embeds", + "add_time_ids", + ] + + def callback_fn(self, pipeline, step_index, timestep, callback_kwargs) -> Dict[str, Any]: + cutoff_step_ratio = self.config.cutoff_step_ratio + cutoff_step_index = self.config.cutoff_step_index + + # Use cutoff_step_index if it's not None, otherwise use cutoff_step_ratio + cutoff_step = ( + cutoff_step_index if cutoff_step_index is not None else int(pipeline.num_timesteps * cutoff_step_ratio) + ) + + if step_index == cutoff_step: + prompt_embeds = callback_kwargs[self.tensor_inputs[0]] + prompt_embeds = prompt_embeds[-1:] # "-1" denotes the embeddings for conditional text tokens. + + add_text_embeds = callback_kwargs[self.tensor_inputs[1]] + add_text_embeds = add_text_embeds[-1:] # "-1" denotes the embeddings for conditional pooled text tokens + + add_time_ids = callback_kwargs[self.tensor_inputs[2]] + add_time_ids = add_time_ids[-1:] # "-1" denotes the embeddings for conditional added time vector + + pipeline._guidance_scale = 0.0 + + callback_kwargs[self.tensor_inputs[0]] = prompt_embeds + callback_kwargs[self.tensor_inputs[1]] = add_text_embeds + callback_kwargs[self.tensor_inputs[2]] = add_time_ids + + return callback_kwargs + + +class SDXLControlnetCFGCutoffCallback(PipelineCallback): + """ + Callback function for the Controlnet Stable Diffusion XL Pipelines. After certain number of steps (set by + `cutoff_step_ratio` or `cutoff_step_index`), this callback will disable the CFG. + + Note: This callback mutates the pipeline by changing the `_guidance_scale` attribute to 0.0 after the cutoff step. + """ + + tensor_inputs = [ + "prompt_embeds", + "add_text_embeds", + "add_time_ids", + "image", + ] + + def callback_fn(self, pipeline, step_index, timestep, callback_kwargs) -> Dict[str, Any]: + cutoff_step_ratio = self.config.cutoff_step_ratio + cutoff_step_index = self.config.cutoff_step_index + + # Use cutoff_step_index if it's not None, otherwise use cutoff_step_ratio + cutoff_step = ( + cutoff_step_index if cutoff_step_index is not None else int(pipeline.num_timesteps * cutoff_step_ratio) + ) + + if step_index == cutoff_step: + prompt_embeds = callback_kwargs[self.tensor_inputs[0]] + prompt_embeds = prompt_embeds[-1:] # "-1" denotes the embeddings for conditional text tokens. + + add_text_embeds = callback_kwargs[self.tensor_inputs[1]] + add_text_embeds = add_text_embeds[-1:] # "-1" denotes the embeddings for conditional pooled text tokens + + add_time_ids = callback_kwargs[self.tensor_inputs[2]] + add_time_ids = add_time_ids[-1:] # "-1" denotes the embeddings for conditional added time vector + + # For Controlnet + image = callback_kwargs[self.tensor_inputs[3]] + image = image[-1:] + + pipeline._guidance_scale = 0.0 + + callback_kwargs[self.tensor_inputs[0]] = prompt_embeds + callback_kwargs[self.tensor_inputs[1]] = add_text_embeds + callback_kwargs[self.tensor_inputs[2]] = add_time_ids + callback_kwargs[self.tensor_inputs[3]] = image + + return callback_kwargs + + +class IPAdapterScaleCutoffCallback(PipelineCallback): + """ + Callback function for any pipeline that inherits `IPAdapterMixin`. After certain number of steps (set by + `cutoff_step_ratio` or `cutoff_step_index`), this callback will set the IP Adapter scale to `0.0`. + + Note: This callback mutates the IP Adapter attention processors by setting the scale to 0.0 after the cutoff step. + """ + + tensor_inputs = [] + + def callback_fn(self, pipeline, step_index, timestep, callback_kwargs) -> Dict[str, Any]: + cutoff_step_ratio = self.config.cutoff_step_ratio + cutoff_step_index = self.config.cutoff_step_index + + # Use cutoff_step_index if it's not None, otherwise use cutoff_step_ratio + cutoff_step = ( + cutoff_step_index if cutoff_step_index is not None else int(pipeline.num_timesteps * cutoff_step_ratio) + ) + + if step_index == cutoff_step: + pipeline.set_ip_adapter_scale(0.0) + return callback_kwargs diff --git a/venv/lib/python3.10/site-packages/diffusers/configuration_utils.py b/venv/lib/python3.10/site-packages/diffusers/configuration_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..f9b652bbc0216a698bb206f31c139a7e3829226a --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/configuration_utils.py @@ -0,0 +1,760 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# 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. +"""ConfigMixin base class and utilities.""" + +import dataclasses +import functools +import importlib +import inspect +import json +import os +import re +from collections import OrderedDict +from pathlib import Path +from typing import Any, Dict, Optional, Tuple, Union + +import numpy as np +from huggingface_hub import DDUFEntry, create_repo, hf_hub_download +from huggingface_hub.utils import ( + EntryNotFoundError, + RepositoryNotFoundError, + RevisionNotFoundError, + validate_hf_hub_args, +) +from requests import HTTPError +from typing_extensions import Self + +from . import __version__ +from .utils import ( + HUGGINGFACE_CO_RESOLVE_ENDPOINT, + DummyObject, + deprecate, + extract_commit_hash, + http_user_agent, + logging, +) + + +logger = logging.get_logger(__name__) + +_re_configuration_file = re.compile(r"config\.(.*)\.json") + + +class FrozenDict(OrderedDict): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + for key, value in self.items(): + setattr(self, key, value) + + self.__frozen = True + + def __delitem__(self, *args, **kwargs): + raise Exception(f"You cannot use ``__delitem__`` on a {self.__class__.__name__} instance.") + + def setdefault(self, *args, **kwargs): + raise Exception(f"You cannot use ``setdefault`` on a {self.__class__.__name__} instance.") + + def pop(self, *args, **kwargs): + raise Exception(f"You cannot use ``pop`` on a {self.__class__.__name__} instance.") + + def update(self, *args, **kwargs): + raise Exception(f"You cannot use ``update`` on a {self.__class__.__name__} instance.") + + def __setattr__(self, name, value): + if hasattr(self, "__frozen") and self.__frozen: + raise Exception(f"You cannot use ``__setattr__`` on a {self.__class__.__name__} instance.") + super().__setattr__(name, value) + + def __setitem__(self, name, value): + if hasattr(self, "__frozen") and self.__frozen: + raise Exception(f"You cannot use ``__setattr__`` on a {self.__class__.__name__} instance.") + super().__setitem__(name, value) + + +class ConfigMixin: + r""" + Base class for all configuration classes. All configuration parameters are stored under `self.config`. Also + provides the [`~ConfigMixin.from_config`] and [`~ConfigMixin.save_config`] methods for loading, downloading, and + saving classes that inherit from [`ConfigMixin`]. + + Class attributes: + - **config_name** (`str`) -- A filename under which the config should stored when calling + [`~ConfigMixin.save_config`] (should be overridden by parent class). + - **ignore_for_config** (`List[str]`) -- A list of attributes that should not be saved in the config (should be + overridden by subclass). + - **has_compatibles** (`bool`) -- Whether the class has compatible classes (should be overridden by subclass). + - **_deprecated_kwargs** (`List[str]`) -- Keyword arguments that are deprecated. Note that the `init` function + should only have a `kwargs` argument if at least one argument is deprecated (should be overridden by + subclass). + """ + + config_name = None + ignore_for_config = [] + has_compatibles = False + + _deprecated_kwargs = [] + + def register_to_config(self, **kwargs): + if self.config_name is None: + raise NotImplementedError(f"Make sure that {self.__class__} has defined a class name `config_name`") + # Special case for `kwargs` used in deprecation warning added to schedulers + # TODO: remove this when we remove the deprecation warning, and the `kwargs` argument, + # or solve in a more general way. + kwargs.pop("kwargs", None) + + if not hasattr(self, "_internal_dict"): + internal_dict = kwargs + else: + previous_dict = dict(self._internal_dict) + internal_dict = {**self._internal_dict, **kwargs} + logger.debug(f"Updating config from {previous_dict} to {internal_dict}") + + self._internal_dict = FrozenDict(internal_dict) + + def __getattr__(self, name: str) -> Any: + """The only reason we overwrite `getattr` here is to gracefully deprecate accessing + config attributes directly. See https://github.com/huggingface/diffusers/pull/3129 + + This function is mostly copied from PyTorch's __getattr__ overwrite: + https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module + """ + + is_in_config = "_internal_dict" in self.__dict__ and hasattr(self.__dict__["_internal_dict"], name) + is_attribute = name in self.__dict__ + + if is_in_config and not is_attribute: + deprecation_message = f"Accessing config attribute `{name}` directly via '{type(self).__name__}' object attribute is deprecated. Please access '{name}' over '{type(self).__name__}'s config object instead, e.g. 'scheduler.config.{name}'." + deprecate("direct config name access", "1.0.0", deprecation_message, standard_warn=False) + return self._internal_dict[name] + + raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'") + + def save_config(self, save_directory: Union[str, os.PathLike], push_to_hub: bool = False, **kwargs): + """ + Save a configuration object to the directory specified in `save_directory` so that it can be reloaded using the + [`~ConfigMixin.from_config`] class method. + + Args: + save_directory (`str` or `os.PathLike`): + Directory where the configuration JSON file is saved (will be created if it does not exist). + push_to_hub (`bool`, *optional*, defaults to `False`): + Whether or not to push your model to the Hugging Face Hub after saving it. You can specify the + repository you want to push to with `repo_id` (will default to the name of `save_directory` in your + namespace). + kwargs (`Dict[str, Any]`, *optional*): + Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method. + """ + if os.path.isfile(save_directory): + raise AssertionError(f"Provided path ({save_directory}) should be a directory, not a file") + + os.makedirs(save_directory, exist_ok=True) + + # If we save using the predefined names, we can load using `from_config` + output_config_file = os.path.join(save_directory, self.config_name) + + self.to_json_file(output_config_file) + logger.info(f"Configuration saved in {output_config_file}") + + if push_to_hub: + commit_message = kwargs.pop("commit_message", None) + private = kwargs.pop("private", None) + create_pr = kwargs.pop("create_pr", False) + token = kwargs.pop("token", None) + repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1]) + repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id + + self._upload_folder( + save_directory, + repo_id, + token=token, + commit_message=commit_message, + create_pr=create_pr, + ) + + @classmethod + def from_config( + cls, config: Union[FrozenDict, Dict[str, Any]] = None, return_unused_kwargs=False, **kwargs + ) -> Union[Self, Tuple[Self, Dict[str, Any]]]: + r""" + Instantiate a Python class from a config dictionary. + + Parameters: + config (`Dict[str, Any]`): + A config dictionary from which the Python class is instantiated. Make sure to only load configuration + files of compatible classes. + return_unused_kwargs (`bool`, *optional*, defaults to `False`): + Whether kwargs that are not consumed by the Python class should be returned or not. + kwargs (remaining dictionary of keyword arguments, *optional*): + Can be used to update the configuration object (after it is loaded) and initiate the Python class. + `**kwargs` are passed directly to the underlying scheduler/model's `__init__` method and eventually + overwrite the same named arguments in `config`. + + Returns: + [`ModelMixin`] or [`SchedulerMixin`]: + A model or scheduler object instantiated from a config dictionary. + + Examples: + + ```python + >>> from diffusers import DDPMScheduler, DDIMScheduler, PNDMScheduler + + >>> # Download scheduler from huggingface.co and cache. + >>> scheduler = DDPMScheduler.from_pretrained("google/ddpm-cifar10-32") + + >>> # Instantiate DDIM scheduler class with same config as DDPM + >>> scheduler = DDIMScheduler.from_config(scheduler.config) + + >>> # Instantiate PNDM scheduler class with same config as DDPM + >>> scheduler = PNDMScheduler.from_config(scheduler.config) + ``` + """ + # <===== TO BE REMOVED WITH DEPRECATION + # TODO(Patrick) - make sure to remove the following lines when config=="model_path" is deprecated + if "pretrained_model_name_or_path" in kwargs: + config = kwargs.pop("pretrained_model_name_or_path") + + if config is None: + raise ValueError("Please make sure to provide a config as the first positional argument.") + # ======> + + if not isinstance(config, dict): + deprecation_message = "It is deprecated to pass a pretrained model name or path to `from_config`." + if "Scheduler" in cls.__name__: + deprecation_message += ( + f"If you were trying to load a scheduler, please use {cls}.from_pretrained(...) instead." + " Otherwise, please make sure to pass a configuration dictionary instead. This functionality will" + " be removed in v1.0.0." + ) + elif "Model" in cls.__name__: + deprecation_message += ( + f"If you were trying to load a model, please use {cls}.load_config(...) followed by" + f" {cls}.from_config(...) instead. Otherwise, please make sure to pass a configuration dictionary" + " instead. This functionality will be removed in v1.0.0." + ) + deprecate("config-passed-as-path", "1.0.0", deprecation_message, standard_warn=False) + config, kwargs = cls.load_config(pretrained_model_name_or_path=config, return_unused_kwargs=True, **kwargs) + + init_dict, unused_kwargs, hidden_dict = cls.extract_init_dict(config, **kwargs) + + # Allow dtype to be specified on initialization + if "dtype" in unused_kwargs: + init_dict["dtype"] = unused_kwargs.pop("dtype") + + # add possible deprecated kwargs + for deprecated_kwarg in cls._deprecated_kwargs: + if deprecated_kwarg in unused_kwargs: + init_dict[deprecated_kwarg] = unused_kwargs.pop(deprecated_kwarg) + + # Return model and optionally state and/or unused_kwargs + model = cls(**init_dict) + + # make sure to also save config parameters that might be used for compatible classes + # update _class_name + if "_class_name" in hidden_dict: + hidden_dict["_class_name"] = cls.__name__ + + model.register_to_config(**hidden_dict) + + # add hidden kwargs of compatible classes to unused_kwargs + unused_kwargs = {**unused_kwargs, **hidden_dict} + + if return_unused_kwargs: + return (model, unused_kwargs) + else: + return model + + @classmethod + def get_config_dict(cls, *args, **kwargs): + deprecation_message = ( + f" The function get_config_dict is deprecated. Please use {cls}.load_config instead. This function will be" + " removed in version v1.0.0" + ) + deprecate("get_config_dict", "1.0.0", deprecation_message, standard_warn=False) + return cls.load_config(*args, **kwargs) + + @classmethod + @validate_hf_hub_args + def load_config( + cls, + pretrained_model_name_or_path: Union[str, os.PathLike], + return_unused_kwargs=False, + return_commit_hash=False, + **kwargs, + ) -> Tuple[Dict[str, Any], Dict[str, Any]]: + r""" + Load a model or scheduler configuration. + + Parameters: + pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*): + Can be either: + + - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on + the Hub. + - A path to a *directory* (for example `./my_model_directory`) containing model weights saved with + [`~ConfigMixin.save_config`]. + + cache_dir (`Union[str, os.PathLike]`, *optional*): + Path to a directory where a downloaded pretrained model configuration is cached if the standard cache + is not used. + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. + output_loading_info(`bool`, *optional*, defaults to `False`): + Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages. + local_files_only (`bool`, *optional*, defaults to `False`): + Whether to only load local model weights and configuration files or not. If set to `True`, the model + won't be downloaded from the Hub. + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from + `diffusers-cli login` (stored in `~/.huggingface`) is used. + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier + allowed by Git. + subfolder (`str`, *optional*, defaults to `""`): + The subfolder location of a model file within a larger model repository on the Hub or locally. + return_unused_kwargs (`bool`, *optional*, defaults to `False): + Whether unused keyword arguments of the config are returned. + return_commit_hash (`bool`, *optional*, defaults to `False): + Whether the `commit_hash` of the loaded configuration are returned. + + Returns: + `dict`: + A dictionary of all the parameters stored in a JSON configuration file. + + """ + cache_dir = kwargs.pop("cache_dir", None) + local_dir = kwargs.pop("local_dir", None) + local_dir_use_symlinks = kwargs.pop("local_dir_use_symlinks", "auto") + force_download = kwargs.pop("force_download", False) + proxies = kwargs.pop("proxies", None) + token = kwargs.pop("token", None) + local_files_only = kwargs.pop("local_files_only", False) + revision = kwargs.pop("revision", None) + _ = kwargs.pop("mirror", None) + subfolder = kwargs.pop("subfolder", None) + user_agent = kwargs.pop("user_agent", {}) + dduf_entries: Optional[Dict[str, DDUFEntry]] = kwargs.pop("dduf_entries", None) + + user_agent = {**user_agent, "file_type": "config"} + user_agent = http_user_agent(user_agent) + + pretrained_model_name_or_path = str(pretrained_model_name_or_path) + + if cls.config_name is None: + raise ValueError( + "`self.config_name` is not defined. Note that one should not load a config from " + "`ConfigMixin`. Please make sure to define `config_name` in a class inheriting from `ConfigMixin`" + ) + # Custom path for now + if dduf_entries: + if subfolder is not None: + raise ValueError( + "DDUF file only allow for 1 level of directory (e.g transformer/model1/model.safetentors is not allowed). " + "Please check the DDUF structure" + ) + config_file = cls._get_config_file_from_dduf(pretrained_model_name_or_path, dduf_entries) + elif os.path.isfile(pretrained_model_name_or_path): + config_file = pretrained_model_name_or_path + elif os.path.isdir(pretrained_model_name_or_path): + if subfolder is not None and os.path.isfile( + os.path.join(pretrained_model_name_or_path, subfolder, cls.config_name) + ): + config_file = os.path.join(pretrained_model_name_or_path, subfolder, cls.config_name) + elif os.path.isfile(os.path.join(pretrained_model_name_or_path, cls.config_name)): + # Load from a PyTorch checkpoint + config_file = os.path.join(pretrained_model_name_or_path, cls.config_name) + else: + raise EnvironmentError( + f"Error no file named {cls.config_name} found in directory {pretrained_model_name_or_path}." + ) + else: + try: + # Load from URL or cache if already cached + config_file = hf_hub_download( + pretrained_model_name_or_path, + filename=cls.config_name, + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + user_agent=user_agent, + subfolder=subfolder, + revision=revision, + local_dir=local_dir, + local_dir_use_symlinks=local_dir_use_symlinks, + ) + except RepositoryNotFoundError: + raise EnvironmentError( + f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier" + " listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a" + " token having permission to this repo with `token` or log in with `huggingface-cli login`." + ) + except RevisionNotFoundError: + raise EnvironmentError( + f"{revision} is not a valid git identifier (branch name, tag name or commit id) that exists for" + " this model name. Check the model page at" + f" 'https://huggingface.co/{pretrained_model_name_or_path}' for available revisions." + ) + except EntryNotFoundError: + raise EnvironmentError( + f"{pretrained_model_name_or_path} does not appear to have a file named {cls.config_name}." + ) + except HTTPError as err: + raise EnvironmentError( + "There was a specific connection error when trying to load" + f" {pretrained_model_name_or_path}:\n{err}" + ) + except ValueError: + raise EnvironmentError( + f"We couldn't connect to '{HUGGINGFACE_CO_RESOLVE_ENDPOINT}' to load this model, couldn't find it" + f" in the cached files and it looks like {pretrained_model_name_or_path} is not the path to a" + f" directory containing a {cls.config_name} file.\nCheckout your internet connection or see how to" + " run the library in offline mode at" + " 'https://huggingface.co/docs/diffusers/installation#offline-mode'." + ) + except EnvironmentError: + raise EnvironmentError( + f"Can't load config for '{pretrained_model_name_or_path}'. If you were trying to load it from " + "'https://huggingface.co/models', make sure you don't have a local directory with the same name. " + f"Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a directory " + f"containing a {cls.config_name} file" + ) + try: + config_dict = cls._dict_from_json_file(config_file, dduf_entries=dduf_entries) + + commit_hash = extract_commit_hash(config_file) + except (json.JSONDecodeError, UnicodeDecodeError): + raise EnvironmentError(f"It looks like the config file at '{config_file}' is not a valid JSON file.") + + if not (return_unused_kwargs or return_commit_hash): + return config_dict + + outputs = (config_dict,) + + if return_unused_kwargs: + outputs += (kwargs,) + + if return_commit_hash: + outputs += (commit_hash,) + + return outputs + + @staticmethod + def _get_init_keys(input_class): + return set(dict(inspect.signature(input_class.__init__).parameters).keys()) + + @classmethod + def extract_init_dict(cls, config_dict, **kwargs): + # Skip keys that were not present in the original config, so default __init__ values were used + used_defaults = config_dict.get("_use_default_values", []) + config_dict = {k: v for k, v in config_dict.items() if k not in used_defaults and k != "_use_default_values"} + + # 0. Copy origin config dict + original_dict = dict(config_dict.items()) + + # 1. Retrieve expected config attributes from __init__ signature + expected_keys = cls._get_init_keys(cls) + expected_keys.remove("self") + # remove general kwargs if present in dict + if "kwargs" in expected_keys: + expected_keys.remove("kwargs") + # remove flax internal keys + if hasattr(cls, "_flax_internal_args"): + for arg in cls._flax_internal_args: + expected_keys.remove(arg) + + # 2. Remove attributes that cannot be expected from expected config attributes + # remove keys to be ignored + if len(cls.ignore_for_config) > 0: + expected_keys = expected_keys - set(cls.ignore_for_config) + + # load diffusers library to import compatible and original scheduler + diffusers_library = importlib.import_module(__name__.split(".")[0]) + + if cls.has_compatibles: + compatible_classes = [c for c in cls._get_compatibles() if not isinstance(c, DummyObject)] + else: + compatible_classes = [] + + expected_keys_comp_cls = set() + for c in compatible_classes: + expected_keys_c = cls._get_init_keys(c) + expected_keys_comp_cls = expected_keys_comp_cls.union(expected_keys_c) + expected_keys_comp_cls = expected_keys_comp_cls - cls._get_init_keys(cls) + config_dict = {k: v for k, v in config_dict.items() if k not in expected_keys_comp_cls} + + # remove attributes from orig class that cannot be expected + orig_cls_name = config_dict.pop("_class_name", cls.__name__) + if ( + isinstance(orig_cls_name, str) + and orig_cls_name != cls.__name__ + and hasattr(diffusers_library, orig_cls_name) + ): + orig_cls = getattr(diffusers_library, orig_cls_name) + unexpected_keys_from_orig = cls._get_init_keys(orig_cls) - expected_keys + config_dict = {k: v for k, v in config_dict.items() if k not in unexpected_keys_from_orig} + elif not isinstance(orig_cls_name, str) and not isinstance(orig_cls_name, (list, tuple)): + raise ValueError( + "Make sure that the `_class_name` is of type string or list of string (for custom pipelines)." + ) + + # remove private attributes + config_dict = {k: v for k, v in config_dict.items() if not k.startswith("_")} + + # remove quantization_config + config_dict = {k: v for k, v in config_dict.items() if k != "quantization_config"} + + # 3. Create keyword arguments that will be passed to __init__ from expected keyword arguments + init_dict = {} + for key in expected_keys: + # if config param is passed to kwarg and is present in config dict + # it should overwrite existing config dict key + if key in kwargs and key in config_dict: + config_dict[key] = kwargs.pop(key) + + if key in kwargs: + # overwrite key + init_dict[key] = kwargs.pop(key) + elif key in config_dict: + # use value from config dict + init_dict[key] = config_dict.pop(key) + + # 4. Give nice warning if unexpected values have been passed + if len(config_dict) > 0: + logger.warning( + f"The config attributes {config_dict} were passed to {cls.__name__}, " + "but are not expected and will be ignored. Please verify your " + f"{cls.config_name} configuration file." + ) + + # 5. Give nice info if config attributes are initialized to default because they have not been passed + passed_keys = set(init_dict.keys()) + if len(expected_keys - passed_keys) > 0: + logger.info( + f"{expected_keys - passed_keys} was not found in config. Values will be initialized to default values." + ) + + # 6. Define unused keyword arguments + unused_kwargs = {**config_dict, **kwargs} + + # 7. Define "hidden" config parameters that were saved for compatible classes + hidden_config_dict = {k: v for k, v in original_dict.items() if k not in init_dict} + + return init_dict, unused_kwargs, hidden_config_dict + + @classmethod + def _dict_from_json_file( + cls, json_file: Union[str, os.PathLike], dduf_entries: Optional[Dict[str, DDUFEntry]] = None + ): + if dduf_entries: + text = dduf_entries[json_file].read_text() + else: + with open(json_file, "r", encoding="utf-8") as reader: + text = reader.read() + return json.loads(text) + + def __repr__(self): + return f"{self.__class__.__name__} {self.to_json_string()}" + + @property + def config(self) -> Dict[str, Any]: + """ + Returns the config of the class as a frozen dictionary + + Returns: + `Dict[str, Any]`: Config of the class. + """ + return self._internal_dict + + def to_json_string(self) -> str: + """ + Serializes the configuration instance to a JSON string. + + Returns: + `str`: + String containing all the attributes that make up the configuration instance in JSON format. + """ + config_dict = self._internal_dict if hasattr(self, "_internal_dict") else {} + config_dict["_class_name"] = self.__class__.__name__ + config_dict["_diffusers_version"] = __version__ + + def to_json_saveable(value): + if isinstance(value, np.ndarray): + value = value.tolist() + elif isinstance(value, Path): + value = value.as_posix() + return value + + if "quantization_config" in config_dict: + config_dict["quantization_config"] = ( + config_dict.quantization_config.to_dict() + if not isinstance(config_dict.quantization_config, dict) + else config_dict.quantization_config + ) + + config_dict = {k: to_json_saveable(v) for k, v in config_dict.items()} + # Don't save "_ignore_files" or "_use_default_values" + config_dict.pop("_ignore_files", None) + config_dict.pop("_use_default_values", None) + # pop the `_pre_quantization_dtype` as torch.dtypes are not serializable. + _ = config_dict.pop("_pre_quantization_dtype", None) + + return json.dumps(config_dict, indent=2, sort_keys=True) + "\n" + + def to_json_file(self, json_file_path: Union[str, os.PathLike]): + """ + Save the configuration instance's parameters to a JSON file. + + Args: + json_file_path (`str` or `os.PathLike`): + Path to the JSON file to save a configuration instance's parameters. + """ + with open(json_file_path, "w", encoding="utf-8") as writer: + writer.write(self.to_json_string()) + + @classmethod + def _get_config_file_from_dduf(cls, pretrained_model_name_or_path: str, dduf_entries: Dict[str, DDUFEntry]): + # paths inside a DDUF file must always be "/" + config_file = ( + cls.config_name + if pretrained_model_name_or_path == "" + else "/".join([pretrained_model_name_or_path, cls.config_name]) + ) + if config_file not in dduf_entries: + raise ValueError( + f"We did not manage to find the file {config_file} in the dduf file. We only have the following files {dduf_entries.keys()}" + ) + return config_file + + +def register_to_config(init): + r""" + Decorator to apply on the init of classes inheriting from [`ConfigMixin`] so that all the arguments are + automatically sent to `self.register_for_config`. To ignore a specific argument accepted by the init but that + shouldn't be registered in the config, use the `ignore_for_config` class variable + + Warning: Once decorated, all private arguments (beginning with an underscore) are trashed and not sent to the init! + """ + + @functools.wraps(init) + def inner_init(self, *args, **kwargs): + # Ignore private kwargs in the init. + init_kwargs = {k: v for k, v in kwargs.items() if not k.startswith("_")} + config_init_kwargs = {k: v for k, v in kwargs.items() if k.startswith("_")} + if not isinstance(self, ConfigMixin): + raise RuntimeError( + f"`@register_for_config` was applied to {self.__class__.__name__} init method, but this class does " + "not inherit from `ConfigMixin`." + ) + + ignore = getattr(self, "ignore_for_config", []) + # Get positional arguments aligned with kwargs + new_kwargs = {} + signature = inspect.signature(init) + parameters = { + name: p.default for i, (name, p) in enumerate(signature.parameters.items()) if i > 0 and name not in ignore + } + for arg, name in zip(args, parameters.keys()): + new_kwargs[name] = arg + + # Then add all kwargs + new_kwargs.update( + { + k: init_kwargs.get(k, default) + for k, default in parameters.items() + if k not in ignore and k not in new_kwargs + } + ) + + # Take note of the parameters that were not present in the loaded config + if len(set(new_kwargs.keys()) - set(init_kwargs)) > 0: + new_kwargs["_use_default_values"] = list(set(new_kwargs.keys()) - set(init_kwargs)) + + new_kwargs = {**config_init_kwargs, **new_kwargs} + getattr(self, "register_to_config")(**new_kwargs) + init(self, *args, **init_kwargs) + + return inner_init + + +def flax_register_to_config(cls): + original_init = cls.__init__ + + @functools.wraps(original_init) + def init(self, *args, **kwargs): + if not isinstance(self, ConfigMixin): + raise RuntimeError( + f"`@register_for_config` was applied to {self.__class__.__name__} init method, but this class does " + "not inherit from `ConfigMixin`." + ) + + # Ignore private kwargs in the init. Retrieve all passed attributes + init_kwargs = dict(kwargs.items()) + + # Retrieve default values + fields = dataclasses.fields(self) + default_kwargs = {} + for field in fields: + # ignore flax specific attributes + if field.name in self._flax_internal_args: + continue + if type(field.default) == dataclasses._MISSING_TYPE: + default_kwargs[field.name] = None + else: + default_kwargs[field.name] = getattr(self, field.name) + + # Make sure init_kwargs override default kwargs + new_kwargs = {**default_kwargs, **init_kwargs} + # dtype should be part of `init_kwargs`, but not `new_kwargs` + if "dtype" in new_kwargs: + new_kwargs.pop("dtype") + + # Get positional arguments aligned with kwargs + for i, arg in enumerate(args): + name = fields[i].name + new_kwargs[name] = arg + + # Take note of the parameters that were not present in the loaded config + if len(set(new_kwargs.keys()) - set(init_kwargs)) > 0: + new_kwargs["_use_default_values"] = list(set(new_kwargs.keys()) - set(init_kwargs)) + + getattr(self, "register_to_config")(**new_kwargs) + original_init(self, *args, **kwargs) + + cls.__init__ = init + return cls + + +class LegacyConfigMixin(ConfigMixin): + r""" + A subclass of `ConfigMixin` to resolve class mapping from legacy classes (like `Transformer2DModel`) to more + pipeline-specific classes (like `DiTTransformer2DModel`). + """ + + @classmethod + def from_config(cls, config: Union[FrozenDict, Dict[str, Any]] = None, return_unused_kwargs=False, **kwargs): + # To prevent dependency import problem. + from .models.model_loading_utils import _fetch_remapped_cls_from_config + + # resolve remapping + remapped_class = _fetch_remapped_cls_from_config(config, cls) + + return remapped_class.from_config(config, return_unused_kwargs, **kwargs) diff --git a/venv/lib/python3.10/site-packages/diffusers/dependency_versions_check.py b/venv/lib/python3.10/site-packages/diffusers/dependency_versions_check.py new file mode 100644 index 0000000000000000000000000000000000000000..e3670b136af45742a64a96b0d68849a2ca0700ed --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/dependency_versions_check.py @@ -0,0 +1,34 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 .dependency_versions_table import deps +from .utils.versions import require_version, require_version_core + + +# define which module versions we always want to check at run time +# (usually the ones defined in `install_requires` in setup.py) +# +# order specific notes: +# - tqdm must be checked before tokenizers + +pkgs_to_check_at_runtime = "python requests filelock numpy".split() +for pkg in pkgs_to_check_at_runtime: + if pkg in deps: + require_version_core(deps[pkg]) + else: + raise ValueError(f"can't find {pkg} in {deps.keys()}, check dependency_versions_table.py") + + +def dep_version_check(pkg, hint=None): + require_version(deps[pkg], hint) diff --git a/venv/lib/python3.10/site-packages/diffusers/dependency_versions_table.py b/venv/lib/python3.10/site-packages/diffusers/dependency_versions_table.py new file mode 100644 index 0000000000000000000000000000000000000000..f35353c49eab7f83e3a2670cb3ff0c1923872cfc --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/dependency_versions_table.py @@ -0,0 +1,53 @@ +# THIS FILE HAS BEEN AUTOGENERATED. To update: +# 1. modify the `_deps` dict in setup.py +# 2. run `make deps_table_update` +deps = { + "Pillow": "Pillow", + "accelerate": "accelerate>=0.31.0", + "compel": "compel==0.1.8", + "datasets": "datasets", + "filelock": "filelock", + "flax": "flax>=0.4.1", + "hf-doc-builder": "hf-doc-builder>=0.3.0", + "huggingface-hub": "huggingface-hub>=0.27.0", + "requests-mock": "requests-mock==1.10.0", + "importlib_metadata": "importlib_metadata", + "invisible-watermark": "invisible-watermark>=0.2.0", + "isort": "isort>=5.5.4", + "jax": "jax>=0.4.1", + "jaxlib": "jaxlib>=0.4.1", + "Jinja2": "Jinja2", + "k-diffusion": "k-diffusion>=0.0.12", + "torchsde": "torchsde", + "note_seq": "note_seq", + "librosa": "librosa", + "numpy": "numpy", + "parameterized": "parameterized", + "peft": "peft>=0.15.0", + "protobuf": "protobuf>=3.20.3,<4", + "pytest": "pytest", + "pytest-timeout": "pytest-timeout", + "pytest-xdist": "pytest-xdist", + "python": "python>=3.8.0", + "ruff": "ruff==0.9.10", + "safetensors": "safetensors>=0.3.1", + "sentencepiece": "sentencepiece>=0.1.91,!=0.1.92", + "GitPython": "GitPython<3.1.19", + "scipy": "scipy", + "onnx": "onnx", + "optimum_quanto": "optimum_quanto>=0.2.6", + "gguf": "gguf>=0.10.0", + "torchao": "torchao>=0.7.0", + "bitsandbytes": "bitsandbytes>=0.43.3", + "regex": "regex!=2019.12.17", + "requests": "requests", + "tensorboard": "tensorboard", + "tiktoken": "tiktoken>=0.7.0", + "torch": "torch>=1.4", + "torchvision": "torchvision", + "transformers": "transformers>=4.41.2", + "urllib3": "urllib3<=2.0.0", + "black": "black", + "phonemizer": "phonemizer", + "opencv-python": "opencv-python", +} diff --git a/venv/lib/python3.10/site-packages/diffusers/image_processor.py b/venv/lib/python3.10/site-packages/diffusers/image_processor.py new file mode 100644 index 0000000000000000000000000000000000000000..6a3cf77a7df77577c895431ec89366305e176fc8 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/image_processor.py @@ -0,0 +1,1319 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 warnings +from typing import List, Optional, Tuple, Union + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from PIL import Image, ImageFilter, ImageOps + +from .configuration_utils import ConfigMixin, register_to_config +from .utils import CONFIG_NAME, PIL_INTERPOLATION, deprecate + + +PipelineImageInput = Union[ + PIL.Image.Image, + np.ndarray, + torch.Tensor, + List[PIL.Image.Image], + List[np.ndarray], + List[torch.Tensor], +] + +PipelineDepthInput = PipelineImageInput + + +def is_valid_image(image) -> bool: + r""" + Checks if the input is a valid image. + + A valid image can be: + - A `PIL.Image.Image`. + - A 2D or 3D `np.ndarray` or `torch.Tensor` (grayscale or color image). + + Args: + image (`Union[PIL.Image.Image, np.ndarray, torch.Tensor]`): + The image to validate. It can be a PIL image, a NumPy array, or a torch tensor. + + Returns: + `bool`: + `True` if the input is a valid image, `False` otherwise. + """ + return isinstance(image, PIL.Image.Image) or isinstance(image, (np.ndarray, torch.Tensor)) and image.ndim in (2, 3) + + +def is_valid_image_imagelist(images): + r""" + Checks if the input is a valid image or list of images. + + The input can be one of the following formats: + - A 4D tensor or numpy array (batch of images). + - A valid single image: `PIL.Image.Image`, 2D `np.ndarray` or `torch.Tensor` (grayscale image), 3D `np.ndarray` or + `torch.Tensor`. + - A list of valid images. + + Args: + images (`Union[np.ndarray, torch.Tensor, PIL.Image.Image, List]`): + The image(s) to check. Can be a batch of images (4D tensor/array), a single image, or a list of valid + images. + + Returns: + `bool`: + `True` if the input is valid, `False` otherwise. + """ + if isinstance(images, (np.ndarray, torch.Tensor)) and images.ndim == 4: + return True + elif is_valid_image(images): + return True + elif isinstance(images, list): + return all(is_valid_image(image) for image in images) + return False + + +class VaeImageProcessor(ConfigMixin): + """ + Image processor for VAE. + + Args: + do_resize (`bool`, *optional*, defaults to `True`): + Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`. Can accept + `height` and `width` arguments from [`image_processor.VaeImageProcessor.preprocess`] method. + vae_scale_factor (`int`, *optional*, defaults to `8`): + VAE scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of this factor. + resample (`str`, *optional*, defaults to `lanczos`): + Resampling filter to use when resizing the image. + do_normalize (`bool`, *optional*, defaults to `True`): + Whether to normalize the image to [-1,1]. + do_binarize (`bool`, *optional*, defaults to `False`): + Whether to binarize the image to 0/1. + do_convert_rgb (`bool`, *optional*, defaults to be `False`): + Whether to convert the images to RGB format. + do_convert_grayscale (`bool`, *optional*, defaults to be `False`): + Whether to convert the images to grayscale format. + """ + + config_name = CONFIG_NAME + + @register_to_config + def __init__( + self, + do_resize: bool = True, + vae_scale_factor: int = 8, + vae_latent_channels: int = 4, + resample: str = "lanczos", + reducing_gap: int = None, + do_normalize: bool = True, + do_binarize: bool = False, + do_convert_rgb: bool = False, + do_convert_grayscale: bool = False, + ): + super().__init__() + if do_convert_rgb and do_convert_grayscale: + raise ValueError( + "`do_convert_rgb` and `do_convert_grayscale` can not both be set to `True`," + " if you intended to convert the image into RGB format, please set `do_convert_grayscale = False`.", + " if you intended to convert the image into grayscale format, please set `do_convert_rgb = False`", + ) + + @staticmethod + def numpy_to_pil(images: np.ndarray) -> List[PIL.Image.Image]: + r""" + Convert a numpy image or a batch of images to a PIL image. + + Args: + images (`np.ndarray`): + The image array to convert to PIL format. + + Returns: + `List[PIL.Image.Image]`: + A list of PIL images. + """ + if images.ndim == 3: + images = images[None, ...] + images = (images * 255).round().astype("uint8") + if images.shape[-1] == 1: + # special case for grayscale (single channel) images + pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images] + else: + pil_images = [Image.fromarray(image) for image in images] + + return pil_images + + @staticmethod + def pil_to_numpy(images: Union[List[PIL.Image.Image], PIL.Image.Image]) -> np.ndarray: + r""" + Convert a PIL image or a list of PIL images to NumPy arrays. + + Args: + images (`PIL.Image.Image` or `List[PIL.Image.Image]`): + The PIL image or list of images to convert to NumPy format. + + Returns: + `np.ndarray`: + A NumPy array representation of the images. + """ + if not isinstance(images, list): + images = [images] + images = [np.array(image).astype(np.float32) / 255.0 for image in images] + images = np.stack(images, axis=0) + + return images + + @staticmethod + def numpy_to_pt(images: np.ndarray) -> torch.Tensor: + r""" + Convert a NumPy image to a PyTorch tensor. + + Args: + images (`np.ndarray`): + The NumPy image array to convert to PyTorch format. + + Returns: + `torch.Tensor`: + A PyTorch tensor representation of the images. + """ + if images.ndim == 3: + images = images[..., None] + + images = torch.from_numpy(images.transpose(0, 3, 1, 2)) + return images + + @staticmethod + def pt_to_numpy(images: torch.Tensor) -> np.ndarray: + r""" + Convert a PyTorch tensor to a NumPy image. + + Args: + images (`torch.Tensor`): + The PyTorch tensor to convert to NumPy format. + + Returns: + `np.ndarray`: + A NumPy array representation of the images. + """ + images = images.cpu().permute(0, 2, 3, 1).float().numpy() + return images + + @staticmethod + def normalize(images: Union[np.ndarray, torch.Tensor]) -> Union[np.ndarray, torch.Tensor]: + r""" + Normalize an image array to [-1,1]. + + Args: + images (`np.ndarray` or `torch.Tensor`): + The image array to normalize. + + Returns: + `np.ndarray` or `torch.Tensor`: + The normalized image array. + """ + return 2.0 * images - 1.0 + + @staticmethod + def denormalize(images: Union[np.ndarray, torch.Tensor]) -> Union[np.ndarray, torch.Tensor]: + r""" + Denormalize an image array to [0,1]. + + Args: + images (`np.ndarray` or `torch.Tensor`): + The image array to denormalize. + + Returns: + `np.ndarray` or `torch.Tensor`: + The denormalized image array. + """ + return (images * 0.5 + 0.5).clamp(0, 1) + + @staticmethod + def convert_to_rgb(image: PIL.Image.Image) -> PIL.Image.Image: + r""" + Converts a PIL image to RGB format. + + Args: + image (`PIL.Image.Image`): + The PIL image to convert to RGB. + + Returns: + `PIL.Image.Image`: + The RGB-converted PIL image. + """ + image = image.convert("RGB") + + return image + + @staticmethod + def convert_to_grayscale(image: PIL.Image.Image) -> PIL.Image.Image: + r""" + Converts a given PIL image to grayscale. + + Args: + image (`PIL.Image.Image`): + The input image to convert. + + Returns: + `PIL.Image.Image`: + The image converted to grayscale. + """ + image = image.convert("L") + + return image + + @staticmethod + def blur(image: PIL.Image.Image, blur_factor: int = 4) -> PIL.Image.Image: + r""" + Applies Gaussian blur to an image. + + Args: + image (`PIL.Image.Image`): + The PIL image to convert to grayscale. + + Returns: + `PIL.Image.Image`: + The grayscale-converted PIL image. + """ + image = image.filter(ImageFilter.GaussianBlur(blur_factor)) + + return image + + @staticmethod + def get_crop_region(mask_image: PIL.Image.Image, width: int, height: int, pad=0): + r""" + Finds a rectangular region that contains all masked ares in an image, and expands region to match the aspect + ratio of the original image; for example, if user drew mask in a 128x32 region, and the dimensions for + processing are 512x512, the region will be expanded to 128x128. + + Args: + mask_image (PIL.Image.Image): Mask image. + width (int): Width of the image to be processed. + height (int): Height of the image to be processed. + pad (int, optional): Padding to be added to the crop region. Defaults to 0. + + Returns: + tuple: (x1, y1, x2, y2) represent a rectangular region that contains all masked ares in an image and + matches the original aspect ratio. + """ + + mask_image = mask_image.convert("L") + mask = np.array(mask_image) + + # 1. find a rectangular region that contains all masked ares in an image + h, w = mask.shape + crop_left = 0 + for i in range(w): + if not (mask[:, i] == 0).all(): + break + crop_left += 1 + + crop_right = 0 + for i in reversed(range(w)): + if not (mask[:, i] == 0).all(): + break + crop_right += 1 + + crop_top = 0 + for i in range(h): + if not (mask[i] == 0).all(): + break + crop_top += 1 + + crop_bottom = 0 + for i in reversed(range(h)): + if not (mask[i] == 0).all(): + break + crop_bottom += 1 + + # 2. add padding to the crop region + x1, y1, x2, y2 = ( + int(max(crop_left - pad, 0)), + int(max(crop_top - pad, 0)), + int(min(w - crop_right + pad, w)), + int(min(h - crop_bottom + pad, h)), + ) + + # 3. expands crop region to match the aspect ratio of the image to be processed + ratio_crop_region = (x2 - x1) / (y2 - y1) + ratio_processing = width / height + + if ratio_crop_region > ratio_processing: + desired_height = (x2 - x1) / ratio_processing + desired_height_diff = int(desired_height - (y2 - y1)) + y1 -= desired_height_diff // 2 + y2 += desired_height_diff - desired_height_diff // 2 + if y2 >= mask_image.height: + diff = y2 - mask_image.height + y2 -= diff + y1 -= diff + if y1 < 0: + y2 -= y1 + y1 -= y1 + if y2 >= mask_image.height: + y2 = mask_image.height + else: + desired_width = (y2 - y1) * ratio_processing + desired_width_diff = int(desired_width - (x2 - x1)) + x1 -= desired_width_diff // 2 + x2 += desired_width_diff - desired_width_diff // 2 + if x2 >= mask_image.width: + diff = x2 - mask_image.width + x2 -= diff + x1 -= diff + if x1 < 0: + x2 -= x1 + x1 -= x1 + if x2 >= mask_image.width: + x2 = mask_image.width + + return x1, y1, x2, y2 + + def _resize_and_fill( + self, + image: PIL.Image.Image, + width: int, + height: int, + ) -> PIL.Image.Image: + r""" + Resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center + the image within the dimensions, filling empty with data from image. + + Args: + image (`PIL.Image.Image`): + The image to resize and fill. + width (`int`): + The width to resize the image to. + height (`int`): + The height to resize the image to. + + Returns: + `PIL.Image.Image`: + The resized and filled image. + """ + + ratio = width / height + src_ratio = image.width / image.height + + src_w = width if ratio < src_ratio else image.width * height // image.height + src_h = height if ratio >= src_ratio else image.height * width // image.width + + resized = image.resize((src_w, src_h), resample=PIL_INTERPOLATION["lanczos"]) + res = Image.new("RGB", (width, height)) + res.paste(resized, box=(width // 2 - src_w // 2, height // 2 - src_h // 2)) + + if ratio < src_ratio: + fill_height = height // 2 - src_h // 2 + if fill_height > 0: + res.paste(resized.resize((width, fill_height), box=(0, 0, width, 0)), box=(0, 0)) + res.paste( + resized.resize((width, fill_height), box=(0, resized.height, width, resized.height)), + box=(0, fill_height + src_h), + ) + elif ratio > src_ratio: + fill_width = width // 2 - src_w // 2 + if fill_width > 0: + res.paste(resized.resize((fill_width, height), box=(0, 0, 0, height)), box=(0, 0)) + res.paste( + resized.resize((fill_width, height), box=(resized.width, 0, resized.width, height)), + box=(fill_width + src_w, 0), + ) + + return res + + def _resize_and_crop( + self, + image: PIL.Image.Image, + width: int, + height: int, + ) -> PIL.Image.Image: + r""" + Resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center + the image within the dimensions, cropping the excess. + + Args: + image (`PIL.Image.Image`): + The image to resize and crop. + width (`int`): + The width to resize the image to. + height (`int`): + The height to resize the image to. + + Returns: + `PIL.Image.Image`: + The resized and cropped image. + """ + ratio = width / height + src_ratio = image.width / image.height + + src_w = width if ratio > src_ratio else image.width * height // image.height + src_h = height if ratio <= src_ratio else image.height * width // image.width + + resized = image.resize((src_w, src_h), resample=PIL_INTERPOLATION["lanczos"]) + res = Image.new("RGB", (width, height)) + res.paste(resized, box=(width // 2 - src_w // 2, height // 2 - src_h // 2)) + return res + + def resize( + self, + image: Union[PIL.Image.Image, np.ndarray, torch.Tensor], + height: int, + width: int, + resize_mode: str = "default", # "default", "fill", "crop" + ) -> Union[PIL.Image.Image, np.ndarray, torch.Tensor]: + """ + Resize image. + + Args: + image (`PIL.Image.Image`, `np.ndarray` or `torch.Tensor`): + The image input, can be a PIL image, numpy array or pytorch tensor. + height (`int`): + The height to resize to. + width (`int`): + The width to resize to. + resize_mode (`str`, *optional*, defaults to `default`): + The resize mode to use, can be one of `default` or `fill`. If `default`, will resize the image to fit + within the specified width and height, and it may not maintaining the original aspect ratio. If `fill`, + will resize the image to fit within the specified width and height, maintaining the aspect ratio, and + then center the image within the dimensions, filling empty with data from image. If `crop`, will resize + the image to fit within the specified width and height, maintaining the aspect ratio, and then center + the image within the dimensions, cropping the excess. Note that resize_mode `fill` and `crop` are only + supported for PIL image input. + + Returns: + `PIL.Image.Image`, `np.ndarray` or `torch.Tensor`: + The resized image. + """ + if resize_mode != "default" and not isinstance(image, PIL.Image.Image): + raise ValueError(f"Only PIL image input is supported for resize_mode {resize_mode}") + if isinstance(image, PIL.Image.Image): + if resize_mode == "default": + image = image.resize( + (width, height), + resample=PIL_INTERPOLATION[self.config.resample], + reducing_gap=self.config.reducing_gap, + ) + elif resize_mode == "fill": + image = self._resize_and_fill(image, width, height) + elif resize_mode == "crop": + image = self._resize_and_crop(image, width, height) + else: + raise ValueError(f"resize_mode {resize_mode} is not supported") + + elif isinstance(image, torch.Tensor): + image = torch.nn.functional.interpolate( + image, + size=(height, width), + ) + elif isinstance(image, np.ndarray): + image = self.numpy_to_pt(image) + image = torch.nn.functional.interpolate( + image, + size=(height, width), + ) + image = self.pt_to_numpy(image) + return image + + def binarize(self, image: PIL.Image.Image) -> PIL.Image.Image: + """ + Create a mask. + + Args: + image (`PIL.Image.Image`): + The image input, should be a PIL image. + + Returns: + `PIL.Image.Image`: + The binarized image. Values less than 0.5 are set to 0, values greater than 0.5 are set to 1. + """ + image[image < 0.5] = 0 + image[image >= 0.5] = 1 + + return image + + def _denormalize_conditionally( + self, images: torch.Tensor, do_denormalize: Optional[List[bool]] = None + ) -> torch.Tensor: + r""" + Denormalize a batch of images based on a condition list. + + Args: + images (`torch.Tensor`): + The input image tensor. + do_denormalize (`Optional[List[bool]`, *optional*, defaults to `None`): + A list of booleans indicating whether to denormalize each image in the batch. If `None`, will use the + value of `do_normalize` in the `VaeImageProcessor` config. + """ + if do_denormalize is None: + return self.denormalize(images) if self.config.do_normalize else images + + return torch.stack( + [self.denormalize(images[i]) if do_denormalize[i] else images[i] for i in range(images.shape[0])] + ) + + def get_default_height_width( + self, + image: Union[PIL.Image.Image, np.ndarray, torch.Tensor], + height: Optional[int] = None, + width: Optional[int] = None, + ) -> Tuple[int, int]: + r""" + Returns the height and width of the image, downscaled to the next integer multiple of `vae_scale_factor`. + + Args: + image (`Union[PIL.Image.Image, np.ndarray, torch.Tensor]`): + The image input, which can be a PIL image, NumPy array, or PyTorch tensor. If it is a NumPy array, it + should have shape `[batch, height, width]` or `[batch, height, width, channels]`. If it is a PyTorch + tensor, it should have shape `[batch, channels, height, width]`. + height (`Optional[int]`, *optional*, defaults to `None`): + The height of the preprocessed image. If `None`, the height of the `image` input will be used. + width (`Optional[int]`, *optional*, defaults to `None`): + The width of the preprocessed image. If `None`, the width of the `image` input will be used. + + Returns: + `Tuple[int, int]`: + A tuple containing the height and width, both resized to the nearest integer multiple of + `vae_scale_factor`. + """ + + if height is None: + if isinstance(image, PIL.Image.Image): + height = image.height + elif isinstance(image, torch.Tensor): + height = image.shape[2] + else: + height = image.shape[1] + + if width is None: + if isinstance(image, PIL.Image.Image): + width = image.width + elif isinstance(image, torch.Tensor): + width = image.shape[3] + else: + width = image.shape[2] + + width, height = ( + x - x % self.config.vae_scale_factor for x in (width, height) + ) # resize to integer multiple of vae_scale_factor + + return height, width + + def preprocess( + self, + image: PipelineImageInput, + height: Optional[int] = None, + width: Optional[int] = None, + resize_mode: str = "default", # "default", "fill", "crop" + crops_coords: Optional[Tuple[int, int, int, int]] = None, + ) -> torch.Tensor: + """ + Preprocess the image input. + + Args: + image (`PipelineImageInput`): + The image input, accepted formats are PIL images, NumPy arrays, PyTorch tensors; Also accept list of + supported formats. + height (`int`, *optional*): + The height in preprocessed image. If `None`, will use the `get_default_height_width()` to get default + height. + width (`int`, *optional*): + The width in preprocessed. If `None`, will use get_default_height_width()` to get the default width. + resize_mode (`str`, *optional*, defaults to `default`): + The resize mode, can be one of `default` or `fill`. If `default`, will resize the image to fit within + the specified width and height, and it may not maintaining the original aspect ratio. If `fill`, will + resize the image to fit within the specified width and height, maintaining the aspect ratio, and then + center the image within the dimensions, filling empty with data from image. If `crop`, will resize the + image to fit within the specified width and height, maintaining the aspect ratio, and then center the + image within the dimensions, cropping the excess. Note that resize_mode `fill` and `crop` are only + supported for PIL image input. + crops_coords (`List[Tuple[int, int, int, int]]`, *optional*, defaults to `None`): + The crop coordinates for each image in the batch. If `None`, will not crop the image. + + Returns: + `torch.Tensor`: + The preprocessed image. + """ + supported_formats = (PIL.Image.Image, np.ndarray, torch.Tensor) + + # Expand the missing dimension for 3-dimensional pytorch tensor or numpy array that represents grayscale image + if self.config.do_convert_grayscale and isinstance(image, (torch.Tensor, np.ndarray)) and image.ndim == 3: + if isinstance(image, torch.Tensor): + # if image is a pytorch tensor could have 2 possible shapes: + # 1. batch x height x width: we should insert the channel dimension at position 1 + # 2. channel x height x width: we should insert batch dimension at position 0, + # however, since both channel and batch dimension has same size 1, it is same to insert at position 1 + # for simplicity, we insert a dimension of size 1 at position 1 for both cases + image = image.unsqueeze(1) + else: + # if it is a numpy array, it could have 2 possible shapes: + # 1. batch x height x width: insert channel dimension on last position + # 2. height x width x channel: insert batch dimension on first position + if image.shape[-1] == 1: + image = np.expand_dims(image, axis=0) + else: + image = np.expand_dims(image, axis=-1) + + if isinstance(image, list) and isinstance(image[0], np.ndarray) and image[0].ndim == 4: + warnings.warn( + "Passing `image` as a list of 4d np.ndarray is deprecated." + "Please concatenate the list along the batch dimension and pass it as a single 4d np.ndarray", + FutureWarning, + ) + image = np.concatenate(image, axis=0) + if isinstance(image, list) and isinstance(image[0], torch.Tensor) and image[0].ndim == 4: + warnings.warn( + "Passing `image` as a list of 4d torch.Tensor is deprecated." + "Please concatenate the list along the batch dimension and pass it as a single 4d torch.Tensor", + FutureWarning, + ) + image = torch.cat(image, axis=0) + + if not is_valid_image_imagelist(image): + raise ValueError( + f"Input is in incorrect format. Currently, we only support {', '.join(str(x) for x in supported_formats)}" + ) + if not isinstance(image, list): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + if crops_coords is not None: + image = [i.crop(crops_coords) for i in image] + if self.config.do_resize: + height, width = self.get_default_height_width(image[0], height, width) + image = [self.resize(i, height, width, resize_mode=resize_mode) for i in image] + if self.config.do_convert_rgb: + image = [self.convert_to_rgb(i) for i in image] + elif self.config.do_convert_grayscale: + image = [self.convert_to_grayscale(i) for i in image] + image = self.pil_to_numpy(image) # to np + image = self.numpy_to_pt(image) # to pt + + elif isinstance(image[0], np.ndarray): + image = np.concatenate(image, axis=0) if image[0].ndim == 4 else np.stack(image, axis=0) + + image = self.numpy_to_pt(image) + + height, width = self.get_default_height_width(image, height, width) + if self.config.do_resize: + image = self.resize(image, height, width) + + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, axis=0) if image[0].ndim == 4 else torch.stack(image, axis=0) + + if self.config.do_convert_grayscale and image.ndim == 3: + image = image.unsqueeze(1) + + channel = image.shape[1] + # don't need any preprocess if the image is latents + if channel == self.config.vae_latent_channels: + return image + + height, width = self.get_default_height_width(image, height, width) + if self.config.do_resize: + image = self.resize(image, height, width) + + # expected range [0,1], normalize to [-1,1] + do_normalize = self.config.do_normalize + if do_normalize and image.min() < 0: + warnings.warn( + "Passing `image` as torch tensor with value range in [-1,1] is deprecated. The expected value range for image tensor is [0,1] " + f"when passing as pytorch tensor or numpy Array. You passed `image` with value range [{image.min()},{image.max()}]", + FutureWarning, + ) + do_normalize = False + if do_normalize: + image = self.normalize(image) + + if self.config.do_binarize: + image = self.binarize(image) + + return image + + def postprocess( + self, + image: torch.Tensor, + output_type: str = "pil", + do_denormalize: Optional[List[bool]] = None, + ) -> Union[PIL.Image.Image, np.ndarray, torch.Tensor]: + """ + Postprocess the image output from tensor to `output_type`. + + Args: + image (`torch.Tensor`): + The image input, should be a pytorch tensor with shape `B x C x H x W`. + output_type (`str`, *optional*, defaults to `pil`): + The output type of the image, can be one of `pil`, `np`, `pt`, `latent`. + do_denormalize (`List[bool]`, *optional*, defaults to `None`): + Whether to denormalize the image to [0,1]. If `None`, will use the value of `do_normalize` in the + `VaeImageProcessor` config. + + Returns: + `PIL.Image.Image`, `np.ndarray` or `torch.Tensor`: + The postprocessed image. + """ + if not isinstance(image, torch.Tensor): + raise ValueError( + f"Input for postprocessing is in incorrect format: {type(image)}. We only support pytorch tensor" + ) + if output_type not in ["latent", "pt", "np", "pil"]: + deprecation_message = ( + f"the output_type {output_type} is outdated and has been set to `np`. Please make sure to set it to one of these instead: " + "`pil`, `np`, `pt`, `latent`" + ) + deprecate("Unsupported output_type", "1.0.0", deprecation_message, standard_warn=False) + output_type = "np" + + if output_type == "latent": + return image + + image = self._denormalize_conditionally(image, do_denormalize) + + if output_type == "pt": + return image + + image = self.pt_to_numpy(image) + + if output_type == "np": + return image + + if output_type == "pil": + return self.numpy_to_pil(image) + + def apply_overlay( + self, + mask: PIL.Image.Image, + init_image: PIL.Image.Image, + image: PIL.Image.Image, + crop_coords: Optional[Tuple[int, int, int, int]] = None, + ) -> PIL.Image.Image: + r""" + Applies an overlay of the mask and the inpainted image on the original image. + + Args: + mask (`PIL.Image.Image`): + The mask image that highlights regions to overlay. + init_image (`PIL.Image.Image`): + The original image to which the overlay is applied. + image (`PIL.Image.Image`): + The image to overlay onto the original. + crop_coords (`Tuple[int, int, int, int]`, *optional*): + Coordinates to crop the image. If provided, the image will be cropped accordingly. + + Returns: + `PIL.Image.Image`: + The final image with the overlay applied. + """ + + width, height = init_image.width, init_image.height + + init_image_masked = PIL.Image.new("RGBa", (width, height)) + init_image_masked.paste(init_image.convert("RGBA").convert("RGBa"), mask=ImageOps.invert(mask.convert("L"))) + + init_image_masked = init_image_masked.convert("RGBA") + + if crop_coords is not None: + x, y, x2, y2 = crop_coords + w = x2 - x + h = y2 - y + base_image = PIL.Image.new("RGBA", (width, height)) + image = self.resize(image, height=h, width=w, resize_mode="crop") + base_image.paste(image, (x, y)) + image = base_image.convert("RGB") + + image = image.convert("RGBA") + image.alpha_composite(init_image_masked) + image = image.convert("RGB") + + return image + + +class VaeImageProcessorLDM3D(VaeImageProcessor): + """ + Image processor for VAE LDM3D. + + Args: + do_resize (`bool`, *optional*, defaults to `True`): + Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`. + vae_scale_factor (`int`, *optional*, defaults to `8`): + VAE scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of this factor. + resample (`str`, *optional*, defaults to `lanczos`): + Resampling filter to use when resizing the image. + do_normalize (`bool`, *optional*, defaults to `True`): + Whether to normalize the image to [-1,1]. + """ + + config_name = CONFIG_NAME + + @register_to_config + def __init__( + self, + do_resize: bool = True, + vae_scale_factor: int = 8, + resample: str = "lanczos", + do_normalize: bool = True, + ): + super().__init__() + + @staticmethod + def numpy_to_pil(images: np.ndarray) -> List[PIL.Image.Image]: + r""" + Convert a NumPy image or a batch of images to a list of PIL images. + + Args: + images (`np.ndarray`): + The input NumPy array of images, which can be a single image or a batch. + + Returns: + `List[PIL.Image.Image]`: + A list of PIL images converted from the input NumPy array. + """ + if images.ndim == 3: + images = images[None, ...] + images = (images * 255).round().astype("uint8") + if images.shape[-1] == 1: + # special case for grayscale (single channel) images + pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images] + else: + pil_images = [Image.fromarray(image[:, :, :3]) for image in images] + + return pil_images + + @staticmethod + def depth_pil_to_numpy(images: Union[List[PIL.Image.Image], PIL.Image.Image]) -> np.ndarray: + r""" + Convert a PIL image or a list of PIL images to NumPy arrays. + + Args: + images (`Union[List[PIL.Image.Image], PIL.Image.Image]`): + The input image or list of images to be converted. + + Returns: + `np.ndarray`: + A NumPy array of the converted images. + """ + if not isinstance(images, list): + images = [images] + + images = [np.array(image).astype(np.float32) / (2**16 - 1) for image in images] + images = np.stack(images, axis=0) + return images + + @staticmethod + def rgblike_to_depthmap(image: Union[np.ndarray, torch.Tensor]) -> Union[np.ndarray, torch.Tensor]: + r""" + Convert an RGB-like depth image to a depth map. + + Args: + image (`Union[np.ndarray, torch.Tensor]`): + The RGB-like depth image to convert. + + Returns: + `Union[np.ndarray, torch.Tensor]`: + The corresponding depth map. + """ + return image[:, :, 1] * 2**8 + image[:, :, 2] + + def numpy_to_depth(self, images: np.ndarray) -> List[PIL.Image.Image]: + r""" + Convert a NumPy depth image or a batch of images to a list of PIL images. + + Args: + images (`np.ndarray`): + The input NumPy array of depth images, which can be a single image or a batch. + + Returns: + `List[PIL.Image.Image]`: + A list of PIL images converted from the input NumPy depth images. + """ + if images.ndim == 3: + images = images[None, ...] + images_depth = images[:, :, :, 3:] + if images.shape[-1] == 6: + images_depth = (images_depth * 255).round().astype("uint8") + pil_images = [ + Image.fromarray(self.rgblike_to_depthmap(image_depth), mode="I;16") for image_depth in images_depth + ] + elif images.shape[-1] == 4: + images_depth = (images_depth * 65535.0).astype(np.uint16) + pil_images = [Image.fromarray(image_depth, mode="I;16") for image_depth in images_depth] + else: + raise Exception("Not supported") + + return pil_images + + def postprocess( + self, + image: torch.Tensor, + output_type: str = "pil", + do_denormalize: Optional[List[bool]] = None, + ) -> Union[PIL.Image.Image, np.ndarray, torch.Tensor]: + """ + Postprocess the image output from tensor to `output_type`. + + Args: + image (`torch.Tensor`): + The image input, should be a pytorch tensor with shape `B x C x H x W`. + output_type (`str`, *optional*, defaults to `pil`): + The output type of the image, can be one of `pil`, `np`, `pt`, `latent`. + do_denormalize (`List[bool]`, *optional*, defaults to `None`): + Whether to denormalize the image to [0,1]. If `None`, will use the value of `do_normalize` in the + `VaeImageProcessor` config. + + Returns: + `PIL.Image.Image`, `np.ndarray` or `torch.Tensor`: + The postprocessed image. + """ + if not isinstance(image, torch.Tensor): + raise ValueError( + f"Input for postprocessing is in incorrect format: {type(image)}. We only support pytorch tensor" + ) + if output_type not in ["latent", "pt", "np", "pil"]: + deprecation_message = ( + f"the output_type {output_type} is outdated and has been set to `np`. Please make sure to set it to one of these instead: " + "`pil`, `np`, `pt`, `latent`" + ) + deprecate("Unsupported output_type", "1.0.0", deprecation_message, standard_warn=False) + output_type = "np" + + image = self._denormalize_conditionally(image, do_denormalize) + + image = self.pt_to_numpy(image) + + if output_type == "np": + if image.shape[-1] == 6: + image_depth = np.stack([self.rgblike_to_depthmap(im[:, :, 3:]) for im in image], axis=0) + else: + image_depth = image[:, :, :, 3:] + return image[:, :, :, :3], image_depth + + if output_type == "pil": + return self.numpy_to_pil(image), self.numpy_to_depth(image) + else: + raise Exception(f"This type {output_type} is not supported") + + def preprocess( + self, + rgb: Union[torch.Tensor, PIL.Image.Image, np.ndarray], + depth: Union[torch.Tensor, PIL.Image.Image, np.ndarray], + height: Optional[int] = None, + width: Optional[int] = None, + target_res: Optional[int] = None, + ) -> torch.Tensor: + r""" + Preprocess the image input. Accepted formats are PIL images, NumPy arrays, or PyTorch tensors. + + Args: + rgb (`Union[torch.Tensor, PIL.Image.Image, np.ndarray]`): + The RGB input image, which can be a single image or a batch. + depth (`Union[torch.Tensor, PIL.Image.Image, np.ndarray]`): + The depth input image, which can be a single image or a batch. + height (`Optional[int]`, *optional*, defaults to `None`): + The desired height of the processed image. If `None`, defaults to the height of the input image. + width (`Optional[int]`, *optional*, defaults to `None`): + The desired width of the processed image. If `None`, defaults to the width of the input image. + target_res (`Optional[int]`, *optional*, defaults to `None`): + Target resolution for resizing the images. If specified, overrides height and width. + + Returns: + `Tuple[torch.Tensor, torch.Tensor]`: + A tuple containing the processed RGB and depth images as PyTorch tensors. + """ + supported_formats = (PIL.Image.Image, np.ndarray, torch.Tensor) + + # Expand the missing dimension for 3-dimensional pytorch tensor or numpy array that represents grayscale image + if self.config.do_convert_grayscale and isinstance(rgb, (torch.Tensor, np.ndarray)) and rgb.ndim == 3: + raise Exception("This is not yet supported") + + if isinstance(rgb, supported_formats): + rgb = [rgb] + depth = [depth] + elif not (isinstance(rgb, list) and all(isinstance(i, supported_formats) for i in rgb)): + raise ValueError( + f"Input is in incorrect format: {[type(i) for i in rgb]}. Currently, we only support {', '.join(supported_formats)}" + ) + + if isinstance(rgb[0], PIL.Image.Image): + if self.config.do_convert_rgb: + raise Exception("This is not yet supported") + # rgb = [self.convert_to_rgb(i) for i in rgb] + # depth = [self.convert_to_depth(i) for i in depth] #TODO define convert_to_depth + if self.config.do_resize or target_res: + height, width = self.get_default_height_width(rgb[0], height, width) if not target_res else target_res + rgb = [self.resize(i, height, width) for i in rgb] + depth = [self.resize(i, height, width) for i in depth] + rgb = self.pil_to_numpy(rgb) # to np + rgb = self.numpy_to_pt(rgb) # to pt + + depth = self.depth_pil_to_numpy(depth) # to np + depth = self.numpy_to_pt(depth) # to pt + + elif isinstance(rgb[0], np.ndarray): + rgb = np.concatenate(rgb, axis=0) if rgb[0].ndim == 4 else np.stack(rgb, axis=0) + rgb = self.numpy_to_pt(rgb) + height, width = self.get_default_height_width(rgb, height, width) + if self.config.do_resize: + rgb = self.resize(rgb, height, width) + + depth = np.concatenate(depth, axis=0) if rgb[0].ndim == 4 else np.stack(depth, axis=0) + depth = self.numpy_to_pt(depth) + height, width = self.get_default_height_width(depth, height, width) + if self.config.do_resize: + depth = self.resize(depth, height, width) + + elif isinstance(rgb[0], torch.Tensor): + raise Exception("This is not yet supported") + # rgb = torch.cat(rgb, axis=0) if rgb[0].ndim == 4 else torch.stack(rgb, axis=0) + + # if self.config.do_convert_grayscale and rgb.ndim == 3: + # rgb = rgb.unsqueeze(1) + + # channel = rgb.shape[1] + + # height, width = self.get_default_height_width(rgb, height, width) + # if self.config.do_resize: + # rgb = self.resize(rgb, height, width) + + # depth = torch.cat(depth, axis=0) if depth[0].ndim == 4 else torch.stack(depth, axis=0) + + # if self.config.do_convert_grayscale and depth.ndim == 3: + # depth = depth.unsqueeze(1) + + # channel = depth.shape[1] + # # don't need any preprocess if the image is latents + # if depth == 4: + # return rgb, depth + + # height, width = self.get_default_height_width(depth, height, width) + # if self.config.do_resize: + # depth = self.resize(depth, height, width) + # expected range [0,1], normalize to [-1,1] + do_normalize = self.config.do_normalize + if rgb.min() < 0 and do_normalize: + warnings.warn( + "Passing `image` as torch tensor with value range in [-1,1] is deprecated. The expected value range for image tensor is [0,1] " + f"when passing as pytorch tensor or numpy Array. You passed `image` with value range [{rgb.min()},{rgb.max()}]", + FutureWarning, + ) + do_normalize = False + + if do_normalize: + rgb = self.normalize(rgb) + depth = self.normalize(depth) + + if self.config.do_binarize: + rgb = self.binarize(rgb) + depth = self.binarize(depth) + + return rgb, depth + + +class IPAdapterMaskProcessor(VaeImageProcessor): + """ + Image processor for IP Adapter image masks. + + Args: + do_resize (`bool`, *optional*, defaults to `True`): + Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`. + vae_scale_factor (`int`, *optional*, defaults to `8`): + VAE scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of this factor. + resample (`str`, *optional*, defaults to `lanczos`): + Resampling filter to use when resizing the image. + do_normalize (`bool`, *optional*, defaults to `False`): + Whether to normalize the image to [-1,1]. + do_binarize (`bool`, *optional*, defaults to `True`): + Whether to binarize the image to 0/1. + do_convert_grayscale (`bool`, *optional*, defaults to be `True`): + Whether to convert the images to grayscale format. + + """ + + config_name = CONFIG_NAME + + @register_to_config + def __init__( + self, + do_resize: bool = True, + vae_scale_factor: int = 8, + resample: str = "lanczos", + do_normalize: bool = False, + do_binarize: bool = True, + do_convert_grayscale: bool = True, + ): + super().__init__( + do_resize=do_resize, + vae_scale_factor=vae_scale_factor, + resample=resample, + do_normalize=do_normalize, + do_binarize=do_binarize, + do_convert_grayscale=do_convert_grayscale, + ) + + @staticmethod + def downsample(mask: torch.Tensor, batch_size: int, num_queries: int, value_embed_dim: int): + """ + Downsamples the provided mask tensor to match the expected dimensions for scaled dot-product attention. If the + aspect ratio of the mask does not match the aspect ratio of the output image, a warning is issued. + + Args: + mask (`torch.Tensor`): + The input mask tensor generated with `IPAdapterMaskProcessor.preprocess()`. + batch_size (`int`): + The batch size. + num_queries (`int`): + The number of queries. + value_embed_dim (`int`): + The dimensionality of the value embeddings. + + Returns: + `torch.Tensor`: + The downsampled mask tensor. + + """ + o_h = mask.shape[1] + o_w = mask.shape[2] + ratio = o_w / o_h + mask_h = int(math.sqrt(num_queries / ratio)) + mask_h = int(mask_h) + int((num_queries % int(mask_h)) != 0) + mask_w = num_queries // mask_h + + mask_downsample = F.interpolate(mask.unsqueeze(0), size=(mask_h, mask_w), mode="bicubic").squeeze(0) + + # Repeat batch_size times + if mask_downsample.shape[0] < batch_size: + mask_downsample = mask_downsample.repeat(batch_size, 1, 1) + + mask_downsample = mask_downsample.view(mask_downsample.shape[0], -1) + + downsampled_area = mask_h * mask_w + # If the output image and the mask do not have the same aspect ratio, tensor shapes will not match + # Pad tensor if downsampled_mask.shape[1] is smaller than num_queries + if downsampled_area < num_queries: + warnings.warn( + "The aspect ratio of the mask does not match the aspect ratio of the output image. " + "Please update your masks or adjust the output size for optimal performance.", + UserWarning, + ) + mask_downsample = F.pad(mask_downsample, (0, num_queries - mask_downsample.shape[1]), value=0.0) + # Discard last embeddings if downsampled_mask.shape[1] is bigger than num_queries + if downsampled_area > num_queries: + warnings.warn( + "The aspect ratio of the mask does not match the aspect ratio of the output image. " + "Please update your masks or adjust the output size for optimal performance.", + UserWarning, + ) + mask_downsample = mask_downsample[:, :num_queries] + + # Repeat last dimension to match SDPA output shape + mask_downsample = mask_downsample.view(mask_downsample.shape[0], mask_downsample.shape[1], 1).repeat( + 1, 1, value_embed_dim + ) + + return mask_downsample + + +class PixArtImageProcessor(VaeImageProcessor): + """ + Image processor for PixArt image resize and crop. + + Args: + do_resize (`bool`, *optional*, defaults to `True`): + Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`. Can accept + `height` and `width` arguments from [`image_processor.VaeImageProcessor.preprocess`] method. + vae_scale_factor (`int`, *optional*, defaults to `8`): + VAE scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of this factor. + resample (`str`, *optional*, defaults to `lanczos`): + Resampling filter to use when resizing the image. + do_normalize (`bool`, *optional*, defaults to `True`): + Whether to normalize the image to [-1,1]. + do_binarize (`bool`, *optional*, defaults to `False`): + Whether to binarize the image to 0/1. + do_convert_rgb (`bool`, *optional*, defaults to be `False`): + Whether to convert the images to RGB format. + do_convert_grayscale (`bool`, *optional*, defaults to be `False`): + Whether to convert the images to grayscale format. + """ + + @register_to_config + def __init__( + self, + do_resize: bool = True, + vae_scale_factor: int = 8, + resample: str = "lanczos", + do_normalize: bool = True, + do_binarize: bool = False, + do_convert_grayscale: bool = False, + ): + super().__init__( + do_resize=do_resize, + vae_scale_factor=vae_scale_factor, + resample=resample, + do_normalize=do_normalize, + do_binarize=do_binarize, + do_convert_grayscale=do_convert_grayscale, + ) + + @staticmethod + def classify_height_width_bin(height: int, width: int, ratios: dict) -> Tuple[int, int]: + r""" + Returns the binned height and width based on the aspect ratio. + + Args: + height (`int`): The height of the image. + width (`int`): The width of the image. + ratios (`dict`): A dictionary where keys are aspect ratios and values are tuples of (height, width). + + Returns: + `Tuple[int, int]`: The closest binned height and width. + """ + ar = float(height / width) + closest_ratio = min(ratios.keys(), key=lambda ratio: abs(float(ratio) - ar)) + default_hw = ratios[closest_ratio] + return int(default_hw[0]), int(default_hw[1]) + + @staticmethod + def resize_and_crop_tensor(samples: torch.Tensor, new_width: int, new_height: int) -> torch.Tensor: + r""" + Resizes and crops a tensor of images to the specified dimensions. + + Args: + samples (`torch.Tensor`): + A tensor of shape (N, C, H, W) where N is the batch size, C is the number of channels, H is the height, + and W is the width. + new_width (`int`): The desired width of the output images. + new_height (`int`): The desired height of the output images. + + Returns: + `torch.Tensor`: A tensor containing the resized and cropped images. + """ + orig_height, orig_width = samples.shape[2], samples.shape[3] + + # Check if resizing is needed + if orig_height != new_height or orig_width != new_width: + ratio = max(new_height / orig_height, new_width / orig_width) + resized_width = int(orig_width * ratio) + resized_height = int(orig_height * ratio) + + # Resize + samples = F.interpolate( + samples, size=(resized_height, resized_width), mode="bilinear", align_corners=False + ) + + # Center Crop + start_x = (resized_width - new_width) // 2 + end_x = start_x + new_width + start_y = (resized_height - new_height) // 2 + end_y = start_y + new_height + samples = samples[:, :, start_y:end_y, start_x:end_x] + + return samples diff --git a/venv/lib/python3.10/site-packages/diffusers/optimization.py b/venv/lib/python3.10/site-packages/diffusers/optimization.py new file mode 100644 index 0000000000000000000000000000000000000000..e0b3576e44265d501e4a9ddcf3cfca64ae3b9bee --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/optimization.py @@ -0,0 +1,361 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# +# 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. +"""PyTorch optimization for diffusion models.""" + +import math +from enum import Enum +from typing import Optional, Union + +from torch.optim import Optimizer +from torch.optim.lr_scheduler import LambdaLR + +from .utils import logging + + +logger = logging.get_logger(__name__) + + +class SchedulerType(Enum): + LINEAR = "linear" + COSINE = "cosine" + COSINE_WITH_RESTARTS = "cosine_with_restarts" + POLYNOMIAL = "polynomial" + CONSTANT = "constant" + CONSTANT_WITH_WARMUP = "constant_with_warmup" + PIECEWISE_CONSTANT = "piecewise_constant" + + +def get_constant_schedule(optimizer: Optimizer, last_epoch: int = -1) -> LambdaLR: + """ + Create a schedule with a constant learning rate, using the learning rate set in optimizer. + + Args: + optimizer ([`~torch.optim.Optimizer`]): + The optimizer for which to schedule the learning rate. + last_epoch (`int`, *optional*, defaults to -1): + The index of the last epoch when resuming training. + + Return: + `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule. + """ + return LambdaLR(optimizer, lambda _: 1, last_epoch=last_epoch) + + +def get_constant_schedule_with_warmup(optimizer: Optimizer, num_warmup_steps: int, last_epoch: int = -1) -> LambdaLR: + """ + Create a schedule with a constant learning rate preceded by a warmup period during which the learning rate + increases linearly between 0 and the initial lr set in the optimizer. + + Args: + optimizer ([`~torch.optim.Optimizer`]): + The optimizer for which to schedule the learning rate. + num_warmup_steps (`int`): + The number of steps for the warmup phase. + last_epoch (`int`, *optional*, defaults to -1): + The index of the last epoch when resuming training. + + Return: + `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule. + """ + + def lr_lambda(current_step: int): + if current_step < num_warmup_steps: + return float(current_step) / float(max(1.0, num_warmup_steps)) + return 1.0 + + return LambdaLR(optimizer, lr_lambda, last_epoch=last_epoch) + + +def get_piecewise_constant_schedule(optimizer: Optimizer, step_rules: str, last_epoch: int = -1) -> LambdaLR: + """ + Create a schedule with a constant learning rate, using the learning rate set in optimizer. + + Args: + optimizer ([`~torch.optim.Optimizer`]): + The optimizer for which to schedule the learning rate. + step_rules (`string`): + The rules for the learning rate. ex: rule_steps="1:10,0.1:20,0.01:30,0.005" it means that the learning rate + if multiple 1 for the first 10 steps, multiple 0.1 for the next 20 steps, multiple 0.01 for the next 30 + steps and multiple 0.005 for the other steps. + last_epoch (`int`, *optional*, defaults to -1): + The index of the last epoch when resuming training. + + Return: + `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule. + """ + + rules_dict = {} + rule_list = step_rules.split(",") + for rule_str in rule_list[:-1]: + value_str, steps_str = rule_str.split(":") + steps = int(steps_str) + value = float(value_str) + rules_dict[steps] = value + last_lr_multiple = float(rule_list[-1]) + + def create_rules_function(rules_dict, last_lr_multiple): + def rule_func(steps: int) -> float: + sorted_steps = sorted(rules_dict.keys()) + for i, sorted_step in enumerate(sorted_steps): + if steps < sorted_step: + return rules_dict[sorted_steps[i]] + return last_lr_multiple + + return rule_func + + rules_func = create_rules_function(rules_dict, last_lr_multiple) + + return LambdaLR(optimizer, rules_func, last_epoch=last_epoch) + + +def get_linear_schedule_with_warmup( + optimizer: Optimizer, num_warmup_steps: int, num_training_steps: int, last_epoch: int = -1 +) -> LambdaLR: + """ + Create a schedule with a learning rate that decreases linearly from the initial lr set in the optimizer to 0, after + a warmup period during which it increases linearly from 0 to the initial lr set in the optimizer. + + Args: + optimizer ([`~torch.optim.Optimizer`]): + The optimizer for which to schedule the learning rate. + num_warmup_steps (`int`): + The number of steps for the warmup phase. + num_training_steps (`int`): + The total number of training steps. + last_epoch (`int`, *optional*, defaults to -1): + The index of the last epoch when resuming training. + + Return: + `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule. + """ + + def lr_lambda(current_step: int): + if current_step < num_warmup_steps: + return float(current_step) / float(max(1, num_warmup_steps)) + return max( + 0.0, float(num_training_steps - current_step) / float(max(1, num_training_steps - num_warmup_steps)) + ) + + return LambdaLR(optimizer, lr_lambda, last_epoch) + + +def get_cosine_schedule_with_warmup( + optimizer: Optimizer, num_warmup_steps: int, num_training_steps: int, num_cycles: float = 0.5, last_epoch: int = -1 +) -> LambdaLR: + """ + Create a schedule with a learning rate that decreases following the values of the cosine function between the + initial lr set in the optimizer to 0, after a warmup period during which it increases linearly between 0 and the + initial lr set in the optimizer. + + Args: + optimizer ([`~torch.optim.Optimizer`]): + The optimizer for which to schedule the learning rate. + num_warmup_steps (`int`): + The number of steps for the warmup phase. + num_training_steps (`int`): + The total number of training steps. + num_periods (`float`, *optional*, defaults to 0.5): + The number of periods of the cosine function in a schedule (the default is to just decrease from the max + value to 0 following a half-cosine). + last_epoch (`int`, *optional*, defaults to -1): + The index of the last epoch when resuming training. + + Return: + `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule. + """ + + def lr_lambda(current_step): + if current_step < num_warmup_steps: + return float(current_step) / float(max(1, num_warmup_steps)) + progress = float(current_step - num_warmup_steps) / float(max(1, num_training_steps - num_warmup_steps)) + return max(0.0, 0.5 * (1.0 + math.cos(math.pi * float(num_cycles) * 2.0 * progress))) + + return LambdaLR(optimizer, lr_lambda, last_epoch) + + +def get_cosine_with_hard_restarts_schedule_with_warmup( + optimizer: Optimizer, num_warmup_steps: int, num_training_steps: int, num_cycles: int = 1, last_epoch: int = -1 +) -> LambdaLR: + """ + Create a schedule with a learning rate that decreases following the values of the cosine function between the + initial lr set in the optimizer to 0, with several hard restarts, after a warmup period during which it increases + linearly between 0 and the initial lr set in the optimizer. + + Args: + optimizer ([`~torch.optim.Optimizer`]): + The optimizer for which to schedule the learning rate. + num_warmup_steps (`int`): + The number of steps for the warmup phase. + num_training_steps (`int`): + The total number of training steps. + num_cycles (`int`, *optional*, defaults to 1): + The number of hard restarts to use. + last_epoch (`int`, *optional*, defaults to -1): + The index of the last epoch when resuming training. + + Return: + `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule. + """ + + def lr_lambda(current_step): + if current_step < num_warmup_steps: + return float(current_step) / float(max(1, num_warmup_steps)) + progress = float(current_step - num_warmup_steps) / float(max(1, num_training_steps - num_warmup_steps)) + if progress >= 1.0: + return 0.0 + return max(0.0, 0.5 * (1.0 + math.cos(math.pi * ((float(num_cycles) * progress) % 1.0)))) + + return LambdaLR(optimizer, lr_lambda, last_epoch) + + +def get_polynomial_decay_schedule_with_warmup( + optimizer: Optimizer, + num_warmup_steps: int, + num_training_steps: int, + lr_end: float = 1e-7, + power: float = 1.0, + last_epoch: int = -1, +) -> LambdaLR: + """ + Create a schedule with a learning rate that decreases as a polynomial decay from the initial lr set in the + optimizer to end lr defined by *lr_end*, after a warmup period during which it increases linearly from 0 to the + initial lr set in the optimizer. + + Args: + optimizer ([`~torch.optim.Optimizer`]): + The optimizer for which to schedule the learning rate. + num_warmup_steps (`int`): + The number of steps for the warmup phase. + num_training_steps (`int`): + The total number of training steps. + lr_end (`float`, *optional*, defaults to 1e-7): + The end LR. + power (`float`, *optional*, defaults to 1.0): + Power factor. + last_epoch (`int`, *optional*, defaults to -1): + The index of the last epoch when resuming training. + + Note: *power* defaults to 1.0 as in the fairseq implementation, which in turn is based on the original BERT + implementation at + https://github.com/google-research/bert/blob/f39e881b169b9d53bea03d2d341b31707a6c052b/optimization.py#L37 + + Return: + `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule. + + """ + + lr_init = optimizer.defaults["lr"] + if not (lr_init > lr_end): + raise ValueError(f"lr_end ({lr_end}) must be smaller than initial lr ({lr_init})") + + def lr_lambda(current_step: int): + if current_step < num_warmup_steps: + return float(current_step) / float(max(1, num_warmup_steps)) + elif current_step > num_training_steps: + return lr_end / lr_init # as LambdaLR multiplies by lr_init + else: + lr_range = lr_init - lr_end + decay_steps = num_training_steps - num_warmup_steps + pct_remaining = 1 - (current_step - num_warmup_steps) / decay_steps + decay = lr_range * pct_remaining**power + lr_end + return decay / lr_init # as LambdaLR multiplies by lr_init + + return LambdaLR(optimizer, lr_lambda, last_epoch) + + +TYPE_TO_SCHEDULER_FUNCTION = { + SchedulerType.LINEAR: get_linear_schedule_with_warmup, + SchedulerType.COSINE: get_cosine_schedule_with_warmup, + SchedulerType.COSINE_WITH_RESTARTS: get_cosine_with_hard_restarts_schedule_with_warmup, + SchedulerType.POLYNOMIAL: get_polynomial_decay_schedule_with_warmup, + SchedulerType.CONSTANT: get_constant_schedule, + SchedulerType.CONSTANT_WITH_WARMUP: get_constant_schedule_with_warmup, + SchedulerType.PIECEWISE_CONSTANT: get_piecewise_constant_schedule, +} + + +def get_scheduler( + name: Union[str, SchedulerType], + optimizer: Optimizer, + step_rules: Optional[str] = None, + num_warmup_steps: Optional[int] = None, + num_training_steps: Optional[int] = None, + num_cycles: int = 1, + power: float = 1.0, + last_epoch: int = -1, +) -> LambdaLR: + """ + Unified API to get any scheduler from its name. + + Args: + name (`str` or `SchedulerType`): + The name of the scheduler to use. + optimizer (`torch.optim.Optimizer`): + The optimizer that will be used during training. + step_rules (`str`, *optional*): + A string representing the step rules to use. This is only used by the `PIECEWISE_CONSTANT` scheduler. + num_warmup_steps (`int`, *optional*): + The number of warmup steps to do. This is not required by all schedulers (hence the argument being + optional), the function will raise an error if it's unset and the scheduler type requires it. + num_training_steps (`int``, *optional*): + The number of training steps to do. This is not required by all schedulers (hence the argument being + optional), the function will raise an error if it's unset and the scheduler type requires it. + num_cycles (`int`, *optional*): + The number of hard restarts used in `COSINE_WITH_RESTARTS` scheduler. + power (`float`, *optional*, defaults to 1.0): + Power factor. See `POLYNOMIAL` scheduler + last_epoch (`int`, *optional*, defaults to -1): + The index of the last epoch when resuming training. + """ + name = SchedulerType(name) + schedule_func = TYPE_TO_SCHEDULER_FUNCTION[name] + if name == SchedulerType.CONSTANT: + return schedule_func(optimizer, last_epoch=last_epoch) + + if name == SchedulerType.PIECEWISE_CONSTANT: + return schedule_func(optimizer, step_rules=step_rules, last_epoch=last_epoch) + + # All other schedulers require `num_warmup_steps` + if num_warmup_steps is None: + raise ValueError(f"{name} requires `num_warmup_steps`, please provide that argument.") + + if name == SchedulerType.CONSTANT_WITH_WARMUP: + return schedule_func(optimizer, num_warmup_steps=num_warmup_steps, last_epoch=last_epoch) + + # All other schedulers require `num_training_steps` + if num_training_steps is None: + raise ValueError(f"{name} requires `num_training_steps`, please provide that argument.") + + if name == SchedulerType.COSINE_WITH_RESTARTS: + return schedule_func( + optimizer, + num_warmup_steps=num_warmup_steps, + num_training_steps=num_training_steps, + num_cycles=num_cycles, + last_epoch=last_epoch, + ) + + if name == SchedulerType.POLYNOMIAL: + return schedule_func( + optimizer, + num_warmup_steps=num_warmup_steps, + num_training_steps=num_training_steps, + power=power, + last_epoch=last_epoch, + ) + + return schedule_func( + optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps, last_epoch=last_epoch + ) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b32d55bd51715ccf4c7a809eb924b577a2072ac9 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/__init__.py @@ -0,0 +1,860 @@ +from typing import TYPE_CHECKING + +from ..utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_flax_available, + is_k_diffusion_available, + is_librosa_available, + is_note_seq_available, + is_onnx_available, + is_opencv_available, + is_sentencepiece_available, + is_torch_available, + is_torch_npu_available, + is_transformers_available, +) + + +# These modules contain pipelines from multiple libraries/frameworks +_dummy_objects = {} +_import_structure = { + "controlnet": [], + "controlnet_hunyuandit": [], + "controlnet_sd3": [], + "controlnet_xs": [], + "deprecated": [], + "latent_diffusion": [], + "ledits_pp": [], + "marigold": [], + "pag": [], + "stable_diffusion": [], + "stable_diffusion_xl": [], +} + +try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_pt_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_pt_objects)) +else: + _import_structure["auto_pipeline"] = [ + "AutoPipelineForImage2Image", + "AutoPipelineForInpainting", + "AutoPipelineForText2Image", + ] + _import_structure["consistency_models"] = ["ConsistencyModelPipeline"] + _import_structure["dance_diffusion"] = ["DanceDiffusionPipeline"] + _import_structure["ddim"] = ["DDIMPipeline"] + _import_structure["ddpm"] = ["DDPMPipeline"] + _import_structure["dit"] = ["DiTPipeline"] + _import_structure["latent_diffusion"].extend(["LDMSuperResolutionPipeline"]) + _import_structure["pipeline_utils"] = [ + "AudioPipelineOutput", + "DiffusionPipeline", + "StableDiffusionMixin", + "ImagePipelineOutput", + ] + _import_structure["deprecated"].extend( + [ + "PNDMPipeline", + "LDMPipeline", + "RePaintPipeline", + "ScoreSdeVePipeline", + "KarrasVePipeline", + ] + ) +try: + if not (is_torch_available() and is_librosa_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_torch_and_librosa_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_librosa_objects)) +else: + _import_structure["deprecated"].extend(["AudioDiffusionPipeline", "Mel"]) + +try: + if not (is_transformers_available() and is_torch_available() and is_note_seq_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_transformers_and_torch_and_note_seq_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_transformers_and_torch_and_note_seq_objects)) +else: + _import_structure["deprecated"].extend( + [ + "MidiProcessor", + "SpectrogramDiffusionPipeline", + ] + ) + +try: + if not (is_torch_available() and is_transformers_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["deprecated"].extend( + [ + "VQDiffusionPipeline", + "AltDiffusionPipeline", + "AltDiffusionImg2ImgPipeline", + "CycleDiffusionPipeline", + "StableDiffusionInpaintPipelineLegacy", + "StableDiffusionPix2PixZeroPipeline", + "StableDiffusionParadigmsPipeline", + "StableDiffusionModelEditingPipeline", + "VersatileDiffusionDualGuidedPipeline", + "VersatileDiffusionImageVariationPipeline", + "VersatileDiffusionPipeline", + "VersatileDiffusionTextToImagePipeline", + ] + ) + _import_structure["allegro"] = ["AllegroPipeline"] + _import_structure["amused"] = ["AmusedImg2ImgPipeline", "AmusedInpaintPipeline", "AmusedPipeline"] + _import_structure["animatediff"] = [ + "AnimateDiffPipeline", + "AnimateDiffControlNetPipeline", + "AnimateDiffSDXLPipeline", + "AnimateDiffSparseControlNetPipeline", + "AnimateDiffVideoToVideoPipeline", + "AnimateDiffVideoToVideoControlNetPipeline", + ] + _import_structure["flux"] = [ + "FluxControlPipeline", + "FluxControlInpaintPipeline", + "FluxControlImg2ImgPipeline", + "FluxControlNetPipeline", + "FluxControlNetImg2ImgPipeline", + "FluxControlNetInpaintPipeline", + "FluxImg2ImgPipeline", + "FluxInpaintPipeline", + "FluxPipeline", + "FluxFillPipeline", + "FluxPriorReduxPipeline", + "ReduxImageEncoder", + ] + _import_structure["audioldm"] = ["AudioLDMPipeline"] + _import_structure["audioldm2"] = [ + "AudioLDM2Pipeline", + "AudioLDM2ProjectionModel", + "AudioLDM2UNet2DConditionModel", + ] + _import_structure["blip_diffusion"] = ["BlipDiffusionPipeline"] + _import_structure["chroma"] = ["ChromaPipeline", "ChromaImg2ImgPipeline"] + _import_structure["cogvideo"] = [ + "CogVideoXPipeline", + "CogVideoXImageToVideoPipeline", + "CogVideoXVideoToVideoPipeline", + "CogVideoXFunControlPipeline", + ] + _import_structure["cogview3"] = ["CogView3PlusPipeline"] + _import_structure["cogview4"] = ["CogView4Pipeline", "CogView4ControlPipeline"] + _import_structure["consisid"] = ["ConsisIDPipeline"] + _import_structure["cosmos"] = [ + "Cosmos2TextToImagePipeline", + "CosmosTextToWorldPipeline", + "CosmosVideoToWorldPipeline", + "Cosmos2VideoToWorldPipeline", + ] + _import_structure["controlnet"].extend( + [ + "BlipDiffusionControlNetPipeline", + "StableDiffusionControlNetImg2ImgPipeline", + "StableDiffusionControlNetInpaintPipeline", + "StableDiffusionControlNetPipeline", + "StableDiffusionXLControlNetImg2ImgPipeline", + "StableDiffusionXLControlNetInpaintPipeline", + "StableDiffusionXLControlNetPipeline", + "StableDiffusionXLControlNetUnionPipeline", + "StableDiffusionXLControlNetUnionInpaintPipeline", + "StableDiffusionXLControlNetUnionImg2ImgPipeline", + ] + ) + _import_structure["pag"].extend( + [ + "StableDiffusionControlNetPAGInpaintPipeline", + "AnimateDiffPAGPipeline", + "KolorsPAGPipeline", + "HunyuanDiTPAGPipeline", + "StableDiffusion3PAGPipeline", + "StableDiffusion3PAGImg2ImgPipeline", + "StableDiffusionPAGPipeline", + "StableDiffusionPAGImg2ImgPipeline", + "StableDiffusionPAGInpaintPipeline", + "StableDiffusionControlNetPAGPipeline", + "StableDiffusionXLPAGPipeline", + "StableDiffusionXLPAGInpaintPipeline", + "StableDiffusionXLControlNetPAGImg2ImgPipeline", + "StableDiffusionXLControlNetPAGPipeline", + "StableDiffusionXLPAGImg2ImgPipeline", + "PixArtSigmaPAGPipeline", + "SanaPAGPipeline", + ] + ) + _import_structure["controlnet_xs"].extend( + [ + "StableDiffusionControlNetXSPipeline", + "StableDiffusionXLControlNetXSPipeline", + ] + ) + _import_structure["controlnet_hunyuandit"].extend( + [ + "HunyuanDiTControlNetPipeline", + ] + ) + _import_structure["controlnet_sd3"].extend( + [ + "StableDiffusion3ControlNetPipeline", + "StableDiffusion3ControlNetInpaintingPipeline", + ] + ) + _import_structure["deepfloyd_if"] = [ + "IFImg2ImgPipeline", + "IFImg2ImgSuperResolutionPipeline", + "IFInpaintingPipeline", + "IFInpaintingSuperResolutionPipeline", + "IFPipeline", + "IFSuperResolutionPipeline", + ] + _import_structure["easyanimate"] = [ + "EasyAnimatePipeline", + "EasyAnimateInpaintPipeline", + "EasyAnimateControlPipeline", + ] + _import_structure["hidream_image"] = ["HiDreamImagePipeline"] + _import_structure["hunyuandit"] = ["HunyuanDiTPipeline"] + _import_structure["hunyuan_video"] = [ + "HunyuanVideoPipeline", + "HunyuanSkyreelsImageToVideoPipeline", + "HunyuanVideoImageToVideoPipeline", + "HunyuanVideoFramepackPipeline", + ] + _import_structure["kandinsky"] = [ + "KandinskyCombinedPipeline", + "KandinskyImg2ImgCombinedPipeline", + "KandinskyImg2ImgPipeline", + "KandinskyInpaintCombinedPipeline", + "KandinskyInpaintPipeline", + "KandinskyPipeline", + "KandinskyPriorPipeline", + ] + _import_structure["kandinsky2_2"] = [ + "KandinskyV22CombinedPipeline", + "KandinskyV22ControlnetImg2ImgPipeline", + "KandinskyV22ControlnetPipeline", + "KandinskyV22Img2ImgCombinedPipeline", + "KandinskyV22Img2ImgPipeline", + "KandinskyV22InpaintCombinedPipeline", + "KandinskyV22InpaintPipeline", + "KandinskyV22Pipeline", + "KandinskyV22PriorEmb2EmbPipeline", + "KandinskyV22PriorPipeline", + ] + _import_structure["kandinsky3"] = [ + "Kandinsky3Img2ImgPipeline", + "Kandinsky3Pipeline", + ] + _import_structure["latent_consistency_models"] = [ + "LatentConsistencyModelImg2ImgPipeline", + "LatentConsistencyModelPipeline", + ] + _import_structure["latent_diffusion"].extend(["LDMTextToImagePipeline"]) + _import_structure["ledits_pp"].extend( + [ + "LEditsPPPipelineStableDiffusion", + "LEditsPPPipelineStableDiffusionXL", + ] + ) + _import_structure["latte"] = ["LattePipeline"] + _import_structure["ltx"] = [ + "LTXPipeline", + "LTXImageToVideoPipeline", + "LTXConditionPipeline", + "LTXLatentUpsamplePipeline", + ] + _import_structure["lumina"] = ["LuminaPipeline", "LuminaText2ImgPipeline"] + _import_structure["lumina2"] = ["Lumina2Pipeline", "Lumina2Text2ImgPipeline"] + _import_structure["marigold"].extend( + [ + "MarigoldDepthPipeline", + "MarigoldIntrinsicsPipeline", + "MarigoldNormalsPipeline", + ] + ) + _import_structure["mochi"] = ["MochiPipeline"] + _import_structure["musicldm"] = ["MusicLDMPipeline"] + _import_structure["omnigen"] = ["OmniGenPipeline"] + _import_structure["visualcloze"] = ["VisualClozePipeline", "VisualClozeGenerationPipeline"] + _import_structure["paint_by_example"] = ["PaintByExamplePipeline"] + _import_structure["pia"] = ["PIAPipeline"] + _import_structure["pixart_alpha"] = ["PixArtAlphaPipeline", "PixArtSigmaPipeline"] + _import_structure["sana"] = [ + "SanaPipeline", + "SanaSprintPipeline", + "SanaControlNetPipeline", + "SanaSprintImg2ImgPipeline", + ] + _import_structure["semantic_stable_diffusion"] = ["SemanticStableDiffusionPipeline"] + _import_structure["shap_e"] = ["ShapEImg2ImgPipeline", "ShapEPipeline"] + _import_structure["stable_audio"] = [ + "StableAudioProjectionModel", + "StableAudioPipeline", + ] + _import_structure["stable_cascade"] = [ + "StableCascadeCombinedPipeline", + "StableCascadeDecoderPipeline", + "StableCascadePriorPipeline", + ] + _import_structure["stable_diffusion"].extend( + [ + "CLIPImageProjection", + "StableDiffusionDepth2ImgPipeline", + "StableDiffusionImageVariationPipeline", + "StableDiffusionImg2ImgPipeline", + "StableDiffusionInpaintPipeline", + "StableDiffusionInstructPix2PixPipeline", + "StableDiffusionLatentUpscalePipeline", + "StableDiffusionPipeline", + "StableDiffusionUpscalePipeline", + "StableUnCLIPImg2ImgPipeline", + "StableUnCLIPPipeline", + "StableDiffusionLDM3DPipeline", + ] + ) + _import_structure["aura_flow"] = ["AuraFlowPipeline"] + _import_structure["stable_diffusion_3"] = [ + "StableDiffusion3Pipeline", + "StableDiffusion3Img2ImgPipeline", + "StableDiffusion3InpaintPipeline", + ] + _import_structure["stable_diffusion_attend_and_excite"] = ["StableDiffusionAttendAndExcitePipeline"] + _import_structure["stable_diffusion_safe"] = ["StableDiffusionPipelineSafe"] + _import_structure["stable_diffusion_sag"] = ["StableDiffusionSAGPipeline"] + _import_structure["stable_diffusion_gligen"] = [ + "StableDiffusionGLIGENPipeline", + "StableDiffusionGLIGENTextImagePipeline", + ] + _import_structure["stable_video_diffusion"] = ["StableVideoDiffusionPipeline"] + _import_structure["stable_diffusion_xl"].extend( + [ + "StableDiffusionXLImg2ImgPipeline", + "StableDiffusionXLInpaintPipeline", + "StableDiffusionXLInstructPix2PixPipeline", + "StableDiffusionXLPipeline", + ] + ) + _import_structure["stable_diffusion_diffedit"] = ["StableDiffusionDiffEditPipeline"] + _import_structure["stable_diffusion_ldm3d"] = ["StableDiffusionLDM3DPipeline"] + _import_structure["stable_diffusion_panorama"] = ["StableDiffusionPanoramaPipeline"] + _import_structure["t2i_adapter"] = [ + "StableDiffusionAdapterPipeline", + "StableDiffusionXLAdapterPipeline", + ] + _import_structure["text_to_video_synthesis"] = [ + "TextToVideoSDPipeline", + "TextToVideoZeroPipeline", + "TextToVideoZeroSDXLPipeline", + "VideoToVideoSDPipeline", + ] + _import_structure["i2vgen_xl"] = ["I2VGenXLPipeline"] + _import_structure["unclip"] = ["UnCLIPImageVariationPipeline", "UnCLIPPipeline"] + _import_structure["unidiffuser"] = [ + "ImageTextPipelineOutput", + "UniDiffuserModel", + "UniDiffuserPipeline", + "UniDiffuserTextDecoder", + ] + _import_structure["wuerstchen"] = [ + "WuerstchenCombinedPipeline", + "WuerstchenDecoderPipeline", + "WuerstchenPriorPipeline", + ] + _import_structure["wan"] = ["WanPipeline", "WanImageToVideoPipeline", "WanVideoToVideoPipeline", "WanVACEPipeline"] +try: + if not is_onnx_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_onnx_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_onnx_objects)) +else: + _import_structure["onnx_utils"] = ["OnnxRuntimeModel"] +try: + if not (is_torch_available() and is_transformers_available() and is_onnx_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_torch_and_transformers_and_onnx_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_onnx_objects)) +else: + _import_structure["stable_diffusion"].extend( + [ + "OnnxStableDiffusionImg2ImgPipeline", + "OnnxStableDiffusionInpaintPipeline", + "OnnxStableDiffusionPipeline", + "OnnxStableDiffusionUpscalePipeline", + "StableDiffusionOnnxPipeline", + ] + ) + +try: + if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import ( + dummy_torch_and_transformers_and_k_diffusion_objects, + ) + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_k_diffusion_objects)) +else: + _import_structure["stable_diffusion_k_diffusion"] = [ + "StableDiffusionKDiffusionPipeline", + "StableDiffusionXLKDiffusionPipeline", + ] + +try: + if not (is_torch_available() and is_transformers_available() and is_sentencepiece_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import ( + dummy_torch_and_transformers_and_sentencepiece_objects, + ) + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_sentencepiece_objects)) +else: + _import_structure["kolors"] = [ + "KolorsPipeline", + "KolorsImg2ImgPipeline", + ] + +try: + if not (is_torch_available() and is_transformers_available() and is_opencv_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import ( + dummy_torch_and_transformers_and_opencv_objects, + ) + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_opencv_objects)) +else: + _import_structure["consisid"] = ["ConsisIDPipeline"] + +try: + if not is_flax_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_flax_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_flax_objects)) +else: + _import_structure["pipeline_flax_utils"] = ["FlaxDiffusionPipeline"] +try: + if not (is_flax_available() and is_transformers_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_flax_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_flax_and_transformers_objects)) +else: + _import_structure["controlnet"].extend(["FlaxStableDiffusionControlNetPipeline"]) + _import_structure["stable_diffusion"].extend( + [ + "FlaxStableDiffusionImg2ImgPipeline", + "FlaxStableDiffusionInpaintPipeline", + "FlaxStableDiffusionPipeline", + ] + ) + _import_structure["stable_diffusion_xl"].extend( + [ + "FlaxStableDiffusionXLPipeline", + ] + ) + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_pt_objects import * # noqa F403 + + else: + from .auto_pipeline import ( + AutoPipelineForImage2Image, + AutoPipelineForInpainting, + AutoPipelineForText2Image, + ) + from .consistency_models import ConsistencyModelPipeline + from .dance_diffusion import DanceDiffusionPipeline + from .ddim import DDIMPipeline + from .ddpm import DDPMPipeline + from .deprecated import KarrasVePipeline, LDMPipeline, PNDMPipeline, RePaintPipeline, ScoreSdeVePipeline + from .dit import DiTPipeline + from .latent_diffusion import LDMSuperResolutionPipeline + from .pipeline_utils import ( + AudioPipelineOutput, + DiffusionPipeline, + ImagePipelineOutput, + StableDiffusionMixin, + ) + + try: + if not (is_torch_available() and is_librosa_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_torch_and_librosa_objects import * + else: + from .deprecated import AudioDiffusionPipeline, Mel + + try: + if not (is_torch_available() and is_transformers_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_torch_and_transformers_objects import * + else: + from .allegro import AllegroPipeline + from .amused import AmusedImg2ImgPipeline, AmusedInpaintPipeline, AmusedPipeline + from .animatediff import ( + AnimateDiffControlNetPipeline, + AnimateDiffPipeline, + AnimateDiffSDXLPipeline, + AnimateDiffSparseControlNetPipeline, + AnimateDiffVideoToVideoControlNetPipeline, + AnimateDiffVideoToVideoPipeline, + ) + from .audioldm import AudioLDMPipeline + from .audioldm2 import ( + AudioLDM2Pipeline, + AudioLDM2ProjectionModel, + AudioLDM2UNet2DConditionModel, + ) + from .aura_flow import AuraFlowPipeline + from .blip_diffusion import BlipDiffusionPipeline + from .chroma import ChromaImg2ImgPipeline, ChromaPipeline + from .cogvideo import ( + CogVideoXFunControlPipeline, + CogVideoXImageToVideoPipeline, + CogVideoXPipeline, + CogVideoXVideoToVideoPipeline, + ) + from .cogview3 import CogView3PlusPipeline + from .cogview4 import CogView4ControlPipeline, CogView4Pipeline + from .controlnet import ( + BlipDiffusionControlNetPipeline, + StableDiffusionControlNetImg2ImgPipeline, + StableDiffusionControlNetInpaintPipeline, + StableDiffusionControlNetPipeline, + StableDiffusionXLControlNetImg2ImgPipeline, + StableDiffusionXLControlNetInpaintPipeline, + StableDiffusionXLControlNetPipeline, + StableDiffusionXLControlNetUnionImg2ImgPipeline, + StableDiffusionXLControlNetUnionInpaintPipeline, + StableDiffusionXLControlNetUnionPipeline, + ) + from .controlnet_hunyuandit import ( + HunyuanDiTControlNetPipeline, + ) + from .controlnet_sd3 import StableDiffusion3ControlNetInpaintingPipeline, StableDiffusion3ControlNetPipeline + from .controlnet_xs import ( + StableDiffusionControlNetXSPipeline, + StableDiffusionXLControlNetXSPipeline, + ) + from .cosmos import ( + Cosmos2TextToImagePipeline, + Cosmos2VideoToWorldPipeline, + CosmosTextToWorldPipeline, + CosmosVideoToWorldPipeline, + ) + from .deepfloyd_if import ( + IFImg2ImgPipeline, + IFImg2ImgSuperResolutionPipeline, + IFInpaintingPipeline, + IFInpaintingSuperResolutionPipeline, + IFPipeline, + IFSuperResolutionPipeline, + ) + from .deprecated import ( + AltDiffusionImg2ImgPipeline, + AltDiffusionPipeline, + CycleDiffusionPipeline, + StableDiffusionInpaintPipelineLegacy, + StableDiffusionModelEditingPipeline, + StableDiffusionParadigmsPipeline, + StableDiffusionPix2PixZeroPipeline, + VersatileDiffusionDualGuidedPipeline, + VersatileDiffusionImageVariationPipeline, + VersatileDiffusionPipeline, + VersatileDiffusionTextToImagePipeline, + VQDiffusionPipeline, + ) + from .easyanimate import ( + EasyAnimateControlPipeline, + EasyAnimateInpaintPipeline, + EasyAnimatePipeline, + ) + from .flux import ( + FluxControlImg2ImgPipeline, + FluxControlInpaintPipeline, + FluxControlNetImg2ImgPipeline, + FluxControlNetInpaintPipeline, + FluxControlNetPipeline, + FluxControlPipeline, + FluxFillPipeline, + FluxImg2ImgPipeline, + FluxInpaintPipeline, + FluxPipeline, + FluxPriorReduxPipeline, + ReduxImageEncoder, + ) + from .hidream_image import HiDreamImagePipeline + from .hunyuan_video import ( + HunyuanSkyreelsImageToVideoPipeline, + HunyuanVideoFramepackPipeline, + HunyuanVideoImageToVideoPipeline, + HunyuanVideoPipeline, + ) + from .hunyuandit import HunyuanDiTPipeline + from .i2vgen_xl import I2VGenXLPipeline + from .kandinsky import ( + KandinskyCombinedPipeline, + KandinskyImg2ImgCombinedPipeline, + KandinskyImg2ImgPipeline, + KandinskyInpaintCombinedPipeline, + KandinskyInpaintPipeline, + KandinskyPipeline, + KandinskyPriorPipeline, + ) + from .kandinsky2_2 import ( + KandinskyV22CombinedPipeline, + KandinskyV22ControlnetImg2ImgPipeline, + KandinskyV22ControlnetPipeline, + KandinskyV22Img2ImgCombinedPipeline, + KandinskyV22Img2ImgPipeline, + KandinskyV22InpaintCombinedPipeline, + KandinskyV22InpaintPipeline, + KandinskyV22Pipeline, + KandinskyV22PriorEmb2EmbPipeline, + KandinskyV22PriorPipeline, + ) + from .kandinsky3 import ( + Kandinsky3Img2ImgPipeline, + Kandinsky3Pipeline, + ) + from .latent_consistency_models import ( + LatentConsistencyModelImg2ImgPipeline, + LatentConsistencyModelPipeline, + ) + from .latent_diffusion import LDMTextToImagePipeline + from .latte import LattePipeline + from .ledits_pp import ( + LEditsPPDiffusionPipelineOutput, + LEditsPPInversionPipelineOutput, + LEditsPPPipelineStableDiffusion, + LEditsPPPipelineStableDiffusionXL, + ) + from .ltx import LTXConditionPipeline, LTXImageToVideoPipeline, LTXLatentUpsamplePipeline, LTXPipeline + from .lumina import LuminaPipeline, LuminaText2ImgPipeline + from .lumina2 import Lumina2Pipeline, Lumina2Text2ImgPipeline + from .marigold import ( + MarigoldDepthPipeline, + MarigoldIntrinsicsPipeline, + MarigoldNormalsPipeline, + ) + from .mochi import MochiPipeline + from .musicldm import MusicLDMPipeline + from .omnigen import OmniGenPipeline + from .pag import ( + AnimateDiffPAGPipeline, + HunyuanDiTPAGPipeline, + KolorsPAGPipeline, + PixArtSigmaPAGPipeline, + SanaPAGPipeline, + StableDiffusion3PAGImg2ImgPipeline, + StableDiffusion3PAGPipeline, + StableDiffusionControlNetPAGInpaintPipeline, + StableDiffusionControlNetPAGPipeline, + StableDiffusionPAGImg2ImgPipeline, + StableDiffusionPAGInpaintPipeline, + StableDiffusionPAGPipeline, + StableDiffusionXLControlNetPAGImg2ImgPipeline, + StableDiffusionXLControlNetPAGPipeline, + StableDiffusionXLPAGImg2ImgPipeline, + StableDiffusionXLPAGInpaintPipeline, + StableDiffusionXLPAGPipeline, + ) + from .paint_by_example import PaintByExamplePipeline + from .pia import PIAPipeline + from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline + from .sana import SanaControlNetPipeline, SanaPipeline, SanaSprintImg2ImgPipeline, SanaSprintPipeline + from .semantic_stable_diffusion import SemanticStableDiffusionPipeline + from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline + from .stable_audio import StableAudioPipeline, StableAudioProjectionModel + from .stable_cascade import ( + StableCascadeCombinedPipeline, + StableCascadeDecoderPipeline, + StableCascadePriorPipeline, + ) + from .stable_diffusion import ( + CLIPImageProjection, + StableDiffusionDepth2ImgPipeline, + StableDiffusionImageVariationPipeline, + StableDiffusionImg2ImgPipeline, + StableDiffusionInpaintPipeline, + StableDiffusionInstructPix2PixPipeline, + StableDiffusionLatentUpscalePipeline, + StableDiffusionPipeline, + StableDiffusionUpscalePipeline, + StableUnCLIPImg2ImgPipeline, + StableUnCLIPPipeline, + ) + from .stable_diffusion_3 import ( + StableDiffusion3Img2ImgPipeline, + StableDiffusion3InpaintPipeline, + StableDiffusion3Pipeline, + ) + from .stable_diffusion_attend_and_excite import StableDiffusionAttendAndExcitePipeline + from .stable_diffusion_diffedit import StableDiffusionDiffEditPipeline + from .stable_diffusion_gligen import StableDiffusionGLIGENPipeline, StableDiffusionGLIGENTextImagePipeline + from .stable_diffusion_ldm3d import StableDiffusionLDM3DPipeline + from .stable_diffusion_panorama import StableDiffusionPanoramaPipeline + from .stable_diffusion_safe import StableDiffusionPipelineSafe + from .stable_diffusion_sag import StableDiffusionSAGPipeline + from .stable_diffusion_xl import ( + StableDiffusionXLImg2ImgPipeline, + StableDiffusionXLInpaintPipeline, + StableDiffusionXLInstructPix2PixPipeline, + StableDiffusionXLPipeline, + ) + from .stable_video_diffusion import StableVideoDiffusionPipeline + from .t2i_adapter import ( + StableDiffusionAdapterPipeline, + StableDiffusionXLAdapterPipeline, + ) + from .text_to_video_synthesis import ( + TextToVideoSDPipeline, + TextToVideoZeroPipeline, + TextToVideoZeroSDXLPipeline, + VideoToVideoSDPipeline, + ) + from .unclip import UnCLIPImageVariationPipeline, UnCLIPPipeline + from .unidiffuser import ( + ImageTextPipelineOutput, + UniDiffuserModel, + UniDiffuserPipeline, + UniDiffuserTextDecoder, + ) + from .visualcloze import VisualClozeGenerationPipeline, VisualClozePipeline + from .wan import WanImageToVideoPipeline, WanPipeline, WanVACEPipeline, WanVideoToVideoPipeline + from .wuerstchen import ( + WuerstchenCombinedPipeline, + WuerstchenDecoderPipeline, + WuerstchenPriorPipeline, + ) + + try: + if not is_onnx_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_onnx_objects import * # noqa F403 + + else: + from .onnx_utils import OnnxRuntimeModel + + try: + if not (is_torch_available() and is_transformers_available() and is_onnx_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_torch_and_transformers_and_onnx_objects import * + else: + from .stable_diffusion import ( + OnnxStableDiffusionImg2ImgPipeline, + OnnxStableDiffusionInpaintPipeline, + OnnxStableDiffusionPipeline, + OnnxStableDiffusionUpscalePipeline, + StableDiffusionOnnxPipeline, + ) + + try: + if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_torch_and_transformers_and_k_diffusion_objects import * + else: + from .stable_diffusion_k_diffusion import ( + StableDiffusionKDiffusionPipeline, + StableDiffusionXLKDiffusionPipeline, + ) + + try: + if not (is_torch_available() and is_transformers_available() and is_sentencepiece_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_torch_and_transformers_and_sentencepiece_objects import * + else: + from .kolors import ( + KolorsImg2ImgPipeline, + KolorsPipeline, + ) + + try: + if not (is_torch_available() and is_transformers_available() and is_opencv_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_torch_and_transformers_and_opencv_objects import * + else: + from .consisid import ConsisIDPipeline + + try: + if not is_flax_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_flax_objects import * # noqa F403 + else: + from .pipeline_flax_utils import FlaxDiffusionPipeline + + try: + if not (is_flax_available() and is_transformers_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_flax_and_transformers_objects import * + else: + from .controlnet import FlaxStableDiffusionControlNetPipeline + from .stable_diffusion import ( + FlaxStableDiffusionImg2ImgPipeline, + FlaxStableDiffusionInpaintPipeline, + FlaxStableDiffusionPipeline, + ) + from .stable_diffusion_xl import ( + FlaxStableDiffusionXLPipeline, + ) + + try: + if not (is_transformers_available() and is_torch_available() and is_note_seq_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_transformers_and_torch_and_note_seq_objects import * # noqa F403 + + else: + from .deprecated import ( + MidiProcessor, + SpectrogramDiffusionPipeline, + ) + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..351156a139d2cca9c2caa4f937e26a13ed944fb0 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/auto_pipeline.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/auto_pipeline.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f08629bc99b97036e9bae10aca8cd4640349c464 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/auto_pipeline.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/free_init_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/free_init_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1101fcf19bc341a23bd049aaf21acff8c7c87ff0 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/free_init_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/free_noise_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/free_noise_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cec013d9a9968c164c4acb870549e579e7f67563 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/free_noise_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/onnx_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/onnx_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bd3ad7cbbb19086658a9a761d4d6a7f0df8876a1 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/onnx_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/pipeline_flax_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/pipeline_flax_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2c99a4c42369b49a44010d6ac59667e5eda8b7c4 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/pipeline_flax_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/pipeline_loading_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/pipeline_loading_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9e1664e1cbf582331f2b372ba717f4da1ef8e871 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/pipeline_loading_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/pipeline_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/pipeline_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..30859d5d33518ee6d66df95c367488d55fa93c62 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/pipeline_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/transformers_loading_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/transformers_loading_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..81f308a00604033b2b7357e674a64bfb1cd80219 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/__pycache__/transformers_loading_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/auto_pipeline.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/auto_pipeline.py new file mode 100644 index 0000000000000000000000000000000000000000..b1a7ffaaea9cef1268e9e793bc702ef2aad68022 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/auto_pipeline.py @@ -0,0 +1,1175 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# +# 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 collections import OrderedDict + +from huggingface_hub.utils import validate_hf_hub_args + +from ..configuration_utils import ConfigMixin +from ..models.controlnets import ControlNetUnionModel +from ..utils import is_sentencepiece_available +from .aura_flow import AuraFlowPipeline +from .chroma import ChromaPipeline +from .cogview3 import CogView3PlusPipeline +from .cogview4 import CogView4ControlPipeline, CogView4Pipeline +from .controlnet import ( + StableDiffusionControlNetImg2ImgPipeline, + StableDiffusionControlNetInpaintPipeline, + StableDiffusionControlNetPipeline, + StableDiffusionXLControlNetImg2ImgPipeline, + StableDiffusionXLControlNetInpaintPipeline, + StableDiffusionXLControlNetPipeline, + StableDiffusionXLControlNetUnionImg2ImgPipeline, + StableDiffusionXLControlNetUnionInpaintPipeline, + StableDiffusionXLControlNetUnionPipeline, +) +from .controlnet_sd3 import ( + StableDiffusion3ControlNetInpaintingPipeline, + StableDiffusion3ControlNetPipeline, +) +from .deepfloyd_if import IFImg2ImgPipeline, IFInpaintingPipeline, IFPipeline +from .flux import ( + FluxControlImg2ImgPipeline, + FluxControlInpaintPipeline, + FluxControlNetImg2ImgPipeline, + FluxControlNetInpaintPipeline, + FluxControlNetPipeline, + FluxControlPipeline, + FluxImg2ImgPipeline, + FluxInpaintPipeline, + FluxPipeline, +) +from .hunyuandit import HunyuanDiTPipeline +from .kandinsky import ( + KandinskyCombinedPipeline, + KandinskyImg2ImgCombinedPipeline, + KandinskyImg2ImgPipeline, + KandinskyInpaintCombinedPipeline, + KandinskyInpaintPipeline, + KandinskyPipeline, +) +from .kandinsky2_2 import ( + KandinskyV22CombinedPipeline, + KandinskyV22Img2ImgCombinedPipeline, + KandinskyV22Img2ImgPipeline, + KandinskyV22InpaintCombinedPipeline, + KandinskyV22InpaintPipeline, + KandinskyV22Pipeline, +) +from .kandinsky3 import Kandinsky3Img2ImgPipeline, Kandinsky3Pipeline +from .latent_consistency_models import LatentConsistencyModelImg2ImgPipeline, LatentConsistencyModelPipeline +from .lumina import LuminaPipeline +from .lumina2 import Lumina2Pipeline +from .pag import ( + HunyuanDiTPAGPipeline, + PixArtSigmaPAGPipeline, + SanaPAGPipeline, + StableDiffusion3PAGImg2ImgPipeline, + StableDiffusion3PAGPipeline, + StableDiffusionControlNetPAGInpaintPipeline, + StableDiffusionControlNetPAGPipeline, + StableDiffusionPAGImg2ImgPipeline, + StableDiffusionPAGInpaintPipeline, + StableDiffusionPAGPipeline, + StableDiffusionXLControlNetPAGImg2ImgPipeline, + StableDiffusionXLControlNetPAGPipeline, + StableDiffusionXLPAGImg2ImgPipeline, + StableDiffusionXLPAGInpaintPipeline, + StableDiffusionXLPAGPipeline, +) +from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline +from .sana import SanaPipeline +from .stable_cascade import StableCascadeCombinedPipeline, StableCascadeDecoderPipeline +from .stable_diffusion import ( + StableDiffusionImg2ImgPipeline, + StableDiffusionInpaintPipeline, + StableDiffusionPipeline, +) +from .stable_diffusion_3 import ( + StableDiffusion3Img2ImgPipeline, + StableDiffusion3InpaintPipeline, + StableDiffusion3Pipeline, +) +from .stable_diffusion_xl import ( + StableDiffusionXLImg2ImgPipeline, + StableDiffusionXLInpaintPipeline, + StableDiffusionXLPipeline, +) +from .wuerstchen import WuerstchenCombinedPipeline, WuerstchenDecoderPipeline + + +AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict( + [ + ("stable-diffusion", StableDiffusionPipeline), + ("stable-diffusion-xl", StableDiffusionXLPipeline), + ("stable-diffusion-3", StableDiffusion3Pipeline), + ("stable-diffusion-3-pag", StableDiffusion3PAGPipeline), + ("if", IFPipeline), + ("hunyuan", HunyuanDiTPipeline), + ("hunyuan-pag", HunyuanDiTPAGPipeline), + ("kandinsky", KandinskyCombinedPipeline), + ("kandinsky22", KandinskyV22CombinedPipeline), + ("kandinsky3", Kandinsky3Pipeline), + ("stable-diffusion-controlnet", StableDiffusionControlNetPipeline), + ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetPipeline), + ("stable-diffusion-xl-controlnet-union", StableDiffusionXLControlNetUnionPipeline), + ("stable-diffusion-3-controlnet", StableDiffusion3ControlNetPipeline), + ("wuerstchen", WuerstchenCombinedPipeline), + ("cascade", StableCascadeCombinedPipeline), + ("lcm", LatentConsistencyModelPipeline), + ("pixart-alpha", PixArtAlphaPipeline), + ("pixart-sigma", PixArtSigmaPipeline), + ("sana", SanaPipeline), + ("sana-pag", SanaPAGPipeline), + ("stable-diffusion-pag", StableDiffusionPAGPipeline), + ("stable-diffusion-controlnet-pag", StableDiffusionControlNetPAGPipeline), + ("stable-diffusion-xl-pag", StableDiffusionXLPAGPipeline), + ("stable-diffusion-xl-controlnet-pag", StableDiffusionXLControlNetPAGPipeline), + ("pixart-sigma-pag", PixArtSigmaPAGPipeline), + ("auraflow", AuraFlowPipeline), + ("flux", FluxPipeline), + ("flux-control", FluxControlPipeline), + ("flux-controlnet", FluxControlNetPipeline), + ("lumina", LuminaPipeline), + ("lumina2", Lumina2Pipeline), + ("chroma", ChromaPipeline), + ("cogview3", CogView3PlusPipeline), + ("cogview4", CogView4Pipeline), + ("cogview4-control", CogView4ControlPipeline), + ] +) + +AUTO_IMAGE2IMAGE_PIPELINES_MAPPING = OrderedDict( + [ + ("stable-diffusion", StableDiffusionImg2ImgPipeline), + ("stable-diffusion-xl", StableDiffusionXLImg2ImgPipeline), + ("stable-diffusion-3", StableDiffusion3Img2ImgPipeline), + ("stable-diffusion-3-pag", StableDiffusion3PAGImg2ImgPipeline), + ("if", IFImg2ImgPipeline), + ("kandinsky", KandinskyImg2ImgCombinedPipeline), + ("kandinsky22", KandinskyV22Img2ImgCombinedPipeline), + ("kandinsky3", Kandinsky3Img2ImgPipeline), + ("stable-diffusion-controlnet", StableDiffusionControlNetImg2ImgPipeline), + ("stable-diffusion-pag", StableDiffusionPAGImg2ImgPipeline), + ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetImg2ImgPipeline), + ("stable-diffusion-xl-controlnet-union", StableDiffusionXLControlNetUnionImg2ImgPipeline), + ("stable-diffusion-xl-pag", StableDiffusionXLPAGImg2ImgPipeline), + ("stable-diffusion-xl-controlnet-pag", StableDiffusionXLControlNetPAGImg2ImgPipeline), + ("lcm", LatentConsistencyModelImg2ImgPipeline), + ("flux", FluxImg2ImgPipeline), + ("flux-controlnet", FluxControlNetImg2ImgPipeline), + ("flux-control", FluxControlImg2ImgPipeline), + ] +) + +AUTO_INPAINT_PIPELINES_MAPPING = OrderedDict( + [ + ("stable-diffusion", StableDiffusionInpaintPipeline), + ("stable-diffusion-xl", StableDiffusionXLInpaintPipeline), + ("stable-diffusion-3", StableDiffusion3InpaintPipeline), + ("if", IFInpaintingPipeline), + ("kandinsky", KandinskyInpaintCombinedPipeline), + ("kandinsky22", KandinskyV22InpaintCombinedPipeline), + ("stable-diffusion-controlnet", StableDiffusionControlNetInpaintPipeline), + ("stable-diffusion-controlnet-pag", StableDiffusionControlNetPAGInpaintPipeline), + ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetInpaintPipeline), + ("stable-diffusion-xl-controlnet-union", StableDiffusionXLControlNetUnionInpaintPipeline), + ("stable-diffusion-3-controlnet", StableDiffusion3ControlNetInpaintingPipeline), + ("stable-diffusion-xl-pag", StableDiffusionXLPAGInpaintPipeline), + ("flux", FluxInpaintPipeline), + ("flux-controlnet", FluxControlNetInpaintPipeline), + ("flux-control", FluxControlInpaintPipeline), + ("stable-diffusion-pag", StableDiffusionPAGInpaintPipeline), + ] +) + +_AUTO_TEXT2IMAGE_DECODER_PIPELINES_MAPPING = OrderedDict( + [ + ("kandinsky", KandinskyPipeline), + ("kandinsky22", KandinskyV22Pipeline), + ("wuerstchen", WuerstchenDecoderPipeline), + ("cascade", StableCascadeDecoderPipeline), + ] +) +_AUTO_IMAGE2IMAGE_DECODER_PIPELINES_MAPPING = OrderedDict( + [ + ("kandinsky", KandinskyImg2ImgPipeline), + ("kandinsky22", KandinskyV22Img2ImgPipeline), + ] +) +_AUTO_INPAINT_DECODER_PIPELINES_MAPPING = OrderedDict( + [ + ("kandinsky", KandinskyInpaintPipeline), + ("kandinsky22", KandinskyV22InpaintPipeline), + ] +) + +if is_sentencepiece_available(): + from .kolors import KolorsImg2ImgPipeline, KolorsPipeline + from .pag import KolorsPAGPipeline + + AUTO_TEXT2IMAGE_PIPELINES_MAPPING["kolors"] = KolorsPipeline + AUTO_TEXT2IMAGE_PIPELINES_MAPPING["kolors-pag"] = KolorsPAGPipeline + AUTO_IMAGE2IMAGE_PIPELINES_MAPPING["kolors"] = KolorsImg2ImgPipeline + +SUPPORTED_TASKS_MAPPINGS = [ + AUTO_TEXT2IMAGE_PIPELINES_MAPPING, + AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, + AUTO_INPAINT_PIPELINES_MAPPING, + _AUTO_TEXT2IMAGE_DECODER_PIPELINES_MAPPING, + _AUTO_IMAGE2IMAGE_DECODER_PIPELINES_MAPPING, + _AUTO_INPAINT_DECODER_PIPELINES_MAPPING, +] + + +def _get_connected_pipeline(pipeline_cls): + # for now connected pipelines can only be loaded from decoder pipelines, such as kandinsky-community/kandinsky-2-2-decoder + if pipeline_cls in _AUTO_TEXT2IMAGE_DECODER_PIPELINES_MAPPING.values(): + return _get_task_class( + AUTO_TEXT2IMAGE_PIPELINES_MAPPING, pipeline_cls.__name__, throw_error_if_not_exist=False + ) + if pipeline_cls in _AUTO_IMAGE2IMAGE_DECODER_PIPELINES_MAPPING.values(): + return _get_task_class( + AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, pipeline_cls.__name__, throw_error_if_not_exist=False + ) + if pipeline_cls in _AUTO_INPAINT_DECODER_PIPELINES_MAPPING.values(): + return _get_task_class(AUTO_INPAINT_PIPELINES_MAPPING, pipeline_cls.__name__, throw_error_if_not_exist=False) + + +def _get_task_class(mapping, pipeline_class_name, throw_error_if_not_exist: bool = True): + def get_model(pipeline_class_name): + for task_mapping in SUPPORTED_TASKS_MAPPINGS: + for model_name, pipeline in task_mapping.items(): + if pipeline.__name__ == pipeline_class_name: + return model_name + + model_name = get_model(pipeline_class_name) + + if model_name is not None: + task_class = mapping.get(model_name, None) + if task_class is not None: + return task_class + + if throw_error_if_not_exist: + raise ValueError(f"AutoPipeline can't find a pipeline linked to {pipeline_class_name} for {model_name}") + + +class AutoPipelineForText2Image(ConfigMixin): + r""" + + [`AutoPipelineForText2Image`] is a generic pipeline class that instantiates a text-to-image pipeline class. The + specific underlying pipeline class is automatically selected from either the + [`~AutoPipelineForText2Image.from_pretrained`] or [`~AutoPipelineForText2Image.from_pipe`] methods. + + This class cannot be instantiated using `__init__()` (throws an error). + + Class attributes: + + - **config_name** (`str`) -- The configuration filename that stores the class and module names of all the + diffusion pipeline's components. + + """ + + config_name = "model_index.json" + + def __init__(self, *args, **kwargs): + raise EnvironmentError( + f"{self.__class__.__name__} is designed to be instantiated " + f"using the `{self.__class__.__name__}.from_pretrained(pretrained_model_name_or_path)` or " + f"`{self.__class__.__name__}.from_pipe(pipeline)` methods." + ) + + @classmethod + @validate_hf_hub_args + def from_pretrained(cls, pretrained_model_or_path, **kwargs): + r""" + Instantiates a text-to-image Pytorch diffusion pipeline from pretrained pipeline weight. + + The from_pretrained() method takes care of returning the correct pipeline class instance by: + 1. Detect the pipeline class of the pretrained_model_or_path based on the _class_name property of its + config object + 2. Find the text-to-image pipeline linked to the pipeline class using pattern matching on pipeline class + name. + + If a `controlnet` argument is passed, it will instantiate a [`StableDiffusionControlNetPipeline`] object. + + The pipeline is set in evaluation mode (`model.eval()`) by default. + + If you get the error message below, you need to finetune the weights for your downstream task: + + ``` + Some weights of UNet2DConditionModel were not initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5 and are newly initialized because the shapes did not match: + - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated + You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference. + ``` + + Parameters: + pretrained_model_or_path (`str` or `os.PathLike`, *optional*): + Can be either: + + - A string, the *repo id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline + hosted on the Hub. + - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights + saved using + [`~DiffusionPipeline.save_pretrained`]. + torch_dtype (`torch.dtype`, *optional*): + Override the default `torch.dtype` and load the model with another dtype. + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + cache_dir (`Union[str, os.PathLike]`, *optional*): + Path to a directory where a downloaded pretrained model configuration is cached if the standard cache + is not used. + + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. + output_loading_info(`bool`, *optional*, defaults to `False`): + Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages. + local_files_only (`bool`, *optional*, defaults to `False`): + Whether to only load local model weights and configuration files or not. If set to `True`, the model + won't be downloaded from the Hub. + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from + `diffusers-cli login` (stored in `~/.huggingface`) is used. + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier + allowed by Git. + custom_revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, or a commit id similar to + `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a + custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub. + mirror (`str`, *optional*): + Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not + guarantee the timeliness or safety of the source, and you should refer to the mirror site for more + information. + device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*): + A map that specifies where each submodule should go. It doesn’t need to be defined for each + parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the + same device. + + Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For + more information about each option see [designing a device + map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map). + max_memory (`Dict`, *optional*): + A dictionary device identifier for the maximum memory. Will default to the maximum memory available for + each GPU and the available CPU RAM if unset. + offload_folder (`str` or `os.PathLike`, *optional*): + The path to offload weights if device_map contains the value `"disk"`. + offload_state_dict (`bool`, *optional*): + If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if + the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True` + when there is some disk offload. + low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`): + Speed up model loading only loading the pretrained weights and not initializing the weights. This also + tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model. + Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this + argument to `True` will raise an error. + use_safetensors (`bool`, *optional*, defaults to `None`): + If set to `None`, the safetensors weights are downloaded if they're available **and** if the + safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors + weights. If set to `False`, safetensors weights are not loaded. + kwargs (remaining dictionary of keyword arguments, *optional*): + Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline + class). The overwritten components are passed directly to the pipelines `__init__` method. See example + below for more information. + variant (`str`, *optional*): + Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when + loading `from_flax`. + + + + To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with + `huggingface-cli login`. + + + + Examples: + + ```py + >>> from diffusers import AutoPipelineForText2Image + + >>> pipeline = AutoPipelineForText2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5") + >>> image = pipeline(prompt).images[0] + ``` + """ + cache_dir = kwargs.pop("cache_dir", None) + force_download = kwargs.pop("force_download", False) + proxies = kwargs.pop("proxies", None) + token = kwargs.pop("token", None) + local_files_only = kwargs.pop("local_files_only", False) + revision = kwargs.pop("revision", None) + + load_config_kwargs = { + "cache_dir": cache_dir, + "force_download": force_download, + "proxies": proxies, + "token": token, + "local_files_only": local_files_only, + "revision": revision, + } + + config = cls.load_config(pretrained_model_or_path, **load_config_kwargs) + orig_class_name = config["_class_name"] + if "ControlPipeline" in orig_class_name: + to_replace = "ControlPipeline" + else: + to_replace = "Pipeline" + + if "controlnet" in kwargs: + if isinstance(kwargs["controlnet"], ControlNetUnionModel): + orig_class_name = config["_class_name"].replace(to_replace, "ControlNetUnionPipeline") + else: + orig_class_name = config["_class_name"].replace(to_replace, "ControlNetPipeline") + if "enable_pag" in kwargs: + enable_pag = kwargs.pop("enable_pag") + if enable_pag: + orig_class_name = orig_class_name.replace(to_replace, "PAGPipeline") + + text_2_image_cls = _get_task_class(AUTO_TEXT2IMAGE_PIPELINES_MAPPING, orig_class_name) + + kwargs = {**load_config_kwargs, **kwargs} + return text_2_image_cls.from_pretrained(pretrained_model_or_path, **kwargs) + + @classmethod + def from_pipe(cls, pipeline, **kwargs): + r""" + Instantiates a text-to-image Pytorch diffusion pipeline from another instantiated diffusion pipeline class. + + The from_pipe() method takes care of returning the correct pipeline class instance by finding the text-to-image + pipeline linked to the pipeline class using pattern matching on pipeline class name. + + All the modules the pipeline contains will be used to initialize the new pipeline without reallocating + additional memory. + + The pipeline is set in evaluation mode (`model.eval()`) by default. + + Parameters: + pipeline (`DiffusionPipeline`): + an instantiated `DiffusionPipeline` object + + ```py + >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image + + >>> pipe_i2i = AutoPipelineForImage2Image.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", requires_safety_checker=False + ... ) + + >>> pipe_t2i = AutoPipelineForText2Image.from_pipe(pipe_i2i) + >>> image = pipe_t2i(prompt).images[0] + ``` + """ + + original_config = dict(pipeline.config) + original_cls_name = pipeline.__class__.__name__ + + # derive the pipeline class to instantiate + text_2_image_cls = _get_task_class(AUTO_TEXT2IMAGE_PIPELINES_MAPPING, original_cls_name) + + if "controlnet" in kwargs: + if kwargs["controlnet"] is not None: + to_replace = "PAGPipeline" if "PAG" in text_2_image_cls.__name__ else "Pipeline" + text_2_image_cls = _get_task_class( + AUTO_TEXT2IMAGE_PIPELINES_MAPPING, + text_2_image_cls.__name__.replace("ControlNet", "").replace(to_replace, "ControlNet" + to_replace), + ) + else: + text_2_image_cls = _get_task_class( + AUTO_TEXT2IMAGE_PIPELINES_MAPPING, + text_2_image_cls.__name__.replace("ControlNet", ""), + ) + + if "enable_pag" in kwargs: + enable_pag = kwargs.pop("enable_pag") + if enable_pag: + text_2_image_cls = _get_task_class( + AUTO_TEXT2IMAGE_PIPELINES_MAPPING, + text_2_image_cls.__name__.replace("PAG", "").replace("Pipeline", "PAGPipeline"), + ) + else: + text_2_image_cls = _get_task_class( + AUTO_TEXT2IMAGE_PIPELINES_MAPPING, + text_2_image_cls.__name__.replace("PAG", ""), + ) + + # define expected module and optional kwargs given the pipeline signature + expected_modules, optional_kwargs = text_2_image_cls._get_signature_keys(text_2_image_cls) + + pretrained_model_name_or_path = original_config.pop("_name_or_path", None) + + # allow users pass modules in `kwargs` to override the original pipeline's components + passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs} + original_class_obj = { + k: pipeline.components[k] + for k, v in pipeline.components.items() + if k in expected_modules and k not in passed_class_obj + } + + # allow users pass optional kwargs to override the original pipelines config attribute + passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs} + original_pipe_kwargs = { + k: original_config[k] + for k, v in original_config.items() + if k in optional_kwargs and k not in passed_pipe_kwargs + } + + # config that were not expected by original pipeline is stored as private attribute + # we will pass them as optional arguments if they can be accepted by the pipeline + additional_pipe_kwargs = [ + k[1:] + for k in original_config.keys() + if k.startswith("_") and k[1:] in optional_kwargs and k[1:] not in passed_pipe_kwargs + ] + for k in additional_pipe_kwargs: + original_pipe_kwargs[k] = original_config.pop(f"_{k}") + + text_2_image_kwargs = {**passed_class_obj, **original_class_obj, **passed_pipe_kwargs, **original_pipe_kwargs} + + # store unused config as private attribute + unused_original_config = { + f"{'' if k.startswith('_') else '_'}{k}": original_config[k] + for k, v in original_config.items() + if k not in text_2_image_kwargs + } + + missing_modules = ( + set(expected_modules) - set(text_2_image_cls._optional_components) - set(text_2_image_kwargs.keys()) + ) + + if len(missing_modules) > 0: + raise ValueError( + f"Pipeline {text_2_image_cls} expected {expected_modules}, but only {set(list(passed_class_obj.keys()) + list(original_class_obj.keys()))} were passed" + ) + + model = text_2_image_cls(**text_2_image_kwargs) + model.register_to_config(_name_or_path=pretrained_model_name_or_path) + model.register_to_config(**unused_original_config) + + return model + + +class AutoPipelineForImage2Image(ConfigMixin): + r""" + + [`AutoPipelineForImage2Image`] is a generic pipeline class that instantiates an image-to-image pipeline class. The + specific underlying pipeline class is automatically selected from either the + [`~AutoPipelineForImage2Image.from_pretrained`] or [`~AutoPipelineForImage2Image.from_pipe`] methods. + + This class cannot be instantiated using `__init__()` (throws an error). + + Class attributes: + + - **config_name** (`str`) -- The configuration filename that stores the class and module names of all the + diffusion pipeline's components. + + """ + + config_name = "model_index.json" + + def __init__(self, *args, **kwargs): + raise EnvironmentError( + f"{self.__class__.__name__} is designed to be instantiated " + f"using the `{self.__class__.__name__}.from_pretrained(pretrained_model_name_or_path)` or " + f"`{self.__class__.__name__}.from_pipe(pipeline)` methods." + ) + + @classmethod + @validate_hf_hub_args + def from_pretrained(cls, pretrained_model_or_path, **kwargs): + r""" + Instantiates a image-to-image Pytorch diffusion pipeline from pretrained pipeline weight. + + The from_pretrained() method takes care of returning the correct pipeline class instance by: + 1. Detect the pipeline class of the pretrained_model_or_path based on the _class_name property of its + config object + 2. Find the image-to-image pipeline linked to the pipeline class using pattern matching on pipeline class + name. + + If a `controlnet` argument is passed, it will instantiate a [`StableDiffusionControlNetImg2ImgPipeline`] + object. + + The pipeline is set in evaluation mode (`model.eval()`) by default. + + If you get the error message below, you need to finetune the weights for your downstream task: + + ``` + Some weights of UNet2DConditionModel were not initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5 and are newly initialized because the shapes did not match: + - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated + You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference. + ``` + + Parameters: + pretrained_model_or_path (`str` or `os.PathLike`, *optional*): + Can be either: + + - A string, the *repo id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline + hosted on the Hub. + - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights + saved using + [`~DiffusionPipeline.save_pretrained`]. + torch_dtype (`str` or `torch.dtype`, *optional*): + Override the default `torch.dtype` and load the model with another dtype. + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + cache_dir (`Union[str, os.PathLike]`, *optional*): + Path to a directory where a downloaded pretrained model configuration is cached if the standard cache + is not used. + + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. + output_loading_info(`bool`, *optional*, defaults to `False`): + Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages. + local_files_only (`bool`, *optional*, defaults to `False`): + Whether to only load local model weights and configuration files or not. If set to `True`, the model + won't be downloaded from the Hub. + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from + `diffusers-cli login` (stored in `~/.huggingface`) is used. + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier + allowed by Git. + custom_revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, or a commit id similar to + `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a + custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub. + mirror (`str`, *optional*): + Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not + guarantee the timeliness or safety of the source, and you should refer to the mirror site for more + information. + device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*): + A map that specifies where each submodule should go. It doesn’t need to be defined for each + parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the + same device. + + Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For + more information about each option see [designing a device + map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map). + max_memory (`Dict`, *optional*): + A dictionary device identifier for the maximum memory. Will default to the maximum memory available for + each GPU and the available CPU RAM if unset. + offload_folder (`str` or `os.PathLike`, *optional*): + The path to offload weights if device_map contains the value `"disk"`. + offload_state_dict (`bool`, *optional*): + If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if + the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True` + when there is some disk offload. + low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`): + Speed up model loading only loading the pretrained weights and not initializing the weights. This also + tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model. + Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this + argument to `True` will raise an error. + use_safetensors (`bool`, *optional*, defaults to `None`): + If set to `None`, the safetensors weights are downloaded if they're available **and** if the + safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors + weights. If set to `False`, safetensors weights are not loaded. + kwargs (remaining dictionary of keyword arguments, *optional*): + Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline + class). The overwritten components are passed directly to the pipelines `__init__` method. See example + below for more information. + variant (`str`, *optional*): + Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when + loading `from_flax`. + + + + To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with + `huggingface-cli login`. + + + + Examples: + + ```py + >>> from diffusers import AutoPipelineForImage2Image + + >>> pipeline = AutoPipelineForImage2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5") + >>> image = pipeline(prompt, image).images[0] + ``` + """ + cache_dir = kwargs.pop("cache_dir", None) + force_download = kwargs.pop("force_download", False) + proxies = kwargs.pop("proxies", None) + token = kwargs.pop("token", None) + local_files_only = kwargs.pop("local_files_only", False) + revision = kwargs.pop("revision", None) + + load_config_kwargs = { + "cache_dir": cache_dir, + "force_download": force_download, + "proxies": proxies, + "token": token, + "local_files_only": local_files_only, + "revision": revision, + } + + config = cls.load_config(pretrained_model_or_path, **load_config_kwargs) + orig_class_name = config["_class_name"] + + # the `orig_class_name` can be: + # `- *Pipeline` (for regular text-to-image checkpoint) + # - `*ControlPipeline` (for Flux tools specific checkpoint) + # `- *Img2ImgPipeline` (for refiner checkpoint) + if "Img2Img" in orig_class_name: + to_replace = "Img2ImgPipeline" + elif "ControlPipeline" in orig_class_name: + to_replace = "ControlPipeline" + else: + to_replace = "Pipeline" + + if "controlnet" in kwargs: + if isinstance(kwargs["controlnet"], ControlNetUnionModel): + orig_class_name = orig_class_name.replace(to_replace, "ControlNetUnion" + to_replace) + else: + orig_class_name = orig_class_name.replace(to_replace, "ControlNet" + to_replace) + if "enable_pag" in kwargs: + enable_pag = kwargs.pop("enable_pag") + if enable_pag: + orig_class_name = orig_class_name.replace(to_replace, "PAG" + to_replace) + + if to_replace == "ControlPipeline": + orig_class_name = orig_class_name.replace(to_replace, "ControlImg2ImgPipeline") + + image_2_image_cls = _get_task_class(AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, orig_class_name) + + kwargs = {**load_config_kwargs, **kwargs} + return image_2_image_cls.from_pretrained(pretrained_model_or_path, **kwargs) + + @classmethod + def from_pipe(cls, pipeline, **kwargs): + r""" + Instantiates a image-to-image Pytorch diffusion pipeline from another instantiated diffusion pipeline class. + + The from_pipe() method takes care of returning the correct pipeline class instance by finding the + image-to-image pipeline linked to the pipeline class using pattern matching on pipeline class name. + + All the modules the pipeline contains will be used to initialize the new pipeline without reallocating + additional memory. + + The pipeline is set in evaluation mode (`model.eval()`) by default. + + Parameters: + pipeline (`DiffusionPipeline`): + an instantiated `DiffusionPipeline` object + + Examples: + + ```py + >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image + + >>> pipe_t2i = AutoPipelineForText2Image.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", requires_safety_checker=False + ... ) + + >>> pipe_i2i = AutoPipelineForImage2Image.from_pipe(pipe_t2i) + >>> image = pipe_i2i(prompt, image).images[0] + ``` + """ + + original_config = dict(pipeline.config) + original_cls_name = pipeline.__class__.__name__ + + # derive the pipeline class to instantiate + image_2_image_cls = _get_task_class(AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, original_cls_name) + + if "controlnet" in kwargs: + if kwargs["controlnet"] is not None: + to_replace = "Img2ImgPipeline" + if "PAG" in image_2_image_cls.__name__: + to_replace = "PAG" + to_replace + image_2_image_cls = _get_task_class( + AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, + image_2_image_cls.__name__.replace("ControlNet", "").replace( + to_replace, "ControlNet" + to_replace + ), + ) + else: + image_2_image_cls = _get_task_class( + AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, + image_2_image_cls.__name__.replace("ControlNet", ""), + ) + + if "enable_pag" in kwargs: + enable_pag = kwargs.pop("enable_pag") + if enable_pag: + image_2_image_cls = _get_task_class( + AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, + image_2_image_cls.__name__.replace("PAG", "").replace("Img2ImgPipeline", "PAGImg2ImgPipeline"), + ) + else: + image_2_image_cls = _get_task_class( + AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, + image_2_image_cls.__name__.replace("PAG", ""), + ) + + # define expected module and optional kwargs given the pipeline signature + expected_modules, optional_kwargs = image_2_image_cls._get_signature_keys(image_2_image_cls) + + pretrained_model_name_or_path = original_config.pop("_name_or_path", None) + + # allow users pass modules in `kwargs` to override the original pipeline's components + passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs} + original_class_obj = { + k: pipeline.components[k] + for k, v in pipeline.components.items() + if k in expected_modules and k not in passed_class_obj + } + + # allow users pass optional kwargs to override the original pipelines config attribute + passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs} + original_pipe_kwargs = { + k: original_config[k] + for k, v in original_config.items() + if k in optional_kwargs and k not in passed_pipe_kwargs + } + + # config attribute that were not expected by original pipeline is stored as its private attribute + # we will pass them as optional arguments if they can be accepted by the pipeline + additional_pipe_kwargs = [ + k[1:] + for k in original_config.keys() + if k.startswith("_") and k[1:] in optional_kwargs and k[1:] not in passed_pipe_kwargs + ] + for k in additional_pipe_kwargs: + original_pipe_kwargs[k] = original_config.pop(f"_{k}") + + image_2_image_kwargs = {**passed_class_obj, **original_class_obj, **passed_pipe_kwargs, **original_pipe_kwargs} + + # store unused config as private attribute + unused_original_config = { + f"{'' if k.startswith('_') else '_'}{k}": original_config[k] + for k, v in original_config.items() + if k not in image_2_image_kwargs + } + + missing_modules = ( + set(expected_modules) - set(image_2_image_cls._optional_components) - set(image_2_image_kwargs.keys()) + ) + + if len(missing_modules) > 0: + raise ValueError( + f"Pipeline {image_2_image_cls} expected {expected_modules}, but only {set(list(passed_class_obj.keys()) + list(original_class_obj.keys()))} were passed" + ) + + model = image_2_image_cls(**image_2_image_kwargs) + model.register_to_config(_name_or_path=pretrained_model_name_or_path) + model.register_to_config(**unused_original_config) + + return model + + +class AutoPipelineForInpainting(ConfigMixin): + r""" + + [`AutoPipelineForInpainting`] is a generic pipeline class that instantiates an inpainting pipeline class. The + specific underlying pipeline class is automatically selected from either the + [`~AutoPipelineForInpainting.from_pretrained`] or [`~AutoPipelineForInpainting.from_pipe`] methods. + + This class cannot be instantiated using `__init__()` (throws an error). + + Class attributes: + + - **config_name** (`str`) -- The configuration filename that stores the class and module names of all the + diffusion pipeline's components. + + """ + + config_name = "model_index.json" + + def __init__(self, *args, **kwargs): + raise EnvironmentError( + f"{self.__class__.__name__} is designed to be instantiated " + f"using the `{self.__class__.__name__}.from_pretrained(pretrained_model_name_or_path)` or " + f"`{self.__class__.__name__}.from_pipe(pipeline)` methods." + ) + + @classmethod + @validate_hf_hub_args + def from_pretrained(cls, pretrained_model_or_path, **kwargs): + r""" + Instantiates a inpainting Pytorch diffusion pipeline from pretrained pipeline weight. + + The from_pretrained() method takes care of returning the correct pipeline class instance by: + 1. Detect the pipeline class of the pretrained_model_or_path based on the _class_name property of its + config object + 2. Find the inpainting pipeline linked to the pipeline class using pattern matching on pipeline class name. + + If a `controlnet` argument is passed, it will instantiate a [`StableDiffusionControlNetInpaintPipeline`] + object. + + The pipeline is set in evaluation mode (`model.eval()`) by default. + + If you get the error message below, you need to finetune the weights for your downstream task: + + ``` + Some weights of UNet2DConditionModel were not initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5 and are newly initialized because the shapes did not match: + - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated + You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference. + ``` + + Parameters: + pretrained_model_or_path (`str` or `os.PathLike`, *optional*): + Can be either: + + - A string, the *repo id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline + hosted on the Hub. + - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights + saved using + [`~DiffusionPipeline.save_pretrained`]. + torch_dtype (`str` or `torch.dtype`, *optional*): + Override the default `torch.dtype` and load the model with another dtype. + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + cache_dir (`Union[str, os.PathLike]`, *optional*): + Path to a directory where a downloaded pretrained model configuration is cached if the standard cache + is not used. + + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. + output_loading_info(`bool`, *optional*, defaults to `False`): + Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages. + local_files_only (`bool`, *optional*, defaults to `False`): + Whether to only load local model weights and configuration files or not. If set to `True`, the model + won't be downloaded from the Hub. + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from + `diffusers-cli login` (stored in `~/.huggingface`) is used. + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier + allowed by Git. + custom_revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, or a commit id similar to + `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a + custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub. + mirror (`str`, *optional*): + Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not + guarantee the timeliness or safety of the source, and you should refer to the mirror site for more + information. + device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*): + A map that specifies where each submodule should go. It doesn’t need to be defined for each + parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the + same device. + + Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For + more information about each option see [designing a device + map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map). + max_memory (`Dict`, *optional*): + A dictionary device identifier for the maximum memory. Will default to the maximum memory available for + each GPU and the available CPU RAM if unset. + offload_folder (`str` or `os.PathLike`, *optional*): + The path to offload weights if device_map contains the value `"disk"`. + offload_state_dict (`bool`, *optional*): + If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if + the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True` + when there is some disk offload. + low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`): + Speed up model loading only loading the pretrained weights and not initializing the weights. This also + tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model. + Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this + argument to `True` will raise an error. + use_safetensors (`bool`, *optional*, defaults to `None`): + If set to `None`, the safetensors weights are downloaded if they're available **and** if the + safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors + weights. If set to `False`, safetensors weights are not loaded. + kwargs (remaining dictionary of keyword arguments, *optional*): + Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline + class). The overwritten components are passed directly to the pipelines `__init__` method. See example + below for more information. + variant (`str`, *optional*): + Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when + loading `from_flax`. + + + + To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with + `huggingface-cli login`. + + + + Examples: + + ```py + >>> from diffusers import AutoPipelineForInpainting + + >>> pipeline = AutoPipelineForInpainting.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5") + >>> image = pipeline(prompt, image=init_image, mask_image=mask_image).images[0] + ``` + """ + cache_dir = kwargs.pop("cache_dir", None) + force_download = kwargs.pop("force_download", False) + proxies = kwargs.pop("proxies", None) + token = kwargs.pop("token", None) + local_files_only = kwargs.pop("local_files_only", False) + revision = kwargs.pop("revision", None) + + load_config_kwargs = { + "cache_dir": cache_dir, + "force_download": force_download, + "proxies": proxies, + "token": token, + "local_files_only": local_files_only, + "revision": revision, + } + + config = cls.load_config(pretrained_model_or_path, **load_config_kwargs) + orig_class_name = config["_class_name"] + + # The `orig_class_name`` can be: + # `- *InpaintPipeline` (for inpaint-specific checkpoint) + # - `*ControlPipeline` (for Flux tools specific checkpoint) + # - or *Pipeline (for regular text-to-image checkpoint) + if "Inpaint" in orig_class_name: + to_replace = "InpaintPipeline" + elif "ControlPipeline" in orig_class_name: + to_replace = "ControlPipeline" + else: + to_replace = "Pipeline" + + if "controlnet" in kwargs: + if isinstance(kwargs["controlnet"], ControlNetUnionModel): + orig_class_name = orig_class_name.replace(to_replace, "ControlNetUnion" + to_replace) + else: + orig_class_name = orig_class_name.replace(to_replace, "ControlNet" + to_replace) + if "enable_pag" in kwargs: + enable_pag = kwargs.pop("enable_pag") + if enable_pag: + orig_class_name = orig_class_name.replace(to_replace, "PAG" + to_replace) + if to_replace == "ControlPipeline": + orig_class_name = orig_class_name.replace(to_replace, "ControlInpaintPipeline") + inpainting_cls = _get_task_class(AUTO_INPAINT_PIPELINES_MAPPING, orig_class_name) + + kwargs = {**load_config_kwargs, **kwargs} + return inpainting_cls.from_pretrained(pretrained_model_or_path, **kwargs) + + @classmethod + def from_pipe(cls, pipeline, **kwargs): + r""" + Instantiates a inpainting Pytorch diffusion pipeline from another instantiated diffusion pipeline class. + + The from_pipe() method takes care of returning the correct pipeline class instance by finding the inpainting + pipeline linked to the pipeline class using pattern matching on pipeline class name. + + All the modules the pipeline class contain will be used to initialize the new pipeline without reallocating + additional memory. + + The pipeline is set in evaluation mode (`model.eval()`) by default. + + Parameters: + pipeline (`DiffusionPipeline`): + an instantiated `DiffusionPipeline` object + + Examples: + + ```py + >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForInpainting + + >>> pipe_t2i = AutoPipelineForText2Image.from_pretrained( + ... "DeepFloyd/IF-I-XL-v1.0", requires_safety_checker=False + ... ) + + >>> pipe_inpaint = AutoPipelineForInpainting.from_pipe(pipe_t2i) + >>> image = pipe_inpaint(prompt, image=init_image, mask_image=mask_image).images[0] + ``` + """ + original_config = dict(pipeline.config) + original_cls_name = pipeline.__class__.__name__ + + # derive the pipeline class to instantiate + inpainting_cls = _get_task_class(AUTO_INPAINT_PIPELINES_MAPPING, original_cls_name) + + if "controlnet" in kwargs: + if kwargs["controlnet"] is not None: + inpainting_cls = _get_task_class( + AUTO_INPAINT_PIPELINES_MAPPING, + inpainting_cls.__name__.replace("ControlNet", "").replace( + "InpaintPipeline", "ControlNetInpaintPipeline" + ), + ) + else: + inpainting_cls = _get_task_class( + AUTO_INPAINT_PIPELINES_MAPPING, + inpainting_cls.__name__.replace("ControlNetInpaintPipeline", "InpaintPipeline"), + ) + + if "enable_pag" in kwargs: + enable_pag = kwargs.pop("enable_pag") + if enable_pag: + inpainting_cls = _get_task_class( + AUTO_INPAINT_PIPELINES_MAPPING, + inpainting_cls.__name__.replace("PAG", "").replace("InpaintPipeline", "PAGInpaintPipeline"), + ) + else: + inpainting_cls = _get_task_class( + AUTO_INPAINT_PIPELINES_MAPPING, + inpainting_cls.__name__.replace("PAGInpaintPipeline", "InpaintPipeline"), + ) + + # define expected module and optional kwargs given the pipeline signature + expected_modules, optional_kwargs = inpainting_cls._get_signature_keys(inpainting_cls) + + pretrained_model_name_or_path = original_config.pop("_name_or_path", None) + + # allow users pass modules in `kwargs` to override the original pipeline's components + passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs} + original_class_obj = { + k: pipeline.components[k] + for k, v in pipeline.components.items() + if k in expected_modules and k not in passed_class_obj + } + + # allow users pass optional kwargs to override the original pipelines config attribute + passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs} + original_pipe_kwargs = { + k: original_config[k] + for k, v in original_config.items() + if k in optional_kwargs and k not in passed_pipe_kwargs + } + + # config that were not expected by original pipeline is stored as private attribute + # we will pass them as optional arguments if they can be accepted by the pipeline + additional_pipe_kwargs = [ + k[1:] + for k in original_config.keys() + if k.startswith("_") and k[1:] in optional_kwargs and k[1:] not in passed_pipe_kwargs + ] + for k in additional_pipe_kwargs: + original_pipe_kwargs[k] = original_config.pop(f"_{k}") + + inpainting_kwargs = {**passed_class_obj, **original_class_obj, **passed_pipe_kwargs, **original_pipe_kwargs} + + # store unused config as private attribute + unused_original_config = { + f"{'' if k.startswith('_') else '_'}{k}": original_config[k] + for k, v in original_config.items() + if k not in inpainting_kwargs + } + + missing_modules = ( + set(expected_modules) - set(inpainting_cls._optional_components) - set(inpainting_kwargs.keys()) + ) + + if len(missing_modules) > 0: + raise ValueError( + f"Pipeline {inpainting_cls} expected {expected_modules}, but only {set(list(passed_class_obj.keys()) + list(original_class_obj.keys()))} were passed" + ) + + model = inpainting_cls(**inpainting_kwargs) + model.register_to_config(_name_or_path=pretrained_model_name_or_path) + model.register_to_config(**unused_original_config) + + return model diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/free_init_utils.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/free_init_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..4495c5ea2683380fee98ad8ddba4959abe8a801e --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/free_init_utils.py @@ -0,0 +1,187 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 typing import Tuple, Union + +import torch +import torch.fft as fft + +from ..utils.torch_utils import randn_tensor + + +class FreeInitMixin: + r"""Mixin class for FreeInit.""" + + def enable_free_init( + self, + num_iters: int = 3, + use_fast_sampling: bool = False, + method: str = "butterworth", + order: int = 4, + spatial_stop_frequency: float = 0.25, + temporal_stop_frequency: float = 0.25, + ): + """Enables the FreeInit mechanism as in https://huggingface.co/papers/2312.07537. + + This implementation has been adapted from the [official repository](https://github.com/TianxingWu/FreeInit). + + Args: + num_iters (`int`, *optional*, defaults to `3`): + Number of FreeInit noise re-initialization iterations. + use_fast_sampling (`bool`, *optional*, defaults to `False`): + Whether or not to speedup sampling procedure at the cost of probably lower quality results. Enables the + "Coarse-to-Fine Sampling" strategy, as mentioned in the paper, if set to `True`. + method (`str`, *optional*, defaults to `butterworth`): + Must be one of `butterworth`, `ideal` or `gaussian` to use as the filtering method for the FreeInit low + pass filter. + order (`int`, *optional*, defaults to `4`): + Order of the filter used in `butterworth` method. Larger values lead to `ideal` method behaviour + whereas lower values lead to `gaussian` method behaviour. + spatial_stop_frequency (`float`, *optional*, defaults to `0.25`): + Normalized stop frequency for spatial dimensions. Must be between 0 to 1. Referred to as `d_s` in the + original implementation. + temporal_stop_frequency (`float`, *optional*, defaults to `0.25`): + Normalized stop frequency for temporal dimensions. Must be between 0 to 1. Referred to as `d_t` in the + original implementation. + """ + self._free_init_num_iters = num_iters + self._free_init_use_fast_sampling = use_fast_sampling + self._free_init_method = method + self._free_init_order = order + self._free_init_spatial_stop_frequency = spatial_stop_frequency + self._free_init_temporal_stop_frequency = temporal_stop_frequency + + def disable_free_init(self): + """Disables the FreeInit mechanism if enabled.""" + self._free_init_num_iters = None + + @property + def free_init_enabled(self): + return hasattr(self, "_free_init_num_iters") and self._free_init_num_iters is not None + + def _get_free_init_freq_filter( + self, + shape: Tuple[int, ...], + device: Union[str, torch.dtype], + filter_type: str, + order: float, + spatial_stop_frequency: float, + temporal_stop_frequency: float, + ) -> torch.Tensor: + r"""Returns the FreeInit filter based on filter type and other input conditions.""" + + time, height, width = shape[-3], shape[-2], shape[-1] + mask = torch.zeros(shape) + + if spatial_stop_frequency == 0 or temporal_stop_frequency == 0: + return mask + + if filter_type == "butterworth": + + def retrieve_mask(x): + return 1 / (1 + (x / spatial_stop_frequency**2) ** order) + elif filter_type == "gaussian": + + def retrieve_mask(x): + return math.exp(-1 / (2 * spatial_stop_frequency**2) * x) + elif filter_type == "ideal": + + def retrieve_mask(x): + return 1 if x <= spatial_stop_frequency * 2 else 0 + else: + raise NotImplementedError("`filter_type` must be one of gaussian, butterworth or ideal") + + for t in range(time): + for h in range(height): + for w in range(width): + d_square = ( + ((spatial_stop_frequency / temporal_stop_frequency) * (2 * t / time - 1)) ** 2 + + (2 * h / height - 1) ** 2 + + (2 * w / width - 1) ** 2 + ) + mask[..., t, h, w] = retrieve_mask(d_square) + + return mask.to(device) + + def _apply_freq_filter(self, x: torch.Tensor, noise: torch.Tensor, low_pass_filter: torch.Tensor) -> torch.Tensor: + r"""Noise reinitialization.""" + # FFT + x_freq = fft.fftn(x, dim=(-3, -2, -1)) + x_freq = fft.fftshift(x_freq, dim=(-3, -2, -1)) + noise_freq = fft.fftn(noise, dim=(-3, -2, -1)) + noise_freq = fft.fftshift(noise_freq, dim=(-3, -2, -1)) + + # frequency mix + high_pass_filter = 1 - low_pass_filter + x_freq_low = x_freq * low_pass_filter + noise_freq_high = noise_freq * high_pass_filter + x_freq_mixed = x_freq_low + noise_freq_high # mix in freq domain + + # IFFT + x_freq_mixed = fft.ifftshift(x_freq_mixed, dim=(-3, -2, -1)) + x_mixed = fft.ifftn(x_freq_mixed, dim=(-3, -2, -1)).real + + return x_mixed + + def _apply_free_init( + self, + latents: torch.Tensor, + free_init_iteration: int, + num_inference_steps: int, + device: torch.device, + dtype: torch.dtype, + generator: torch.Generator, + ): + if free_init_iteration == 0: + self._free_init_initial_noise = latents.detach().clone() + else: + latent_shape = latents.shape + + free_init_filter_shape = (1, *latent_shape[1:]) + free_init_freq_filter = self._get_free_init_freq_filter( + shape=free_init_filter_shape, + device=device, + filter_type=self._free_init_method, + order=self._free_init_order, + spatial_stop_frequency=self._free_init_spatial_stop_frequency, + temporal_stop_frequency=self._free_init_temporal_stop_frequency, + ) + + current_diffuse_timestep = self.scheduler.config.num_train_timesteps - 1 + diffuse_timesteps = torch.full((latent_shape[0],), current_diffuse_timestep).long() + + z_t = self.scheduler.add_noise( + original_samples=latents, noise=self._free_init_initial_noise, timesteps=diffuse_timesteps.to(device) + ).to(dtype=torch.float32) + + z_rand = randn_tensor( + shape=latent_shape, + generator=generator, + device=device, + dtype=torch.float32, + ) + latents = self._apply_freq_filter(z_t, z_rand, low_pass_filter=free_init_freq_filter) + latents = latents.to(dtype) + + # Coarse-to-Fine Sampling for faster inference (can lead to lower quality) + if self._free_init_use_fast_sampling: + num_inference_steps = max( + 1, int(num_inference_steps / self._free_init_num_iters * (free_init_iteration + 1)) + ) + + if num_inference_steps > 0: + self.scheduler.set_timesteps(num_inference_steps, device=device) + + return latents, self.scheduler.timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/free_noise_utils.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/free_noise_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..2910afaf237bada3528628316fa5634946aa38a0 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/free_noise_utils.py @@ -0,0 +1,596 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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, Dict, List, Optional, Tuple, Union + +import torch +import torch.nn as nn + +from ..models.attention import BasicTransformerBlock, FreeNoiseTransformerBlock +from ..models.resnet import Downsample2D, ResnetBlock2D, Upsample2D +from ..models.transformers.transformer_2d import Transformer2DModel +from ..models.unets.unet_motion_model import ( + AnimateDiffTransformer3D, + CrossAttnDownBlockMotion, + DownBlockMotion, + UpBlockMotion, +) +from ..pipelines.pipeline_utils import DiffusionPipeline +from ..utils import logging +from ..utils.torch_utils import randn_tensor + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class SplitInferenceModule(nn.Module): + r""" + A wrapper module class that splits inputs along a specified dimension before performing a forward pass. + + This module is useful when you need to perform inference on large tensors in a memory-efficient way by breaking + them into smaller chunks, processing each chunk separately, and then reassembling the results. + + Args: + module (`nn.Module`): + The underlying PyTorch module that will be applied to each chunk of split inputs. + split_size (`int`, defaults to `1`): + The size of each chunk after splitting the input tensor. + split_dim (`int`, defaults to `0`): + The dimension along which the input tensors are split. + input_kwargs_to_split (`List[str]`, defaults to `["hidden_states"]`): + A list of keyword arguments (strings) that represent the input tensors to be split. + + Workflow: + 1. The keyword arguments specified in `input_kwargs_to_split` are split into smaller chunks using + `torch.split()` along the dimension `split_dim` and with a chunk size of `split_size`. + 2. The `module` is invoked once for each split with both the split inputs and any unchanged arguments + that were passed. + 3. The output tensors from each split are concatenated back together along `split_dim` before returning. + + Example: + ```python + >>> import torch + >>> import torch.nn as nn + + >>> model = nn.Linear(1000, 1000) + >>> split_module = SplitInferenceModule(model, split_size=2, split_dim=0, input_kwargs_to_split=["input"]) + + >>> input_tensor = torch.randn(42, 1000) + >>> # Will split the tensor into 21 slices of shape [2, 1000]. + >>> output = split_module(input=input_tensor) + ``` + + It is also possible to nest `SplitInferenceModule` across different split dimensions for more complex + multi-dimensional splitting. + """ + + def __init__( + self, + module: nn.Module, + split_size: int = 1, + split_dim: int = 0, + input_kwargs_to_split: List[str] = ["hidden_states"], + ) -> None: + super().__init__() + + self.module = module + self.split_size = split_size + self.split_dim = split_dim + self.input_kwargs_to_split = set(input_kwargs_to_split) + + def forward(self, *args, **kwargs) -> Union[torch.Tensor, Tuple[torch.Tensor]]: + r"""Forward method for the `SplitInferenceModule`. + + This method processes the input by splitting specified keyword arguments along a given dimension, running the + underlying module on each split, and then concatenating the results. The splitting is controlled by the + `split_size` and `split_dim` parameters specified during initialization. + + Args: + *args (`Any`): + Positional arguments that are passed directly to the `module` without modification. + **kwargs (`Dict[str, torch.Tensor]`): + Keyword arguments passed to the underlying `module`. Only keyword arguments whose names match the + entries in `input_kwargs_to_split` and are of type `torch.Tensor` will be split. The remaining keyword + arguments are passed unchanged. + + Returns: + `Union[torch.Tensor, Tuple[torch.Tensor]]`: + The outputs obtained from `SplitInferenceModule` are the same as if the underlying module was inferred + without it. + - If the underlying module returns a single tensor, the result will be a single concatenated tensor + along the same `split_dim` after processing all splits. + - If the underlying module returns a tuple of tensors, each element of the tuple will be concatenated + along the `split_dim` across all splits, and the final result will be a tuple of concatenated tensors. + """ + split_inputs = {} + + # 1. Split inputs that were specified during initialization and also present in passed kwargs + for key in list(kwargs.keys()): + if key not in self.input_kwargs_to_split or not torch.is_tensor(kwargs[key]): + continue + split_inputs[key] = torch.split(kwargs[key], self.split_size, self.split_dim) + kwargs.pop(key) + + # 2. Invoke forward pass across each split + results = [] + for split_input in zip(*split_inputs.values()): + inputs = dict(zip(split_inputs.keys(), split_input)) + inputs.update(kwargs) + + intermediate_tensor_or_tensor_tuple = self.module(*args, **inputs) + results.append(intermediate_tensor_or_tensor_tuple) + + # 3. Concatenate split restuls to obtain final outputs + if isinstance(results[0], torch.Tensor): + return torch.cat(results, dim=self.split_dim) + elif isinstance(results[0], tuple): + return tuple([torch.cat(x, dim=self.split_dim) for x in zip(*results)]) + else: + raise ValueError( + "In order to use the SplitInferenceModule, it is necessary for the underlying `module` to either return a torch.Tensor or a tuple of torch.Tensor's." + ) + + +class AnimateDiffFreeNoiseMixin: + r"""Mixin class for [FreeNoise](https://huggingface.co/papers/2310.15169).""" + + def _enable_free_noise_in_block(self, block: Union[CrossAttnDownBlockMotion, DownBlockMotion, UpBlockMotion]): + r"""Helper function to enable FreeNoise in transformer blocks.""" + + for motion_module in block.motion_modules: + num_transformer_blocks = len(motion_module.transformer_blocks) + + for i in range(num_transformer_blocks): + if isinstance(motion_module.transformer_blocks[i], FreeNoiseTransformerBlock): + motion_module.transformer_blocks[i].set_free_noise_properties( + self._free_noise_context_length, + self._free_noise_context_stride, + self._free_noise_weighting_scheme, + ) + else: + assert isinstance(motion_module.transformer_blocks[i], BasicTransformerBlock) + basic_transfomer_block = motion_module.transformer_blocks[i] + + motion_module.transformer_blocks[i] = FreeNoiseTransformerBlock( + dim=basic_transfomer_block.dim, + num_attention_heads=basic_transfomer_block.num_attention_heads, + attention_head_dim=basic_transfomer_block.attention_head_dim, + dropout=basic_transfomer_block.dropout, + cross_attention_dim=basic_transfomer_block.cross_attention_dim, + activation_fn=basic_transfomer_block.activation_fn, + attention_bias=basic_transfomer_block.attention_bias, + only_cross_attention=basic_transfomer_block.only_cross_attention, + double_self_attention=basic_transfomer_block.double_self_attention, + positional_embeddings=basic_transfomer_block.positional_embeddings, + num_positional_embeddings=basic_transfomer_block.num_positional_embeddings, + context_length=self._free_noise_context_length, + context_stride=self._free_noise_context_stride, + weighting_scheme=self._free_noise_weighting_scheme, + ).to(device=self.device, dtype=self.dtype) + + motion_module.transformer_blocks[i].load_state_dict( + basic_transfomer_block.state_dict(), strict=True + ) + motion_module.transformer_blocks[i].set_chunk_feed_forward( + basic_transfomer_block._chunk_size, basic_transfomer_block._chunk_dim + ) + + def _disable_free_noise_in_block(self, block: Union[CrossAttnDownBlockMotion, DownBlockMotion, UpBlockMotion]): + r"""Helper function to disable FreeNoise in transformer blocks.""" + + for motion_module in block.motion_modules: + num_transformer_blocks = len(motion_module.transformer_blocks) + + for i in range(num_transformer_blocks): + if isinstance(motion_module.transformer_blocks[i], FreeNoiseTransformerBlock): + free_noise_transfomer_block = motion_module.transformer_blocks[i] + + motion_module.transformer_blocks[i] = BasicTransformerBlock( + dim=free_noise_transfomer_block.dim, + num_attention_heads=free_noise_transfomer_block.num_attention_heads, + attention_head_dim=free_noise_transfomer_block.attention_head_dim, + dropout=free_noise_transfomer_block.dropout, + cross_attention_dim=free_noise_transfomer_block.cross_attention_dim, + activation_fn=free_noise_transfomer_block.activation_fn, + attention_bias=free_noise_transfomer_block.attention_bias, + only_cross_attention=free_noise_transfomer_block.only_cross_attention, + double_self_attention=free_noise_transfomer_block.double_self_attention, + positional_embeddings=free_noise_transfomer_block.positional_embeddings, + num_positional_embeddings=free_noise_transfomer_block.num_positional_embeddings, + ).to(device=self.device, dtype=self.dtype) + + motion_module.transformer_blocks[i].load_state_dict( + free_noise_transfomer_block.state_dict(), strict=True + ) + motion_module.transformer_blocks[i].set_chunk_feed_forward( + free_noise_transfomer_block._chunk_size, free_noise_transfomer_block._chunk_dim + ) + + def _check_inputs_free_noise( + self, + prompt, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + num_frames, + ) -> None: + if not isinstance(prompt, (str, dict)): + raise ValueError(f"Expected `prompt` to have type `str` or `dict` but found {type(prompt)=}") + + if negative_prompt is not None: + if not isinstance(negative_prompt, (str, dict)): + raise ValueError( + f"Expected `negative_prompt` to have type `str` or `dict` but found {type(negative_prompt)=}" + ) + + if prompt_embeds is not None or negative_prompt_embeds is not None: + raise ValueError("`prompt_embeds` and `negative_prompt_embeds` is not supported in FreeNoise yet.") + + frame_indices = [isinstance(x, int) for x in prompt.keys()] + frame_prompts = [isinstance(x, str) for x in prompt.values()] + min_frame = min(list(prompt.keys())) + max_frame = max(list(prompt.keys())) + + if not all(frame_indices): + raise ValueError("Expected integer keys in `prompt` dict for FreeNoise.") + if not all(frame_prompts): + raise ValueError("Expected str values in `prompt` dict for FreeNoise.") + if min_frame != 0: + raise ValueError("The minimum frame index in `prompt` dict must be 0 as a starting prompt is necessary.") + if max_frame >= num_frames: + raise ValueError( + f"The maximum frame index in `prompt` dict must be lesser than {num_frames=} and follow 0-based indexing." + ) + + def _encode_prompt_free_noise( + self, + prompt: Union[str, Dict[int, str]], + num_frames: int, + device: torch.device, + num_videos_per_prompt: int, + do_classifier_free_guidance: bool, + negative_prompt: Optional[Union[str, Dict[int, str]]] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ) -> torch.Tensor: + if negative_prompt is None: + negative_prompt = "" + + # Ensure that we have a dictionary of prompts + if isinstance(prompt, str): + prompt = {0: prompt} + if isinstance(negative_prompt, str): + negative_prompt = {0: negative_prompt} + + self._check_inputs_free_noise(prompt, negative_prompt, prompt_embeds, negative_prompt_embeds, num_frames) + + # Sort the prompts based on frame indices + prompt = dict(sorted(prompt.items())) + negative_prompt = dict(sorted(negative_prompt.items())) + + # Ensure that we have a prompt for the last frame index + prompt[num_frames - 1] = prompt[list(prompt.keys())[-1]] + negative_prompt[num_frames - 1] = negative_prompt[list(negative_prompt.keys())[-1]] + + frame_indices = list(prompt.keys()) + frame_prompts = list(prompt.values()) + frame_negative_indices = list(negative_prompt.keys()) + frame_negative_prompts = list(negative_prompt.values()) + + # Generate and interpolate positive prompts + prompt_embeds, _ = self.encode_prompt( + prompt=frame_prompts, + device=device, + num_images_per_prompt=num_videos_per_prompt, + do_classifier_free_guidance=False, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + lora_scale=lora_scale, + clip_skip=clip_skip, + ) + + shape = (num_frames, *prompt_embeds.shape[1:]) + prompt_interpolation_embeds = prompt_embeds.new_zeros(shape) + + for i in range(len(frame_indices) - 1): + start_frame = frame_indices[i] + end_frame = frame_indices[i + 1] + start_tensor = prompt_embeds[i].unsqueeze(0) + end_tensor = prompt_embeds[i + 1].unsqueeze(0) + + prompt_interpolation_embeds[start_frame : end_frame + 1] = self._free_noise_prompt_interpolation_callback( + start_frame, end_frame, start_tensor, end_tensor + ) + + # Generate and interpolate negative prompts + negative_prompt_embeds = None + negative_prompt_interpolation_embeds = None + + if do_classifier_free_guidance: + _, negative_prompt_embeds = self.encode_prompt( + prompt=[""] * len(frame_negative_prompts), + device=device, + num_images_per_prompt=num_videos_per_prompt, + do_classifier_free_guidance=True, + negative_prompt=frame_negative_prompts, + prompt_embeds=None, + negative_prompt_embeds=None, + lora_scale=lora_scale, + clip_skip=clip_skip, + ) + + negative_prompt_interpolation_embeds = negative_prompt_embeds.new_zeros(shape) + + for i in range(len(frame_negative_indices) - 1): + start_frame = frame_negative_indices[i] + end_frame = frame_negative_indices[i + 1] + start_tensor = negative_prompt_embeds[i].unsqueeze(0) + end_tensor = negative_prompt_embeds[i + 1].unsqueeze(0) + + negative_prompt_interpolation_embeds[start_frame : end_frame + 1] = ( + self._free_noise_prompt_interpolation_callback(start_frame, end_frame, start_tensor, end_tensor) + ) + + prompt_embeds = prompt_interpolation_embeds + negative_prompt_embeds = negative_prompt_interpolation_embeds + + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds, negative_prompt_embeds + + def _prepare_latents_free_noise( + self, + batch_size: int, + num_channels_latents: int, + num_frames: int, + height: int, + width: int, + dtype: torch.dtype, + device: torch.device, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.Tensor] = None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + context_num_frames = ( + self._free_noise_context_length if self._free_noise_context_length == "repeat_context" else num_frames + ) + + shape = ( + batch_size, + num_channels_latents, + context_num_frames, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + if self._free_noise_noise_type == "random": + return latents + else: + if latents.size(2) == num_frames: + return latents + elif latents.size(2) != self._free_noise_context_length: + raise ValueError( + f"You have passed `latents` as a parameter to FreeNoise. The expected number of frames is either {num_frames} or {self._free_noise_context_length}, but found {latents.size(2)}" + ) + latents = latents.to(device) + + if self._free_noise_noise_type == "shuffle_context": + for i in range(self._free_noise_context_length, num_frames, self._free_noise_context_stride): + # ensure window is within bounds + window_start = max(0, i - self._free_noise_context_length) + window_end = min(num_frames, window_start + self._free_noise_context_stride) + window_length = window_end - window_start + + if window_length == 0: + break + + indices = torch.LongTensor(list(range(window_start, window_end))) + shuffled_indices = indices[torch.randperm(window_length, generator=generator)] + + current_start = i + current_end = min(num_frames, current_start + window_length) + if current_end == current_start + window_length: + # batch of frames perfectly fits the window + latents[:, :, current_start:current_end] = latents[:, :, shuffled_indices] + else: + # handle the case where the last batch of frames does not fit perfectly with the window + prefix_length = current_end - current_start + shuffled_indices = shuffled_indices[:prefix_length] + latents[:, :, current_start:current_end] = latents[:, :, shuffled_indices] + + elif self._free_noise_noise_type == "repeat_context": + num_repeats = (num_frames + self._free_noise_context_length - 1) // self._free_noise_context_length + latents = torch.cat([latents] * num_repeats, dim=2) + + latents = latents[:, :, :num_frames] + return latents + + def _lerp( + self, start_index: int, end_index: int, start_tensor: torch.Tensor, end_tensor: torch.Tensor + ) -> torch.Tensor: + num_indices = end_index - start_index + 1 + interpolated_tensors = [] + + for i in range(num_indices): + alpha = i / (num_indices - 1) + interpolated_tensor = (1 - alpha) * start_tensor + alpha * end_tensor + interpolated_tensors.append(interpolated_tensor) + + interpolated_tensors = torch.cat(interpolated_tensors) + return interpolated_tensors + + def enable_free_noise( + self, + context_length: Optional[int] = 16, + context_stride: int = 4, + weighting_scheme: str = "pyramid", + noise_type: str = "shuffle_context", + prompt_interpolation_callback: Optional[ + Callable[[DiffusionPipeline, int, int, torch.Tensor, torch.Tensor], torch.Tensor] + ] = None, + ) -> None: + r""" + Enable long video generation using FreeNoise. + + Args: + context_length (`int`, defaults to `16`, *optional*): + The number of video frames to process at once. It's recommended to set this to the maximum frames the + Motion Adapter was trained with (usually 16/24/32). If `None`, the default value from the motion + adapter config is used. + context_stride (`int`, *optional*): + Long videos are generated by processing many frames. FreeNoise processes these frames in sliding + windows of size `context_length`. Context stride allows you to specify how many frames to skip between + each window. For example, a context length of 16 and context stride of 4 would process 24 frames as: + [0, 15], [4, 19], [8, 23] (0-based indexing) + weighting_scheme (`str`, defaults to `pyramid`): + Weighting scheme for averaging latents after accumulation in FreeNoise blocks. The following weighting + schemes are supported currently: + - "flat" + Performs weighting averaging with a flat weight pattern: [1, 1, 1, 1, 1]. + - "pyramid" + Performs weighted averaging with a pyramid like weight pattern: [1, 2, 3, 2, 1]. + - "delayed_reverse_sawtooth" + Performs weighted averaging with low weights for earlier frames and high-to-low weights for + later frames: [0.01, 0.01, 3, 2, 1]. + noise_type (`str`, defaults to "shuffle_context"): + Must be one of ["shuffle_context", "repeat_context", "random"]. + - "shuffle_context" + Shuffles a fixed batch of `context_length` latents to create a final latent of size + `num_frames`. This is usually the best setting for most generation scenarios. However, there + might be visible repetition noticeable in the kinds of motion/animation generated. + - "repeated_context" + Repeats a fixed batch of `context_length` latents to create a final latent of size + `num_frames`. + - "random" + The final latents are random without any repetition. + """ + + allowed_weighting_scheme = ["flat", "pyramid", "delayed_reverse_sawtooth"] + allowed_noise_type = ["shuffle_context", "repeat_context", "random"] + + if context_length > self.motion_adapter.config.motion_max_seq_length: + logger.warning( + f"You have set {context_length=} which is greater than {self.motion_adapter.config.motion_max_seq_length=}. This can lead to bad generation results." + ) + if weighting_scheme not in allowed_weighting_scheme: + raise ValueError( + f"The parameter `weighting_scheme` must be one of {allowed_weighting_scheme}, but got {weighting_scheme=}" + ) + if noise_type not in allowed_noise_type: + raise ValueError(f"The parameter `noise_type` must be one of {allowed_noise_type}, but got {noise_type=}") + + self._free_noise_context_length = context_length or self.motion_adapter.config.motion_max_seq_length + self._free_noise_context_stride = context_stride + self._free_noise_weighting_scheme = weighting_scheme + self._free_noise_noise_type = noise_type + self._free_noise_prompt_interpolation_callback = prompt_interpolation_callback or self._lerp + + if hasattr(self.unet.mid_block, "motion_modules"): + blocks = [*self.unet.down_blocks, self.unet.mid_block, *self.unet.up_blocks] + else: + blocks = [*self.unet.down_blocks, *self.unet.up_blocks] + + for block in blocks: + self._enable_free_noise_in_block(block) + + def disable_free_noise(self) -> None: + r"""Disable the FreeNoise sampling mechanism.""" + self._free_noise_context_length = None + + if hasattr(self.unet.mid_block, "motion_modules"): + blocks = [*self.unet.down_blocks, self.unet.mid_block, *self.unet.up_blocks] + else: + blocks = [*self.unet.down_blocks, *self.unet.up_blocks] + + blocks = [*self.unet.down_blocks, self.unet.mid_block, *self.unet.up_blocks] + for block in blocks: + self._disable_free_noise_in_block(block) + + def _enable_split_inference_motion_modules_( + self, motion_modules: List[AnimateDiffTransformer3D], spatial_split_size: int + ) -> None: + for motion_module in motion_modules: + motion_module.proj_in = SplitInferenceModule(motion_module.proj_in, spatial_split_size, 0, ["input"]) + + for i in range(len(motion_module.transformer_blocks)): + motion_module.transformer_blocks[i] = SplitInferenceModule( + motion_module.transformer_blocks[i], + spatial_split_size, + 0, + ["hidden_states", "encoder_hidden_states"], + ) + + motion_module.proj_out = SplitInferenceModule(motion_module.proj_out, spatial_split_size, 0, ["input"]) + + def _enable_split_inference_attentions_( + self, attentions: List[Transformer2DModel], temporal_split_size: int + ) -> None: + for i in range(len(attentions)): + attentions[i] = SplitInferenceModule( + attentions[i], temporal_split_size, 0, ["hidden_states", "encoder_hidden_states"] + ) + + def _enable_split_inference_resnets_(self, resnets: List[ResnetBlock2D], temporal_split_size: int) -> None: + for i in range(len(resnets)): + resnets[i] = SplitInferenceModule(resnets[i], temporal_split_size, 0, ["input_tensor", "temb"]) + + def _enable_split_inference_samplers_( + self, samplers: Union[List[Downsample2D], List[Upsample2D]], temporal_split_size: int + ) -> None: + for i in range(len(samplers)): + samplers[i] = SplitInferenceModule(samplers[i], temporal_split_size, 0, ["hidden_states"]) + + def enable_free_noise_split_inference(self, spatial_split_size: int = 256, temporal_split_size: int = 16) -> None: + r""" + Enable FreeNoise memory optimizations by utilizing + [`~diffusers.pipelines.free_noise_utils.SplitInferenceModule`] across different intermediate modeling blocks. + + Args: + spatial_split_size (`int`, defaults to `256`): + The split size across spatial dimensions for internal blocks. This is used in facilitating split + inference across the effective batch dimension (`[B x H x W, F, C]`) of intermediate tensors in motion + modeling blocks. + temporal_split_size (`int`, defaults to `16`): + The split size across temporal dimensions for internal blocks. This is used in facilitating split + inference across the effective batch dimension (`[B x F, H x W, C]`) of intermediate tensors in spatial + attention, resnets, downsampling and upsampling blocks. + """ + # TODO(aryan): Discuss on what's the best way to provide more control to users + blocks = [*self.unet.down_blocks, self.unet.mid_block, *self.unet.up_blocks] + for block in blocks: + if getattr(block, "motion_modules", None) is not None: + self._enable_split_inference_motion_modules_(block.motion_modules, spatial_split_size) + if getattr(block, "attentions", None) is not None: + self._enable_split_inference_attentions_(block.attentions, temporal_split_size) + if getattr(block, "resnets", None) is not None: + self._enable_split_inference_resnets_(block.resnets, temporal_split_size) + if getattr(block, "downsamplers", None) is not None: + self._enable_split_inference_samplers_(block.downsamplers, temporal_split_size) + if getattr(block, "upsamplers", None) is not None: + self._enable_split_inference_samplers_(block.upsamplers, temporal_split_size) + + @property + def free_noise_enabled(self): + return hasattr(self, "_free_noise_context_length") and self._free_noise_context_length is not None diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/onnx_utils.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/onnx_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..74e9f0b97800ce798f14941f6b2c0fb4c2e532fa --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/onnx_utils.py @@ -0,0 +1,230 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# 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 os +import shutil +from pathlib import Path +from typing import Optional, Union + +import numpy as np +from huggingface_hub import hf_hub_download +from huggingface_hub.utils import validate_hf_hub_args + +from ..utils import ONNX_EXTERNAL_WEIGHTS_NAME, ONNX_WEIGHTS_NAME, is_onnx_available, logging + + +if is_onnx_available(): + import onnxruntime as ort + + +logger = logging.get_logger(__name__) + +ORT_TO_NP_TYPE = { + "tensor(bool)": np.bool_, + "tensor(int8)": np.int8, + "tensor(uint8)": np.uint8, + "tensor(int16)": np.int16, + "tensor(uint16)": np.uint16, + "tensor(int32)": np.int32, + "tensor(uint32)": np.uint32, + "tensor(int64)": np.int64, + "tensor(uint64)": np.uint64, + "tensor(float16)": np.float16, + "tensor(float)": np.float32, + "tensor(double)": np.float64, +} + + +class OnnxRuntimeModel: + def __init__(self, model=None, **kwargs): + logger.info("`diffusers.OnnxRuntimeModel` is experimental and might change in the future.") + self.model = model + self.model_save_dir = kwargs.get("model_save_dir", None) + self.latest_model_name = kwargs.get("latest_model_name", ONNX_WEIGHTS_NAME) + + def __call__(self, **kwargs): + inputs = {k: np.array(v) for k, v in kwargs.items()} + return self.model.run(None, inputs) + + @staticmethod + def load_model(path: Union[str, Path], provider=None, sess_options=None, provider_options=None): + """ + Loads an ONNX Inference session with an ExecutionProvider. Default provider is `CPUExecutionProvider` + + Arguments: + path (`str` or `Path`): + Directory from which to load + provider(`str`, *optional*): + Onnxruntime execution provider to use for loading the model, defaults to `CPUExecutionProvider` + """ + if provider is None: + logger.info("No onnxruntime provider specified, using CPUExecutionProvider") + provider = "CPUExecutionProvider" + + if provider_options is None: + provider_options = [] + elif not isinstance(provider_options, list): + provider_options = [provider_options] + + return ort.InferenceSession( + path, providers=[provider], sess_options=sess_options, provider_options=provider_options + ) + + def _save_pretrained(self, save_directory: Union[str, Path], file_name: Optional[str] = None, **kwargs): + """ + Save a model and its configuration file to a directory, so that it can be re-loaded using the + [`~optimum.onnxruntime.modeling_ort.ORTModel.from_pretrained`] class method. It will always save the + latest_model_name. + + Arguments: + save_directory (`str` or `Path`): + Directory where to save the model file. + file_name(`str`, *optional*): + Overwrites the default model file name from `"model.onnx"` to `file_name`. This allows you to save the + model with a different name. + """ + model_file_name = file_name if file_name is not None else ONNX_WEIGHTS_NAME + + src_path = self.model_save_dir.joinpath(self.latest_model_name) + dst_path = Path(save_directory).joinpath(model_file_name) + try: + shutil.copyfile(src_path, dst_path) + except shutil.SameFileError: + pass + + # copy external weights (for models >2GB) + src_path = self.model_save_dir.joinpath(ONNX_EXTERNAL_WEIGHTS_NAME) + if src_path.exists(): + dst_path = Path(save_directory).joinpath(ONNX_EXTERNAL_WEIGHTS_NAME) + try: + shutil.copyfile(src_path, dst_path) + except shutil.SameFileError: + pass + + def save_pretrained( + self, + save_directory: Union[str, os.PathLike], + **kwargs, + ): + """ + Save a model to a directory, so that it can be re-loaded using the [`~OnnxModel.from_pretrained`] class + method.: + + Arguments: + save_directory (`str` or `os.PathLike`): + Directory to which to save. Will be created if it doesn't exist. + """ + if os.path.isfile(save_directory): + logger.error(f"Provided path ({save_directory}) should be a directory, not a file") + return + + os.makedirs(save_directory, exist_ok=True) + + # saving model weights/files + self._save_pretrained(save_directory, **kwargs) + + @classmethod + @validate_hf_hub_args + def _from_pretrained( + cls, + model_id: Union[str, Path], + token: Optional[Union[bool, str, None]] = None, + revision: Optional[Union[str, None]] = None, + force_download: bool = False, + cache_dir: Optional[str] = None, + file_name: Optional[str] = None, + provider: Optional[str] = None, + sess_options: Optional["ort.SessionOptions"] = None, + **kwargs, + ): + """ + Load a model from a directory or the HF Hub. + + Arguments: + model_id (`str` or `Path`): + Directory from which to load + token (`str` or `bool`): + Is needed to load models from a private or gated repository + revision (`str`): + Revision is the specific model version to use. It can be a branch name, a tag name, or a commit id + cache_dir (`Union[str, Path]`, *optional*): + Path to a directory in which a downloaded pretrained model configuration should be cached if the + standard cache should not be used. + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + file_name(`str`): + Overwrites the default model file name from `"model.onnx"` to `file_name`. This allows you to load + different model files from the same repository or directory. + provider(`str`): + The ONNX runtime provider, e.g. `CPUExecutionProvider` or `CUDAExecutionProvider`. + kwargs (`Dict`, *optional*): + kwargs will be passed to the model during initialization + """ + model_file_name = file_name if file_name is not None else ONNX_WEIGHTS_NAME + # load model from local directory + if os.path.isdir(model_id): + model = OnnxRuntimeModel.load_model( + Path(model_id, model_file_name).as_posix(), + provider=provider, + sess_options=sess_options, + provider_options=kwargs.pop("provider_options"), + ) + kwargs["model_save_dir"] = Path(model_id) + # load model from hub + else: + # download model + model_cache_path = hf_hub_download( + repo_id=model_id, + filename=model_file_name, + token=token, + revision=revision, + cache_dir=cache_dir, + force_download=force_download, + ) + kwargs["model_save_dir"] = Path(model_cache_path).parent + kwargs["latest_model_name"] = Path(model_cache_path).name + model = OnnxRuntimeModel.load_model( + model_cache_path, + provider=provider, + sess_options=sess_options, + provider_options=kwargs.pop("provider_options"), + ) + return cls(model=model, **kwargs) + + @classmethod + @validate_hf_hub_args + def from_pretrained( + cls, + model_id: Union[str, Path], + force_download: bool = True, + token: Optional[str] = None, + cache_dir: Optional[str] = None, + **model_kwargs, + ): + revision = None + if len(str(model_id).split("@")) == 2: + model_id, revision = model_id.split("@") + + return cls._from_pretrained( + model_id=model_id, + revision=revision, + cache_dir=cache_dir, + force_download=force_download, + token=token, + **model_kwargs, + ) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/pipeline_flax_utils.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/pipeline_flax_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..7c5ac89602da649fc2d03a540ffd28740b1687b6 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/pipeline_flax_utils.py @@ -0,0 +1,610 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# 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 importlib +import inspect +import os +from typing import Any, Dict, List, Optional, Union + +import flax +import numpy as np +import PIL.Image +from flax.core.frozen_dict import FrozenDict +from huggingface_hub import create_repo, snapshot_download +from huggingface_hub.utils import validate_hf_hub_args +from PIL import Image +from tqdm.auto import tqdm + +from ..configuration_utils import ConfigMixin +from ..models.modeling_flax_utils import FLAX_WEIGHTS_NAME, FlaxModelMixin +from ..schedulers.scheduling_utils_flax import SCHEDULER_CONFIG_NAME, FlaxSchedulerMixin +from ..utils import ( + CONFIG_NAME, + BaseOutput, + PushToHubMixin, + http_user_agent, + is_transformers_available, + logging, +) + + +if is_transformers_available(): + from transformers import FlaxPreTrainedModel + +INDEX_FILE = "diffusion_flax_model.bin" + + +logger = logging.get_logger(__name__) + + +LOADABLE_CLASSES = { + "diffusers": { + "FlaxModelMixin": ["save_pretrained", "from_pretrained"], + "FlaxSchedulerMixin": ["save_pretrained", "from_pretrained"], + "FlaxDiffusionPipeline": ["save_pretrained", "from_pretrained"], + }, + "transformers": { + "PreTrainedTokenizer": ["save_pretrained", "from_pretrained"], + "PreTrainedTokenizerFast": ["save_pretrained", "from_pretrained"], + "FlaxPreTrainedModel": ["save_pretrained", "from_pretrained"], + "FeatureExtractionMixin": ["save_pretrained", "from_pretrained"], + "ProcessorMixin": ["save_pretrained", "from_pretrained"], + "ImageProcessingMixin": ["save_pretrained", "from_pretrained"], + }, +} + +ALL_IMPORTABLE_CLASSES = {} +for library in LOADABLE_CLASSES: + ALL_IMPORTABLE_CLASSES.update(LOADABLE_CLASSES[library]) + + +def import_flax_or_no_model(module, class_name): + try: + # 1. First make sure that if a Flax object is present, import this one + class_obj = getattr(module, "Flax" + class_name) + except AttributeError: + # 2. If this doesn't work, it's not a model and we don't append "Flax" + class_obj = getattr(module, class_name) + except AttributeError: + raise ValueError(f"Neither Flax{class_name} nor {class_name} exist in {module}") + + return class_obj + + +@flax.struct.dataclass +class FlaxImagePipelineOutput(BaseOutput): + """ + Output class for image pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] + + +class FlaxDiffusionPipeline(ConfigMixin, PushToHubMixin): + r""" + Base class for Flax-based pipelines. + + [`FlaxDiffusionPipeline`] stores all components (models, schedulers, and processors) for diffusion pipelines and + provides methods for loading, downloading and saving models. It also includes methods to: + + - enable/disable the progress bar for the denoising iteration + + Class attributes: + + - **config_name** ([`str`]) -- The configuration filename that stores the class and module names of all the + diffusion pipeline's components. + """ + + config_name = "model_index.json" + + def register_modules(self, **kwargs): + # import it here to avoid circular import + from diffusers import pipelines + + for name, module in kwargs.items(): + if module is None: + register_dict = {name: (None, None)} + else: + # retrieve library + library = module.__module__.split(".")[0] + + # check if the module is a pipeline module + pipeline_dir = module.__module__.split(".")[-2] + path = module.__module__.split(".") + is_pipeline_module = pipeline_dir in path and hasattr(pipelines, pipeline_dir) + + # if library is not in LOADABLE_CLASSES, then it is a custom module. + # Or if it's a pipeline module, then the module is inside the pipeline + # folder so we set the library to module name. + if library not in LOADABLE_CLASSES or is_pipeline_module: + library = pipeline_dir + + # retrieve class_name + class_name = module.__class__.__name__ + + register_dict = {name: (library, class_name)} + + # save model index config + self.register_to_config(**register_dict) + + # set models + setattr(self, name, module) + + def save_pretrained( + self, + save_directory: Union[str, os.PathLike], + params: Union[Dict, FrozenDict], + push_to_hub: bool = False, + **kwargs, + ): + # TODO: handle inference_state + """ + Save all saveable variables of the pipeline to a directory. A pipeline variable can be saved and loaded if its + class implements both a save and loading method. The pipeline is easily reloaded using the + [`~FlaxDiffusionPipeline.from_pretrained`] class method. + + Arguments: + save_directory (`str` or `os.PathLike`): + Directory to which to save. Will be created if it doesn't exist. + push_to_hub (`bool`, *optional*, defaults to `False`): + Whether or not to push your model to the Hugging Face model hub after saving it. You can specify the + repository you want to push to with `repo_id` (will default to the name of `save_directory` in your + namespace). + kwargs (`Dict[str, Any]`, *optional*): + Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method. + """ + self.save_config(save_directory) + + model_index_dict = dict(self.config) + model_index_dict.pop("_class_name") + model_index_dict.pop("_diffusers_version") + model_index_dict.pop("_module", None) + + if push_to_hub: + commit_message = kwargs.pop("commit_message", None) + private = kwargs.pop("private", None) + create_pr = kwargs.pop("create_pr", False) + token = kwargs.pop("token", None) + repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1]) + repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id + + for pipeline_component_name in model_index_dict.keys(): + sub_model = getattr(self, pipeline_component_name) + if sub_model is None: + # edge case for saving a pipeline with safety_checker=None + continue + + model_cls = sub_model.__class__ + + save_method_name = None + # search for the model's base class in LOADABLE_CLASSES + for library_name, library_classes in LOADABLE_CLASSES.items(): + library = importlib.import_module(library_name) + for base_class, save_load_methods in library_classes.items(): + class_candidate = getattr(library, base_class, None) + if class_candidate is not None and issubclass(model_cls, class_candidate): + # if we found a suitable base class in LOADABLE_CLASSES then grab its save method + save_method_name = save_load_methods[0] + break + if save_method_name is not None: + break + + save_method = getattr(sub_model, save_method_name) + expects_params = "params" in set(inspect.signature(save_method).parameters.keys()) + + if expects_params: + save_method( + os.path.join(save_directory, pipeline_component_name), params=params[pipeline_component_name] + ) + else: + save_method(os.path.join(save_directory, pipeline_component_name)) + + if push_to_hub: + self._upload_folder( + save_directory, + repo_id, + token=token, + commit_message=commit_message, + create_pr=create_pr, + ) + + @classmethod + @validate_hf_hub_args + def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs): + r""" + Instantiate a Flax-based diffusion pipeline from pretrained pipeline weights. + + The pipeline is set in evaluation mode (`model.eval()) by default and dropout modules are deactivated. + + If you get the error message below, you need to finetune the weights for your downstream task: + + ``` + Some weights of FlaxUNet2DConditionModel were not initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5 and are newly initialized because the shapes did not match: + ``` + + Parameters: + pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*): + Can be either: + + - A string, the *repo id* (for example `stable-diffusion-v1-5/stable-diffusion-v1-5`) of a + pretrained pipeline hosted on the Hub. + - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved + using [`~FlaxDiffusionPipeline.save_pretrained`]. + dtype (`jnp.dtype`, *optional*): + Override the default `jnp.dtype` and load the model under this dtype. + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. + output_loading_info(`bool`, *optional*, defaults to `False`): + Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages. + local_files_only (`bool`, *optional*, defaults to `False`): + Whether to only load local model weights and configuration files or not. If set to `True`, the model + won't be downloaded from the Hub. + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from + `diffusers-cli login` (stored in `~/.huggingface`) is used. + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier + allowed by Git. + mirror (`str`, *optional*): + Mirror source to resolve accessibility issues if you're downloading a model in China. We do not + guarantee the timeliness or safety of the source, and you should refer to the mirror site for more + information. + kwargs (remaining dictionary of keyword arguments, *optional*): + Can be used to overwrite load and saveable variables (the pipeline components) of the specific pipeline + class. The overwritten components are passed directly to the pipelines `__init__` method. + + + + To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in with + `huggingface-cli login`. + + + + Examples: + + ```py + >>> from diffusers import FlaxDiffusionPipeline + + >>> # Download pipeline from huggingface.co and cache. + >>> # Requires to be logged in to Hugging Face hub, + >>> # see more in [the documentation](https://huggingface.co/docs/hub/security-tokens) + >>> pipeline, params = FlaxDiffusionPipeline.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", + ... variant="bf16", + ... dtype=jnp.bfloat16, + ... ) + + >>> # Download pipeline, but use a different scheduler + >>> from diffusers import FlaxDPMSolverMultistepScheduler + + >>> model_id = "stable-diffusion-v1-5/stable-diffusion-v1-5" + >>> dpmpp, dpmpp_state = FlaxDPMSolverMultistepScheduler.from_pretrained( + ... model_id, + ... subfolder="scheduler", + ... ) + + >>> dpm_pipe, dpm_params = FlaxStableDiffusionPipeline.from_pretrained( + ... model_id, variant="bf16", dtype=jnp.bfloat16, scheduler=dpmpp + ... ) + >>> dpm_params["scheduler"] = dpmpp_state + ``` + """ + cache_dir = kwargs.pop("cache_dir", None) + proxies = kwargs.pop("proxies", None) + local_files_only = kwargs.pop("local_files_only", False) + token = kwargs.pop("token", None) + revision = kwargs.pop("revision", None) + from_pt = kwargs.pop("from_pt", False) + use_memory_efficient_attention = kwargs.pop("use_memory_efficient_attention", False) + split_head_dim = kwargs.pop("split_head_dim", False) + dtype = kwargs.pop("dtype", None) + + # 1. Download the checkpoints and configs + # use snapshot download here to get it working from from_pretrained + if not os.path.isdir(pretrained_model_name_or_path): + config_dict = cls.load_config( + pretrained_model_name_or_path, + cache_dir=cache_dir, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + ) + # make sure we only download sub-folders and `diffusers` filenames + folder_names = [k for k in config_dict.keys() if not k.startswith("_")] + allow_patterns = [os.path.join(k, "*") for k in folder_names] + allow_patterns += [FLAX_WEIGHTS_NAME, SCHEDULER_CONFIG_NAME, CONFIG_NAME, cls.config_name] + + ignore_patterns = ["*.bin", "*.safetensors"] if not from_pt else [] + ignore_patterns += ["*.onnx", "*.onnx_data", "*.xml", "*.pb"] + + if cls != FlaxDiffusionPipeline: + requested_pipeline_class = cls.__name__ + else: + requested_pipeline_class = config_dict.get("_class_name", cls.__name__) + requested_pipeline_class = ( + requested_pipeline_class + if requested_pipeline_class.startswith("Flax") + else "Flax" + requested_pipeline_class + ) + + user_agent = {"pipeline_class": requested_pipeline_class} + user_agent = http_user_agent(user_agent) + + # download all allow_patterns + cached_folder = snapshot_download( + pretrained_model_name_or_path, + cache_dir=cache_dir, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + allow_patterns=allow_patterns, + ignore_patterns=ignore_patterns, + user_agent=user_agent, + ) + else: + cached_folder = pretrained_model_name_or_path + + config_dict = cls.load_config(cached_folder) + + # 2. Load the pipeline class, if using custom module then load it from the hub + # if we load from explicit class, let's use it + if cls != FlaxDiffusionPipeline: + pipeline_class = cls + else: + diffusers_module = importlib.import_module(cls.__module__.split(".")[0]) + class_name = ( + config_dict["_class_name"] + if config_dict["_class_name"].startswith("Flax") + else "Flax" + config_dict["_class_name"] + ) + pipeline_class = getattr(diffusers_module, class_name) + + # some modules can be passed directly to the init + # in this case they are already instantiated in `kwargs` + # extract them here + expected_modules, optional_kwargs = cls._get_signature_keys(pipeline_class) + passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs} + passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs} + + init_dict, unused_kwargs, _ = pipeline_class.extract_init_dict(config_dict, **kwargs) + + # define init kwargs + init_kwargs = {k: init_dict.pop(k) for k in optional_kwargs if k in init_dict} + init_kwargs = {**init_kwargs, **passed_pipe_kwargs} + + # remove `null` components + def load_module(name, value): + if value[0] is None: + return False + if name in passed_class_obj and passed_class_obj[name] is None: + return False + return True + + init_dict = {k: v for k, v in init_dict.items() if load_module(k, v)} + + # Throw nice warnings / errors for fast accelerate loading + if len(unused_kwargs) > 0: + logger.warning( + f"Keyword arguments {unused_kwargs} are not expected by {pipeline_class.__name__} and will be ignored." + ) + + # inference_params + params = {} + + # import it here to avoid circular import + from diffusers import pipelines + + # 3. Load each module in the pipeline + for name, (library_name, class_name) in init_dict.items(): + if class_name is None: + # edge case for when the pipeline was saved with safety_checker=None + init_kwargs[name] = None + continue + + is_pipeline_module = hasattr(pipelines, library_name) + loaded_sub_model = None + sub_model_should_be_defined = True + + # if the model is in a pipeline module, then we load it from the pipeline + if name in passed_class_obj: + # 1. check that passed_class_obj has correct parent class + if not is_pipeline_module: + library = importlib.import_module(library_name) + class_obj = getattr(library, class_name) + importable_classes = LOADABLE_CLASSES[library_name] + class_candidates = {c: getattr(library, c, None) for c in importable_classes.keys()} + + expected_class_obj = None + for class_name, class_candidate in class_candidates.items(): + if class_candidate is not None and issubclass(class_obj, class_candidate): + expected_class_obj = class_candidate + + if not issubclass(passed_class_obj[name].__class__, expected_class_obj): + raise ValueError( + f"{passed_class_obj[name]} is of type: {type(passed_class_obj[name])}, but should be" + f" {expected_class_obj}" + ) + elif passed_class_obj[name] is None: + logger.warning( + f"You have passed `None` for {name} to disable its functionality in {pipeline_class}. Note" + f" that this might lead to problems when using {pipeline_class} and is not recommended." + ) + sub_model_should_be_defined = False + else: + logger.warning( + f"You have passed a non-standard module {passed_class_obj[name]}. We cannot verify whether it" + " has the correct type" + ) + + # set passed class object + loaded_sub_model = passed_class_obj[name] + elif is_pipeline_module: + pipeline_module = getattr(pipelines, library_name) + class_obj = import_flax_or_no_model(pipeline_module, class_name) + + importable_classes = ALL_IMPORTABLE_CLASSES + class_candidates = dict.fromkeys(importable_classes.keys(), class_obj) + else: + # else we just import it from the library. + library = importlib.import_module(library_name) + class_obj = import_flax_or_no_model(library, class_name) + + importable_classes = LOADABLE_CLASSES[library_name] + class_candidates = {c: getattr(library, c, None) for c in importable_classes.keys()} + + if loaded_sub_model is None and sub_model_should_be_defined: + load_method_name = None + for class_name, class_candidate in class_candidates.items(): + if class_candidate is not None and issubclass(class_obj, class_candidate): + load_method_name = importable_classes[class_name][1] + + load_method = getattr(class_obj, load_method_name) + + # check if the module is in a subdirectory + if os.path.isdir(os.path.join(cached_folder, name)): + loadable_folder = os.path.join(cached_folder, name) + else: + loaded_sub_model = cached_folder + + if issubclass(class_obj, FlaxModelMixin): + loaded_sub_model, loaded_params = load_method( + loadable_folder, + from_pt=from_pt, + use_memory_efficient_attention=use_memory_efficient_attention, + split_head_dim=split_head_dim, + dtype=dtype, + ) + params[name] = loaded_params + elif is_transformers_available() and issubclass(class_obj, FlaxPreTrainedModel): + if from_pt: + # TODO(Suraj): Fix this in Transformers. We should be able to use `_do_init=False` here + loaded_sub_model = load_method(loadable_folder, from_pt=from_pt) + loaded_params = loaded_sub_model.params + del loaded_sub_model._params + else: + loaded_sub_model, loaded_params = load_method(loadable_folder, _do_init=False) + params[name] = loaded_params + elif issubclass(class_obj, FlaxSchedulerMixin): + loaded_sub_model, scheduler_state = load_method(loadable_folder) + params[name] = scheduler_state + else: + loaded_sub_model = load_method(loadable_folder) + + init_kwargs[name] = loaded_sub_model # UNet(...), # DiffusionSchedule(...) + + # 4. Potentially add passed objects if expected + missing_modules = set(expected_modules) - set(init_kwargs.keys()) + passed_modules = list(passed_class_obj.keys()) + + if len(missing_modules) > 0 and missing_modules <= set(passed_modules): + for module in missing_modules: + init_kwargs[module] = passed_class_obj.get(module, None) + elif len(missing_modules) > 0: + passed_modules = set(list(init_kwargs.keys()) + list(passed_class_obj.keys())) - optional_kwargs + raise ValueError( + f"Pipeline {pipeline_class} expected {expected_modules}, but only {passed_modules} were passed." + ) + + model = pipeline_class(**init_kwargs, dtype=dtype) + return model, params + + @classmethod + def _get_signature_keys(cls, obj): + parameters = inspect.signature(obj.__init__).parameters + required_parameters = {k: v for k, v in parameters.items() if v.default == inspect._empty} + optional_parameters = set({k for k, v in parameters.items() if v.default != inspect._empty}) + expected_modules = set(required_parameters.keys()) - {"self"} + + return expected_modules, optional_parameters + + @property + def components(self) -> Dict[str, Any]: + r""" + + The `self.components` property can be useful to run different pipelines with the same weights and + configurations to not have to re-allocate memory. + + Examples: + + ```py + >>> from diffusers import ( + ... FlaxStableDiffusionPipeline, + ... FlaxStableDiffusionImg2ImgPipeline, + ... ) + + >>> text2img = FlaxStableDiffusionPipeline.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", variant="bf16", dtype=jnp.bfloat16 + ... ) + >>> img2img = FlaxStableDiffusionImg2ImgPipeline(**text2img.components) + ``` + + Returns: + A dictionary containing all the modules needed to initialize the pipeline. + """ + expected_modules, optional_parameters = self._get_signature_keys(self) + components = { + k: getattr(self, k) for k in self.config.keys() if not k.startswith("_") and k not in optional_parameters + } + + if set(components.keys()) != expected_modules: + raise ValueError( + f"{self} has been incorrectly initialized or {self.__class__} is incorrectly implemented. Expected" + f" {expected_modules} to be defined, but {components} are defined." + ) + + return components + + @staticmethod + def numpy_to_pil(images): + """ + Convert a NumPy image or a batch of images to a PIL image. + """ + if images.ndim == 3: + images = images[None, ...] + images = (images * 255).round().astype("uint8") + if images.shape[-1] == 1: + # special case for grayscale (single channel) images + pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images] + else: + pil_images = [Image.fromarray(image) for image in images] + + return pil_images + + # TODO: make it compatible with jax.lax + def progress_bar(self, iterable): + if not hasattr(self, "_progress_bar_config"): + self._progress_bar_config = {} + elif not isinstance(self._progress_bar_config, dict): + raise ValueError( + f"`self._progress_bar_config` should be of type `dict`, but is {type(self._progress_bar_config)}." + ) + + return tqdm(iterable, **self._progress_bar_config) + + def set_progress_bar_config(self, **kwargs): + self._progress_bar_config = kwargs diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/pipeline_loading_utils.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/pipeline_loading_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..d1c2c2adb4c30d00bdc382c17044ce3431c0ba5b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/pipeline_loading_utils.py @@ -0,0 +1,1156 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# +# 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 importlib +import os +import re +import warnings +from pathlib import Path +from typing import Any, Callable, Dict, List, Optional, Union + +import requests +import torch +from huggingface_hub import DDUFEntry, ModelCard, model_info, snapshot_download +from huggingface_hub.utils import OfflineModeIsEnabled, validate_hf_hub_args +from packaging import version +from requests.exceptions import HTTPError + +from .. import __version__ +from ..utils import ( + FLAX_WEIGHTS_NAME, + ONNX_EXTERNAL_WEIGHTS_NAME, + ONNX_WEIGHTS_NAME, + SAFETENSORS_WEIGHTS_NAME, + WEIGHTS_NAME, + deprecate, + get_class_from_dynamic_module, + is_accelerate_available, + is_peft_available, + is_transformers_available, + is_transformers_version, + logging, +) +from ..utils.torch_utils import is_compiled_module +from .transformers_loading_utils import _load_tokenizer_from_dduf, _load_transformers_model_from_dduf + + +if is_transformers_available(): + import transformers + from transformers import PreTrainedModel, PreTrainedTokenizerBase + from transformers.utils import FLAX_WEIGHTS_NAME as TRANSFORMERS_FLAX_WEIGHTS_NAME + from transformers.utils import SAFE_WEIGHTS_NAME as TRANSFORMERS_SAFE_WEIGHTS_NAME + from transformers.utils import WEIGHTS_NAME as TRANSFORMERS_WEIGHTS_NAME + +if is_accelerate_available(): + import accelerate + from accelerate import dispatch_model + from accelerate.hooks import remove_hook_from_module + from accelerate.utils import compute_module_sizes, get_max_memory + + +INDEX_FILE = "diffusion_pytorch_model.bin" +CUSTOM_PIPELINE_FILE_NAME = "pipeline.py" +DUMMY_MODULES_FOLDER = "diffusers.utils" +TRANSFORMERS_DUMMY_MODULES_FOLDER = "transformers.utils" +CONNECTED_PIPES_KEYS = ["prior"] + +logger = logging.get_logger(__name__) + +LOADABLE_CLASSES = { + "diffusers": { + "ModelMixin": ["save_pretrained", "from_pretrained"], + "SchedulerMixin": ["save_pretrained", "from_pretrained"], + "DiffusionPipeline": ["save_pretrained", "from_pretrained"], + "OnnxRuntimeModel": ["save_pretrained", "from_pretrained"], + }, + "transformers": { + "PreTrainedTokenizer": ["save_pretrained", "from_pretrained"], + "PreTrainedTokenizerFast": ["save_pretrained", "from_pretrained"], + "PreTrainedModel": ["save_pretrained", "from_pretrained"], + "FeatureExtractionMixin": ["save_pretrained", "from_pretrained"], + "ProcessorMixin": ["save_pretrained", "from_pretrained"], + "ImageProcessingMixin": ["save_pretrained", "from_pretrained"], + }, + "onnxruntime.training": { + "ORTModule": ["save_pretrained", "from_pretrained"], + }, +} + +ALL_IMPORTABLE_CLASSES = {} +for library in LOADABLE_CLASSES: + ALL_IMPORTABLE_CLASSES.update(LOADABLE_CLASSES[library]) + + +def is_safetensors_compatible(filenames, passed_components=None, folder_names=None, variant=None) -> bool: + """ + Checking for safetensors compatibility: + - The model is safetensors compatible only if there is a safetensors file for each model component present in + filenames. + + Converting default pytorch serialized filenames to safetensors serialized filenames: + - For models from the diffusers library, just replace the ".bin" extension with ".safetensors" + - For models from the transformers library, the filename changes from "pytorch_model" to "model", and the ".bin" + extension is replaced with ".safetensors" + """ + weight_names = [ + WEIGHTS_NAME, + SAFETENSORS_WEIGHTS_NAME, + FLAX_WEIGHTS_NAME, + ONNX_WEIGHTS_NAME, + ONNX_EXTERNAL_WEIGHTS_NAME, + ] + + if is_transformers_available(): + weight_names += [TRANSFORMERS_WEIGHTS_NAME, TRANSFORMERS_SAFE_WEIGHTS_NAME, TRANSFORMERS_FLAX_WEIGHTS_NAME] + + # model_pytorch, diffusion_model_pytorch, ... + weight_prefixes = [w.split(".")[0] for w in weight_names] + # .bin, .safetensors, ... + weight_suffixs = [w.split(".")[-1] for w in weight_names] + # -00001-of-00002 + transformers_index_format = r"\d{5}-of-\d{5}" + # `diffusion_pytorch_model.bin` as well as `model-00001-of-00002.safetensors` + variant_file_re = re.compile( + rf"({'|'.join(weight_prefixes)})\.({variant}|{variant}-{transformers_index_format})\.({'|'.join(weight_suffixs)})$" + ) + non_variant_file_re = re.compile( + rf"({'|'.join(weight_prefixes)})(-{transformers_index_format})?\.({'|'.join(weight_suffixs)})$" + ) + + passed_components = passed_components or [] + if folder_names: + filenames = {f for f in filenames if os.path.split(f)[0] in folder_names} + + # extract all components of the pipeline and their associated files + components = {} + for filename in filenames: + if not len(filename.split("/")) == 2: + continue + + component, component_filename = filename.split("/") + if component in passed_components: + continue + + components.setdefault(component, []) + components[component].append(component_filename) + + # If there are no component folders check the main directory for safetensors files + filtered_filenames = set() + if not components: + if variant is not None: + filtered_filenames = filter_with_regex(filenames, variant_file_re) + + # If no variant filenames exist check if non-variant files are available + if not filtered_filenames: + filtered_filenames = filter_with_regex(filenames, non_variant_file_re) + return any(".safetensors" in filename for filename in filtered_filenames) + + # iterate over all files of a component + # check if safetensor files exist for that component + for component, component_filenames in components.items(): + matches = [] + filtered_component_filenames = set() + # if variant is provided check if the variant of the safetensors exists + if variant is not None: + filtered_component_filenames = filter_with_regex(component_filenames, variant_file_re) + + # if variant safetensor files do not exist check for non-variants + if not filtered_component_filenames: + filtered_component_filenames = filter_with_regex(component_filenames, non_variant_file_re) + for component_filename in filtered_component_filenames: + filename, extension = os.path.splitext(component_filename) + + match_exists = extension == ".safetensors" + matches.append(match_exists) + + if not any(matches): + return False + + return True + + +def filter_model_files(filenames): + """Filter model repo files for just files/folders that contain model weights""" + weight_names = [ + WEIGHTS_NAME, + SAFETENSORS_WEIGHTS_NAME, + FLAX_WEIGHTS_NAME, + ONNX_WEIGHTS_NAME, + ONNX_EXTERNAL_WEIGHTS_NAME, + ] + + if is_transformers_available(): + weight_names += [TRANSFORMERS_WEIGHTS_NAME, TRANSFORMERS_SAFE_WEIGHTS_NAME, TRANSFORMERS_FLAX_WEIGHTS_NAME] + + allowed_extensions = [wn.split(".")[-1] for wn in weight_names] + + return [f for f in filenames if any(f.endswith(extension) for extension in allowed_extensions)] + + +def filter_with_regex(filenames, pattern_re): + return {f for f in filenames if pattern_re.match(f.split("/")[-1]) is not None} + + +def variant_compatible_siblings(filenames, variant=None, ignore_patterns=None) -> Union[List[os.PathLike], str]: + weight_names = [ + WEIGHTS_NAME, + SAFETENSORS_WEIGHTS_NAME, + FLAX_WEIGHTS_NAME, + ONNX_WEIGHTS_NAME, + ONNX_EXTERNAL_WEIGHTS_NAME, + ] + + if is_transformers_available(): + weight_names += [TRANSFORMERS_WEIGHTS_NAME, TRANSFORMERS_SAFE_WEIGHTS_NAME, TRANSFORMERS_FLAX_WEIGHTS_NAME] + + # model_pytorch, diffusion_model_pytorch, ... + weight_prefixes = [w.split(".")[0] for w in weight_names] + # .bin, .safetensors, ... + weight_suffixs = [w.split(".")[-1] for w in weight_names] + # -00001-of-00002 + transformers_index_format = r"\d{5}-of-\d{5}" + + if variant is not None: + # `diffusion_pytorch_model.fp16.bin` as well as `model.fp16-00001-of-00002.safetensors` + variant_file_re = re.compile( + rf"({'|'.join(weight_prefixes)})\.({variant}|{variant}-{transformers_index_format})\.({'|'.join(weight_suffixs)})$" + ) + # `text_encoder/pytorch_model.bin.index.fp16.json` + variant_index_re = re.compile( + rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.index\.{variant}\.json$" + ) + legacy_variant_file_re = re.compile(rf".*-{transformers_index_format}\.{variant}\.[a-z]+$") + legacy_variant_index_re = re.compile( + rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.{variant}\.index\.json$" + ) + + # `diffusion_pytorch_model.bin` as well as `model-00001-of-00002.safetensors` + non_variant_file_re = re.compile( + rf"({'|'.join(weight_prefixes)})(-{transformers_index_format})?\.({'|'.join(weight_suffixs)})$" + ) + # `text_encoder/pytorch_model.bin.index.json` + non_variant_index_re = re.compile(rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.index\.json") + + def filter_for_compatible_extensions(filenames, ignore_patterns=None): + if not ignore_patterns: + return filenames + + # ignore patterns uses glob style patterns e.g *.safetensors but we're only + # interested in the extension name + return {f for f in filenames if not any(f.endswith(pat.lstrip("*.")) for pat in ignore_patterns)} + + # Group files by component + components = {} + for filename in filenames: + if not len(filename.split("/")) == 2: + components.setdefault("", []).append(filename) + continue + + component, _ = filename.split("/") + components.setdefault(component, []).append(filename) + + usable_filenames = set() + variant_filenames = set() + for component, component_filenames in components.items(): + component_filenames = filter_for_compatible_extensions(component_filenames, ignore_patterns=ignore_patterns) + + component_variants = set() + component_legacy_variants = set() + component_non_variants = set() + if variant is not None: + component_variants = filter_with_regex(component_filenames, variant_file_re) + component_variant_index_files = filter_with_regex(component_filenames, variant_index_re) + + component_legacy_variants = filter_with_regex(component_filenames, legacy_variant_file_re) + component_legacy_variant_index_files = filter_with_regex(component_filenames, legacy_variant_index_re) + + if component_variants or component_legacy_variants: + variant_filenames.update( + component_variants | component_variant_index_files + if component_variants + else component_legacy_variants | component_legacy_variant_index_files + ) + + else: + component_non_variants = filter_with_regex(component_filenames, non_variant_file_re) + component_variant_index_files = filter_with_regex(component_filenames, non_variant_index_re) + + usable_filenames.update(component_non_variants | component_variant_index_files) + + usable_filenames.update(variant_filenames) + + if len(variant_filenames) == 0 and variant is not None: + error_message = f"You are trying to load model files of the `variant={variant}`, but no such modeling files are available. " + raise ValueError(error_message) + + if len(variant_filenames) > 0 and usable_filenames != variant_filenames: + logger.warning( + f"\nA mixture of {variant} and non-{variant} filenames will be loaded.\nLoaded {variant} filenames:\n" + f"[{', '.join(variant_filenames)}]\nLoaded non-{variant} filenames:\n" + f"[{', '.join(usable_filenames - variant_filenames)}\nIf this behavior is not " + f"expected, please check your folder structure." + ) + + return usable_filenames, variant_filenames + + +@validate_hf_hub_args +def warn_deprecated_model_variant(pretrained_model_name_or_path, token, variant, revision, model_filenames): + info = model_info( + pretrained_model_name_or_path, + token=token, + revision=None, + ) + filenames = {sibling.rfilename for sibling in info.siblings} + comp_model_filenames, _ = variant_compatible_siblings(filenames, variant=revision) + comp_model_filenames = [".".join(f.split(".")[:1] + f.split(".")[2:]) for f in comp_model_filenames] + + if set(model_filenames).issubset(set(comp_model_filenames)): + warnings.warn( + f"You are loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'` even though you can load it via `variant=`{revision}`. Loading model variants via `revision='{revision}'` is deprecated and will be removed in diffusers v1. Please use `variant='{revision}'` instead.", + FutureWarning, + ) + else: + warnings.warn( + f"You are loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'`. This behavior is deprecated and will be removed in diffusers v1. One should use `variant='{revision}'` instead. However, it appears that {pretrained_model_name_or_path} currently does not have the required variant filenames in the 'main' branch. \n The Diffusers team and community would be very grateful if you could open an issue: https://github.com/huggingface/diffusers/issues/new with the title '{pretrained_model_name_or_path} is missing {revision} files' so that the correct variant file can be added.", + FutureWarning, + ) + + +def _unwrap_model(model): + """Unwraps a model.""" + if is_compiled_module(model): + model = model._orig_mod + + if is_peft_available(): + from peft import PeftModel + + if isinstance(model, PeftModel): + model = model.base_model.model + + return model + + +def maybe_raise_or_warn( + library_name, library, class_name, importable_classes, passed_class_obj, name, is_pipeline_module +): + """Simple helper method to raise or warn in case incorrect module has been passed""" + if not is_pipeline_module: + library = importlib.import_module(library_name) + class_obj = getattr(library, class_name) + class_candidates = {c: getattr(library, c, None) for c in importable_classes.keys()} + + expected_class_obj = None + for class_name, class_candidate in class_candidates.items(): + if class_candidate is not None and issubclass(class_obj, class_candidate): + expected_class_obj = class_candidate + + # Dynamo wraps the original model in a private class. + # I didn't find a public API to get the original class. + sub_model = passed_class_obj[name] + unwrapped_sub_model = _unwrap_model(sub_model) + model_cls = unwrapped_sub_model.__class__ + + if not issubclass(model_cls, expected_class_obj): + raise ValueError(f"{passed_class_obj[name]} is of type: {model_cls}, but should be {expected_class_obj}") + else: + logger.warning( + f"You have passed a non-standard module {passed_class_obj[name]}. We cannot verify whether it" + " has the correct type" + ) + + +def get_class_obj_and_candidates( + library_name, class_name, importable_classes, pipelines, is_pipeline_module, component_name=None, cache_dir=None +): + """Simple helper method to retrieve class object of module as well as potential parent class objects""" + component_folder = os.path.join(cache_dir, component_name) if component_name and cache_dir else None + + if is_pipeline_module: + pipeline_module = getattr(pipelines, library_name) + + class_obj = getattr(pipeline_module, class_name) + class_candidates = dict.fromkeys(importable_classes.keys(), class_obj) + elif component_folder and os.path.isfile(os.path.join(component_folder, library_name + ".py")): + # load custom component + class_obj = get_class_from_dynamic_module( + component_folder, module_file=library_name + ".py", class_name=class_name + ) + class_candidates = dict.fromkeys(importable_classes.keys(), class_obj) + else: + # else we just import it from the library. + library = importlib.import_module(library_name) + + class_obj = getattr(library, class_name) + class_candidates = {c: getattr(library, c, None) for c in importable_classes.keys()} + + return class_obj, class_candidates + + +def _get_custom_pipeline_class( + custom_pipeline, + repo_id=None, + hub_revision=None, + class_name=None, + cache_dir=None, + revision=None, +): + if custom_pipeline.endswith(".py"): + path = Path(custom_pipeline) + # decompose into folder & file + file_name = path.name + custom_pipeline = path.parent.absolute() + elif repo_id is not None: + file_name = f"{custom_pipeline}.py" + custom_pipeline = repo_id + else: + file_name = CUSTOM_PIPELINE_FILE_NAME + + if repo_id is not None and hub_revision is not None: + # if we load the pipeline code from the Hub + # make sure to overwrite the `revision` + revision = hub_revision + + return get_class_from_dynamic_module( + custom_pipeline, + module_file=file_name, + class_name=class_name, + cache_dir=cache_dir, + revision=revision, + ) + + +def _get_pipeline_class( + class_obj, + config=None, + load_connected_pipeline=False, + custom_pipeline=None, + repo_id=None, + hub_revision=None, + class_name=None, + cache_dir=None, + revision=None, +): + if custom_pipeline is not None: + return _get_custom_pipeline_class( + custom_pipeline, + repo_id=repo_id, + hub_revision=hub_revision, + class_name=class_name, + cache_dir=cache_dir, + revision=revision, + ) + + if class_obj.__name__ != "DiffusionPipeline": + return class_obj + + diffusers_module = importlib.import_module(class_obj.__module__.split(".")[0]) + class_name = class_name or config["_class_name"] + if not class_name: + raise ValueError( + "The class name could not be found in the configuration file. Please make sure to pass the correct `class_name`." + ) + + class_name = class_name[4:] if class_name.startswith("Flax") else class_name + + pipeline_cls = getattr(diffusers_module, class_name) + + if load_connected_pipeline: + from .auto_pipeline import _get_connected_pipeline + + connected_pipeline_cls = _get_connected_pipeline(pipeline_cls) + if connected_pipeline_cls is not None: + logger.info( + f"Loading connected pipeline {connected_pipeline_cls.__name__} instead of {pipeline_cls.__name__} as specified via `load_connected_pipeline=True`" + ) + else: + logger.info(f"{pipeline_cls.__name__} has no connected pipeline class. Loading {pipeline_cls.__name__}.") + + pipeline_cls = connected_pipeline_cls or pipeline_cls + + return pipeline_cls + + +def _load_empty_model( + library_name: str, + class_name: str, + importable_classes: List[Any], + pipelines: Any, + is_pipeline_module: bool, + name: str, + torch_dtype: Union[str, torch.dtype], + cached_folder: Union[str, os.PathLike], + **kwargs, +): + # retrieve class objects. + class_obj, _ = get_class_obj_and_candidates( + library_name, + class_name, + importable_classes, + pipelines, + is_pipeline_module, + component_name=name, + cache_dir=cached_folder, + ) + + if is_transformers_available(): + transformers_version = version.parse(version.parse(transformers.__version__).base_version) + else: + transformers_version = "N/A" + + # Determine library. + is_transformers_model = ( + is_transformers_available() + and issubclass(class_obj, PreTrainedModel) + and transformers_version >= version.parse("4.20.0") + ) + diffusers_module = importlib.import_module(__name__.split(".")[0]) + is_diffusers_model = issubclass(class_obj, diffusers_module.ModelMixin) + + model = None + config_path = cached_folder + user_agent = { + "diffusers": __version__, + "file_type": "model", + "framework": "pytorch", + } + + if is_diffusers_model: + # Load config and then the model on meta. + config, unused_kwargs, commit_hash = class_obj.load_config( + os.path.join(config_path, name), + cache_dir=cached_folder, + return_unused_kwargs=True, + return_commit_hash=True, + force_download=kwargs.pop("force_download", False), + proxies=kwargs.pop("proxies", None), + local_files_only=kwargs.pop("local_files_only", False), + token=kwargs.pop("token", None), + revision=kwargs.pop("revision", None), + subfolder=kwargs.pop("subfolder", None), + user_agent=user_agent, + ) + with accelerate.init_empty_weights(): + model = class_obj.from_config(config, **unused_kwargs) + elif is_transformers_model: + config_class = getattr(class_obj, "config_class", None) + if config_class is None: + raise ValueError("`config_class` cannot be None. Please double-check the model.") + + config = config_class.from_pretrained( + cached_folder, + subfolder=name, + force_download=kwargs.pop("force_download", False), + proxies=kwargs.pop("proxies", None), + local_files_only=kwargs.pop("local_files_only", False), + token=kwargs.pop("token", None), + revision=kwargs.pop("revision", None), + user_agent=user_agent, + ) + with accelerate.init_empty_weights(): + model = class_obj(config) + + if model is not None: + model = model.to(dtype=torch_dtype) + return model + + +def _assign_components_to_devices( + module_sizes: Dict[str, float], device_memory: Dict[str, float], device_mapping_strategy: str = "balanced" +): + device_ids = list(device_memory.keys()) + device_cycle = device_ids + device_ids[::-1] + device_memory = device_memory.copy() + + device_id_component_mapping = {} + current_device_index = 0 + for component in module_sizes: + device_id = device_cycle[current_device_index % len(device_cycle)] + component_memory = module_sizes[component] + curr_device_memory = device_memory[device_id] + + # If the GPU doesn't fit the current component offload to the CPU. + if component_memory > curr_device_memory: + device_id_component_mapping["cpu"] = [component] + else: + if device_id not in device_id_component_mapping: + device_id_component_mapping[device_id] = [component] + else: + device_id_component_mapping[device_id].append(component) + + # Update the device memory. + device_memory[device_id] -= component_memory + current_device_index += 1 + + return device_id_component_mapping + + +def _get_final_device_map(device_map, pipeline_class, passed_class_obj, init_dict, library, max_memory, **kwargs): + # To avoid circular import problem. + from diffusers import pipelines + + torch_dtype = kwargs.get("torch_dtype", torch.float32) + + # Load each module in the pipeline on a meta device so that we can derive the device map. + init_empty_modules = {} + for name, (library_name, class_name) in init_dict.items(): + if class_name.startswith("Flax"): + raise ValueError("Flax pipelines are not supported with `device_map`.") + + # Define all importable classes + is_pipeline_module = hasattr(pipelines, library_name) + importable_classes = ALL_IMPORTABLE_CLASSES + loaded_sub_model = None + + # Use passed sub model or load class_name from library_name + if name in passed_class_obj: + # if the model is in a pipeline module, then we load it from the pipeline + # check that passed_class_obj has correct parent class + maybe_raise_or_warn( + library_name, + library, + class_name, + importable_classes, + passed_class_obj, + name, + is_pipeline_module, + ) + with accelerate.init_empty_weights(): + loaded_sub_model = passed_class_obj[name] + + else: + sub_model_dtype = ( + torch_dtype.get(name, torch_dtype.get("default", torch.float32)) + if isinstance(torch_dtype, dict) + else torch_dtype + ) + loaded_sub_model = _load_empty_model( + library_name=library_name, + class_name=class_name, + importable_classes=importable_classes, + pipelines=pipelines, + is_pipeline_module=is_pipeline_module, + pipeline_class=pipeline_class, + name=name, + torch_dtype=sub_model_dtype, + cached_folder=kwargs.get("cached_folder", None), + force_download=kwargs.get("force_download", None), + proxies=kwargs.get("proxies", None), + local_files_only=kwargs.get("local_files_only", None), + token=kwargs.get("token", None), + revision=kwargs.get("revision", None), + ) + + if loaded_sub_model is not None: + init_empty_modules[name] = loaded_sub_model + + # determine device map + # Obtain a sorted dictionary for mapping the model-level components + # to their sizes. + module_sizes = { + module_name: compute_module_sizes( + module, + dtype=torch_dtype.get(module_name, torch_dtype.get("default", torch.float32)) + if isinstance(torch_dtype, dict) + else torch_dtype, + )[""] + for module_name, module in init_empty_modules.items() + if isinstance(module, torch.nn.Module) + } + module_sizes = dict(sorted(module_sizes.items(), key=lambda item: item[1], reverse=True)) + + # Obtain maximum memory available per device (GPUs only). + max_memory = get_max_memory(max_memory) + max_memory = dict(sorted(max_memory.items(), key=lambda item: item[1], reverse=True)) + max_memory = {k: v for k, v in max_memory.items() if k != "cpu"} + + # Obtain a dictionary mapping the model-level components to the available + # devices based on the maximum memory and the model sizes. + final_device_map = None + if len(max_memory) > 0: + device_id_component_mapping = _assign_components_to_devices( + module_sizes, max_memory, device_mapping_strategy=device_map + ) + + # Obtain the final device map, e.g., `{"unet": 0, "text_encoder": 1, "vae": 1, ...}` + final_device_map = {} + for device_id, components in device_id_component_mapping.items(): + for component in components: + final_device_map[component] = device_id + + return final_device_map + + +def load_sub_model( + library_name: str, + class_name: str, + importable_classes: List[Any], + pipelines: Any, + is_pipeline_module: bool, + pipeline_class: Any, + torch_dtype: torch.dtype, + provider: Any, + sess_options: Any, + device_map: Optional[Union[Dict[str, torch.device], str]], + max_memory: Optional[Dict[Union[int, str], Union[int, str]]], + offload_folder: Optional[Union[str, os.PathLike]], + offload_state_dict: bool, + model_variants: Dict[str, str], + name: str, + from_flax: bool, + variant: str, + low_cpu_mem_usage: bool, + cached_folder: Union[str, os.PathLike], + use_safetensors: bool, + dduf_entries: Optional[Dict[str, DDUFEntry]], + provider_options: Any, + quantization_config: Optional[Any] = None, +): + """Helper method to load the module `name` from `library_name` and `class_name`""" + from ..quantizers import PipelineQuantizationConfig + + # retrieve class candidates + + class_obj, class_candidates = get_class_obj_and_candidates( + library_name, + class_name, + importable_classes, + pipelines, + is_pipeline_module, + component_name=name, + cache_dir=cached_folder, + ) + + load_method_name = None + # retrieve load method name + for class_name, class_candidate in class_candidates.items(): + if class_candidate is not None and issubclass(class_obj, class_candidate): + load_method_name = importable_classes[class_name][1] + + # if load method name is None, then we have a dummy module -> raise Error + if load_method_name is None: + none_module = class_obj.__module__ + is_dummy_path = none_module.startswith(DUMMY_MODULES_FOLDER) or none_module.startswith( + TRANSFORMERS_DUMMY_MODULES_FOLDER + ) + if is_dummy_path and "dummy" in none_module: + # call class_obj for nice error message of missing requirements + class_obj() + + raise ValueError( + f"The component {class_obj} of {pipeline_class} cannot be loaded as it does not seem to have" + f" any of the loading methods defined in {ALL_IMPORTABLE_CLASSES}." + ) + + load_method = _get_load_method(class_obj, load_method_name, is_dduf=dduf_entries is not None) + + # add kwargs to loading method + diffusers_module = importlib.import_module(__name__.split(".")[0]) + loading_kwargs = {} + if issubclass(class_obj, torch.nn.Module): + loading_kwargs["torch_dtype"] = torch_dtype + if issubclass(class_obj, diffusers_module.OnnxRuntimeModel): + loading_kwargs["provider"] = provider + loading_kwargs["sess_options"] = sess_options + loading_kwargs["provider_options"] = provider_options + + is_diffusers_model = issubclass(class_obj, diffusers_module.ModelMixin) + + if is_transformers_available(): + transformers_version = version.parse(version.parse(transformers.__version__).base_version) + else: + transformers_version = "N/A" + + is_transformers_model = ( + is_transformers_available() + and issubclass(class_obj, PreTrainedModel) + and transformers_version >= version.parse("4.20.0") + ) + + # When loading a transformers model, if the device_map is None, the weights will be initialized as opposed to diffusers. + # To make default loading faster we set the `low_cpu_mem_usage=low_cpu_mem_usage` flag which is `True` by default. + # This makes sure that the weights won't be initialized which significantly speeds up loading. + if is_diffusers_model or is_transformers_model: + loading_kwargs["device_map"] = device_map + loading_kwargs["max_memory"] = max_memory + loading_kwargs["offload_folder"] = offload_folder + loading_kwargs["offload_state_dict"] = offload_state_dict + loading_kwargs["variant"] = model_variants.pop(name, None) + loading_kwargs["use_safetensors"] = use_safetensors + + if from_flax: + loading_kwargs["from_flax"] = True + + # the following can be deleted once the minimum required `transformers` version + # is higher than 4.27 + if ( + is_transformers_model + and loading_kwargs["variant"] is not None + and transformers_version < version.parse("4.27.0") + ): + raise ImportError( + f"When passing `variant='{variant}'`, please make sure to upgrade your `transformers` version to at least 4.27.0.dev0" + ) + elif is_transformers_model and loading_kwargs["variant"] is None: + loading_kwargs.pop("variant") + + # if `from_flax` and model is transformer model, can currently not load with `low_cpu_mem_usage` + if not (from_flax and is_transformers_model): + loading_kwargs["low_cpu_mem_usage"] = low_cpu_mem_usage + else: + loading_kwargs["low_cpu_mem_usage"] = False + + if ( + quantization_config is not None + and isinstance(quantization_config, PipelineQuantizationConfig) + and issubclass(class_obj, torch.nn.Module) + ): + model_quant_config = quantization_config._resolve_quant_config( + is_diffusers=is_diffusers_model, module_name=name + ) + if model_quant_config is not None: + loading_kwargs["quantization_config"] = model_quant_config + + # check if the module is in a subdirectory + if dduf_entries: + loading_kwargs["dduf_entries"] = dduf_entries + loaded_sub_model = load_method(name, **loading_kwargs) + elif os.path.isdir(os.path.join(cached_folder, name)): + loaded_sub_model = load_method(os.path.join(cached_folder, name), **loading_kwargs) + else: + # else load from the root directory + loaded_sub_model = load_method(cached_folder, **loading_kwargs) + + if isinstance(loaded_sub_model, torch.nn.Module) and isinstance(device_map, dict): + # remove hooks + remove_hook_from_module(loaded_sub_model, recurse=True) + needs_offloading_to_cpu = device_map[""] == "cpu" + + if needs_offloading_to_cpu: + dispatch_model( + loaded_sub_model, + state_dict=loaded_sub_model.state_dict(), + device_map=device_map, + force_hooks=True, + main_device=0, + ) + else: + dispatch_model(loaded_sub_model, device_map=device_map, force_hooks=True) + + return loaded_sub_model + + +def _get_load_method(class_obj: object, load_method_name: str, is_dduf: bool) -> Callable: + """ + Return the method to load the sub model. + + In practice, this method will return the `"from_pretrained"` (or `load_method_name`) method of the class object + except if loading from a DDUF checkpoint. In that case, transformers models and tokenizers have a specific loading + method that we need to use. + """ + if is_dduf: + if issubclass(class_obj, PreTrainedTokenizerBase): + return lambda *args, **kwargs: _load_tokenizer_from_dduf(class_obj, *args, **kwargs) + if issubclass(class_obj, PreTrainedModel): + return lambda *args, **kwargs: _load_transformers_model_from_dduf(class_obj, *args, **kwargs) + return getattr(class_obj, load_method_name) + + +def _fetch_class_library_tuple(module): + # import it here to avoid circular import + diffusers_module = importlib.import_module(__name__.split(".")[0]) + pipelines = getattr(diffusers_module, "pipelines") + + # register the config from the original module, not the dynamo compiled one + not_compiled_module = _unwrap_model(module) + library = not_compiled_module.__module__.split(".")[0] + + # check if the module is a pipeline module + module_path_items = not_compiled_module.__module__.split(".") + pipeline_dir = module_path_items[-2] if len(module_path_items) > 2 else None + + path = not_compiled_module.__module__.split(".") + is_pipeline_module = pipeline_dir in path and hasattr(pipelines, pipeline_dir) + + # if library is not in LOADABLE_CLASSES, then it is a custom module. + # Or if it's a pipeline module, then the module is inside the pipeline + # folder so we set the library to module name. + if is_pipeline_module: + library = pipeline_dir + elif library not in LOADABLE_CLASSES: + library = not_compiled_module.__module__ + + # retrieve class_name + class_name = not_compiled_module.__class__.__name__ + + return (library, class_name) + + +def _identify_model_variants(folder: str, variant: str, config: dict) -> dict: + model_variants = {} + if variant is not None: + for sub_folder in os.listdir(folder): + folder_path = os.path.join(folder, sub_folder) + is_folder = os.path.isdir(folder_path) and sub_folder in config + variant_exists = is_folder and any(p.split(".")[1].startswith(variant) for p in os.listdir(folder_path)) + if variant_exists: + model_variants[sub_folder] = variant + return model_variants + + +def _resolve_custom_pipeline_and_cls(folder, config, custom_pipeline): + custom_class_name = None + if os.path.isfile(os.path.join(folder, f"{custom_pipeline}.py")): + custom_pipeline = os.path.join(folder, f"{custom_pipeline}.py") + elif isinstance(config["_class_name"], (list, tuple)) and os.path.isfile( + os.path.join(folder, f"{config['_class_name'][0]}.py") + ): + custom_pipeline = os.path.join(folder, f"{config['_class_name'][0]}.py") + custom_class_name = config["_class_name"][1] + + return custom_pipeline, custom_class_name + + +def _maybe_raise_warning_for_inpainting(pipeline_class, pretrained_model_name_or_path: str, config: dict): + if pipeline_class.__name__ == "StableDiffusionInpaintPipeline" and version.parse( + version.parse(config["_diffusers_version"]).base_version + ) <= version.parse("0.5.1"): + from diffusers import StableDiffusionInpaintPipeline, StableDiffusionInpaintPipelineLegacy + + pipeline_class = StableDiffusionInpaintPipelineLegacy + + deprecation_message = ( + "You are using a legacy checkpoint for inpainting with Stable Diffusion, therefore we are loading the" + f" {StableDiffusionInpaintPipelineLegacy} class instead of {StableDiffusionInpaintPipeline}. For" + " better inpainting results, we strongly suggest using Stable Diffusion's official inpainting" + " checkpoint: https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-inpainting instead or adapting your" + f" checkpoint {pretrained_model_name_or_path} to the format of" + " https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-inpainting. Note that we do not actively maintain" + " the {StableDiffusionInpaintPipelineLegacy} class and will likely remove it in version 1.0.0." + ) + deprecate("StableDiffusionInpaintPipelineLegacy", "1.0.0", deprecation_message, standard_warn=False) + + +def _update_init_kwargs_with_connected_pipeline( + init_kwargs: dict, passed_pipe_kwargs: dict, passed_class_objs: dict, folder: str, **pipeline_loading_kwargs +) -> dict: + from .pipeline_utils import DiffusionPipeline + + modelcard = ModelCard.load(os.path.join(folder, "README.md")) + connected_pipes = {prefix: getattr(modelcard.data, prefix, [None])[0] for prefix in CONNECTED_PIPES_KEYS} + + # We don't scheduler argument to match the existing logic: + # https://github.com/huggingface/diffusers/blob/867e0c919e1aa7ef8b03c8eb1460f4f875a683ae/src/diffusers/pipelines/pipeline_utils.py#L906C13-L925C14 + pipeline_loading_kwargs_cp = pipeline_loading_kwargs.copy() + if pipeline_loading_kwargs_cp is not None and len(pipeline_loading_kwargs_cp) >= 1: + for k in pipeline_loading_kwargs: + if "scheduler" in k: + _ = pipeline_loading_kwargs_cp.pop(k) + + def get_connected_passed_kwargs(prefix): + connected_passed_class_obj = { + k.replace(f"{prefix}_", ""): w for k, w in passed_class_objs.items() if k.split("_")[0] == prefix + } + connected_passed_pipe_kwargs = { + k.replace(f"{prefix}_", ""): w for k, w in passed_pipe_kwargs.items() if k.split("_")[0] == prefix + } + + connected_passed_kwargs = {**connected_passed_class_obj, **connected_passed_pipe_kwargs} + return connected_passed_kwargs + + connected_pipes = { + prefix: DiffusionPipeline.from_pretrained( + repo_id, **pipeline_loading_kwargs_cp, **get_connected_passed_kwargs(prefix) + ) + for prefix, repo_id in connected_pipes.items() + if repo_id is not None + } + + for prefix, connected_pipe in connected_pipes.items(): + # add connected pipes to `init_kwargs` with _, e.g. "prior_text_encoder" + init_kwargs.update( + {"_".join([prefix, name]): component for name, component in connected_pipe.components.items()} + ) + + return init_kwargs + + +def _get_custom_components_and_folders( + pretrained_model_name: str, + config_dict: Dict[str, Any], + filenames: Optional[List[str]] = None, + variant_filenames: Optional[List[str]] = None, + variant: Optional[str] = None, +): + config_dict = config_dict.copy() + + # retrieve all folder_names that contain relevant files + folder_names = [k for k, v in config_dict.items() if isinstance(v, list) and k != "_class_name"] + + diffusers_module = importlib.import_module(__name__.split(".")[0]) + pipelines = getattr(diffusers_module, "pipelines") + + # optionally create a custom component <> custom file mapping + custom_components = {} + for component in folder_names: + module_candidate = config_dict[component][0] + + if module_candidate is None or not isinstance(module_candidate, str): + continue + + # We compute candidate file path on the Hub. Do not use `os.path.join`. + candidate_file = f"{component}/{module_candidate}.py" + + if candidate_file in filenames: + custom_components[component] = module_candidate + elif module_candidate not in LOADABLE_CLASSES and not hasattr(pipelines, module_candidate): + raise ValueError( + f"{candidate_file} as defined in `model_index.json` does not exist in {pretrained_model_name} and is not a module in 'diffusers/pipelines'." + ) + + return custom_components, folder_names + + +def _get_ignore_patterns( + passed_components, + model_folder_names: List[str], + model_filenames: List[str], + use_safetensors: bool, + from_flax: bool, + allow_pickle: bool, + use_onnx: bool, + is_onnx: bool, + variant: Optional[str] = None, +) -> List[str]: + if ( + use_safetensors + and not allow_pickle + and not is_safetensors_compatible( + model_filenames, passed_components=passed_components, folder_names=model_folder_names, variant=variant + ) + ): + raise EnvironmentError( + f"Could not find the necessary `safetensors` weights in {model_filenames} (variant={variant})" + ) + + if from_flax: + ignore_patterns = ["*.bin", "*.safetensors", "*.onnx", "*.pb"] + + elif use_safetensors and is_safetensors_compatible( + model_filenames, passed_components=passed_components, folder_names=model_folder_names, variant=variant + ): + ignore_patterns = ["*.bin", "*.msgpack"] + + use_onnx = use_onnx if use_onnx is not None else is_onnx + if not use_onnx: + ignore_patterns += ["*.onnx", "*.pb"] + + else: + ignore_patterns = ["*.safetensors", "*.msgpack"] + + use_onnx = use_onnx if use_onnx is not None else is_onnx + if not use_onnx: + ignore_patterns += ["*.onnx", "*.pb"] + + return ignore_patterns + + +def _download_dduf_file( + pretrained_model_name: str, + dduf_file: str, + pipeline_class_name: str, + cache_dir: str, + proxies: str, + local_files_only: bool, + token: str, + revision: str, +): + model_info_call_error = None + if not local_files_only: + try: + info = model_info(pretrained_model_name, token=token, revision=revision) + except (HTTPError, OfflineModeIsEnabled, requests.ConnectionError) as e: + logger.warning(f"Couldn't connect to the Hub: {e}.\nWill try to load from local cache.") + local_files_only = True + model_info_call_error = e # save error to reraise it if model is not cached locally + + if ( + not local_files_only + and dduf_file is not None + and dduf_file not in (sibling.rfilename for sibling in info.siblings) + ): + raise ValueError(f"Requested {dduf_file} file is not available in {pretrained_model_name}.") + + try: + user_agent = {"pipeline_class": pipeline_class_name, "dduf": True} + cached_folder = snapshot_download( + pretrained_model_name, + cache_dir=cache_dir, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + allow_patterns=[dduf_file], + user_agent=user_agent, + ) + return cached_folder + except FileNotFoundError: + # Means we tried to load pipeline with `local_files_only=True` but the files have not been found in local cache. + # This can happen in two cases: + # 1. If the user passed `local_files_only=True` => we raise the error directly + # 2. If we forced `local_files_only=True` when `model_info` failed => we raise the initial error + if model_info_call_error is None: + # 1. user passed `local_files_only=True` + raise + else: + # 2. we forced `local_files_only=True` when `model_info` failed + raise EnvironmentError( + f"Cannot load model {pretrained_model_name}: model is not cached locally and an error occurred" + " while trying to fetch metadata from the Hub. Please check out the root cause in the stacktrace" + " above." + ) from model_info_call_error + + +def _maybe_raise_error_for_incorrect_transformers(config_dict): + has_transformers_component = False + for k in config_dict: + if isinstance(config_dict[k], list): + has_transformers_component = config_dict[k][0] == "transformers" + if has_transformers_component: + break + if has_transformers_component and not is_transformers_version(">", "4.47.1"): + raise ValueError("Please upgrade your `transformers` installation to the latest version to use DDUF.") + + +def _maybe_warn_for_wrong_component_in_quant_config(pipe_init_dict, quant_config): + if quant_config is None: + return + + actual_pipe_components = set(pipe_init_dict.keys()) + missing = "" + quant_components = None + if getattr(quant_config, "components_to_quantize", None) is not None: + quant_components = set(quant_config.components_to_quantize) + elif getattr(quant_config, "quant_mapping", None) is not None and isinstance(quant_config.quant_mapping, dict): + quant_components = set(quant_config.quant_mapping.keys()) + + if quant_components and not quant_components.issubset(actual_pipe_components): + missing = quant_components - actual_pipe_components + + if missing: + logger.warning( + f"The following components in the quantization config {missing} will be ignored " + "as they do not belong to the underlying pipeline. Acceptable values for the pipeline " + f"components are: {', '.join(actual_pipe_components)}." + ) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/pipeline_utils.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/pipeline_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..2c03811b51ab3b3db3a9194d0a8bbcaba42745df --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/pipeline_utils.py @@ -0,0 +1,2147 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# 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 fnmatch +import importlib +import inspect +import os +import re +import sys +from dataclasses import dataclass +from pathlib import Path +from typing import Any, Callable, Dict, List, Optional, Union, get_args, get_origin + +import numpy as np +import PIL.Image +import requests +import torch +from huggingface_hub import ( + DDUFEntry, + ModelCard, + create_repo, + hf_hub_download, + model_info, + read_dduf_file, + snapshot_download, +) +from huggingface_hub.utils import OfflineModeIsEnabled, validate_hf_hub_args +from packaging import version +from requests.exceptions import HTTPError +from tqdm.auto import tqdm +from typing_extensions import Self + +from .. import __version__ +from ..configuration_utils import ConfigMixin +from ..models import AutoencoderKL +from ..models.attention_processor import FusedAttnProcessor2_0 +from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, ModelMixin +from ..quantizers import PipelineQuantizationConfig +from ..quantizers.bitsandbytes.utils import _check_bnb_status +from ..schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME +from ..utils import ( + CONFIG_NAME, + DEPRECATED_REVISION_ARGS, + BaseOutput, + PushToHubMixin, + _get_detailed_type, + _is_valid_type, + is_accelerate_available, + is_accelerate_version, + is_hpu_available, + is_torch_npu_available, + is_torch_version, + is_transformers_version, + logging, + numpy_to_pil, +) +from ..utils.hub_utils import _check_legacy_sharding_variant_format, load_or_create_model_card, populate_model_card +from ..utils.torch_utils import empty_device_cache, get_device, is_compiled_module + + +if is_torch_npu_available(): + import torch_npu # noqa: F401 + +from .pipeline_loading_utils import ( + ALL_IMPORTABLE_CLASSES, + CONNECTED_PIPES_KEYS, + CUSTOM_PIPELINE_FILE_NAME, + LOADABLE_CLASSES, + _download_dduf_file, + _fetch_class_library_tuple, + _get_custom_components_and_folders, + _get_custom_pipeline_class, + _get_final_device_map, + _get_ignore_patterns, + _get_pipeline_class, + _identify_model_variants, + _maybe_raise_error_for_incorrect_transformers, + _maybe_raise_warning_for_inpainting, + _maybe_warn_for_wrong_component_in_quant_config, + _resolve_custom_pipeline_and_cls, + _unwrap_model, + _update_init_kwargs_with_connected_pipeline, + filter_model_files, + load_sub_model, + maybe_raise_or_warn, + variant_compatible_siblings, + warn_deprecated_model_variant, +) + + +if is_accelerate_available(): + import accelerate + + +LIBRARIES = [] +for library in LOADABLE_CLASSES: + LIBRARIES.append(library) + +SUPPORTED_DEVICE_MAP = ["balanced"] + +logger = logging.get_logger(__name__) + + +@dataclass +class ImagePipelineOutput(BaseOutput): + """ + Output class for image pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] + + +@dataclass +class AudioPipelineOutput(BaseOutput): + """ + Output class for audio pipelines. + + Args: + audios (`np.ndarray`) + List of denoised audio samples of a NumPy array of shape `(batch_size, num_channels, sample_rate)`. + """ + + audios: np.ndarray + + +class DeprecatedPipelineMixin: + """ + A mixin that can be used to mark a pipeline as deprecated. + + Pipelines inheriting from this mixin will raise a warning when instantiated, indicating that they are deprecated + and won't receive updates past the specified version. Tests will be skipped for pipelines that inherit from this + mixin. + + Example usage: + ```python + class MyDeprecatedPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + _last_supported_version = "0.20.0" + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + ``` + """ + + # Override this in the inheriting class to specify the last version that will support this pipeline + _last_supported_version = None + + def __init__(self, *args, **kwargs): + # Get the class name for the warning message + class_name = self.__class__.__name__ + + # Get the last supported version or use the current version if not specified + version_info = getattr(self.__class__, "_last_supported_version", __version__) + + # Raise a warning that this pipeline is deprecated + logger.warning( + f"The {class_name} has been deprecated and will not receive bug fixes or feature updates after Diffusers version {version_info}. " + ) + + # Call the parent class's __init__ method + super().__init__(*args, **kwargs) + + +class DiffusionPipeline(ConfigMixin, PushToHubMixin): + r""" + Base class for all pipelines. + + [`DiffusionPipeline`] stores all components (models, schedulers, and processors) for diffusion pipelines and + provides methods for loading, downloading and saving models. It also includes methods to: + + - move all PyTorch modules to the device of your choice + - enable/disable the progress bar for the denoising iteration + + Class attributes: + + - **config_name** (`str`) -- The configuration filename that stores the class and module names of all the + diffusion pipeline's components. + - **_optional_components** (`List[str]`) -- List of all optional components that don't have to be passed to the + pipeline to function (should be overridden by subclasses). + """ + + config_name = "model_index.json" + model_cpu_offload_seq = None + hf_device_map = None + _optional_components = [] + _exclude_from_cpu_offload = [] + _load_connected_pipes = False + _is_onnx = False + + def register_modules(self, **kwargs): + for name, module in kwargs.items(): + # retrieve library + if module is None or isinstance(module, (tuple, list)) and module[0] is None: + register_dict = {name: (None, None)} + else: + library, class_name = _fetch_class_library_tuple(module) + register_dict = {name: (library, class_name)} + + # save model index config + self.register_to_config(**register_dict) + + # set models + setattr(self, name, module) + + def __setattr__(self, name: str, value: Any): + if name in self.__dict__ and hasattr(self.config, name): + # We need to overwrite the config if name exists in config + if isinstance(getattr(self.config, name), (tuple, list)): + if value is not None and self.config[name][0] is not None: + class_library_tuple = _fetch_class_library_tuple(value) + else: + class_library_tuple = (None, None) + + self.register_to_config(**{name: class_library_tuple}) + else: + self.register_to_config(**{name: value}) + + super().__setattr__(name, value) + + def save_pretrained( + self, + save_directory: Union[str, os.PathLike], + safe_serialization: bool = True, + variant: Optional[str] = None, + max_shard_size: Optional[Union[int, str]] = None, + push_to_hub: bool = False, + **kwargs, + ): + """ + Save all saveable variables of the pipeline to a directory. A pipeline variable can be saved and loaded if its + class implements both a save and loading method. The pipeline is easily reloaded using the + [`~DiffusionPipeline.from_pretrained`] class method. + + Arguments: + save_directory (`str` or `os.PathLike`): + Directory to save a pipeline to. Will be created if it doesn't exist. + safe_serialization (`bool`, *optional*, defaults to `True`): + Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`. + variant (`str`, *optional*): + If specified, weights are saved in the format `pytorch_model..bin`. + max_shard_size (`int` or `str`, defaults to `None`): + The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size + lower than this size. If expressed as a string, needs to be digits followed by a unit (like `"5GB"`). + If expressed as an integer, the unit is bytes. Note that this limit will be decreased after a certain + period of time (starting from Oct 2024) to allow users to upgrade to the latest version of `diffusers`. + This is to establish a common default size for this argument across different libraries in the Hugging + Face ecosystem (`transformers`, and `accelerate`, for example). + push_to_hub (`bool`, *optional*, defaults to `False`): + Whether or not to push your model to the Hugging Face model hub after saving it. You can specify the + repository you want to push to with `repo_id` (will default to the name of `save_directory` in your + namespace). + + kwargs (`Dict[str, Any]`, *optional*): + Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method. + """ + model_index_dict = dict(self.config) + model_index_dict.pop("_class_name", None) + model_index_dict.pop("_diffusers_version", None) + model_index_dict.pop("_module", None) + model_index_dict.pop("_name_or_path", None) + + if push_to_hub: + commit_message = kwargs.pop("commit_message", None) + private = kwargs.pop("private", None) + create_pr = kwargs.pop("create_pr", False) + token = kwargs.pop("token", None) + repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1]) + repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id + + expected_modules, optional_kwargs = self._get_signature_keys(self) + + def is_saveable_module(name, value): + if name not in expected_modules: + return False + if name in self._optional_components and value[0] is None: + return False + return True + + model_index_dict = {k: v for k, v in model_index_dict.items() if is_saveable_module(k, v)} + for pipeline_component_name in model_index_dict.keys(): + sub_model = getattr(self, pipeline_component_name) + model_cls = sub_model.__class__ + + # Dynamo wraps the original model in a private class. + # I didn't find a public API to get the original class. + if is_compiled_module(sub_model): + sub_model = _unwrap_model(sub_model) + model_cls = sub_model.__class__ + + save_method_name = None + # search for the model's base class in LOADABLE_CLASSES + for library_name, library_classes in LOADABLE_CLASSES.items(): + if library_name in sys.modules: + library = importlib.import_module(library_name) + else: + logger.info( + f"{library_name} is not installed. Cannot save {pipeline_component_name} as {library_classes} from {library_name}" + ) + + for base_class, save_load_methods in library_classes.items(): + class_candidate = getattr(library, base_class, None) + if class_candidate is not None and issubclass(model_cls, class_candidate): + # if we found a suitable base class in LOADABLE_CLASSES then grab its save method + save_method_name = save_load_methods[0] + break + if save_method_name is not None: + break + + if save_method_name is None: + logger.warning( + f"self.{pipeline_component_name}={sub_model} of type {type(sub_model)} cannot be saved." + ) + # make sure that unsaveable components are not tried to be loaded afterward + self.register_to_config(**{pipeline_component_name: (None, None)}) + continue + + save_method = getattr(sub_model, save_method_name) + + # Call the save method with the argument safe_serialization only if it's supported + save_method_signature = inspect.signature(save_method) + save_method_accept_safe = "safe_serialization" in save_method_signature.parameters + save_method_accept_variant = "variant" in save_method_signature.parameters + save_method_accept_max_shard_size = "max_shard_size" in save_method_signature.parameters + + save_kwargs = {} + if save_method_accept_safe: + save_kwargs["safe_serialization"] = safe_serialization + if save_method_accept_variant: + save_kwargs["variant"] = variant + if save_method_accept_max_shard_size and max_shard_size is not None: + # max_shard_size is expected to not be None in ModelMixin + save_kwargs["max_shard_size"] = max_shard_size + + save_method(os.path.join(save_directory, pipeline_component_name), **save_kwargs) + + # finally save the config + self.save_config(save_directory) + + if push_to_hub: + # Create a new empty model card and eventually tag it + model_card = load_or_create_model_card(repo_id, token=token, is_pipeline=True) + model_card = populate_model_card(model_card) + model_card.save(os.path.join(save_directory, "README.md")) + + self._upload_folder( + save_directory, + repo_id, + token=token, + commit_message=commit_message, + create_pr=create_pr, + ) + + def to(self, *args, **kwargs) -> Self: + r""" + Performs Pipeline dtype and/or device conversion. A torch.dtype and torch.device are inferred from the + arguments of `self.to(*args, **kwargs).` + + + + If the pipeline already has the correct torch.dtype and torch.device, then it is returned as is. Otherwise, + the returned pipeline is a copy of self with the desired torch.dtype and torch.device. + + + + + Here are the ways to call `to`: + + - `to(dtype, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the specified + [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype) + - `to(device, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the specified + [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device) + - `to(device=None, dtype=None, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the + specified [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device) and + [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype) + + Arguments: + dtype (`torch.dtype`, *optional*): + Returns a pipeline with the specified + [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype) + device (`torch.Device`, *optional*): + Returns a pipeline with the specified + [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device) + silence_dtype_warnings (`str`, *optional*, defaults to `False`): + Whether to omit warnings if the target `dtype` is not compatible with the target `device`. + + Returns: + [`DiffusionPipeline`]: The pipeline converted to specified `dtype` and/or `dtype`. + """ + dtype = kwargs.pop("dtype", None) + device = kwargs.pop("device", None) + silence_dtype_warnings = kwargs.pop("silence_dtype_warnings", False) + + dtype_arg = None + device_arg = None + if len(args) == 1: + if isinstance(args[0], torch.dtype): + dtype_arg = args[0] + else: + device_arg = torch.device(args[0]) if args[0] is not None else None + elif len(args) == 2: + if isinstance(args[0], torch.dtype): + raise ValueError( + "When passing two arguments, make sure the first corresponds to `device` and the second to `dtype`." + ) + device_arg = torch.device(args[0]) if args[0] is not None else None + dtype_arg = args[1] + elif len(args) > 2: + raise ValueError("Please make sure to pass at most two arguments (`device` and `dtype`) `.to(...)`") + + if dtype is not None and dtype_arg is not None: + raise ValueError( + "You have passed `dtype` both as an argument and as a keyword argument. Please only pass one of the two." + ) + + dtype = dtype or dtype_arg + + if device is not None and device_arg is not None: + raise ValueError( + "You have passed `device` both as an argument and as a keyword argument. Please only pass one of the two." + ) + + device = device or device_arg + device_type = torch.device(device).type if device is not None else None + pipeline_has_bnb = any(any((_check_bnb_status(module))) for _, module in self.components.items()) + + # throw warning if pipeline is in "offloaded"-mode but user tries to manually set to GPU. + def module_is_sequentially_offloaded(module): + if not is_accelerate_available() or is_accelerate_version("<", "0.14.0"): + return False + + _, _, is_loaded_in_8bit_bnb = _check_bnb_status(module) + + if is_loaded_in_8bit_bnb: + return False + + return hasattr(module, "_hf_hook") and ( + isinstance(module._hf_hook, accelerate.hooks.AlignDevicesHook) + or hasattr(module._hf_hook, "hooks") + and isinstance(module._hf_hook.hooks[0], accelerate.hooks.AlignDevicesHook) + ) + + def module_is_offloaded(module): + if not is_accelerate_available() or is_accelerate_version("<", "0.17.0.dev0"): + return False + + return hasattr(module, "_hf_hook") and isinstance(module._hf_hook, accelerate.hooks.CpuOffload) + + # .to("cuda") would raise an error if the pipeline is sequentially offloaded, so we raise our own to make it clearer + pipeline_is_sequentially_offloaded = any( + module_is_sequentially_offloaded(module) for _, module in self.components.items() + ) + + is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1 + if is_pipeline_device_mapped: + raise ValueError( + "It seems like you have activated a device mapping strategy on the pipeline which doesn't allow explicit device placement using `to()`. You can call `reset_device_map()` to remove the existing device map from the pipeline." + ) + + if device_type in ["cuda", "xpu"]: + if pipeline_is_sequentially_offloaded and not pipeline_has_bnb: + raise ValueError( + "It seems like you have activated sequential model offloading by calling `enable_sequential_cpu_offload`, but are now attempting to move the pipeline to GPU. This is not compatible with offloading. Please, move your pipeline `.to('cpu')` or consider removing the move altogether if you use sequential offloading." + ) + # PR: https://github.com/huggingface/accelerate/pull/3223/ + elif pipeline_has_bnb and is_accelerate_version("<", "1.1.0.dev0"): + raise ValueError( + "You are trying to call `.to('cuda')` on a pipeline that has models quantized with `bitsandbytes`. Your current `accelerate` installation does not support it. Please upgrade the installation." + ) + + # Display a warning in this case (the operation succeeds but the benefits are lost) + pipeline_is_offloaded = any(module_is_offloaded(module) for _, module in self.components.items()) + if pipeline_is_offloaded and device_type in ["cuda", "xpu"]: + logger.warning( + f"It seems like you have activated model offloading by calling `enable_model_cpu_offload`, but are now manually moving the pipeline to GPU. It is strongly recommended against doing so as memory gains from offloading are likely to be lost. Offloading automatically takes care of moving the individual components {', '.join(self.components.keys())} to GPU when needed. To make sure offloading works as expected, you should consider moving the pipeline back to CPU: `pipeline.to('cpu')` or removing the move altogether if you use offloading." + ) + + # Enable generic support for Intel Gaudi accelerator using GPU/HPU migration + if device_type == "hpu" and kwargs.pop("hpu_migration", True) and is_hpu_available(): + os.environ["PT_HPU_GPU_MIGRATION"] = "1" + logger.debug("Environment variable set: PT_HPU_GPU_MIGRATION=1") + + import habana_frameworks.torch # noqa: F401 + + # HPU hardware check + if not (hasattr(torch, "hpu") and torch.hpu.is_available()): + raise ValueError("You are trying to call `.to('hpu')` but HPU device is unavailable.") + + os.environ["PT_HPU_MAX_COMPOUND_OP_SIZE"] = "1" + logger.debug("Environment variable set: PT_HPU_MAX_COMPOUND_OP_SIZE=1") + + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [m for m in modules if isinstance(m, torch.nn.Module)] + + is_offloaded = pipeline_is_offloaded or pipeline_is_sequentially_offloaded + for module in modules: + _, is_loaded_in_4bit_bnb, is_loaded_in_8bit_bnb = _check_bnb_status(module) + is_group_offloaded = self._maybe_raise_error_if_group_offload_active(module=module) + + if (is_loaded_in_4bit_bnb or is_loaded_in_8bit_bnb) and dtype is not None: + logger.warning( + f"The module '{module.__class__.__name__}' has been loaded in `bitsandbytes` {'4bit' if is_loaded_in_4bit_bnb else '8bit'} and conversion to {dtype} is not supported. Module is still in {'4bit' if is_loaded_in_4bit_bnb else '8bit'} precision." + ) + + if is_loaded_in_8bit_bnb and device is not None: + logger.warning( + f"The module '{module.__class__.__name__}' has been loaded in `bitsandbytes` 8bit and moving it to {device} via `.to()` is not supported. Module is still on {module.device}." + ) + + # Note: we also handle this at the ModelMixin level. The reason for doing it here too is that modeling + # components can be from outside diffusers too, but still have group offloading enabled. + if ( + self._maybe_raise_error_if_group_offload_active(raise_error=False, module=module) + and device is not None + ): + logger.warning( + f"The module '{module.__class__.__name__}' is group offloaded and moving it to {device} via `.to()` is not supported." + ) + + # This can happen for `transformer` models. CPU placement was added in + # https://github.com/huggingface/transformers/pull/33122. So, we guard this accordingly. + if is_loaded_in_4bit_bnb and device is not None and is_transformers_version(">", "4.44.0"): + module.to(device=device) + elif not is_loaded_in_4bit_bnb and not is_loaded_in_8bit_bnb and not is_group_offloaded: + module.to(device, dtype) + + if ( + module.dtype == torch.float16 + and str(device) in ["cpu"] + and not silence_dtype_warnings + and not is_offloaded + ): + logger.warning( + "Pipelines loaded with `dtype=torch.float16` cannot run with `cpu` device. It" + " is not recommended to move them to `cpu` as running them will fail. Please make" + " sure to use an accelerator to run the pipeline in inference, due to the lack of" + " support for`float16` operations on this device in PyTorch. Please, remove the" + " `torch_dtype=torch.float16` argument, or use another device for inference." + ) + return self + + @property + def device(self) -> torch.device: + r""" + Returns: + `torch.device`: The torch device on which the pipeline is located. + """ + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [m for m in modules if isinstance(m, torch.nn.Module)] + + for module in modules: + return module.device + + return torch.device("cpu") + + @property + def dtype(self) -> torch.dtype: + r""" + Returns: + `torch.dtype`: The torch dtype on which the pipeline is located. + """ + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [m for m in modules if isinstance(m, torch.nn.Module)] + + for module in modules: + return module.dtype + + return torch.float32 + + @classmethod + @validate_hf_hub_args + def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs) -> Self: + r""" + Instantiate a PyTorch diffusion pipeline from pretrained pipeline weights. + + The pipeline is set in evaluation mode (`model.eval()`) by default. + + If you get the error message below, you need to finetune the weights for your downstream task: + + ``` + Some weights of UNet2DConditionModel were not initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5 and are newly initialized because the shapes did not match: + - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated + You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference. + ``` + + Parameters: + pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*): + Can be either: + + - A string, the *repo id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline + hosted on the Hub. + - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights + saved using + [`~DiffusionPipeline.save_pretrained`]. + - A path to a *directory* (for example `./my_pipeline_directory/`) containing a dduf file + torch_dtype (`torch.dtype` or `dict[str, Union[str, torch.dtype]]`, *optional*): + Override the default `torch.dtype` and load the model with another dtype. To load submodels with + different dtype pass a `dict` (for example `{'transformer': torch.bfloat16, 'vae': torch.float16}`). + Set the default dtype for unspecified components with `default` (for example `{'transformer': + torch.bfloat16, 'default': torch.float16}`). If a component is not specified and no default is set, + `torch.float32` is used. + custom_pipeline (`str`, *optional*): + + + + 🧪 This is an experimental feature and may change in the future. + + + + Can be either: + + - A string, the *repo id* (for example `hf-internal-testing/diffusers-dummy-pipeline`) of a custom + pipeline hosted on the Hub. The repository must contain a file called pipeline.py that defines + the custom pipeline. + - A string, the *file name* of a community pipeline hosted on GitHub under + [Community](https://github.com/huggingface/diffusers/tree/main/examples/community). Valid file + names must match the file name and not the pipeline script (`clip_guided_stable_diffusion` + instead of `clip_guided_stable_diffusion.py`). Community pipelines are always loaded from the + current main branch of GitHub. + - A path to a directory (`./my_pipeline_directory/`) containing a custom pipeline. The directory + must contain a file called `pipeline.py` that defines the custom pipeline. + + For more information on how to load and create custom pipelines, please have a look at [Loading and + Adding Custom + Pipelines](https://huggingface.co/docs/diffusers/using-diffusers/custom_pipeline_overview) + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + cache_dir (`Union[str, os.PathLike]`, *optional*): + Path to a directory where a downloaded pretrained model configuration is cached if the standard cache + is not used. + + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. + output_loading_info(`bool`, *optional*, defaults to `False`): + Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages. + local_files_only (`bool`, *optional*, defaults to `False`): + Whether to only load local model weights and configuration files or not. If set to `True`, the model + won't be downloaded from the Hub. + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from + `diffusers-cli login` (stored in `~/.huggingface`) is used. + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier + allowed by Git. + custom_revision (`str`, *optional*): + The specific model version to use. It can be a branch name, a tag name, or a commit id similar to + `revision` when loading a custom pipeline from the Hub. Defaults to the latest stable 🤗 Diffusers + version. + mirror (`str`, *optional*): + Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not + guarantee the timeliness or safety of the source, and you should refer to the mirror site for more + information. + device_map (`str`, *optional*): + Strategy that dictates how the different components of a pipeline should be placed on available + devices. Currently, only "balanced" `device_map` is supported. Check out + [this](https://huggingface.co/docs/diffusers/main/en/tutorials/inference_with_big_models#device-placement) + to know more. + max_memory (`Dict`, *optional*): + A dictionary device identifier for the maximum memory. Will default to the maximum memory available for + each GPU and the available CPU RAM if unset. + offload_folder (`str` or `os.PathLike`, *optional*): + The path to offload weights if device_map contains the value `"disk"`. + offload_state_dict (`bool`, *optional*): + If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if + the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True` + when there is some disk offload. + low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`): + Speed up model loading only loading the pretrained weights and not initializing the weights. This also + tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model. + Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this + argument to `True` will raise an error. + use_safetensors (`bool`, *optional*, defaults to `None`): + If set to `None`, the safetensors weights are downloaded if they're available **and** if the + safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors + weights. If set to `False`, safetensors weights are not loaded. + use_onnx (`bool`, *optional*, defaults to `None`): + If set to `True`, ONNX weights will always be downloaded if present. If set to `False`, ONNX weights + will never be downloaded. By default `use_onnx` defaults to the `_is_onnx` class attribute which is + `False` for non-ONNX pipelines and `True` for ONNX pipelines. ONNX weights include both files ending + with `.onnx` and `.pb`. + kwargs (remaining dictionary of keyword arguments, *optional*): + Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline + class). The overwritten components are passed directly to the pipelines `__init__` method. See example + below for more information. + variant (`str`, *optional*): + Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when + loading `from_flax`. + dduf_file(`str`, *optional*): + Load weights from the specified dduf file. + + + + To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with + `huggingface-cli login`. + + + + Examples: + + ```py + >>> from diffusers import DiffusionPipeline + + >>> # Download pipeline from huggingface.co and cache. + >>> pipeline = DiffusionPipeline.from_pretrained("CompVis/ldm-text2im-large-256") + + >>> # Download pipeline that requires an authorization token + >>> # For more information on access tokens, please refer to this section + >>> # of the documentation](https://huggingface.co/docs/hub/security-tokens) + >>> pipeline = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5") + + >>> # Use a different scheduler + >>> from diffusers import LMSDiscreteScheduler + + >>> scheduler = LMSDiscreteScheduler.from_config(pipeline.scheduler.config) + >>> pipeline.scheduler = scheduler + ``` + """ + # Copy the kwargs to re-use during loading connected pipeline. + kwargs_copied = kwargs.copy() + + cache_dir = kwargs.pop("cache_dir", None) + force_download = kwargs.pop("force_download", False) + proxies = kwargs.pop("proxies", None) + local_files_only = kwargs.pop("local_files_only", None) + token = kwargs.pop("token", None) + revision = kwargs.pop("revision", None) + from_flax = kwargs.pop("from_flax", False) + torch_dtype = kwargs.pop("torch_dtype", None) + custom_pipeline = kwargs.pop("custom_pipeline", None) + custom_revision = kwargs.pop("custom_revision", None) + provider = kwargs.pop("provider", None) + sess_options = kwargs.pop("sess_options", None) + provider_options = kwargs.pop("provider_options", None) + device_map = kwargs.pop("device_map", None) + max_memory = kwargs.pop("max_memory", None) + offload_folder = kwargs.pop("offload_folder", None) + offload_state_dict = kwargs.pop("offload_state_dict", None) + low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT) + variant = kwargs.pop("variant", None) + dduf_file = kwargs.pop("dduf_file", None) + use_safetensors = kwargs.pop("use_safetensors", None) + use_onnx = kwargs.pop("use_onnx", None) + load_connected_pipeline = kwargs.pop("load_connected_pipeline", False) + quantization_config = kwargs.pop("quantization_config", None) + + if torch_dtype is not None and not isinstance(torch_dtype, dict) and not isinstance(torch_dtype, torch.dtype): + torch_dtype = torch.float32 + logger.warning( + f"Passed `torch_dtype` {torch_dtype} is not a `torch.dtype`. Defaulting to `torch.float32`." + ) + + if low_cpu_mem_usage and not is_accelerate_available(): + low_cpu_mem_usage = False + logger.warning( + "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the" + " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install" + " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip" + " install accelerate\n```\n." + ) + + if quantization_config is not None and not isinstance(quantization_config, PipelineQuantizationConfig): + raise ValueError("`quantization_config` must be an instance of `PipelineQuantizationConfig`.") + + if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"): + raise NotImplementedError( + "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set" + " `low_cpu_mem_usage=False`." + ) + + if device_map is not None and not is_torch_version(">=", "1.9.0"): + raise NotImplementedError( + "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set" + " `device_map=None`." + ) + + if device_map is not None and not is_accelerate_available(): + raise NotImplementedError( + "Using `device_map` requires the `accelerate` library. Please install it using: `pip install accelerate`." + ) + + if device_map is not None and not isinstance(device_map, str): + raise ValueError("`device_map` must be a string.") + + if device_map is not None and device_map not in SUPPORTED_DEVICE_MAP: + raise NotImplementedError( + f"{device_map} not supported. Supported strategies are: {', '.join(SUPPORTED_DEVICE_MAP)}" + ) + + if device_map is not None and device_map in SUPPORTED_DEVICE_MAP: + if is_accelerate_version("<", "0.28.0"): + raise NotImplementedError("Device placement requires `accelerate` version `0.28.0` or later.") + + if low_cpu_mem_usage is False and device_map is not None: + raise ValueError( + f"You cannot set `low_cpu_mem_usage` to False while using device_map={device_map} for loading and" + " dispatching. Please make sure to set `low_cpu_mem_usage=True`." + ) + + if dduf_file: + if custom_pipeline: + raise NotImplementedError("Custom pipelines are not supported with DDUF at the moment.") + if load_connected_pipeline: + raise NotImplementedError("Connected pipelines are not supported with DDUF at the moment.") + + # 1. Download the checkpoints and configs + # use snapshot download here to get it working from from_pretrained + if not os.path.isdir(pretrained_model_name_or_path): + if pretrained_model_name_or_path.count("/") > 1: + raise ValueError( + f'The provided pretrained_model_name_or_path "{pretrained_model_name_or_path}"' + " is neither a valid local path nor a valid repo id. Please check the parameter." + ) + cached_folder = cls.download( + pretrained_model_name_or_path, + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + from_flax=from_flax, + use_safetensors=use_safetensors, + use_onnx=use_onnx, + custom_pipeline=custom_pipeline, + custom_revision=custom_revision, + variant=variant, + dduf_file=dduf_file, + load_connected_pipeline=load_connected_pipeline, + **kwargs, + ) + else: + cached_folder = pretrained_model_name_or_path + + # The variant filenames can have the legacy sharding checkpoint format that we check and throw + # a warning if detected. + if variant is not None and _check_legacy_sharding_variant_format(folder=cached_folder, variant=variant): + warn_msg = ( + f"Warning: The repository contains sharded checkpoints for variant '{variant}' maybe in a deprecated format. " + "Please check your files carefully:\n\n" + "- Correct format example: diffusion_pytorch_model.fp16-00003-of-00003.safetensors\n" + "- Deprecated format example: diffusion_pytorch_model-00001-of-00002.fp16.safetensors\n\n" + "If you find any files in the deprecated format:\n" + "1. Remove all existing checkpoint files for this variant.\n" + "2. Re-obtain the correct files by running `save_pretrained()`.\n\n" + "This will ensure you're using the most up-to-date and compatible checkpoint format." + ) + logger.warning(warn_msg) + + dduf_entries = None + if dduf_file: + dduf_file_path = os.path.join(cached_folder, dduf_file) + dduf_entries = read_dduf_file(dduf_file_path) + # The reader contains already all the files needed, no need to check it again + cached_folder = "" + + config_dict = cls.load_config(cached_folder, dduf_entries=dduf_entries) + + if dduf_file: + _maybe_raise_error_for_incorrect_transformers(config_dict) + + # pop out "_ignore_files" as it is only needed for download + config_dict.pop("_ignore_files", None) + + # 2. Define which model components should load variants + # We retrieve the information by matching whether variant model checkpoints exist in the subfolders. + # Example: `diffusion_pytorch_model.safetensors` -> `diffusion_pytorch_model.fp16.safetensors` + # with variant being `"fp16"`. + model_variants = _identify_model_variants(folder=cached_folder, variant=variant, config=config_dict) + if len(model_variants) == 0 and variant is not None: + error_message = f"You are trying to load the model files of the `variant={variant}`, but no such modeling files are available." + raise ValueError(error_message) + + # 3. Load the pipeline class, if using custom module then load it from the hub + # if we load from explicit class, let's use it + custom_pipeline, custom_class_name = _resolve_custom_pipeline_and_cls( + folder=cached_folder, config=config_dict, custom_pipeline=custom_pipeline + ) + pipeline_class = _get_pipeline_class( + cls, + config=config_dict, + load_connected_pipeline=load_connected_pipeline, + custom_pipeline=custom_pipeline, + class_name=custom_class_name, + cache_dir=cache_dir, + revision=custom_revision, + ) + + if device_map is not None and pipeline_class._load_connected_pipes: + raise NotImplementedError("`device_map` is not yet supported for connected pipelines.") + + # DEPRECATED: To be removed in 1.0.0 + # we are deprecating the `StableDiffusionInpaintPipelineLegacy` pipeline which gets loaded + # when a user requests for a `StableDiffusionInpaintPipeline` with `diffusers` version being <= 0.5.1. + _maybe_raise_warning_for_inpainting( + pipeline_class=pipeline_class, + pretrained_model_name_or_path=pretrained_model_name_or_path, + config=config_dict, + ) + + # 4. Define expected modules given pipeline signature + # and define non-None initialized modules (=`init_kwargs`) + + # some modules can be passed directly to the init + # in this case they are already instantiated in `kwargs` + # extract them here + expected_modules, optional_kwargs = cls._get_signature_keys(pipeline_class) + expected_types = pipeline_class._get_signature_types() + passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs} + passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs} + init_dict, unused_kwargs, _ = pipeline_class.extract_init_dict(config_dict, **kwargs) + + # define init kwargs and make sure that optional component modules are filtered out + init_kwargs = { + k: init_dict.pop(k) + for k in optional_kwargs + if k in init_dict and k not in pipeline_class._optional_components + } + init_kwargs = {**init_kwargs, **passed_pipe_kwargs} + + # remove `null` components + def load_module(name, value): + if value[0] is None: + return False + if name in passed_class_obj and passed_class_obj[name] is None: + return False + return True + + init_dict = {k: v for k, v in init_dict.items() if load_module(k, v)} + + # Special case: safety_checker must be loaded separately when using `from_flax` + if from_flax and "safety_checker" in init_dict and "safety_checker" not in passed_class_obj: + raise NotImplementedError( + "The safety checker cannot be automatically loaded when loading weights `from_flax`." + " Please, pass `safety_checker=None` to `from_pretrained`, and load the safety checker" + " separately if you need it." + ) + + # 5. Throw nice warnings / errors for fast accelerate loading + if len(unused_kwargs) > 0: + logger.warning( + f"Keyword arguments {unused_kwargs} are not expected by {pipeline_class.__name__} and will be ignored." + ) + + # import it here to avoid circular import + from diffusers import pipelines + + # 6. device map delegation + final_device_map = None + if device_map is not None: + final_device_map = _get_final_device_map( + device_map=device_map, + pipeline_class=pipeline_class, + passed_class_obj=passed_class_obj, + init_dict=init_dict, + library=library, + max_memory=max_memory, + torch_dtype=torch_dtype, + cached_folder=cached_folder, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + ) + + # 7. Load each module in the pipeline + current_device_map = None + _maybe_warn_for_wrong_component_in_quant_config(init_dict, quantization_config) + for name, (library_name, class_name) in logging.tqdm(init_dict.items(), desc="Loading pipeline components..."): + # 7.1 device_map shenanigans + if final_device_map is not None and len(final_device_map) > 0: + component_device = final_device_map.get(name, None) + if component_device is not None: + current_device_map = {"": component_device} + else: + current_device_map = None + + # 7.2 - now that JAX/Flax is an official framework of the library, we might load from Flax names + class_name = class_name[4:] if class_name.startswith("Flax") else class_name + + # 7.3 Define all importable classes + is_pipeline_module = hasattr(pipelines, library_name) + importable_classes = ALL_IMPORTABLE_CLASSES + loaded_sub_model = None + + # 7.4 Use passed sub model or load class_name from library_name + if name in passed_class_obj: + # if the model is in a pipeline module, then we load it from the pipeline + # check that passed_class_obj has correct parent class + maybe_raise_or_warn( + library_name, library, class_name, importable_classes, passed_class_obj, name, is_pipeline_module + ) + + loaded_sub_model = passed_class_obj[name] + else: + # load sub model + sub_model_dtype = ( + torch_dtype.get(name, torch_dtype.get("default", torch.float32)) + if isinstance(torch_dtype, dict) + else torch_dtype + ) + loaded_sub_model = load_sub_model( + library_name=library_name, + class_name=class_name, + importable_classes=importable_classes, + pipelines=pipelines, + is_pipeline_module=is_pipeline_module, + pipeline_class=pipeline_class, + torch_dtype=sub_model_dtype, + provider=provider, + sess_options=sess_options, + device_map=current_device_map, + max_memory=max_memory, + offload_folder=offload_folder, + offload_state_dict=offload_state_dict, + model_variants=model_variants, + name=name, + from_flax=from_flax, + variant=variant, + low_cpu_mem_usage=low_cpu_mem_usage, + cached_folder=cached_folder, + use_safetensors=use_safetensors, + dduf_entries=dduf_entries, + provider_options=provider_options, + quantization_config=quantization_config, + ) + logger.info( + f"Loaded {name} as {class_name} from `{name}` subfolder of {pretrained_model_name_or_path}." + ) + + init_kwargs[name] = loaded_sub_model # UNet(...), # DiffusionSchedule(...) + + # 8. Handle connected pipelines. + if pipeline_class._load_connected_pipes and os.path.isfile(os.path.join(cached_folder, "README.md")): + init_kwargs = _update_init_kwargs_with_connected_pipeline( + init_kwargs=init_kwargs, + passed_pipe_kwargs=passed_pipe_kwargs, + passed_class_objs=passed_class_obj, + folder=cached_folder, + **kwargs_copied, + ) + + # 9. Potentially add passed objects if expected + missing_modules = set(expected_modules) - set(init_kwargs.keys()) + passed_modules = list(passed_class_obj.keys()) + optional_modules = pipeline_class._optional_components + if len(missing_modules) > 0 and missing_modules <= set(passed_modules + optional_modules): + for module in missing_modules: + init_kwargs[module] = passed_class_obj.get(module, None) + elif len(missing_modules) > 0: + passed_modules = set(list(init_kwargs.keys()) + list(passed_class_obj.keys())) - set(optional_kwargs) + raise ValueError( + f"Pipeline {pipeline_class} expected {expected_modules}, but only {passed_modules} were passed." + ) + + # 10. Type checking init arguments + for kw, arg in init_kwargs.items(): + # Too complex to validate with type annotation alone + if "scheduler" in kw: + continue + # Many tokenizer annotations don't include its "Fast" variant, so skip this + # e.g T5Tokenizer but not T5TokenizerFast + elif "tokenizer" in kw: + continue + elif ( + arg is not None # Skip if None + and not expected_types[kw] == (inspect.Signature.empty,) # Skip if no type annotations + and not _is_valid_type(arg, expected_types[kw]) # Check type + ): + logger.warning(f"Expected types for {kw}: {expected_types[kw]}, got {_get_detailed_type(arg)}.") + + # 11. Instantiate the pipeline + model = pipeline_class(**init_kwargs) + + # 12. Save where the model was instantiated from + model.register_to_config(_name_or_path=pretrained_model_name_or_path) + if device_map is not None: + setattr(model, "hf_device_map", final_device_map) + return model + + @property + def name_or_path(self) -> str: + return getattr(self.config, "_name_or_path", None) + + @property + def _execution_device(self): + r""" + Returns the device on which the pipeline's models will be executed. After calling + [`~DiffusionPipeline.enable_sequential_cpu_offload`] the execution device can only be inferred from + Accelerate's module hooks. + """ + from ..hooks.group_offloading import _get_group_onload_device + + # When apply group offloading at the leaf_level, we're in the same situation as accelerate's sequential + # offloading. We need to return the onload device of the group offloading hooks so that the intermediates + # required for computation (latents, prompt embeddings, etc.) can be created on the correct device. + for name, model in self.components.items(): + if not isinstance(model, torch.nn.Module): + continue + try: + return _get_group_onload_device(model) + except ValueError: + pass + + for name, model in self.components.items(): + if not isinstance(model, torch.nn.Module) or name in self._exclude_from_cpu_offload: + continue + + if not hasattr(model, "_hf_hook"): + return self.device + for module in model.modules(): + if ( + hasattr(module, "_hf_hook") + and hasattr(module._hf_hook, "execution_device") + and module._hf_hook.execution_device is not None + ): + return torch.device(module._hf_hook.execution_device) + return self.device + + def remove_all_hooks(self): + r""" + Removes all hooks that were added when using `enable_sequential_cpu_offload` or `enable_model_cpu_offload`. + """ + for _, model in self.components.items(): + if isinstance(model, torch.nn.Module) and hasattr(model, "_hf_hook"): + accelerate.hooks.remove_hook_from_module(model, recurse=True) + self._all_hooks = [] + + def enable_model_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None): + r""" + Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared + to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the accelerator when its + `forward` method is called, and the model remains in accelerator until the next model runs. Memory savings are + lower than with `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution + of the `unet`. + + Arguments: + gpu_id (`int`, *optional*): + The ID of the accelerator that shall be used in inference. If not specified, it will default to 0. + device (`torch.Device` or `str`, *optional*, defaults to None): + The PyTorch device type of the accelerator that shall be used in inference. If not specified, it will + automatically detect the available accelerator and use. + """ + self._maybe_raise_error_if_group_offload_active(raise_error=True) + + is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1 + if is_pipeline_device_mapped: + raise ValueError( + "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_model_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_model_cpu_offload()`." + ) + + if self.model_cpu_offload_seq is None: + raise ValueError( + "Model CPU offload cannot be enabled because no `model_cpu_offload_seq` class attribute is set." + ) + + if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"): + from accelerate import cpu_offload_with_hook + else: + raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.") + + self.remove_all_hooks() + + if device is None: + device = get_device() + if device == "cpu": + raise RuntimeError("`enable_model_cpu_offload` requires accelerator, but not found") + + torch_device = torch.device(device) + device_index = torch_device.index + + if gpu_id is not None and device_index is not None: + raise ValueError( + f"You have passed both `gpu_id`={gpu_id} and an index as part of the passed device `device`={device}" + f"Cannot pass both. Please make sure to either not define `gpu_id` or not pass the index as part of the device: `device`={torch_device.type}" + ) + + # _offload_gpu_id should be set to passed gpu_id (or id in passed `device`) or default to previously set id or default to 0 + self._offload_gpu_id = gpu_id or torch_device.index or getattr(self, "_offload_gpu_id", 0) + + device_type = torch_device.type + device = torch.device(f"{device_type}:{self._offload_gpu_id}") + self._offload_device = device + + self.to("cpu", silence_dtype_warnings=True) + empty_device_cache(device.type) + + all_model_components = {k: v for k, v in self.components.items() if isinstance(v, torch.nn.Module)} + + self._all_hooks = [] + hook = None + for model_str in self.model_cpu_offload_seq.split("->"): + model = all_model_components.pop(model_str, None) + + if not isinstance(model, torch.nn.Module): + continue + + # This is because the model would already be placed on a CUDA device. + _, _, is_loaded_in_8bit_bnb = _check_bnb_status(model) + if is_loaded_in_8bit_bnb: + logger.info( + f"Skipping the hook placement for the {model.__class__.__name__} as it is loaded in `bitsandbytes` 8bit." + ) + continue + + _, hook = cpu_offload_with_hook(model, device, prev_module_hook=hook) + self._all_hooks.append(hook) + + # CPU offload models that are not in the seq chain unless they are explicitly excluded + # these models will stay on CPU until maybe_free_model_hooks is called + # some models cannot be in the seq chain because they are iteratively called, such as controlnet + for name, model in all_model_components.items(): + if not isinstance(model, torch.nn.Module): + continue + + if name in self._exclude_from_cpu_offload: + model.to(device) + else: + _, hook = cpu_offload_with_hook(model, device) + self._all_hooks.append(hook) + + def maybe_free_model_hooks(self): + r""" + Method that performs the following: + - Offloads all components. + - Removes all model hooks that were added when using `enable_model_cpu_offload`, and then applies them again. + In case the model has not been offloaded, this function is a no-op. + - Resets stateful diffusers hooks of denoiser components if they were added with + [`~hooks.HookRegistry.register_hook`]. + + Make sure to add this function to the end of the `__call__` function of your pipeline so that it functions + correctly when applying `enable_model_cpu_offload`. + """ + for component in self.components.values(): + if hasattr(component, "_reset_stateful_cache"): + component._reset_stateful_cache() + + if not hasattr(self, "_all_hooks") or len(self._all_hooks) == 0: + # `enable_model_cpu_offload` has not be called, so silently do nothing + return + + # make sure the model is in the same state as before calling it + self.enable_model_cpu_offload(device=getattr(self, "_offload_device", "cuda")) + + def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None): + r""" + Offloads all models to CPU using 🤗 Accelerate, significantly reducing memory usage. When called, the state + dicts of all `torch.nn.Module` components (except those in `self._exclude_from_cpu_offload`) are saved to CPU + and then moved to `torch.device('meta')` and loaded to accelerator only when their specific submodule has its + `forward` method called. Offloading happens on a submodule basis. Memory savings are higher than with + `enable_model_cpu_offload`, but performance is lower. + + Arguments: + gpu_id (`int`, *optional*): + The ID of the accelerator that shall be used in inference. If not specified, it will default to 0. + device (`torch.Device` or `str`, *optional*, defaults to None): + The PyTorch device type of the accelerator that shall be used in inference. If not specified, it will + automatically detect the available accelerator and use. + """ + self._maybe_raise_error_if_group_offload_active(raise_error=True) + + if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"): + from accelerate import cpu_offload + else: + raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher") + self.remove_all_hooks() + + is_pipeline_device_mapped = self.hf_device_map is not None and len(self.hf_device_map) > 1 + if is_pipeline_device_mapped: + raise ValueError( + "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_sequential_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_sequential_cpu_offload()`." + ) + + if device is None: + device = get_device() + if device == "cpu": + raise RuntimeError("`enable_sequential_cpu_offload` requires accelerator, but not found") + + torch_device = torch.device(device) + device_index = torch_device.index + + if gpu_id is not None and device_index is not None: + raise ValueError( + f"You have passed both `gpu_id`={gpu_id} and an index as part of the passed device `device`={device}" + f"Cannot pass both. Please make sure to either not define `gpu_id` or not pass the index as part of the device: `device`={torch_device.type}" + ) + + # _offload_gpu_id should be set to passed gpu_id (or id in passed `device`) or default to previously set id or default to 0 + self._offload_gpu_id = gpu_id or torch_device.index or getattr(self, "_offload_gpu_id", 0) + + device_type = torch_device.type + device = torch.device(f"{device_type}:{self._offload_gpu_id}") + self._offload_device = device + + if self.device.type != "cpu": + orig_device_type = self.device.type + self.to("cpu", silence_dtype_warnings=True) + empty_device_cache(orig_device_type) + + for name, model in self.components.items(): + if not isinstance(model, torch.nn.Module): + continue + + if name in self._exclude_from_cpu_offload: + model.to(device) + else: + # make sure to offload buffers if not all high level weights + # are of type nn.Module + offload_buffers = len(model._parameters) > 0 + cpu_offload(model, device, offload_buffers=offload_buffers) + + def reset_device_map(self): + r""" + Resets the device maps (if any) to None. + """ + if self.hf_device_map is None: + return + else: + self.remove_all_hooks() + for name, component in self.components.items(): + if isinstance(component, torch.nn.Module): + component.to("cpu") + self.hf_device_map = None + + @classmethod + @validate_hf_hub_args + def download(cls, pretrained_model_name, **kwargs) -> Union[str, os.PathLike]: + r""" + Download and cache a PyTorch diffusion pipeline from pretrained pipeline weights. + + Parameters: + pretrained_model_name (`str` or `os.PathLike`, *optional*): + A string, the *repository id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline + hosted on the Hub. + custom_pipeline (`str`, *optional*): + Can be either: + + - A string, the *repository id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained + pipeline hosted on the Hub. The repository must contain a file called `pipeline.py` that defines + the custom pipeline. + + - A string, the *file name* of a community pipeline hosted on GitHub under + [Community](https://github.com/huggingface/diffusers/tree/main/examples/community). Valid file + names must match the file name and not the pipeline script (`clip_guided_stable_diffusion` + instead of `clip_guided_stable_diffusion.py`). Community pipelines are always loaded from the + current `main` branch of GitHub. + + - A path to a *directory* (`./my_pipeline_directory/`) containing a custom pipeline. The directory + must contain a file called `pipeline.py` that defines the custom pipeline. + + + + 🧪 This is an experimental feature and may change in the future. + + + + For more information on how to load and create custom pipelines, take a look at [How to contribute a + community pipeline](https://huggingface.co/docs/diffusers/main/en/using-diffusers/contribute_pipeline). + + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. + output_loading_info(`bool`, *optional*, defaults to `False`): + Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages. + local_files_only (`bool`, *optional*, defaults to `False`): + Whether to only load local model weights and configuration files or not. If set to `True`, the model + won't be downloaded from the Hub. + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from + `diffusers-cli login` (stored in `~/.huggingface`) is used. + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier + allowed by Git. + custom_revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, or a commit id similar to + `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a + custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub. + mirror (`str`, *optional*): + Mirror source to resolve accessibility issues if you're downloading a model in China. We do not + guarantee the timeliness or safety of the source, and you should refer to the mirror site for more + information. + variant (`str`, *optional*): + Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when + loading `from_flax`. + dduf_file(`str`, *optional*): + Load weights from the specified DDUF file. + use_safetensors (`bool`, *optional*, defaults to `None`): + If set to `None`, the safetensors weights are downloaded if they're available **and** if the + safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors + weights. If set to `False`, safetensors weights are not loaded. + use_onnx (`bool`, *optional*, defaults to `False`): + If set to `True`, ONNX weights will always be downloaded if present. If set to `False`, ONNX weights + will never be downloaded. By default `use_onnx` defaults to the `_is_onnx` class attribute which is + `False` for non-ONNX pipelines and `True` for ONNX pipelines. ONNX weights include both files ending + with `.onnx` and `.pb`. + trust_remote_code (`bool`, *optional*, defaults to `False`): + Whether or not to allow for custom pipelines and components defined on the Hub in their own files. This + option should only be set to `True` for repositories you trust and in which you have read the code, as + it will execute code present on the Hub on your local machine. + + Returns: + `os.PathLike`: + A path to the downloaded pipeline. + + + + To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in with + `huggingface-cli login`. + + + + """ + cache_dir = kwargs.pop("cache_dir", None) + force_download = kwargs.pop("force_download", False) + proxies = kwargs.pop("proxies", None) + local_files_only = kwargs.pop("local_files_only", None) + token = kwargs.pop("token", None) + revision = kwargs.pop("revision", None) + from_flax = kwargs.pop("from_flax", False) + custom_pipeline = kwargs.pop("custom_pipeline", None) + custom_revision = kwargs.pop("custom_revision", None) + variant = kwargs.pop("variant", None) + use_safetensors = kwargs.pop("use_safetensors", None) + use_onnx = kwargs.pop("use_onnx", None) + load_connected_pipeline = kwargs.pop("load_connected_pipeline", False) + trust_remote_code = kwargs.pop("trust_remote_code", False) + dduf_file: Optional[Dict[str, DDUFEntry]] = kwargs.pop("dduf_file", None) + + if dduf_file: + if custom_pipeline: + raise NotImplementedError("Custom pipelines are not supported with DDUF at the moment.") + if load_connected_pipeline: + raise NotImplementedError("Connected pipelines are not supported with DDUF at the moment.") + return _download_dduf_file( + pretrained_model_name=pretrained_model_name, + dduf_file=dduf_file, + pipeline_class_name=cls.__name__, + cache_dir=cache_dir, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + ) + + allow_pickle = True if (use_safetensors is None or use_safetensors is False) else False + use_safetensors = use_safetensors if use_safetensors is not None else True + + allow_patterns = None + ignore_patterns = None + + model_info_call_error: Optional[Exception] = None + if not local_files_only: + try: + info = model_info(pretrained_model_name, token=token, revision=revision) + except (HTTPError, OfflineModeIsEnabled, requests.ConnectionError) as e: + logger.warning(f"Couldn't connect to the Hub: {e}.\nWill try to load from local cache.") + local_files_only = True + model_info_call_error = e # save error to reraise it if model is not cached locally + + if not local_files_only: + config_file = hf_hub_download( + pretrained_model_name, + cls.config_name, + cache_dir=cache_dir, + revision=revision, + proxies=proxies, + force_download=force_download, + token=token, + ) + config_dict = cls._dict_from_json_file(config_file) + ignore_filenames = config_dict.pop("_ignore_files", []) + + filenames = {sibling.rfilename for sibling in info.siblings} + if variant is not None and _check_legacy_sharding_variant_format(filenames=filenames, variant=variant): + warn_msg = ( + f"Warning: The repository contains sharded checkpoints for variant '{variant}' maybe in a deprecated format. " + "Please check your files carefully:\n\n" + "- Correct format example: diffusion_pytorch_model.fp16-00003-of-00003.safetensors\n" + "- Deprecated format example: diffusion_pytorch_model-00001-of-00002.fp16.safetensors\n\n" + "If you find any files in the deprecated format:\n" + "1. Remove all existing checkpoint files for this variant.\n" + "2. Re-obtain the correct files by running `save_pretrained()`.\n\n" + "This will ensure you're using the most up-to-date and compatible checkpoint format." + ) + logger.warning(warn_msg) + + filenames = set(filenames) - set(ignore_filenames) + if revision in DEPRECATED_REVISION_ARGS and version.parse( + version.parse(__version__).base_version + ) >= version.parse("0.22.0"): + warn_deprecated_model_variant(pretrained_model_name, token, variant, revision, filenames) + + custom_components, folder_names = _get_custom_components_and_folders( + pretrained_model_name, config_dict, filenames, variant + ) + custom_class_name = None + if custom_pipeline is None and isinstance(config_dict["_class_name"], (list, tuple)): + custom_pipeline = config_dict["_class_name"][0] + custom_class_name = config_dict["_class_name"][1] + + load_pipe_from_hub = custom_pipeline is not None and f"{custom_pipeline}.py" in filenames + load_components_from_hub = len(custom_components) > 0 + + if load_pipe_from_hub and not trust_remote_code: + raise ValueError( + f"The repository for {pretrained_model_name} contains custom code in {custom_pipeline}.py which must be executed to correctly " + f"load the model. You can inspect the repository content at https://hf.co/{pretrained_model_name}/blob/main/{custom_pipeline}.py.\n" + f"Please pass the argument `trust_remote_code=True` to allow custom code to be run." + ) + + if load_components_from_hub and not trust_remote_code: + raise ValueError( + f"The repository for {pretrained_model_name} contains custom code in {'.py, '.join([os.path.join(k, v) for k, v in custom_components.items()])} which must be executed to correctly " + f"load the model. You can inspect the repository content at {', '.join([f'https://hf.co/{pretrained_model_name}/{k}/{v}.py' for k, v in custom_components.items()])}.\n" + f"Please pass the argument `trust_remote_code=True` to allow custom code to be run." + ) + + # retrieve passed components that should not be downloaded + pipeline_class = _get_pipeline_class( + cls, + config_dict, + load_connected_pipeline=load_connected_pipeline, + custom_pipeline=custom_pipeline, + repo_id=pretrained_model_name if load_pipe_from_hub else None, + hub_revision=revision, + class_name=custom_class_name, + cache_dir=cache_dir, + revision=custom_revision, + ) + expected_components, _ = cls._get_signature_keys(pipeline_class) + passed_components = [k for k in expected_components if k in kwargs] + + # retrieve the names of the folders containing model weights + model_folder_names = { + os.path.split(f)[0] for f in filter_model_files(filenames) if os.path.split(f)[0] in folder_names + } + # retrieve all patterns that should not be downloaded and error out when needed + ignore_patterns = _get_ignore_patterns( + passed_components, + model_folder_names, + filenames, + use_safetensors, + from_flax, + allow_pickle, + use_onnx, + pipeline_class._is_onnx, + variant, + ) + + model_filenames, variant_filenames = variant_compatible_siblings( + filenames, variant=variant, ignore_patterns=ignore_patterns + ) + + # all filenames compatible with variant will be added + allow_patterns = list(model_filenames) + + # allow all patterns from non-model folders + # this enables downloading schedulers, tokenizers, ... + allow_patterns += [f"{k}/*" for k in folder_names if k not in model_folder_names] + # add custom component files + allow_patterns += [f"{k}/{f}.py" for k, f in custom_components.items()] + # add custom pipeline file + allow_patterns += [f"{custom_pipeline}.py"] if f"{custom_pipeline}.py" in filenames else [] + # also allow downloading config.json files with the model + allow_patterns += [os.path.join(k, "config.json") for k in model_folder_names] + allow_patterns += [ + SCHEDULER_CONFIG_NAME, + CONFIG_NAME, + cls.config_name, + CUSTOM_PIPELINE_FILE_NAME, + ] + + # Don't download any objects that are passed + allow_patterns = [ + p for p in allow_patterns if not (len(p.split("/")) == 2 and p.split("/")[0] in passed_components) + ] + + if pipeline_class._load_connected_pipes: + allow_patterns.append("README.md") + + # Don't download index files of forbidden patterns either + ignore_patterns = ignore_patterns + [f"{i}.index.*json" for i in ignore_patterns] + re_ignore_pattern = [re.compile(fnmatch.translate(p)) for p in ignore_patterns] + re_allow_pattern = [re.compile(fnmatch.translate(p)) for p in allow_patterns] + + expected_files = [f for f in filenames if not any(p.match(f) for p in re_ignore_pattern)] + expected_files = [f for f in expected_files if any(p.match(f) for p in re_allow_pattern)] + + snapshot_folder = Path(config_file).parent + pipeline_is_cached = all((snapshot_folder / f).is_file() for f in expected_files) + + if pipeline_is_cached and not force_download: + # if the pipeline is cached, we can directly return it + # else call snapshot_download + return snapshot_folder + + user_agent = {"pipeline_class": cls.__name__} + if custom_pipeline is not None and not custom_pipeline.endswith(".py"): + user_agent["custom_pipeline"] = custom_pipeline + + # download all allow_patterns - ignore_patterns + try: + cached_folder = snapshot_download( + pretrained_model_name, + cache_dir=cache_dir, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + allow_patterns=allow_patterns, + ignore_patterns=ignore_patterns, + user_agent=user_agent, + ) + + cls_name = cls.load_config(os.path.join(cached_folder, "model_index.json")).get("_class_name", None) + cls_name = cls_name[4:] if isinstance(cls_name, str) and cls_name.startswith("Flax") else cls_name + + diffusers_module = importlib.import_module(__name__.split(".")[0]) + pipeline_class = getattr(diffusers_module, cls_name, None) if isinstance(cls_name, str) else None + + if pipeline_class is not None and pipeline_class._load_connected_pipes: + modelcard = ModelCard.load(os.path.join(cached_folder, "README.md")) + connected_pipes = sum([getattr(modelcard.data, k, []) for k in CONNECTED_PIPES_KEYS], []) + for connected_pipe_repo_id in connected_pipes: + download_kwargs = { + "cache_dir": cache_dir, + "force_download": force_download, + "proxies": proxies, + "local_files_only": local_files_only, + "token": token, + "variant": variant, + "use_safetensors": use_safetensors, + } + DiffusionPipeline.download(connected_pipe_repo_id, **download_kwargs) + + return cached_folder + + except FileNotFoundError: + # Means we tried to load pipeline with `local_files_only=True` but the files have not been found in local cache. + # This can happen in two cases: + # 1. If the user passed `local_files_only=True` => we raise the error directly + # 2. If we forced `local_files_only=True` when `model_info` failed => we raise the initial error + if model_info_call_error is None: + # 1. user passed `local_files_only=True` + raise + else: + # 2. we forced `local_files_only=True` when `model_info` failed + raise EnvironmentError( + f"Cannot load model {pretrained_model_name}: model is not cached locally and an error occurred" + " while trying to fetch metadata from the Hub. Please check out the root cause in the stacktrace" + " above." + ) from model_info_call_error + + @classmethod + def _get_signature_keys(cls, obj): + parameters = inspect.signature(obj.__init__).parameters + required_parameters = {k: v for k, v in parameters.items() if v.default == inspect._empty} + optional_parameters = set({k for k, v in parameters.items() if v.default != inspect._empty}) + expected_modules = set(required_parameters.keys()) - {"self"} + + optional_names = list(optional_parameters) + for name in optional_names: + if name in cls._optional_components: + expected_modules.add(name) + optional_parameters.remove(name) + + return sorted(expected_modules), sorted(optional_parameters) + + @classmethod + def _get_signature_types(cls): + signature_types = {} + for k, v in inspect.signature(cls.__init__).parameters.items(): + if inspect.isclass(v.annotation): + signature_types[k] = (v.annotation,) + elif get_origin(v.annotation) == Union: + signature_types[k] = get_args(v.annotation) + elif get_origin(v.annotation) in [List, Dict, list, dict]: + signature_types[k] = (v.annotation,) + else: + logger.warning(f"cannot get type annotation for Parameter {k} of {cls}.") + return signature_types + + @property + def components(self) -> Dict[str, Any]: + r""" + The `self.components` property can be useful to run different pipelines with the same weights and + configurations without reallocating additional memory. + + Returns (`dict`): + A dictionary containing all the modules needed to initialize the pipeline. + + Examples: + + ```py + >>> from diffusers import ( + ... StableDiffusionPipeline, + ... StableDiffusionImg2ImgPipeline, + ... StableDiffusionInpaintPipeline, + ... ) + + >>> text2img = StableDiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5") + >>> img2img = StableDiffusionImg2ImgPipeline(**text2img.components) + >>> inpaint = StableDiffusionInpaintPipeline(**text2img.components) + ``` + """ + expected_modules, optional_parameters = self._get_signature_keys(self) + components = { + k: getattr(self, k) for k in self.config.keys() if not k.startswith("_") and k not in optional_parameters + } + + actual = sorted(set(components.keys())) + expected = sorted(expected_modules) + if actual != expected: + raise ValueError( + f"{self} has been incorrectly initialized or {self.__class__} is incorrectly implemented. Expected" + f" {expected} to be defined, but {actual} are defined." + ) + + return components + + @staticmethod + def numpy_to_pil(images): + """ + Convert a NumPy image or a batch of images to a PIL image. + """ + return numpy_to_pil(images) + + @torch.compiler.disable + def progress_bar(self, iterable=None, total=None): + if not hasattr(self, "_progress_bar_config"): + self._progress_bar_config = {} + elif not isinstance(self._progress_bar_config, dict): + raise ValueError( + f"`self._progress_bar_config` should be of type `dict`, but is {type(self._progress_bar_config)}." + ) + + if iterable is not None: + return tqdm(iterable, **self._progress_bar_config) + elif total is not None: + return tqdm(total=total, **self._progress_bar_config) + else: + raise ValueError("Either `total` or `iterable` has to be defined.") + + def set_progress_bar_config(self, **kwargs): + self._progress_bar_config = kwargs + + def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None): + r""" + Enable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/). When this + option is enabled, you should observe lower GPU memory usage and a potential speed up during inference. Speed + up during training is not guaranteed. + + + + ⚠️ When memory efficient attention and sliced attention are both enabled, memory efficient attention takes + precedent. + + + + Parameters: + attention_op (`Callable`, *optional*): + Override the default `None` operator for use as `op` argument to the + [`memory_efficient_attention()`](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.memory_efficient_attention) + function of xFormers. + + Examples: + + ```py + >>> import torch + >>> from diffusers import DiffusionPipeline + >>> from xformers.ops import MemoryEfficientAttentionFlashAttentionOp + + >>> pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16) + >>> pipe = pipe.to("cuda") + >>> pipe.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp) + >>> # Workaround for not accepting attention shape using VAE for Flash Attention + >>> pipe.vae.enable_xformers_memory_efficient_attention(attention_op=None) + ``` + """ + self.set_use_memory_efficient_attention_xformers(True, attention_op) + + def disable_xformers_memory_efficient_attention(self): + r""" + Disable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/). + """ + self.set_use_memory_efficient_attention_xformers(False) + + def set_use_memory_efficient_attention_xformers( + self, valid: bool, attention_op: Optional[Callable] = None + ) -> None: + # Recursively walk through all the children. + # Any children which exposes the set_use_memory_efficient_attention_xformers method + # gets the message + def fn_recursive_set_mem_eff(module: torch.nn.Module): + if hasattr(module, "set_use_memory_efficient_attention_xformers"): + module.set_use_memory_efficient_attention_xformers(valid, attention_op) + + for child in module.children(): + fn_recursive_set_mem_eff(child) + + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [m for m in modules if isinstance(m, torch.nn.Module)] + + for module in modules: + fn_recursive_set_mem_eff(module) + + def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"): + r""" + Enable sliced attention computation. When this option is enabled, the attention module splits the input tensor + in slices to compute attention in several steps. For more than one attention head, the computation is performed + sequentially over each head. This is useful to save some memory in exchange for a small speed decrease. + + + + ⚠️ Don't enable attention slicing if you're already using `scaled_dot_product_attention` (SDPA) from PyTorch + 2.0 or xFormers. These attention computations are already very memory efficient so you won't need to enable + this function. If you enable attention slicing with SDPA or xFormers, it can lead to serious slow downs! + + + + Args: + slice_size (`str` or `int`, *optional*, defaults to `"auto"`): + When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If + `"max"`, maximum amount of memory will be saved by running only one slice at a time. If a number is + provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim` + must be a multiple of `slice_size`. + + Examples: + + ```py + >>> import torch + >>> from diffusers import StableDiffusionPipeline + + >>> pipe = StableDiffusionPipeline.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", + ... torch_dtype=torch.float16, + ... use_safetensors=True, + ... ) + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> pipe.enable_attention_slicing() + >>> image = pipe(prompt).images[0] + ``` + """ + self.set_attention_slice(slice_size) + + def disable_attention_slicing(self): + r""" + Disable sliced attention computation. If `enable_attention_slicing` was previously called, attention is + computed in one step. + """ + # set slice_size = `None` to disable `attention slicing` + self.enable_attention_slicing(None) + + def set_attention_slice(self, slice_size: Optional[int]): + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [m for m in modules if isinstance(m, torch.nn.Module) and hasattr(m, "set_attention_slice")] + + for module in modules: + module.set_attention_slice(slice_size) + + @classmethod + def from_pipe(cls, pipeline, **kwargs): + r""" + Create a new pipeline from a given pipeline. This method is useful to create a new pipeline from the existing + pipeline components without reallocating additional memory. + + Arguments: + pipeline (`DiffusionPipeline`): + The pipeline from which to create a new pipeline. + + Returns: + `DiffusionPipeline`: + A new pipeline with the same weights and configurations as `pipeline`. + + Examples: + + ```py + >>> from diffusers import StableDiffusionPipeline, StableDiffusionSAGPipeline + + >>> pipe = StableDiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5") + >>> new_pipe = StableDiffusionSAGPipeline.from_pipe(pipe) + ``` + """ + + original_config = dict(pipeline.config) + torch_dtype = kwargs.pop("torch_dtype", torch.float32) + + # derive the pipeline class to instantiate + custom_pipeline = kwargs.pop("custom_pipeline", None) + custom_revision = kwargs.pop("custom_revision", None) + + if custom_pipeline is not None: + pipeline_class = _get_custom_pipeline_class(custom_pipeline, revision=custom_revision) + else: + pipeline_class = cls + + expected_modules, optional_kwargs = cls._get_signature_keys(pipeline_class) + # true_optional_modules are optional components with default value in signature so it is ok not to pass them to `__init__` + # e.g. `image_encoder` for StableDiffusionPipeline + parameters = inspect.signature(cls.__init__).parameters + true_optional_modules = set( + {k for k, v in parameters.items() if v.default != inspect._empty and k in expected_modules} + ) + + # get the class of each component based on its type hint + # e.g. {"unet": UNet2DConditionModel, "text_encoder": CLIPTextMode} + component_types = pipeline_class._get_signature_types() + + pretrained_model_name_or_path = original_config.pop("_name_or_path", None) + # allow users pass modules in `kwargs` to override the original pipeline's components + passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs} + + original_class_obj = {} + for name, component in pipeline.components.items(): + if name in expected_modules and name not in passed_class_obj: + # for model components, we will not switch over if the class does not matches the type hint in the new pipeline's signature + if ( + not isinstance(component, ModelMixin) + or type(component) in component_types[name] + or (component is None and name in cls._optional_components) + ): + original_class_obj[name] = component + else: + logger.warning( + f"component {name} is not switched over to new pipeline because type does not match the expected." + f" {name} is {type(component)} while the new pipeline expect {component_types[name]}." + f" please pass the component of the correct type to the new pipeline. `from_pipe(..., {name}={name})`" + ) + + # allow users pass optional kwargs to override the original pipelines config attribute + passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs} + original_pipe_kwargs = { + k: original_config[k] + for k in original_config.keys() + if k in optional_kwargs and k not in passed_pipe_kwargs + } + + # config attribute that were not expected by pipeline is stored as its private attribute + # (i.e. when the original pipeline was also instantiated with `from_pipe` from another pipeline that has this config) + # in this case, we will pass them as optional arguments if they can be accepted by the new pipeline + additional_pipe_kwargs = [ + k[1:] + for k in original_config.keys() + if k.startswith("_") and k[1:] in optional_kwargs and k[1:] not in passed_pipe_kwargs + ] + for k in additional_pipe_kwargs: + original_pipe_kwargs[k] = original_config.pop(f"_{k}") + + pipeline_kwargs = { + **passed_class_obj, + **original_class_obj, + **passed_pipe_kwargs, + **original_pipe_kwargs, + **kwargs, + } + + # store unused config as private attribute in the new pipeline + unused_original_config = { + f"{'' if k.startswith('_') else '_'}{k}": v for k, v in original_config.items() if k not in pipeline_kwargs + } + + missing_modules = ( + set(expected_modules) + - set(pipeline._optional_components) + - set(pipeline_kwargs.keys()) + - set(true_optional_modules) + ) + + if len(missing_modules) > 0: + raise ValueError( + f"Pipeline {pipeline_class} expected {expected_modules}, but only {set(list(passed_class_obj.keys()) + list(original_class_obj.keys()))} were passed" + ) + + new_pipeline = pipeline_class(**pipeline_kwargs) + if pretrained_model_name_or_path is not None: + new_pipeline.register_to_config(_name_or_path=pretrained_model_name_or_path) + new_pipeline.register_to_config(**unused_original_config) + + if torch_dtype is not None: + new_pipeline.to(dtype=torch_dtype) + + return new_pipeline + + def _maybe_raise_error_if_group_offload_active( + self, raise_error: bool = False, module: Optional[torch.nn.Module] = None + ) -> bool: + from ..hooks.group_offloading import _is_group_offload_enabled + + components = self.components.values() if module is None else [module] + components = [component for component in components if isinstance(component, torch.nn.Module)] + for component in components: + if _is_group_offload_enabled(component): + if raise_error: + raise ValueError( + "You are trying to apply model/sequential CPU offloading to a pipeline that contains components " + "with group offloading enabled. This is not supported. Please disable group offloading for " + "components of the pipeline to use other offloading methods." + ) + return True + return False + + +class StableDiffusionMixin: + r""" + Helper for DiffusionPipeline with vae and unet.(mainly for LDM such as stable diffusion) + """ + + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + def enable_freeu(self, s1: float, s2: float, b1: float, b2: float): + r"""Enables the FreeU mechanism as in https://huggingface.co/papers/2309.11497. + + The suffixes after the scaling factors represent the stages where they are being applied. + + Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values + that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL. + + Args: + s1 (`float`): + Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to + mitigate "oversmoothing effect" in the enhanced denoising process. + s2 (`float`): + Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to + mitigate "oversmoothing effect" in the enhanced denoising process. + b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. + b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. + """ + if not hasattr(self, "unet"): + raise ValueError("The pipeline must have `unet` for using FreeU.") + self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2) + + def disable_freeu(self): + """Disables the FreeU mechanism if enabled.""" + self.unet.disable_freeu() + + def fuse_qkv_projections(self, unet: bool = True, vae: bool = True): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value) + are fused. For cross-attention modules, key and value projection matrices are fused. + + + + This API is 🧪 experimental. + + + + Args: + unet (`bool`, defaults to `True`): To apply fusion on the UNet. + vae (`bool`, defaults to `True`): To apply fusion on the VAE. + """ + self.fusing_unet = False + self.fusing_vae = False + + if unet: + self.fusing_unet = True + self.unet.fuse_qkv_projections() + self.unet.set_attn_processor(FusedAttnProcessor2_0()) + + if vae: + if not isinstance(self.vae, AutoencoderKL): + raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.") + + self.fusing_vae = True + self.vae.fuse_qkv_projections() + self.vae.set_attn_processor(FusedAttnProcessor2_0()) + + def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True): + """Disable QKV projection fusion if enabled. + + + + This API is 🧪 experimental. + + + + Args: + unet (`bool`, defaults to `True`): To apply fusion on the UNet. + vae (`bool`, defaults to `True`): To apply fusion on the VAE. + + """ + if unet: + if not self.fusing_unet: + logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.") + else: + self.unet.unfuse_qkv_projections() + self.fusing_unet = False + + if vae: + if not self.fusing_vae: + logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.") + else: + self.vae.unfuse_qkv_projections() + self.fusing_vae = False diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b05a3ce2a7c9d7b388d422268b77f306419e40a6 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__init__.py @@ -0,0 +1,195 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_flax_available, + is_k_diffusion_available, + is_k_diffusion_version, + is_onnx_available, + is_torch_available, + is_transformers_available, + is_transformers_version, +) + + +_dummy_objects = {} +_additional_imports = {} +_import_structure = {"pipeline_output": ["StableDiffusionPipelineOutput"]} + +if is_transformers_available() and is_flax_available(): + _import_structure["pipeline_output"].extend(["FlaxStableDiffusionPipelineOutput"]) +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["clip_image_project_model"] = ["CLIPImageProjection"] + _import_structure["pipeline_stable_diffusion"] = ["StableDiffusionPipeline"] + _import_structure["pipeline_stable_diffusion_img2img"] = ["StableDiffusionImg2ImgPipeline"] + _import_structure["pipeline_stable_diffusion_inpaint"] = ["StableDiffusionInpaintPipeline"] + _import_structure["pipeline_stable_diffusion_instruct_pix2pix"] = ["StableDiffusionInstructPix2PixPipeline"] + _import_structure["pipeline_stable_diffusion_latent_upscale"] = ["StableDiffusionLatentUpscalePipeline"] + _import_structure["pipeline_stable_diffusion_upscale"] = ["StableDiffusionUpscalePipeline"] + _import_structure["pipeline_stable_unclip"] = ["StableUnCLIPPipeline"] + _import_structure["pipeline_stable_unclip_img2img"] = ["StableUnCLIPImg2ImgPipeline"] + _import_structure["safety_checker"] = ["StableDiffusionSafetyChecker"] + _import_structure["stable_unclip_image_normalizer"] = ["StableUnCLIPImageNormalizer"] +try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import ( + StableDiffusionImageVariationPipeline, + ) + + _dummy_objects.update({"StableDiffusionImageVariationPipeline": StableDiffusionImageVariationPipeline}) +else: + _import_structure["pipeline_stable_diffusion_image_variation"] = ["StableDiffusionImageVariationPipeline"] +try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.26.0")): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import ( + StableDiffusionDepth2ImgPipeline, + ) + + _dummy_objects.update( + { + "StableDiffusionDepth2ImgPipeline": StableDiffusionDepth2ImgPipeline, + } + ) +else: + _import_structure["pipeline_stable_diffusion_depth2img"] = ["StableDiffusionDepth2ImgPipeline"] + +try: + if not (is_transformers_available() and is_onnx_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_onnx_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_onnx_objects)) +else: + _import_structure["pipeline_onnx_stable_diffusion"] = [ + "OnnxStableDiffusionPipeline", + "StableDiffusionOnnxPipeline", + ] + _import_structure["pipeline_onnx_stable_diffusion_img2img"] = ["OnnxStableDiffusionImg2ImgPipeline"] + _import_structure["pipeline_onnx_stable_diffusion_inpaint"] = ["OnnxStableDiffusionInpaintPipeline"] + _import_structure["pipeline_onnx_stable_diffusion_inpaint_legacy"] = ["OnnxStableDiffusionInpaintPipelineLegacy"] + _import_structure["pipeline_onnx_stable_diffusion_upscale"] = ["OnnxStableDiffusionUpscalePipeline"] + +if is_transformers_available() and is_flax_available(): + from ...schedulers.scheduling_pndm_flax import PNDMSchedulerState + + _additional_imports.update({"PNDMSchedulerState": PNDMSchedulerState}) + _import_structure["pipeline_flax_stable_diffusion"] = ["FlaxStableDiffusionPipeline"] + _import_structure["pipeline_flax_stable_diffusion_img2img"] = ["FlaxStableDiffusionImg2ImgPipeline"] + _import_structure["pipeline_flax_stable_diffusion_inpaint"] = ["FlaxStableDiffusionInpaintPipeline"] + _import_structure["safety_checker_flax"] = ["FlaxStableDiffusionSafetyChecker"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + + else: + from .clip_image_project_model import CLIPImageProjection + from .pipeline_stable_diffusion import ( + StableDiffusionPipeline, + StableDiffusionPipelineOutput, + ) + from .pipeline_stable_diffusion_img2img import StableDiffusionImg2ImgPipeline + from .pipeline_stable_diffusion_inpaint import StableDiffusionInpaintPipeline + from .pipeline_stable_diffusion_instruct_pix2pix import ( + StableDiffusionInstructPix2PixPipeline, + ) + from .pipeline_stable_diffusion_latent_upscale import ( + StableDiffusionLatentUpscalePipeline, + ) + from .pipeline_stable_diffusion_upscale import StableDiffusionUpscalePipeline + from .pipeline_stable_unclip import StableUnCLIPPipeline + from .pipeline_stable_unclip_img2img import StableUnCLIPImg2ImgPipeline + from .safety_checker import StableDiffusionSafetyChecker + from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer + + try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import ( + StableDiffusionImageVariationPipeline, + ) + else: + from .pipeline_stable_diffusion_image_variation import ( + StableDiffusionImageVariationPipeline, + ) + + try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.26.0")): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import StableDiffusionDepth2ImgPipeline + else: + from .pipeline_stable_diffusion_depth2img import ( + StableDiffusionDepth2ImgPipeline, + ) + + try: + if not (is_transformers_available() and is_onnx_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_onnx_objects import * + else: + from .pipeline_onnx_stable_diffusion import ( + OnnxStableDiffusionPipeline, + StableDiffusionOnnxPipeline, + ) + from .pipeline_onnx_stable_diffusion_img2img import ( + OnnxStableDiffusionImg2ImgPipeline, + ) + from .pipeline_onnx_stable_diffusion_inpaint import ( + OnnxStableDiffusionInpaintPipeline, + ) + from .pipeline_onnx_stable_diffusion_upscale import ( + OnnxStableDiffusionUpscalePipeline, + ) + + try: + if not (is_transformers_available() and is_flax_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_flax_objects import * + else: + from .pipeline_flax_stable_diffusion import FlaxStableDiffusionPipeline + from .pipeline_flax_stable_diffusion_img2img import ( + FlaxStableDiffusionImg2ImgPipeline, + ) + from .pipeline_flax_stable_diffusion_inpaint import ( + FlaxStableDiffusionInpaintPipeline, + ) + from .pipeline_output import FlaxStableDiffusionPipelineOutput + from .safety_checker_flax import FlaxStableDiffusionSafetyChecker + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) + for name, value in _additional_imports.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__pycache__/pipeline_stable_unclip_img2img.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__pycache__/pipeline_stable_unclip_img2img.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..43ab2d3482a313e4e243f34e216d0a5d01f97d77 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__pycache__/pipeline_stable_unclip_img2img.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__pycache__/safety_checker.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__pycache__/safety_checker.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..353bccce9b68210a838566c36c67811c37cb81b2 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__pycache__/safety_checker.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__pycache__/safety_checker_flax.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__pycache__/safety_checker_flax.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0be189bc7c7a94df948c18428b6e5f1dedec72d9 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__pycache__/safety_checker_flax.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__pycache__/stable_unclip_image_normalizer.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__pycache__/stable_unclip_image_normalizer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a684feb69ed2b5cfce050e772ff274cb3b2081bb Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/__pycache__/stable_unclip_image_normalizer.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/clip_image_project_model.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/clip_image_project_model.py new file mode 100644 index 0000000000000000000000000000000000000000..30dd90242d07d249e62e976c148a9d2cde02271e --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/clip_image_project_model.py @@ -0,0 +1,29 @@ +# Copyright 2025 The GLIGEN Authors and HuggingFace Team. All rights reserved. +# +# 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 torch import nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models.modeling_utils import ModelMixin + + +class CLIPImageProjection(ModelMixin, ConfigMixin): + @register_to_config + def __init__(self, hidden_size: int = 768): + super().__init__() + self.hidden_size = hidden_size + self.project = nn.Linear(self.hidden_size, self.hidden_size, bias=False) + + def forward(self, x): + return self.project(x) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/convert_from_ckpt.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/convert_from_ckpt.py new file mode 100644 index 0000000000000000000000000000000000000000..6c0221d2092a46dfac6129ab5bbe783c753ab97c --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/convert_from_ckpt.py @@ -0,0 +1,1877 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# +# 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. +"""Conversion script for the Stable Diffusion checkpoints.""" + +import re +from contextlib import nullcontext +from io import BytesIO +from typing import Dict, Optional, Union + +import requests +import torch +import yaml +from transformers import ( + AutoFeatureExtractor, + BertTokenizerFast, + CLIPImageProcessor, + CLIPTextConfig, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionConfig, + CLIPVisionModelWithProjection, +) + +from ...models import ( + AutoencoderKL, + ControlNetModel, + PriorTransformer, + UNet2DConditionModel, +) +from ...schedulers import ( + DDIMScheduler, + DDPMScheduler, + DPMSolverMultistepScheduler, + EulerAncestralDiscreteScheduler, + EulerDiscreteScheduler, + HeunDiscreteScheduler, + LMSDiscreteScheduler, + PNDMScheduler, + UnCLIPScheduler, +) +from ...utils import is_accelerate_available, logging +from ...utils.constants import DIFFUSERS_REQUEST_TIMEOUT +from ...utils.torch_utils import get_device +from ..latent_diffusion.pipeline_latent_diffusion import LDMBertConfig, LDMBertModel +from ..paint_by_example import PaintByExampleImageEncoder +from ..pipeline_utils import DiffusionPipeline +from .safety_checker import StableDiffusionSafetyChecker +from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer + + +if is_accelerate_available(): + from accelerate import init_empty_weights + from accelerate.utils import set_module_tensor_to_device + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def shave_segments(path, n_shave_prefix_segments=1): + """ + Removes segments. Positive values shave the first segments, negative shave the last segments. + """ + if n_shave_prefix_segments >= 0: + return ".".join(path.split(".")[n_shave_prefix_segments:]) + else: + return ".".join(path.split(".")[:n_shave_prefix_segments]) + + +def renew_resnet_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside resnets to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item.replace("in_layers.0", "norm1") + new_item = new_item.replace("in_layers.2", "conv1") + + new_item = new_item.replace("out_layers.0", "norm2") + new_item = new_item.replace("out_layers.3", "conv2") + + new_item = new_item.replace("emb_layers.1", "time_emb_proj") + new_item = new_item.replace("skip_connection", "conv_shortcut") + + new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def renew_vae_resnet_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside resnets to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item + + new_item = new_item.replace("nin_shortcut", "conv_shortcut") + new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def renew_attention_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside attentions to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item + + # new_item = new_item.replace('norm.weight', 'group_norm.weight') + # new_item = new_item.replace('norm.bias', 'group_norm.bias') + + # new_item = new_item.replace('proj_out.weight', 'proj_attn.weight') + # new_item = new_item.replace('proj_out.bias', 'proj_attn.bias') + + # new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def renew_vae_attention_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside attentions to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item + + new_item = new_item.replace("norm.weight", "group_norm.weight") + new_item = new_item.replace("norm.bias", "group_norm.bias") + + new_item = new_item.replace("q.weight", "to_q.weight") + new_item = new_item.replace("q.bias", "to_q.bias") + + new_item = new_item.replace("k.weight", "to_k.weight") + new_item = new_item.replace("k.bias", "to_k.bias") + + new_item = new_item.replace("v.weight", "to_v.weight") + new_item = new_item.replace("v.bias", "to_v.bias") + + new_item = new_item.replace("proj_out.weight", "to_out.0.weight") + new_item = new_item.replace("proj_out.bias", "to_out.0.bias") + + new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def assign_to_checkpoint( + paths, checkpoint, old_checkpoint, attention_paths_to_split=None, additional_replacements=None, config=None +): + """ + This does the final conversion step: take locally converted weights and apply a global renaming to them. It splits + attention layers, and takes into account additional replacements that may arise. + + Assigns the weights to the new checkpoint. + """ + assert isinstance(paths, list), "Paths should be a list of dicts containing 'old' and 'new' keys." + + # Splits the attention layers into three variables. + if attention_paths_to_split is not None: + for path, path_map in attention_paths_to_split.items(): + old_tensor = old_checkpoint[path] + channels = old_tensor.shape[0] // 3 + + target_shape = (-1, channels) if len(old_tensor.shape) == 3 else (-1) + + num_heads = old_tensor.shape[0] // config["num_head_channels"] // 3 + + old_tensor = old_tensor.reshape((num_heads, 3 * channels // num_heads) + old_tensor.shape[1:]) + query, key, value = old_tensor.split(channels // num_heads, dim=1) + + checkpoint[path_map["query"]] = query.reshape(target_shape) + checkpoint[path_map["key"]] = key.reshape(target_shape) + checkpoint[path_map["value"]] = value.reshape(target_shape) + + for path in paths: + new_path = path["new"] + + # These have already been assigned + if attention_paths_to_split is not None and new_path in attention_paths_to_split: + continue + + # Global renaming happens here + new_path = new_path.replace("middle_block.0", "mid_block.resnets.0") + new_path = new_path.replace("middle_block.1", "mid_block.attentions.0") + new_path = new_path.replace("middle_block.2", "mid_block.resnets.1") + + if additional_replacements is not None: + for replacement in additional_replacements: + new_path = new_path.replace(replacement["old"], replacement["new"]) + + # proj_attn.weight has to be converted from conv 1D to linear + is_attn_weight = "proj_attn.weight" in new_path or ("attentions" in new_path and "to_" in new_path) + shape = old_checkpoint[path["old"]].shape + if is_attn_weight and len(shape) == 3: + checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0] + elif is_attn_weight and len(shape) == 4: + checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0, 0] + else: + checkpoint[new_path] = old_checkpoint[path["old"]] + + +def conv_attn_to_linear(checkpoint): + keys = list(checkpoint.keys()) + attn_keys = ["query.weight", "key.weight", "value.weight"] + for key in keys: + if ".".join(key.split(".")[-2:]) in attn_keys: + if checkpoint[key].ndim > 2: + checkpoint[key] = checkpoint[key][:, :, 0, 0] + elif "proj_attn.weight" in key: + if checkpoint[key].ndim > 2: + checkpoint[key] = checkpoint[key][:, :, 0] + + +def create_unet_diffusers_config(original_config, image_size: int, controlnet=False): + """ + Creates a config for the diffusers based on the config of the LDM model. + """ + if controlnet: + unet_params = original_config["model"]["params"]["control_stage_config"]["params"] + else: + if ( + "unet_config" in original_config["model"]["params"] + and original_config["model"]["params"]["unet_config"] is not None + ): + unet_params = original_config["model"]["params"]["unet_config"]["params"] + else: + unet_params = original_config["model"]["params"]["network_config"]["params"] + + vae_params = original_config["model"]["params"]["first_stage_config"]["params"]["ddconfig"] + + block_out_channels = [unet_params["model_channels"] * mult for mult in unet_params["channel_mult"]] + + down_block_types = [] + resolution = 1 + for i in range(len(block_out_channels)): + block_type = "CrossAttnDownBlock2D" if resolution in unet_params["attention_resolutions"] else "DownBlock2D" + down_block_types.append(block_type) + if i != len(block_out_channels) - 1: + resolution *= 2 + + up_block_types = [] + for i in range(len(block_out_channels)): + block_type = "CrossAttnUpBlock2D" if resolution in unet_params["attention_resolutions"] else "UpBlock2D" + up_block_types.append(block_type) + resolution //= 2 + + if unet_params["transformer_depth"] is not None: + transformer_layers_per_block = ( + unet_params["transformer_depth"] + if isinstance(unet_params["transformer_depth"], int) + else list(unet_params["transformer_depth"]) + ) + else: + transformer_layers_per_block = 1 + + vae_scale_factor = 2 ** (len(vae_params["ch_mult"]) - 1) + + head_dim = unet_params["num_heads"] if "num_heads" in unet_params else None + use_linear_projection = ( + unet_params["use_linear_in_transformer"] if "use_linear_in_transformer" in unet_params else False + ) + if use_linear_projection: + # stable diffusion 2-base-512 and 2-768 + if head_dim is None: + head_dim_mult = unet_params["model_channels"] // unet_params["num_head_channels"] + head_dim = [head_dim_mult * c for c in list(unet_params["channel_mult"])] + + class_embed_type = None + addition_embed_type = None + addition_time_embed_dim = None + projection_class_embeddings_input_dim = None + context_dim = None + + if unet_params["context_dim"] is not None: + context_dim = ( + unet_params["context_dim"] + if isinstance(unet_params["context_dim"], int) + else unet_params["context_dim"][0] + ) + + if "num_classes" in unet_params: + if unet_params["num_classes"] == "sequential": + if context_dim in [2048, 1280]: + # SDXL + addition_embed_type = "text_time" + addition_time_embed_dim = 256 + else: + class_embed_type = "projection" + assert "adm_in_channels" in unet_params + projection_class_embeddings_input_dim = unet_params["adm_in_channels"] + + config = { + "sample_size": image_size // vae_scale_factor, + "in_channels": unet_params["in_channels"], + "down_block_types": tuple(down_block_types), + "block_out_channels": tuple(block_out_channels), + "layers_per_block": unet_params["num_res_blocks"], + "cross_attention_dim": context_dim, + "attention_head_dim": head_dim, + "use_linear_projection": use_linear_projection, + "class_embed_type": class_embed_type, + "addition_embed_type": addition_embed_type, + "addition_time_embed_dim": addition_time_embed_dim, + "projection_class_embeddings_input_dim": projection_class_embeddings_input_dim, + "transformer_layers_per_block": transformer_layers_per_block, + } + + if "disable_self_attentions" in unet_params: + config["only_cross_attention"] = unet_params["disable_self_attentions"] + + if "num_classes" in unet_params and isinstance(unet_params["num_classes"], int): + config["num_class_embeds"] = unet_params["num_classes"] + + if controlnet: + config["conditioning_channels"] = unet_params["hint_channels"] + else: + config["out_channels"] = unet_params["out_channels"] + config["up_block_types"] = tuple(up_block_types) + + return config + + +def create_vae_diffusers_config(original_config, image_size: int): + """ + Creates a config for the diffusers based on the config of the LDM model. + """ + vae_params = original_config["model"]["params"]["first_stage_config"]["params"]["ddconfig"] + _ = original_config["model"]["params"]["first_stage_config"]["params"]["embed_dim"] + + block_out_channels = [vae_params["ch"] * mult for mult in vae_params["ch_mult"]] + down_block_types = [ + "DownEncoderBlock2D" if image_size // 2**i not in vae_params["attn_resolutions"] else "AttnDownEncoderBlock2D" + for i, _ in enumerate(block_out_channels) + ] + up_block_types = [ + "UpDecoderBlock2D" if image_size // 2**i not in vae_params["attn_resolutions"] else "AttnUpDecoderBlock2D" + for i, _ in enumerate(block_out_channels) + ][::-1] + + config = { + "sample_size": image_size, + "in_channels": vae_params["in_channels"], + "out_channels": vae_params["out_ch"], + "down_block_types": tuple(down_block_types), + "up_block_types": tuple(up_block_types), + "block_out_channels": tuple(block_out_channels), + "latent_channels": vae_params["z_channels"], + "layers_per_block": vae_params["num_res_blocks"], + } + return config + + +def create_diffusers_schedular(original_config): + schedular = DDIMScheduler( + num_train_timesteps=original_config["model"]["params"]["timesteps"], + beta_start=original_config["model"]["params"]["linear_start"], + beta_end=original_config["model"]["params"]["linear_end"], + beta_schedule="scaled_linear", + ) + return schedular + + +def create_ldm_bert_config(original_config): + bert_params = original_config["model"]["params"]["cond_stage_config"]["params"] + config = LDMBertConfig( + d_model=bert_params.n_embed, + encoder_layers=bert_params.n_layer, + encoder_ffn_dim=bert_params.n_embed * 4, + ) + return config + + +def convert_ldm_unet_checkpoint( + checkpoint, config, path=None, extract_ema=False, controlnet=False, skip_extract_state_dict=False +): + """ + Takes a state dict and a config, and returns a converted checkpoint. + """ + + if skip_extract_state_dict: + unet_state_dict = checkpoint + else: + # extract state_dict for UNet + unet_state_dict = {} + keys = list(checkpoint.keys()) + + if controlnet: + unet_key = "control_model." + else: + unet_key = "model.diffusion_model." + + # at least a 100 parameters have to start with `model_ema` in order for the checkpoint to be EMA + if sum(k.startswith("model_ema") for k in keys) > 100 and extract_ema: + logger.warning(f"Checkpoint {path} has both EMA and non-EMA weights.") + logger.warning( + "In this conversion only the EMA weights are extracted. If you want to instead extract the non-EMA" + " weights (useful to continue fine-tuning), please make sure to remove the `--extract_ema` flag." + ) + for key in keys: + if key.startswith("model.diffusion_model"): + flat_ema_key = "model_ema." + "".join(key.split(".")[1:]) + unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(flat_ema_key) + else: + if sum(k.startswith("model_ema") for k in keys) > 100: + logger.warning( + "In this conversion only the non-EMA weights are extracted. If you want to instead extract the EMA" + " weights (usually better for inference), please make sure to add the `--extract_ema` flag." + ) + + for key in keys: + if key.startswith(unet_key): + unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(key) + + new_checkpoint = {} + + new_checkpoint["time_embedding.linear_1.weight"] = unet_state_dict["time_embed.0.weight"] + new_checkpoint["time_embedding.linear_1.bias"] = unet_state_dict["time_embed.0.bias"] + new_checkpoint["time_embedding.linear_2.weight"] = unet_state_dict["time_embed.2.weight"] + new_checkpoint["time_embedding.linear_2.bias"] = unet_state_dict["time_embed.2.bias"] + + if config["class_embed_type"] is None: + # No parameters to port + ... + elif config["class_embed_type"] == "timestep" or config["class_embed_type"] == "projection": + new_checkpoint["class_embedding.linear_1.weight"] = unet_state_dict["label_emb.0.0.weight"] + new_checkpoint["class_embedding.linear_1.bias"] = unet_state_dict["label_emb.0.0.bias"] + new_checkpoint["class_embedding.linear_2.weight"] = unet_state_dict["label_emb.0.2.weight"] + new_checkpoint["class_embedding.linear_2.bias"] = unet_state_dict["label_emb.0.2.bias"] + else: + raise NotImplementedError(f"Not implemented `class_embed_type`: {config['class_embed_type']}") + + if config["addition_embed_type"] == "text_time": + new_checkpoint["add_embedding.linear_1.weight"] = unet_state_dict["label_emb.0.0.weight"] + new_checkpoint["add_embedding.linear_1.bias"] = unet_state_dict["label_emb.0.0.bias"] + new_checkpoint["add_embedding.linear_2.weight"] = unet_state_dict["label_emb.0.2.weight"] + new_checkpoint["add_embedding.linear_2.bias"] = unet_state_dict["label_emb.0.2.bias"] + + # Relevant to StableDiffusionUpscalePipeline + if "num_class_embeds" in config: + if (config["num_class_embeds"] is not None) and ("label_emb.weight" in unet_state_dict): + new_checkpoint["class_embedding.weight"] = unet_state_dict["label_emb.weight"] + + new_checkpoint["conv_in.weight"] = unet_state_dict["input_blocks.0.0.weight"] + new_checkpoint["conv_in.bias"] = unet_state_dict["input_blocks.0.0.bias"] + + if not controlnet: + new_checkpoint["conv_norm_out.weight"] = unet_state_dict["out.0.weight"] + new_checkpoint["conv_norm_out.bias"] = unet_state_dict["out.0.bias"] + new_checkpoint["conv_out.weight"] = unet_state_dict["out.2.weight"] + new_checkpoint["conv_out.bias"] = unet_state_dict["out.2.bias"] + + # Retrieves the keys for the input blocks only + num_input_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "input_blocks" in layer}) + input_blocks = { + layer_id: [key for key in unet_state_dict if f"input_blocks.{layer_id}" in key] + for layer_id in range(num_input_blocks) + } + + # Retrieves the keys for the middle blocks only + num_middle_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "middle_block" in layer}) + middle_blocks = { + layer_id: [key for key in unet_state_dict if f"middle_block.{layer_id}" in key] + for layer_id in range(num_middle_blocks) + } + + # Retrieves the keys for the output blocks only + num_output_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "output_blocks" in layer}) + output_blocks = { + layer_id: [key for key in unet_state_dict if f"output_blocks.{layer_id}" in key] + for layer_id in range(num_output_blocks) + } + + for i in range(1, num_input_blocks): + block_id = (i - 1) // (config["layers_per_block"] + 1) + layer_in_block_id = (i - 1) % (config["layers_per_block"] + 1) + + resnets = [ + key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key + ] + attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key] + + if f"input_blocks.{i}.0.op.weight" in unet_state_dict: + new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = unet_state_dict.pop( + f"input_blocks.{i}.0.op.weight" + ) + new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = unet_state_dict.pop( + f"input_blocks.{i}.0.op.bias" + ) + + paths = renew_resnet_paths(resnets) + meta_path = {"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"} + assign_to_checkpoint( + paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + + if len(attentions): + paths = renew_attention_paths(attentions) + + meta_path = {"old": f"input_blocks.{i}.1", "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}"} + assign_to_checkpoint( + paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + + resnet_0 = middle_blocks[0] + attentions = middle_blocks[1] + resnet_1 = middle_blocks[2] + + resnet_0_paths = renew_resnet_paths(resnet_0) + assign_to_checkpoint(resnet_0_paths, new_checkpoint, unet_state_dict, config=config) + + resnet_1_paths = renew_resnet_paths(resnet_1) + assign_to_checkpoint(resnet_1_paths, new_checkpoint, unet_state_dict, config=config) + + attentions_paths = renew_attention_paths(attentions) + meta_path = {"old": "middle_block.1", "new": "mid_block.attentions.0"} + assign_to_checkpoint( + attentions_paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + + for i in range(num_output_blocks): + block_id = i // (config["layers_per_block"] + 1) + layer_in_block_id = i % (config["layers_per_block"] + 1) + output_block_layers = [shave_segments(name, 2) for name in output_blocks[i]] + output_block_list = {} + + for layer in output_block_layers: + layer_id, layer_name = layer.split(".")[0], shave_segments(layer, 1) + if layer_id in output_block_list: + output_block_list[layer_id].append(layer_name) + else: + output_block_list[layer_id] = [layer_name] + + if len(output_block_list) > 1: + resnets = [key for key in output_blocks[i] if f"output_blocks.{i}.0" in key] + attentions = [key for key in output_blocks[i] if f"output_blocks.{i}.1" in key] + + resnet_0_paths = renew_resnet_paths(resnets) + paths = renew_resnet_paths(resnets) + + meta_path = {"old": f"output_blocks.{i}.0", "new": f"up_blocks.{block_id}.resnets.{layer_in_block_id}"} + assign_to_checkpoint( + paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + + output_block_list = {k: sorted(v) for k, v in sorted(output_block_list.items())} + if ["conv.bias", "conv.weight"] in output_block_list.values(): + index = list(output_block_list.values()).index(["conv.bias", "conv.weight"]) + new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = unet_state_dict[ + f"output_blocks.{i}.{index}.conv.weight" + ] + new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = unet_state_dict[ + f"output_blocks.{i}.{index}.conv.bias" + ] + + # Clear attentions as they have been attributed above. + if len(attentions) == 2: + attentions = [] + + if len(attentions): + paths = renew_attention_paths(attentions) + meta_path = { + "old": f"output_blocks.{i}.1", + "new": f"up_blocks.{block_id}.attentions.{layer_in_block_id}", + } + assign_to_checkpoint( + paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + else: + resnet_0_paths = renew_resnet_paths(output_block_layers, n_shave_prefix_segments=1) + for path in resnet_0_paths: + old_path = ".".join(["output_blocks", str(i), path["old"]]) + new_path = ".".join(["up_blocks", str(block_id), "resnets", str(layer_in_block_id), path["new"]]) + + new_checkpoint[new_path] = unet_state_dict[old_path] + + if controlnet: + # conditioning embedding + + orig_index = 0 + + new_checkpoint["controlnet_cond_embedding.conv_in.weight"] = unet_state_dict.pop( + f"input_hint_block.{orig_index}.weight" + ) + new_checkpoint["controlnet_cond_embedding.conv_in.bias"] = unet_state_dict.pop( + f"input_hint_block.{orig_index}.bias" + ) + + orig_index += 2 + + diffusers_index = 0 + + while diffusers_index < 6: + new_checkpoint[f"controlnet_cond_embedding.blocks.{diffusers_index}.weight"] = unet_state_dict.pop( + f"input_hint_block.{orig_index}.weight" + ) + new_checkpoint[f"controlnet_cond_embedding.blocks.{diffusers_index}.bias"] = unet_state_dict.pop( + f"input_hint_block.{orig_index}.bias" + ) + diffusers_index += 1 + orig_index += 2 + + new_checkpoint["controlnet_cond_embedding.conv_out.weight"] = unet_state_dict.pop( + f"input_hint_block.{orig_index}.weight" + ) + new_checkpoint["controlnet_cond_embedding.conv_out.bias"] = unet_state_dict.pop( + f"input_hint_block.{orig_index}.bias" + ) + + # down blocks + for i in range(num_input_blocks): + new_checkpoint[f"controlnet_down_blocks.{i}.weight"] = unet_state_dict.pop(f"zero_convs.{i}.0.weight") + new_checkpoint[f"controlnet_down_blocks.{i}.bias"] = unet_state_dict.pop(f"zero_convs.{i}.0.bias") + + # mid block + new_checkpoint["controlnet_mid_block.weight"] = unet_state_dict.pop("middle_block_out.0.weight") + new_checkpoint["controlnet_mid_block.bias"] = unet_state_dict.pop("middle_block_out.0.bias") + + return new_checkpoint + + +def convert_ldm_vae_checkpoint(checkpoint, config): + # extract state dict for VAE + vae_state_dict = {} + keys = list(checkpoint.keys()) + vae_key = "first_stage_model." if any(k.startswith("first_stage_model.") for k in keys) else "" + for key in keys: + if key.startswith(vae_key): + vae_state_dict[key.replace(vae_key, "")] = checkpoint.get(key) + + new_checkpoint = {} + + new_checkpoint["encoder.conv_in.weight"] = vae_state_dict["encoder.conv_in.weight"] + new_checkpoint["encoder.conv_in.bias"] = vae_state_dict["encoder.conv_in.bias"] + new_checkpoint["encoder.conv_out.weight"] = vae_state_dict["encoder.conv_out.weight"] + new_checkpoint["encoder.conv_out.bias"] = vae_state_dict["encoder.conv_out.bias"] + new_checkpoint["encoder.conv_norm_out.weight"] = vae_state_dict["encoder.norm_out.weight"] + new_checkpoint["encoder.conv_norm_out.bias"] = vae_state_dict["encoder.norm_out.bias"] + + new_checkpoint["decoder.conv_in.weight"] = vae_state_dict["decoder.conv_in.weight"] + new_checkpoint["decoder.conv_in.bias"] = vae_state_dict["decoder.conv_in.bias"] + new_checkpoint["decoder.conv_out.weight"] = vae_state_dict["decoder.conv_out.weight"] + new_checkpoint["decoder.conv_out.bias"] = vae_state_dict["decoder.conv_out.bias"] + new_checkpoint["decoder.conv_norm_out.weight"] = vae_state_dict["decoder.norm_out.weight"] + new_checkpoint["decoder.conv_norm_out.bias"] = vae_state_dict["decoder.norm_out.bias"] + + new_checkpoint["quant_conv.weight"] = vae_state_dict["quant_conv.weight"] + new_checkpoint["quant_conv.bias"] = vae_state_dict["quant_conv.bias"] + new_checkpoint["post_quant_conv.weight"] = vae_state_dict["post_quant_conv.weight"] + new_checkpoint["post_quant_conv.bias"] = vae_state_dict["post_quant_conv.bias"] + + # Retrieves the keys for the encoder down blocks only + num_down_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "encoder.down" in layer}) + down_blocks = { + layer_id: [key for key in vae_state_dict if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks) + } + + # Retrieves the keys for the decoder up blocks only + num_up_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "decoder.up" in layer}) + up_blocks = { + layer_id: [key for key in vae_state_dict if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks) + } + + for i in range(num_down_blocks): + resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key] + + if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict: + new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.pop( + f"encoder.down.{i}.downsample.conv.weight" + ) + new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = vae_state_dict.pop( + f"encoder.down.{i}.downsample.conv.bias" + ) + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_resnets = [key for key in vae_state_dict if "encoder.mid.block" in key] + num_mid_res_blocks = 2 + for i in range(1, num_mid_res_blocks + 1): + resnets = [key for key in mid_resnets if f"encoder.mid.block_{i}" in key] + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_attentions = [key for key in vae_state_dict if "encoder.mid.attn" in key] + paths = renew_vae_attention_paths(mid_attentions) + meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + conv_attn_to_linear(new_checkpoint) + + for i in range(num_up_blocks): + block_id = num_up_blocks - 1 - i + resnets = [ + key for key in up_blocks[block_id] if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key + ] + + if f"decoder.up.{block_id}.upsample.conv.weight" in vae_state_dict: + new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.weight"] = vae_state_dict[ + f"decoder.up.{block_id}.upsample.conv.weight" + ] + new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.bias"] = vae_state_dict[ + f"decoder.up.{block_id}.upsample.conv.bias" + ] + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"up.{block_id}.block", "new": f"up_blocks.{i}.resnets"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_resnets = [key for key in vae_state_dict if "decoder.mid.block" in key] + num_mid_res_blocks = 2 + for i in range(1, num_mid_res_blocks + 1): + resnets = [key for key in mid_resnets if f"decoder.mid.block_{i}" in key] + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_attentions = [key for key in vae_state_dict if "decoder.mid.attn" in key] + paths = renew_vae_attention_paths(mid_attentions) + meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + conv_attn_to_linear(new_checkpoint) + return new_checkpoint + + +def convert_ldm_bert_checkpoint(checkpoint, config): + def _copy_attn_layer(hf_attn_layer, pt_attn_layer): + hf_attn_layer.q_proj.weight.data = pt_attn_layer.to_q.weight + hf_attn_layer.k_proj.weight.data = pt_attn_layer.to_k.weight + hf_attn_layer.v_proj.weight.data = pt_attn_layer.to_v.weight + + hf_attn_layer.out_proj.weight = pt_attn_layer.to_out.weight + hf_attn_layer.out_proj.bias = pt_attn_layer.to_out.bias + + def _copy_linear(hf_linear, pt_linear): + hf_linear.weight = pt_linear.weight + hf_linear.bias = pt_linear.bias + + def _copy_layer(hf_layer, pt_layer): + # copy layer norms + _copy_linear(hf_layer.self_attn_layer_norm, pt_layer[0][0]) + _copy_linear(hf_layer.final_layer_norm, pt_layer[1][0]) + + # copy attn + _copy_attn_layer(hf_layer.self_attn, pt_layer[0][1]) + + # copy MLP + pt_mlp = pt_layer[1][1] + _copy_linear(hf_layer.fc1, pt_mlp.net[0][0]) + _copy_linear(hf_layer.fc2, pt_mlp.net[2]) + + def _copy_layers(hf_layers, pt_layers): + for i, hf_layer in enumerate(hf_layers): + if i != 0: + i += i + pt_layer = pt_layers[i : i + 2] + _copy_layer(hf_layer, pt_layer) + + hf_model = LDMBertModel(config).eval() + + # copy embeds + hf_model.model.embed_tokens.weight = checkpoint.transformer.token_emb.weight + hf_model.model.embed_positions.weight.data = checkpoint.transformer.pos_emb.emb.weight + + # copy layer norm + _copy_linear(hf_model.model.layer_norm, checkpoint.transformer.norm) + + # copy hidden layers + _copy_layers(hf_model.model.layers, checkpoint.transformer.attn_layers.layers) + + _copy_linear(hf_model.to_logits, checkpoint.transformer.to_logits) + + return hf_model + + +def convert_ldm_clip_checkpoint(checkpoint, local_files_only=False, text_encoder=None): + if text_encoder is None: + config_name = "openai/clip-vit-large-patch14" + try: + config = CLIPTextConfig.from_pretrained(config_name, local_files_only=local_files_only) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the configuration in the following path: 'openai/clip-vit-large-patch14'." + ) + + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + text_model = CLIPTextModel(config) + else: + text_model = text_encoder + + keys = list(checkpoint.keys()) + + text_model_dict = {} + + remove_prefixes = ["cond_stage_model.transformer", "conditioner.embedders.0.transformer"] + + for key in keys: + for prefix in remove_prefixes: + if key.startswith(prefix): + text_model_dict[key[len(prefix + ".") :]] = checkpoint[key] + + if is_accelerate_available(): + for param_name, param in text_model_dict.items(): + set_module_tensor_to_device(text_model, param_name, "cpu", value=param) + else: + if not (hasattr(text_model, "embeddings") and hasattr(text_model.embeddings.position_ids)): + text_model_dict.pop("text_model.embeddings.position_ids", None) + + text_model.load_state_dict(text_model_dict) + + return text_model + + +textenc_conversion_lst = [ + ("positional_embedding", "text_model.embeddings.position_embedding.weight"), + ("token_embedding.weight", "text_model.embeddings.token_embedding.weight"), + ("ln_final.weight", "text_model.final_layer_norm.weight"), + ("ln_final.bias", "text_model.final_layer_norm.bias"), + ("text_projection", "text_projection.weight"), +] +textenc_conversion_map = {x[0]: x[1] for x in textenc_conversion_lst} + +textenc_transformer_conversion_lst = [ + # (stable-diffusion, HF Diffusers) + ("resblocks.", "text_model.encoder.layers."), + ("ln_1", "layer_norm1"), + ("ln_2", "layer_norm2"), + (".c_fc.", ".fc1."), + (".c_proj.", ".fc2."), + (".attn", ".self_attn"), + ("ln_final.", "transformer.text_model.final_layer_norm."), + ("token_embedding.weight", "transformer.text_model.embeddings.token_embedding.weight"), + ("positional_embedding", "transformer.text_model.embeddings.position_embedding.weight"), +] +protected = {re.escape(x[0]): x[1] for x in textenc_transformer_conversion_lst} +textenc_pattern = re.compile("|".join(protected.keys())) + + +def convert_paint_by_example_checkpoint(checkpoint, local_files_only=False): + config = CLIPVisionConfig.from_pretrained("openai/clip-vit-large-patch14", local_files_only=local_files_only) + model = PaintByExampleImageEncoder(config) + + keys = list(checkpoint.keys()) + + text_model_dict = {} + + for key in keys: + if key.startswith("cond_stage_model.transformer"): + text_model_dict[key[len("cond_stage_model.transformer.") :]] = checkpoint[key] + + # load clip vision + model.model.load_state_dict(text_model_dict) + + # load mapper + keys_mapper = { + k[len("cond_stage_model.mapper.res") :]: v + for k, v in checkpoint.items() + if k.startswith("cond_stage_model.mapper") + } + + MAPPING = { + "attn.c_qkv": ["attn1.to_q", "attn1.to_k", "attn1.to_v"], + "attn.c_proj": ["attn1.to_out.0"], + "ln_1": ["norm1"], + "ln_2": ["norm3"], + "mlp.c_fc": ["ff.net.0.proj"], + "mlp.c_proj": ["ff.net.2"], + } + + mapped_weights = {} + for key, value in keys_mapper.items(): + prefix = key[: len("blocks.i")] + suffix = key.split(prefix)[-1].split(".")[-1] + name = key.split(prefix)[-1].split(suffix)[0][1:-1] + mapped_names = MAPPING[name] + + num_splits = len(mapped_names) + for i, mapped_name in enumerate(mapped_names): + new_name = ".".join([prefix, mapped_name, suffix]) + shape = value.shape[0] // num_splits + mapped_weights[new_name] = value[i * shape : (i + 1) * shape] + + model.mapper.load_state_dict(mapped_weights) + + # load final layer norm + model.final_layer_norm.load_state_dict( + { + "bias": checkpoint["cond_stage_model.final_ln.bias"], + "weight": checkpoint["cond_stage_model.final_ln.weight"], + } + ) + + # load final proj + model.proj_out.load_state_dict( + { + "bias": checkpoint["proj_out.bias"], + "weight": checkpoint["proj_out.weight"], + } + ) + + # load uncond vector + model.uncond_vector.data = torch.nn.Parameter(checkpoint["learnable_vector"]) + return model + + +def convert_open_clip_checkpoint( + checkpoint, + config_name, + prefix="cond_stage_model.model.", + has_projection=False, + local_files_only=False, + **config_kwargs, +): + # text_model = CLIPTextModel.from_pretrained("stabilityai/stable-diffusion-2", subfolder="text_encoder") + # text_model = CLIPTextModelWithProjection.from_pretrained( + # "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", projection_dim=1280 + # ) + try: + config = CLIPTextConfig.from_pretrained(config_name, **config_kwargs, local_files_only=local_files_only) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the configuration in the following path: '{config_name}'." + ) + + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + text_model = CLIPTextModelWithProjection(config) if has_projection else CLIPTextModel(config) + + keys = list(checkpoint.keys()) + + keys_to_ignore = [] + if config_name == "stabilityai/stable-diffusion-2" and config.num_hidden_layers == 23: + # make sure to remove all keys > 22 + keys_to_ignore += [k for k in keys if k.startswith("cond_stage_model.model.transformer.resblocks.23")] + keys_to_ignore += ["cond_stage_model.model.text_projection"] + + text_model_dict = {} + + if prefix + "text_projection" in checkpoint: + d_model = int(checkpoint[prefix + "text_projection"].shape[0]) + else: + d_model = 1024 + + text_model_dict["text_model.embeddings.position_ids"] = text_model.text_model.embeddings.get_buffer("position_ids") + + for key in keys: + if key in keys_to_ignore: + continue + if key[len(prefix) :] in textenc_conversion_map: + if key.endswith("text_projection"): + value = checkpoint[key].T.contiguous() + else: + value = checkpoint[key] + + text_model_dict[textenc_conversion_map[key[len(prefix) :]]] = value + + if key.startswith(prefix + "transformer."): + new_key = key[len(prefix + "transformer.") :] + if new_key.endswith(".in_proj_weight"): + new_key = new_key[: -len(".in_proj_weight")] + new_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], new_key) + text_model_dict[new_key + ".q_proj.weight"] = checkpoint[key][:d_model, :] + text_model_dict[new_key + ".k_proj.weight"] = checkpoint[key][d_model : d_model * 2, :] + text_model_dict[new_key + ".v_proj.weight"] = checkpoint[key][d_model * 2 :, :] + elif new_key.endswith(".in_proj_bias"): + new_key = new_key[: -len(".in_proj_bias")] + new_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], new_key) + text_model_dict[new_key + ".q_proj.bias"] = checkpoint[key][:d_model] + text_model_dict[new_key + ".k_proj.bias"] = checkpoint[key][d_model : d_model * 2] + text_model_dict[new_key + ".v_proj.bias"] = checkpoint[key][d_model * 2 :] + else: + new_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], new_key) + + text_model_dict[new_key] = checkpoint[key] + + if is_accelerate_available(): + for param_name, param in text_model_dict.items(): + set_module_tensor_to_device(text_model, param_name, "cpu", value=param) + else: + if not (hasattr(text_model, "embeddings") and hasattr(text_model.embeddings.position_ids)): + text_model_dict.pop("text_model.embeddings.position_ids", None) + + text_model.load_state_dict(text_model_dict) + + return text_model + + +def stable_unclip_image_encoder(original_config, local_files_only=False): + """ + Returns the image processor and clip image encoder for the img2img unclip pipeline. + + We currently know of two types of stable unclip models which separately use the clip and the openclip image + encoders. + """ + + image_embedder_config = original_config["model"]["params"]["embedder_config"] + + sd_clip_image_embedder_class = image_embedder_config["target"] + sd_clip_image_embedder_class = sd_clip_image_embedder_class.split(".")[-1] + + if sd_clip_image_embedder_class == "ClipImageEmbedder": + clip_model_name = image_embedder_config.params.model + + if clip_model_name == "ViT-L/14": + feature_extractor = CLIPImageProcessor() + image_encoder = CLIPVisionModelWithProjection.from_pretrained( + "openai/clip-vit-large-patch14", local_files_only=local_files_only + ) + else: + raise NotImplementedError(f"Unknown CLIP checkpoint name in stable diffusion checkpoint {clip_model_name}") + + elif sd_clip_image_embedder_class == "FrozenOpenCLIPImageEmbedder": + feature_extractor = CLIPImageProcessor() + image_encoder = CLIPVisionModelWithProjection.from_pretrained( + "laion/CLIP-ViT-H-14-laion2B-s32B-b79K", local_files_only=local_files_only + ) + else: + raise NotImplementedError( + f"Unknown CLIP image embedder class in stable diffusion checkpoint {sd_clip_image_embedder_class}" + ) + + return feature_extractor, image_encoder + + +def stable_unclip_image_noising_components( + original_config, clip_stats_path: Optional[str] = None, device: Optional[str] = None +): + """ + Returns the noising components for the img2img and txt2img unclip pipelines. + + Converts the stability noise augmentor into + 1. a `StableUnCLIPImageNormalizer` for holding the CLIP stats + 2. a `DDPMScheduler` for holding the noise schedule + + If the noise augmentor config specifies a clip stats path, the `clip_stats_path` must be provided. + """ + noise_aug_config = original_config["model"]["params"]["noise_aug_config"] + noise_aug_class = noise_aug_config["target"] + noise_aug_class = noise_aug_class.split(".")[-1] + + if noise_aug_class == "CLIPEmbeddingNoiseAugmentation": + noise_aug_config = noise_aug_config.params + embedding_dim = noise_aug_config.timestep_dim + max_noise_level = noise_aug_config.noise_schedule_config.timesteps + beta_schedule = noise_aug_config.noise_schedule_config.beta_schedule + + image_normalizer = StableUnCLIPImageNormalizer(embedding_dim=embedding_dim) + image_noising_scheduler = DDPMScheduler(num_train_timesteps=max_noise_level, beta_schedule=beta_schedule) + + if "clip_stats_path" in noise_aug_config: + if clip_stats_path is None: + raise ValueError("This stable unclip config requires a `clip_stats_path`") + + clip_mean, clip_std = torch.load(clip_stats_path, map_location=device) + clip_mean = clip_mean[None, :] + clip_std = clip_std[None, :] + + clip_stats_state_dict = { + "mean": clip_mean, + "std": clip_std, + } + + image_normalizer.load_state_dict(clip_stats_state_dict) + else: + raise NotImplementedError(f"Unknown noise augmentor class: {noise_aug_class}") + + return image_normalizer, image_noising_scheduler + + +def convert_controlnet_checkpoint( + checkpoint, + original_config, + checkpoint_path, + image_size, + upcast_attention, + extract_ema, + use_linear_projection=None, + cross_attention_dim=None, +): + ctrlnet_config = create_unet_diffusers_config(original_config, image_size=image_size, controlnet=True) + ctrlnet_config["upcast_attention"] = upcast_attention + + ctrlnet_config.pop("sample_size") + + if use_linear_projection is not None: + ctrlnet_config["use_linear_projection"] = use_linear_projection + + if cross_attention_dim is not None: + ctrlnet_config["cross_attention_dim"] = cross_attention_dim + + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + controlnet = ControlNetModel(**ctrlnet_config) + + # Some controlnet ckpt files are distributed independently from the rest of the + # model components i.e. https://huggingface.co/thibaud/controlnet-sd21/ + if "time_embed.0.weight" in checkpoint: + skip_extract_state_dict = True + else: + skip_extract_state_dict = False + + converted_ctrl_checkpoint = convert_ldm_unet_checkpoint( + checkpoint, + ctrlnet_config, + path=checkpoint_path, + extract_ema=extract_ema, + controlnet=True, + skip_extract_state_dict=skip_extract_state_dict, + ) + + if is_accelerate_available(): + for param_name, param in converted_ctrl_checkpoint.items(): + set_module_tensor_to_device(controlnet, param_name, "cpu", value=param) + else: + controlnet.load_state_dict(converted_ctrl_checkpoint) + + return controlnet + + +def download_from_original_stable_diffusion_ckpt( + checkpoint_path_or_dict: Union[str, Dict[str, torch.Tensor]], + original_config_file: str = None, + image_size: Optional[int] = None, + prediction_type: str = None, + model_type: str = None, + extract_ema: bool = False, + scheduler_type: str = "pndm", + num_in_channels: Optional[int] = None, + upcast_attention: Optional[bool] = None, + device: str = None, + from_safetensors: bool = False, + stable_unclip: Optional[str] = None, + stable_unclip_prior: Optional[str] = None, + clip_stats_path: Optional[str] = None, + controlnet: Optional[bool] = None, + adapter: Optional[bool] = None, + load_safety_checker: bool = True, + safety_checker: Optional[StableDiffusionSafetyChecker] = None, + feature_extractor: Optional[AutoFeatureExtractor] = None, + pipeline_class: DiffusionPipeline = None, + local_files_only=False, + vae_path=None, + vae=None, + text_encoder=None, + text_encoder_2=None, + tokenizer=None, + tokenizer_2=None, + config_files=None, +) -> DiffusionPipeline: + """ + Load a Stable Diffusion pipeline object from a CompVis-style `.ckpt`/`.safetensors` file and (ideally) a `.yaml` + config file. + + Although many of the arguments can be automatically inferred, some of these rely on brittle checks against the + global step count, which will likely fail for models that have undergone further fine-tuning. Therefore, it is + recommended that you override the default values and/or supply an `original_config_file` wherever possible. + + Args: + checkpoint_path_or_dict (`str` or `dict`): Path to `.ckpt` file, or the state dict. + original_config_file (`str`): + Path to `.yaml` config file corresponding to the original architecture. If `None`, will be automatically + inferred by looking for a key that only exists in SD2.0 models. + image_size (`int`, *optional*, defaults to 512): + The image size that the model was trained on. Use 512 for Stable Diffusion v1.X and Stable Diffusion v2 + Base. Use 768 for Stable Diffusion v2. + prediction_type (`str`, *optional*): + The prediction type that the model was trained on. Use `'epsilon'` for Stable Diffusion v1.X and Stable + Diffusion v2 Base. Use `'v_prediction'` for Stable Diffusion v2. + num_in_channels (`int`, *optional*, defaults to None): + The number of input channels. If `None`, it will be automatically inferred. + scheduler_type (`str`, *optional*, defaults to 'pndm'): + Type of scheduler to use. Should be one of `["pndm", "lms", "heun", "euler", "euler-ancestral", "dpm", + "ddim"]`. + model_type (`str`, *optional*, defaults to `None`): + The pipeline type. `None` to automatically infer, or one of `["FrozenOpenCLIPEmbedder", + "FrozenCLIPEmbedder", "PaintByExample"]`. + is_img2img (`bool`, *optional*, defaults to `False`): + Whether the model should be loaded as an img2img pipeline. + extract_ema (`bool`, *optional*, defaults to `False`): Only relevant for + checkpoints that have both EMA and non-EMA weights. Whether to extract the EMA weights or not. Defaults to + `False`. Pass `True` to extract the EMA weights. EMA weights usually yield higher quality images for + inference. Non-EMA weights are usually better to continue fine-tuning. + upcast_attention (`bool`, *optional*, defaults to `None`): + Whether the attention computation should always be upcasted. This is necessary when running stable + diffusion 2.1. + device (`str`, *optional*, defaults to `None`): + The device to use. Pass `None` to determine automatically. + from_safetensors (`str`, *optional*, defaults to `False`): + If `checkpoint_path` is in `safetensors` format, load checkpoint with safetensors instead of PyTorch. + load_safety_checker (`bool`, *optional*, defaults to `True`): + Whether to load the safety checker or not. Defaults to `True`. + safety_checker (`StableDiffusionSafetyChecker`, *optional*, defaults to `None`): + Safety checker to use. If this parameter is `None`, the function will load a new instance of + [StableDiffusionSafetyChecker] by itself, if needed. + feature_extractor (`AutoFeatureExtractor`, *optional*, defaults to `None`): + Feature extractor to use. If this parameter is `None`, the function will load a new instance of + [AutoFeatureExtractor] by itself, if needed. + pipeline_class (`str`, *optional*, defaults to `None`): + The pipeline class to use. Pass `None` to determine automatically. + local_files_only (`bool`, *optional*, defaults to `False`): + Whether or not to only look at local files (i.e., do not try to download the model). + vae (`AutoencoderKL`, *optional*, defaults to `None`): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. If + this parameter is `None`, the function will load a new instance of [CLIP] by itself, if needed. + text_encoder (`CLIPTextModel`, *optional*, defaults to `None`): + An instance of [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel) + to use, specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) + variant. If this parameter is `None`, the function will load a new instance of [CLIP] by itself, if needed. + tokenizer (`CLIPTokenizer`, *optional*, defaults to `None`): + An instance of + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer) + to use. If this parameter is `None`, the function will load a new instance of [CLIPTokenizer] by itself, if + needed. + config_files (`Dict[str, str]`, *optional*, defaults to `None`): + A dictionary mapping from config file names to their contents. If this parameter is `None`, the function + will load the config files by itself, if needed. Valid keys are: + - `v1`: Config file for Stable Diffusion v1 + - `v2`: Config file for Stable Diffusion v2 + - `xl`: Config file for Stable Diffusion XL + - `xl_refiner`: Config file for Stable Diffusion XL Refiner + return: A StableDiffusionPipeline object representing the passed-in `.ckpt`/`.safetensors` file. + """ + + # import pipelines here to avoid circular import error when using from_single_file method + from diffusers import ( + LDMTextToImagePipeline, + PaintByExamplePipeline, + StableDiffusionControlNetPipeline, + StableDiffusionInpaintPipeline, + StableDiffusionPipeline, + StableDiffusionUpscalePipeline, + StableDiffusionXLControlNetInpaintPipeline, + StableDiffusionXLImg2ImgPipeline, + StableDiffusionXLInpaintPipeline, + StableDiffusionXLPipeline, + StableUnCLIPImg2ImgPipeline, + StableUnCLIPPipeline, + ) + + if prediction_type == "v-prediction": + prediction_type = "v_prediction" + + if isinstance(checkpoint_path_or_dict, str): + if from_safetensors: + from safetensors.torch import load_file as safe_load + + checkpoint = safe_load(checkpoint_path_or_dict, device="cpu") + else: + if device is None: + device = get_device() + checkpoint = torch.load(checkpoint_path_or_dict, map_location=device) + else: + checkpoint = torch.load(checkpoint_path_or_dict, map_location=device) + elif isinstance(checkpoint_path_or_dict, dict): + checkpoint = checkpoint_path_or_dict + + # Sometimes models don't have the global_step item + if "global_step" in checkpoint: + global_step = checkpoint["global_step"] + else: + logger.debug("global_step key not found in model") + global_step = None + + # NOTE: this while loop isn't great but this controlnet checkpoint has one additional + # "state_dict" key https://huggingface.co/thibaud/controlnet-canny-sd21 + while "state_dict" in checkpoint: + checkpoint = checkpoint["state_dict"] + + if original_config_file is None: + key_name_v2_1 = "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight" + key_name_sd_xl_base = "conditioner.embedders.1.model.transformer.resblocks.9.mlp.c_proj.bias" + key_name_sd_xl_refiner = "conditioner.embedders.0.model.transformer.resblocks.9.mlp.c_proj.bias" + is_upscale = pipeline_class == StableDiffusionUpscalePipeline + + config_url = None + + # model_type = "v1" + if config_files is not None and "v1" in config_files: + original_config_file = config_files["v1"] + else: + config_url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/configs/stable-diffusion/v1-inference.yaml" + + if key_name_v2_1 in checkpoint and checkpoint[key_name_v2_1].shape[-1] == 1024: + # model_type = "v2" + if config_files is not None and "v2" in config_files: + original_config_file = config_files["v2"] + else: + config_url = "https://raw.githubusercontent.com/Stability-AI/stablediffusion/main/configs/stable-diffusion/v2-inference-v.yaml" + if global_step == 110000: + # v2.1 needs to upcast attention + upcast_attention = True + elif key_name_sd_xl_base in checkpoint: + # only base xl has two text embedders + if config_files is not None and "xl" in config_files: + original_config_file = config_files["xl"] + else: + config_url = "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_base.yaml" + elif key_name_sd_xl_refiner in checkpoint: + # only refiner xl has embedder and one text embedders + if config_files is not None and "xl_refiner" in config_files: + original_config_file = config_files["xl_refiner"] + else: + config_url = "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_refiner.yaml" + + if is_upscale: + config_url = "https://raw.githubusercontent.com/Stability-AI/stablediffusion/main/configs/stable-diffusion/x4-upscaling.yaml" + + if config_url is not None: + original_config_file = BytesIO(requests.get(config_url, timeout=DIFFUSERS_REQUEST_TIMEOUT).content) + else: + with open(original_config_file, "r") as f: + original_config_file = f.read() + else: + with open(original_config_file, "r") as f: + original_config_file = f.read() + + original_config = yaml.safe_load(original_config_file) + + # Convert the text model. + if ( + model_type is None + and "cond_stage_config" in original_config["model"]["params"] + and original_config["model"]["params"]["cond_stage_config"] is not None + ): + model_type = original_config["model"]["params"]["cond_stage_config"]["target"].split(".")[-1] + logger.debug(f"no `model_type` given, `model_type` inferred as: {model_type}") + elif model_type is None and original_config["model"]["params"]["network_config"] is not None: + if original_config["model"]["params"]["network_config"]["params"]["context_dim"] == 2048: + model_type = "SDXL" + else: + model_type = "SDXL-Refiner" + if image_size is None: + image_size = 1024 + + if pipeline_class is None: + # Check if we have a SDXL or SD model and initialize default pipeline + if model_type not in ["SDXL", "SDXL-Refiner"]: + pipeline_class = StableDiffusionPipeline if not controlnet else StableDiffusionControlNetPipeline + else: + pipeline_class = StableDiffusionXLPipeline if model_type == "SDXL" else StableDiffusionXLImg2ImgPipeline + + if num_in_channels is None and pipeline_class in [ + StableDiffusionInpaintPipeline, + StableDiffusionXLInpaintPipeline, + StableDiffusionXLControlNetInpaintPipeline, + ]: + num_in_channels = 9 + if num_in_channels is None and pipeline_class == StableDiffusionUpscalePipeline: + num_in_channels = 7 + elif num_in_channels is None: + num_in_channels = 4 + + if "unet_config" in original_config["model"]["params"]: + original_config["model"]["params"]["unet_config"]["params"]["in_channels"] = num_in_channels + elif "network_config" in original_config["model"]["params"]: + original_config["model"]["params"]["network_config"]["params"]["in_channels"] = num_in_channels + + if ( + "parameterization" in original_config["model"]["params"] + and original_config["model"]["params"]["parameterization"] == "v" + ): + if prediction_type is None: + # NOTE: For stable diffusion 2 base it is recommended to pass `prediction_type=="epsilon"` + # as it relies on a brittle global step parameter here + prediction_type = "epsilon" if global_step == 875000 else "v_prediction" + if image_size is None: + # NOTE: For stable diffusion 2 base one has to pass `image_size==512` + # as it relies on a brittle global step parameter here + image_size = 512 if global_step == 875000 else 768 + else: + if prediction_type is None: + prediction_type = "epsilon" + if image_size is None: + image_size = 512 + + if controlnet is None and "control_stage_config" in original_config["model"]["params"]: + path = checkpoint_path_or_dict if isinstance(checkpoint_path_or_dict, str) else "" + controlnet = convert_controlnet_checkpoint( + checkpoint, original_config, path, image_size, upcast_attention, extract_ema + ) + + if "timesteps" in original_config["model"]["params"]: + num_train_timesteps = original_config["model"]["params"]["timesteps"] + else: + num_train_timesteps = 1000 + + if model_type in ["SDXL", "SDXL-Refiner"]: + scheduler_dict = { + "beta_schedule": "scaled_linear", + "beta_start": 0.00085, + "beta_end": 0.012, + "interpolation_type": "linear", + "num_train_timesteps": num_train_timesteps, + "prediction_type": "epsilon", + "sample_max_value": 1.0, + "set_alpha_to_one": False, + "skip_prk_steps": True, + "steps_offset": 1, + "timestep_spacing": "leading", + } + scheduler = EulerDiscreteScheduler.from_config(scheduler_dict) + scheduler_type = "euler" + else: + if "linear_start" in original_config["model"]["params"]: + beta_start = original_config["model"]["params"]["linear_start"] + else: + beta_start = 0.02 + + if "linear_end" in original_config["model"]["params"]: + beta_end = original_config["model"]["params"]["linear_end"] + else: + beta_end = 0.085 + scheduler = DDIMScheduler( + beta_end=beta_end, + beta_schedule="scaled_linear", + beta_start=beta_start, + num_train_timesteps=num_train_timesteps, + steps_offset=1, + clip_sample=False, + set_alpha_to_one=False, + prediction_type=prediction_type, + ) + # make sure scheduler works correctly with DDIM + scheduler.register_to_config(clip_sample=False) + + if scheduler_type == "pndm": + config = dict(scheduler.config) + config["skip_prk_steps"] = True + scheduler = PNDMScheduler.from_config(config) + elif scheduler_type == "lms": + scheduler = LMSDiscreteScheduler.from_config(scheduler.config) + elif scheduler_type == "heun": + scheduler = HeunDiscreteScheduler.from_config(scheduler.config) + elif scheduler_type == "euler": + scheduler = EulerDiscreteScheduler.from_config(scheduler.config) + elif scheduler_type == "euler-ancestral": + scheduler = EulerAncestralDiscreteScheduler.from_config(scheduler.config) + elif scheduler_type == "dpm": + scheduler = DPMSolverMultistepScheduler.from_config(scheduler.config) + elif scheduler_type == "ddim": + scheduler = scheduler + else: + raise ValueError(f"Scheduler of type {scheduler_type} doesn't exist!") + + if pipeline_class == StableDiffusionUpscalePipeline: + image_size = original_config["model"]["params"]["unet_config"]["params"]["image_size"] + + # Convert the UNet2DConditionModel model. + unet_config = create_unet_diffusers_config(original_config, image_size=image_size) + unet_config["upcast_attention"] = upcast_attention + + path = checkpoint_path_or_dict if isinstance(checkpoint_path_or_dict, str) else "" + converted_unet_checkpoint = convert_ldm_unet_checkpoint( + checkpoint, unet_config, path=path, extract_ema=extract_ema + ) + + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + unet = UNet2DConditionModel(**unet_config) + + if is_accelerate_available(): + if model_type not in ["SDXL", "SDXL-Refiner"]: # SBM Delay this. + for param_name, param in converted_unet_checkpoint.items(): + set_module_tensor_to_device(unet, param_name, "cpu", value=param) + else: + unet.load_state_dict(converted_unet_checkpoint) + + # Convert the VAE model. + if vae_path is None and vae is None: + vae_config = create_vae_diffusers_config(original_config, image_size=image_size) + converted_vae_checkpoint = convert_ldm_vae_checkpoint(checkpoint, vae_config) + + if ( + "model" in original_config + and "params" in original_config["model"] + and "scale_factor" in original_config["model"]["params"] + ): + vae_scaling_factor = original_config["model"]["params"]["scale_factor"] + else: + vae_scaling_factor = 0.18215 # default SD scaling factor + + vae_config["scaling_factor"] = vae_scaling_factor + + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + vae = AutoencoderKL(**vae_config) + + if is_accelerate_available(): + for param_name, param in converted_vae_checkpoint.items(): + set_module_tensor_to_device(vae, param_name, "cpu", value=param) + else: + vae.load_state_dict(converted_vae_checkpoint) + elif vae is None: + vae = AutoencoderKL.from_pretrained(vae_path, local_files_only=local_files_only) + + if model_type == "FrozenOpenCLIPEmbedder": + config_name = "stabilityai/stable-diffusion-2" + config_kwargs = {"subfolder": "text_encoder"} + + if text_encoder is None: + text_model = convert_open_clip_checkpoint( + checkpoint, config_name, local_files_only=local_files_only, **config_kwargs + ) + else: + text_model = text_encoder + + try: + tokenizer = CLIPTokenizer.from_pretrained( + "stabilityai/stable-diffusion-2", subfolder="tokenizer", local_files_only=local_files_only + ) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'stabilityai/stable-diffusion-2'." + ) + + if stable_unclip is None: + if controlnet: + pipe = pipeline_class( + vae=vae, + text_encoder=text_model, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + controlnet=controlnet, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + if hasattr(pipe, "requires_safety_checker"): + pipe.requires_safety_checker = False + + elif pipeline_class == StableDiffusionUpscalePipeline: + scheduler = DDIMScheduler.from_pretrained( + "stabilityai/stable-diffusion-x4-upscaler", subfolder="scheduler" + ) + low_res_scheduler = DDPMScheduler.from_pretrained( + "stabilityai/stable-diffusion-x4-upscaler", subfolder="low_res_scheduler" + ) + + pipe = pipeline_class( + vae=vae, + text_encoder=text_model, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + low_res_scheduler=low_res_scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + + else: + pipe = pipeline_class( + vae=vae, + text_encoder=text_model, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + if hasattr(pipe, "requires_safety_checker"): + pipe.requires_safety_checker = False + + else: + image_normalizer, image_noising_scheduler = stable_unclip_image_noising_components( + original_config, clip_stats_path=clip_stats_path, device=device + ) + + if stable_unclip == "img2img": + feature_extractor, image_encoder = stable_unclip_image_encoder(original_config) + + pipe = StableUnCLIPImg2ImgPipeline( + # image encoding components + feature_extractor=feature_extractor, + image_encoder=image_encoder, + # image noising components + image_normalizer=image_normalizer, + image_noising_scheduler=image_noising_scheduler, + # regular denoising components + tokenizer=tokenizer, + text_encoder=text_model, + unet=unet, + scheduler=scheduler, + # vae + vae=vae, + ) + elif stable_unclip == "txt2img": + if stable_unclip_prior is None or stable_unclip_prior == "karlo": + karlo_model = "kakaobrain/karlo-v1-alpha" + prior = PriorTransformer.from_pretrained( + karlo_model, subfolder="prior", local_files_only=local_files_only + ) + + try: + prior_tokenizer = CLIPTokenizer.from_pretrained( + "openai/clip-vit-large-patch14", local_files_only=local_files_only + ) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'openai/clip-vit-large-patch14'." + ) + prior_text_model = CLIPTextModelWithProjection.from_pretrained( + "openai/clip-vit-large-patch14", local_files_only=local_files_only + ) + + prior_scheduler = UnCLIPScheduler.from_pretrained( + karlo_model, subfolder="prior_scheduler", local_files_only=local_files_only + ) + prior_scheduler = DDPMScheduler.from_config(prior_scheduler.config) + else: + raise NotImplementedError(f"unknown prior for stable unclip model: {stable_unclip_prior}") + + pipe = StableUnCLIPPipeline( + # prior components + prior_tokenizer=prior_tokenizer, + prior_text_encoder=prior_text_model, + prior=prior, + prior_scheduler=prior_scheduler, + # image noising components + image_normalizer=image_normalizer, + image_noising_scheduler=image_noising_scheduler, + # regular denoising components + tokenizer=tokenizer, + text_encoder=text_model, + unet=unet, + scheduler=scheduler, + # vae + vae=vae, + ) + else: + raise NotImplementedError(f"unknown `stable_unclip` type: {stable_unclip}") + elif model_type == "PaintByExample": + vision_model = convert_paint_by_example_checkpoint(checkpoint) + try: + tokenizer = CLIPTokenizer.from_pretrained( + "openai/clip-vit-large-patch14", local_files_only=local_files_only + ) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'openai/clip-vit-large-patch14'." + ) + try: + feature_extractor = AutoFeatureExtractor.from_pretrained( + "CompVis/stable-diffusion-safety-checker", local_files_only=local_files_only + ) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the feature_extractor in the following path: 'CompVis/stable-diffusion-safety-checker'." + ) + pipe = PaintByExamplePipeline( + vae=vae, + image_encoder=vision_model, + unet=unet, + scheduler=scheduler, + safety_checker=None, + feature_extractor=feature_extractor, + ) + elif model_type == "FrozenCLIPEmbedder": + text_model = convert_ldm_clip_checkpoint( + checkpoint, local_files_only=local_files_only, text_encoder=text_encoder + ) + try: + tokenizer = ( + CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14", local_files_only=local_files_only) + if tokenizer is None + else tokenizer + ) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'openai/clip-vit-large-patch14'." + ) + + if load_safety_checker: + safety_checker = StableDiffusionSafetyChecker.from_pretrained( + "CompVis/stable-diffusion-safety-checker", local_files_only=local_files_only + ) + feature_extractor = AutoFeatureExtractor.from_pretrained( + "CompVis/stable-diffusion-safety-checker", local_files_only=local_files_only + ) + + if controlnet: + pipe = pipeline_class( + vae=vae, + text_encoder=text_model, + tokenizer=tokenizer, + unet=unet, + controlnet=controlnet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + else: + pipe = pipeline_class( + vae=vae, + text_encoder=text_model, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + elif model_type in ["SDXL", "SDXL-Refiner"]: + is_refiner = model_type == "SDXL-Refiner" + + if (is_refiner is False) and (tokenizer is None): + try: + tokenizer = CLIPTokenizer.from_pretrained( + "openai/clip-vit-large-patch14", local_files_only=local_files_only + ) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'openai/clip-vit-large-patch14'." + ) + + if (is_refiner is False) and (text_encoder is None): + text_encoder = convert_ldm_clip_checkpoint(checkpoint, local_files_only=local_files_only) + + if tokenizer_2 is None: + try: + tokenizer_2 = CLIPTokenizer.from_pretrained( + "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", pad_token="!", local_files_only=local_files_only + ) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'laion/CLIP-ViT-bigG-14-laion2B-39B-b160k' with `pad_token` set to '!'." + ) + + if text_encoder_2 is None: + config_name = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k" + config_kwargs = {"projection_dim": 1280} + prefix = "conditioner.embedders.0.model." if is_refiner else "conditioner.embedders.1.model." + + text_encoder_2 = convert_open_clip_checkpoint( + checkpoint, + config_name, + prefix=prefix, + has_projection=True, + local_files_only=local_files_only, + **config_kwargs, + ) + + if is_accelerate_available(): # SBM Now move model to cpu. + for param_name, param in converted_unet_checkpoint.items(): + set_module_tensor_to_device(unet, param_name, "cpu", value=param) + + if controlnet: + pipe = pipeline_class( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + unet=unet, + controlnet=controlnet, + scheduler=scheduler, + force_zeros_for_empty_prompt=True, + ) + elif adapter: + pipe = pipeline_class( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + unet=unet, + adapter=adapter, + scheduler=scheduler, + force_zeros_for_empty_prompt=True, + ) + + else: + pipeline_kwargs = { + "vae": vae, + "text_encoder": text_encoder, + "tokenizer": tokenizer, + "text_encoder_2": text_encoder_2, + "tokenizer_2": tokenizer_2, + "unet": unet, + "scheduler": scheduler, + } + + if (pipeline_class == StableDiffusionXLImg2ImgPipeline) or ( + pipeline_class == StableDiffusionXLInpaintPipeline + ): + pipeline_kwargs.update({"requires_aesthetics_score": is_refiner}) + + if is_refiner: + pipeline_kwargs.update({"force_zeros_for_empty_prompt": False}) + + pipe = pipeline_class(**pipeline_kwargs) + else: + text_config = create_ldm_bert_config(original_config) + text_model = convert_ldm_bert_checkpoint(checkpoint, text_config) + tokenizer = BertTokenizerFast.from_pretrained("bert-base-uncased", local_files_only=local_files_only) + pipe = LDMTextToImagePipeline(vqvae=vae, bert=text_model, tokenizer=tokenizer, unet=unet, scheduler=scheduler) + + return pipe + + +def download_controlnet_from_original_ckpt( + checkpoint_path: str, + original_config_file: str, + image_size: int = 512, + extract_ema: bool = False, + num_in_channels: Optional[int] = None, + upcast_attention: Optional[bool] = None, + device: str = None, + from_safetensors: bool = False, + use_linear_projection: Optional[bool] = None, + cross_attention_dim: Optional[bool] = None, +) -> DiffusionPipeline: + if from_safetensors: + from safetensors import safe_open + + checkpoint = {} + with safe_open(checkpoint_path, framework="pt", device="cpu") as f: + for key in f.keys(): + checkpoint[key] = f.get_tensor(key) + else: + if device is None: + device = get_device() + checkpoint = torch.load(checkpoint_path, map_location=device) + else: + checkpoint = torch.load(checkpoint_path, map_location=device) + + # NOTE: this while loop isn't great but this controlnet checkpoint has one additional + # "state_dict" key https://huggingface.co/thibaud/controlnet-canny-sd21 + while "state_dict" in checkpoint: + checkpoint = checkpoint["state_dict"] + + with open(original_config_file, "r") as f: + original_config_file = f.read() + original_config = yaml.safe_load(original_config_file) + + if num_in_channels is not None: + original_config["model"]["params"]["unet_config"]["params"]["in_channels"] = num_in_channels + + if "control_stage_config" not in original_config["model"]["params"]: + raise ValueError("`control_stage_config` not present in original config") + + controlnet = convert_controlnet_checkpoint( + checkpoint, + original_config, + checkpoint_path, + image_size, + upcast_attention, + extract_ema, + use_linear_projection=use_linear_projection, + cross_attention_dim=cross_attention_dim, + ) + + return controlnet diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..1afa7698da7c15e890c2832a9bd10b9675d96c2c --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion.py @@ -0,0 +1,477 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 warnings +from functools import partial +from typing import Dict, List, Optional, Union + +import jax +import jax.numpy as jnp +import numpy as np +from flax.core.frozen_dict import FrozenDict +from flax.jax_utils import unreplicate +from flax.training.common_utils import shard +from packaging import version +from PIL import Image +from transformers import CLIPImageProcessor, CLIPTokenizer, FlaxCLIPTextModel + +from ...models import FlaxAutoencoderKL, FlaxUNet2DConditionModel +from ...schedulers import ( + FlaxDDIMScheduler, + FlaxDPMSolverMultistepScheduler, + FlaxLMSDiscreteScheduler, + FlaxPNDMScheduler, +) +from ...utils import deprecate, logging, replace_example_docstring +from ..pipeline_flax_utils import FlaxDiffusionPipeline +from .pipeline_output import FlaxStableDiffusionPipelineOutput +from .safety_checker_flax import FlaxStableDiffusionSafetyChecker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +# Set to True to use python for loop instead of jax.fori_loop for easier debugging +DEBUG = False + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import jax + >>> import numpy as np + >>> from flax.jax_utils import replicate + >>> from flax.training.common_utils import shard + + >>> from diffusers import FlaxStableDiffusionPipeline + + >>> pipeline, params = FlaxStableDiffusionPipeline.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", variant="bf16", dtype=jax.numpy.bfloat16 + ... ) + + >>> prompt = "a photo of an astronaut riding a horse on mars" + + >>> prng_seed = jax.random.PRNGKey(0) + >>> num_inference_steps = 50 + + >>> num_samples = jax.device_count() + >>> prompt = num_samples * [prompt] + >>> prompt_ids = pipeline.prepare_inputs(prompt) + # shard inputs and rng + + >>> params = replicate(params) + >>> prng_seed = jax.random.split(prng_seed, jax.device_count()) + >>> prompt_ids = shard(prompt_ids) + + >>> images = pipeline(prompt_ids, params, prng_seed, num_inference_steps, jit=True).images + >>> images = pipeline.numpy_to_pil(np.asarray(images.reshape((num_samples,) + images.shape[-3:]))) + ``` +""" + + +class FlaxStableDiffusionPipeline(FlaxDiffusionPipeline): + r""" + Flax-based pipeline for text-to-image generation using Stable Diffusion. + + This model inherits from [`FlaxDiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`FlaxAutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.FlaxCLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`FlaxUNet2DConditionModel`]): + A `FlaxUNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`FlaxDDIMScheduler`], [`FlaxLMSDiscreteScheduler`], [`FlaxPNDMScheduler`], or + [`FlaxDPMSolverMultistepScheduler`]. + safety_checker ([`FlaxStableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + def __init__( + self, + vae: FlaxAutoencoderKL, + text_encoder: FlaxCLIPTextModel, + tokenizer: CLIPTokenizer, + unet: FlaxUNet2DConditionModel, + scheduler: Union[ + FlaxDDIMScheduler, FlaxPNDMScheduler, FlaxLMSDiscreteScheduler, FlaxDPMSolverMultistepScheduler + ], + safety_checker: FlaxStableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + dtype: jnp.dtype = jnp.float32, + ): + super().__init__() + self.dtype = dtype + + if safety_checker is None: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + + def prepare_inputs(self, prompt: Union[str, List[str]]): + if not isinstance(prompt, (str, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + text_input = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="np", + ) + return text_input.input_ids + + def _get_has_nsfw_concepts(self, features, params): + has_nsfw_concepts = self.safety_checker(features, params) + return has_nsfw_concepts + + def _run_safety_checker(self, images, safety_model_params, jit=False): + # safety_model_params should already be replicated when jit is True + pil_images = [Image.fromarray(image) for image in images] + features = self.feature_extractor(pil_images, return_tensors="np").pixel_values + + if jit: + features = shard(features) + has_nsfw_concepts = _p_get_has_nsfw_concepts(self, features, safety_model_params) + has_nsfw_concepts = unshard(has_nsfw_concepts) + safety_model_params = unreplicate(safety_model_params) + else: + has_nsfw_concepts = self._get_has_nsfw_concepts(features, safety_model_params) + + images_was_copied = False + for idx, has_nsfw_concept in enumerate(has_nsfw_concepts): + if has_nsfw_concept: + if not images_was_copied: + images_was_copied = True + images = images.copy() + + images[idx] = np.zeros(images[idx].shape, dtype=np.uint8) # black image + + if any(has_nsfw_concepts): + warnings.warn( + "Potential NSFW content was detected in one or more images. A black image will be returned" + " instead. Try again with a different prompt and/or seed." + ) + + return images, has_nsfw_concepts + + def _generate( + self, + prompt_ids: jnp.array, + params: Union[Dict, FrozenDict], + prng_seed: jax.Array, + num_inference_steps: int, + height: int, + width: int, + guidance_scale: float, + latents: Optional[jnp.ndarray] = None, + neg_prompt_ids: Optional[jnp.ndarray] = None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + # get prompt text embeddings + prompt_embeds = self.text_encoder(prompt_ids, params=params["text_encoder"])[0] + + # TODO: currently it is assumed `do_classifier_free_guidance = guidance_scale > 1.0` + # implement this conditional `do_classifier_free_guidance = guidance_scale > 1.0` + batch_size = prompt_ids.shape[0] + + max_length = prompt_ids.shape[-1] + + if neg_prompt_ids is None: + uncond_input = self.tokenizer( + [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="np" + ).input_ids + else: + uncond_input = neg_prompt_ids + negative_prompt_embeds = self.text_encoder(uncond_input, params=params["text_encoder"])[0] + context = jnp.concatenate([negative_prompt_embeds, prompt_embeds]) + + # Ensure model output will be `float32` before going into the scheduler + guidance_scale = jnp.array([guidance_scale], dtype=jnp.float32) + + latents_shape = ( + batch_size, + self.unet.config.in_channels, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + if latents is None: + latents = jax.random.normal(prng_seed, shape=latents_shape, dtype=jnp.float32) + else: + if latents.shape != latents_shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}") + + def loop_body(step, args): + latents, scheduler_state = args + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + latents_input = jnp.concatenate([latents] * 2) + + t = jnp.array(scheduler_state.timesteps, dtype=jnp.int32)[step] + timestep = jnp.broadcast_to(t, latents_input.shape[0]) + + latents_input = self.scheduler.scale_model_input(scheduler_state, latents_input, t) + + # predict the noise residual + noise_pred = self.unet.apply( + {"params": params["unet"]}, + jnp.array(latents_input), + jnp.array(timestep, dtype=jnp.int32), + encoder_hidden_states=context, + ).sample + # perform guidance + noise_pred_uncond, noise_prediction_text = jnp.split(noise_pred, 2, axis=0) + noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents, scheduler_state = self.scheduler.step(scheduler_state, noise_pred, t, latents).to_tuple() + return latents, scheduler_state + + scheduler_state = self.scheduler.set_timesteps( + params["scheduler"], num_inference_steps=num_inference_steps, shape=latents.shape + ) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * params["scheduler"].init_noise_sigma + + if DEBUG: + # run with python for loop + for i in range(num_inference_steps): + latents, scheduler_state = loop_body(i, (latents, scheduler_state)) + else: + latents, _ = jax.lax.fori_loop(0, num_inference_steps, loop_body, (latents, scheduler_state)) + + # scale and decode the image latents with vae + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.apply({"params": params["vae"]}, latents, method=self.vae.decode).sample + + image = (image / 2 + 0.5).clip(0, 1).transpose(0, 2, 3, 1) + return image + + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt_ids: jnp.array, + params: Union[Dict, FrozenDict], + prng_seed: jax.Array, + num_inference_steps: int = 50, + height: Optional[int] = None, + width: Optional[int] = None, + guidance_scale: Union[float, jnp.ndarray] = 7.5, + latents: jnp.ndarray = None, + neg_prompt_ids: jnp.ndarray = None, + return_dict: bool = True, + jit: bool = False, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + latents (`jnp.ndarray`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + array is generated by sampling using the supplied random `generator`. + jit (`bool`, defaults to `False`): + Whether to run `pmap` versions of the generation and safety scoring functions. + + + + This argument exists because `__call__` is not yet end-to-end pmap-able. It will be removed in a + future release. + + + + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] instead of + a plain tuple. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] is + returned, otherwise a `tuple` is returned where the first element is a list with the generated images + and the second element is a list of `bool`s indicating whether the corresponding generated image + contains "not-safe-for-work" (nsfw) content. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + if isinstance(guidance_scale, float): + # Convert to a tensor so each device gets a copy. Follow the prompt_ids for + # shape information, as they may be sharded (when `jit` is `True`), or not. + guidance_scale = jnp.array([guidance_scale] * prompt_ids.shape[0]) + if len(prompt_ids.shape) > 2: + # Assume sharded + guidance_scale = guidance_scale[:, None] + + if jit: + images = _p_generate( + self, + prompt_ids, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + neg_prompt_ids, + ) + else: + images = self._generate( + prompt_ids, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + neg_prompt_ids, + ) + + if self.safety_checker is not None: + safety_params = params["safety_checker"] + images_uint8_casted = (images * 255).round().astype("uint8") + num_devices, batch_size = images.shape[:2] + + images_uint8_casted = np.asarray(images_uint8_casted).reshape(num_devices * batch_size, height, width, 3) + images_uint8_casted, has_nsfw_concept = self._run_safety_checker(images_uint8_casted, safety_params, jit) + images = np.asarray(images).copy() + + # block images + if any(has_nsfw_concept): + for i, is_nsfw in enumerate(has_nsfw_concept): + if is_nsfw: + images[i, 0] = np.asarray(images_uint8_casted[i]) + + images = images.reshape(num_devices, batch_size, height, width, 3) + else: + images = np.asarray(images) + has_nsfw_concept = False + + if not return_dict: + return (images, has_nsfw_concept) + + return FlaxStableDiffusionPipelineOutput(images=images, nsfw_content_detected=has_nsfw_concept) + + +# Static argnums are pipe, num_inference_steps, height, width. A change would trigger recompilation. +# Non-static args are (sharded) input tensors mapped over their first dimension (hence, `0`). +@partial( + jax.pmap, + in_axes=(None, 0, 0, 0, None, None, None, 0, 0, 0), + static_broadcasted_argnums=(0, 4, 5, 6), +) +def _p_generate( + pipe, + prompt_ids, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + neg_prompt_ids, +): + return pipe._generate( + prompt_ids, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + neg_prompt_ids, + ) + + +@partial(jax.pmap, static_broadcasted_argnums=(0,)) +def _p_get_has_nsfw_concepts(pipe, features, params): + return pipe._get_has_nsfw_concepts(features, params) + + +def unshard(x: jnp.ndarray): + # einops.rearrange(x, 'd b ... -> (d b) ...') + num_devices, batch_size = x.shape[:2] + rest = x.shape[2:] + return x.reshape(num_devices * batch_size, *rest) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_img2img.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..78e3ba239c4e6ff35a194ff64b601a7cbe584149 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_img2img.py @@ -0,0 +1,532 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 warnings +from functools import partial +from typing import Dict, List, Optional, Union + +import jax +import jax.numpy as jnp +import numpy as np +from flax.core.frozen_dict import FrozenDict +from flax.jax_utils import unreplicate +from flax.training.common_utils import shard +from PIL import Image +from transformers import CLIPImageProcessor, CLIPTokenizer, FlaxCLIPTextModel + +from ...models import FlaxAutoencoderKL, FlaxUNet2DConditionModel +from ...schedulers import ( + FlaxDDIMScheduler, + FlaxDPMSolverMultistepScheduler, + FlaxLMSDiscreteScheduler, + FlaxPNDMScheduler, +) +from ...utils import PIL_INTERPOLATION, logging, replace_example_docstring +from ..pipeline_flax_utils import FlaxDiffusionPipeline +from .pipeline_output import FlaxStableDiffusionPipelineOutput +from .safety_checker_flax import FlaxStableDiffusionSafetyChecker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +# Set to True to use python for loop instead of jax.fori_loop for easier debugging +DEBUG = False + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import jax + >>> import numpy as np + >>> import jax.numpy as jnp + >>> from flax.jax_utils import replicate + >>> from flax.training.common_utils import shard + >>> import requests + >>> from io import BytesIO + >>> from PIL import Image + >>> from diffusers import FlaxStableDiffusionImg2ImgPipeline + + + >>> def create_key(seed=0): + ... return jax.random.PRNGKey(seed) + + + >>> rng = create_key(0) + + >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + >>> response = requests.get(url) + >>> init_img = Image.open(BytesIO(response.content)).convert("RGB") + >>> init_img = init_img.resize((768, 512)) + + >>> prompts = "A fantasy landscape, trending on artstation" + + >>> pipeline, params = FlaxStableDiffusionImg2ImgPipeline.from_pretrained( + ... "CompVis/stable-diffusion-v1-4", + ... revision="flax", + ... dtype=jnp.bfloat16, + ... ) + + >>> num_samples = jax.device_count() + >>> rng = jax.random.split(rng, jax.device_count()) + >>> prompt_ids, processed_image = pipeline.prepare_inputs( + ... prompt=[prompts] * num_samples, image=[init_img] * num_samples + ... ) + >>> p_params = replicate(params) + >>> prompt_ids = shard(prompt_ids) + >>> processed_image = shard(processed_image) + + >>> output = pipeline( + ... prompt_ids=prompt_ids, + ... image=processed_image, + ... params=p_params, + ... prng_seed=rng, + ... strength=0.75, + ... num_inference_steps=50, + ... jit=True, + ... height=512, + ... width=768, + ... ).images + + >>> output_images = pipeline.numpy_to_pil(np.asarray(output.reshape((num_samples,) + output.shape[-3:]))) + ``` +""" + + +class FlaxStableDiffusionImg2ImgPipeline(FlaxDiffusionPipeline): + r""" + Flax-based pipeline for text-guided image-to-image generation using Stable Diffusion. + + This model inherits from [`FlaxDiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`FlaxAutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.FlaxCLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`FlaxUNet2DConditionModel`]): + A `FlaxUNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`FlaxDDIMScheduler`], [`FlaxLMSDiscreteScheduler`], [`FlaxPNDMScheduler`], or + [`FlaxDPMSolverMultistepScheduler`]. + safety_checker ([`FlaxStableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + def __init__( + self, + vae: FlaxAutoencoderKL, + text_encoder: FlaxCLIPTextModel, + tokenizer: CLIPTokenizer, + unet: FlaxUNet2DConditionModel, + scheduler: Union[ + FlaxDDIMScheduler, FlaxPNDMScheduler, FlaxLMSDiscreteScheduler, FlaxDPMSolverMultistepScheduler + ], + safety_checker: FlaxStableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + dtype: jnp.dtype = jnp.float32, + ): + super().__init__() + self.dtype = dtype + + if safety_checker is None: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + + def prepare_inputs(self, prompt: Union[str, List[str]], image: Union[Image.Image, List[Image.Image]]): + if not isinstance(prompt, (str, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if not isinstance(image, (Image.Image, list)): + raise ValueError(f"image has to be of type `PIL.Image.Image` or list but is {type(image)}") + + if isinstance(image, Image.Image): + image = [image] + + processed_images = jnp.concatenate([preprocess(img, jnp.float32) for img in image]) + + text_input = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="np", + ) + return text_input.input_ids, processed_images + + def _get_has_nsfw_concepts(self, features, params): + has_nsfw_concepts = self.safety_checker(features, params) + return has_nsfw_concepts + + def _run_safety_checker(self, images, safety_model_params, jit=False): + # safety_model_params should already be replicated when jit is True + pil_images = [Image.fromarray(image) for image in images] + features = self.feature_extractor(pil_images, return_tensors="np").pixel_values + + if jit: + features = shard(features) + has_nsfw_concepts = _p_get_has_nsfw_concepts(self, features, safety_model_params) + has_nsfw_concepts = unshard(has_nsfw_concepts) + safety_model_params = unreplicate(safety_model_params) + else: + has_nsfw_concepts = self._get_has_nsfw_concepts(features, safety_model_params) + + images_was_copied = False + for idx, has_nsfw_concept in enumerate(has_nsfw_concepts): + if has_nsfw_concept: + if not images_was_copied: + images_was_copied = True + images = images.copy() + + images[idx] = np.zeros(images[idx].shape, dtype=np.uint8) # black image + + if any(has_nsfw_concepts): + warnings.warn( + "Potential NSFW content was detected in one or more images. A black image will be returned" + " instead. Try again with a different prompt and/or seed." + ) + + return images, has_nsfw_concepts + + def get_timestep_start(self, num_inference_steps, strength): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + + return t_start + + def _generate( + self, + prompt_ids: jnp.ndarray, + image: jnp.ndarray, + params: Union[Dict, FrozenDict], + prng_seed: jax.Array, + start_timestep: int, + num_inference_steps: int, + height: int, + width: int, + guidance_scale: float, + noise: Optional[jnp.ndarray] = None, + neg_prompt_ids: Optional[jnp.ndarray] = None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + # get prompt text embeddings + prompt_embeds = self.text_encoder(prompt_ids, params=params["text_encoder"])[0] + + # TODO: currently it is assumed `do_classifier_free_guidance = guidance_scale > 1.0` + # implement this conditional `do_classifier_free_guidance = guidance_scale > 1.0` + batch_size = prompt_ids.shape[0] + + max_length = prompt_ids.shape[-1] + + if neg_prompt_ids is None: + uncond_input = self.tokenizer( + [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="np" + ).input_ids + else: + uncond_input = neg_prompt_ids + negative_prompt_embeds = self.text_encoder(uncond_input, params=params["text_encoder"])[0] + context = jnp.concatenate([negative_prompt_embeds, prompt_embeds]) + + latents_shape = ( + batch_size, + self.unet.config.in_channels, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + if noise is None: + noise = jax.random.normal(prng_seed, shape=latents_shape, dtype=jnp.float32) + else: + if noise.shape != latents_shape: + raise ValueError(f"Unexpected latents shape, got {noise.shape}, expected {latents_shape}") + + # Create init_latents + init_latent_dist = self.vae.apply({"params": params["vae"]}, image, method=self.vae.encode).latent_dist + init_latents = init_latent_dist.sample(key=prng_seed).transpose((0, 3, 1, 2)) + init_latents = self.vae.config.scaling_factor * init_latents + + def loop_body(step, args): + latents, scheduler_state = args + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + latents_input = jnp.concatenate([latents] * 2) + + t = jnp.array(scheduler_state.timesteps, dtype=jnp.int32)[step] + timestep = jnp.broadcast_to(t, latents_input.shape[0]) + + latents_input = self.scheduler.scale_model_input(scheduler_state, latents_input, t) + + # predict the noise residual + noise_pred = self.unet.apply( + {"params": params["unet"]}, + jnp.array(latents_input), + jnp.array(timestep, dtype=jnp.int32), + encoder_hidden_states=context, + ).sample + # perform guidance + noise_pred_uncond, noise_prediction_text = jnp.split(noise_pred, 2, axis=0) + noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents, scheduler_state = self.scheduler.step(scheduler_state, noise_pred, t, latents).to_tuple() + return latents, scheduler_state + + scheduler_state = self.scheduler.set_timesteps( + params["scheduler"], num_inference_steps=num_inference_steps, shape=latents_shape + ) + + latent_timestep = scheduler_state.timesteps[start_timestep : start_timestep + 1].repeat(batch_size) + + latents = self.scheduler.add_noise(params["scheduler"], init_latents, noise, latent_timestep) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * params["scheduler"].init_noise_sigma + + if DEBUG: + # run with python for loop + for i in range(start_timestep, num_inference_steps): + latents, scheduler_state = loop_body(i, (latents, scheduler_state)) + else: + latents, _ = jax.lax.fori_loop(start_timestep, num_inference_steps, loop_body, (latents, scheduler_state)) + + # scale and decode the image latents with vae + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.apply({"params": params["vae"]}, latents, method=self.vae.decode).sample + + image = (image / 2 + 0.5).clip(0, 1).transpose(0, 2, 3, 1) + return image + + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt_ids: jnp.ndarray, + image: jnp.ndarray, + params: Union[Dict, FrozenDict], + prng_seed: jax.Array, + strength: float = 0.8, + num_inference_steps: int = 50, + height: Optional[int] = None, + width: Optional[int] = None, + guidance_scale: Union[float, jnp.ndarray] = 7.5, + noise: jnp.ndarray = None, + neg_prompt_ids: jnp.ndarray = None, + return_dict: bool = True, + jit: bool = False, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt_ids (`jnp.ndarray`): + The prompt or prompts to guide image generation. + image (`jnp.ndarray`): + Array representing an image batch to be used as the starting point. + params (`Dict` or `FrozenDict`): + Dictionary containing the model parameters/weights. + prng_seed (`jax.Array` or `jax.Array`): + Array containing random number generator key. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter is modulated by `strength`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + noise (`jnp.ndarray`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. The array is generated by + sampling using the supplied random `generator`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] instead of + a plain tuple. + jit (`bool`, defaults to `False`): + Whether to run `pmap` versions of the generation and safety scoring functions. + + + + This argument exists because `__call__` is not yet end-to-end pmap-able. It will be removed in a + future release. + + + + Examples: + + Returns: + [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] is + returned, otherwise a `tuple` is returned where the first element is a list with the generated images + and the second element is a list of `bool`s indicating whether the corresponding generated image + contains "not-safe-for-work" (nsfw) content. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + if isinstance(guidance_scale, float): + # Convert to a tensor so each device gets a copy. Follow the prompt_ids for + # shape information, as they may be sharded (when `jit` is `True`), or not. + guidance_scale = jnp.array([guidance_scale] * prompt_ids.shape[0]) + if len(prompt_ids.shape) > 2: + # Assume sharded + guidance_scale = guidance_scale[:, None] + + start_timestep = self.get_timestep_start(num_inference_steps, strength) + + if jit: + images = _p_generate( + self, + prompt_ids, + image, + params, + prng_seed, + start_timestep, + num_inference_steps, + height, + width, + guidance_scale, + noise, + neg_prompt_ids, + ) + else: + images = self._generate( + prompt_ids, + image, + params, + prng_seed, + start_timestep, + num_inference_steps, + height, + width, + guidance_scale, + noise, + neg_prompt_ids, + ) + + if self.safety_checker is not None: + safety_params = params["safety_checker"] + images_uint8_casted = (images * 255).round().astype("uint8") + num_devices, batch_size = images.shape[:2] + + images_uint8_casted = np.asarray(images_uint8_casted).reshape(num_devices * batch_size, height, width, 3) + images_uint8_casted, has_nsfw_concept = self._run_safety_checker(images_uint8_casted, safety_params, jit) + images = np.asarray(images) + + # block images + if any(has_nsfw_concept): + for i, is_nsfw in enumerate(has_nsfw_concept): + if is_nsfw: + images[i] = np.asarray(images_uint8_casted[i]) + + images = images.reshape(num_devices, batch_size, height, width, 3) + else: + images = np.asarray(images) + has_nsfw_concept = False + + if not return_dict: + return (images, has_nsfw_concept) + + return FlaxStableDiffusionPipelineOutput(images=images, nsfw_content_detected=has_nsfw_concept) + + +# Static argnums are pipe, start_timestep, num_inference_steps, height, width. A change would trigger recompilation. +# Non-static args are (sharded) input tensors mapped over their first dimension (hence, `0`). +@partial( + jax.pmap, + in_axes=(None, 0, 0, 0, 0, None, None, None, None, 0, 0, 0), + static_broadcasted_argnums=(0, 5, 6, 7, 8), +) +def _p_generate( + pipe, + prompt_ids, + image, + params, + prng_seed, + start_timestep, + num_inference_steps, + height, + width, + guidance_scale, + noise, + neg_prompt_ids, +): + return pipe._generate( + prompt_ids, + image, + params, + prng_seed, + start_timestep, + num_inference_steps, + height, + width, + guidance_scale, + noise, + neg_prompt_ids, + ) + + +@partial(jax.pmap, static_broadcasted_argnums=(0,)) +def _p_get_has_nsfw_concepts(pipe, features, params): + return pipe._get_has_nsfw_concepts(features, params) + + +def unshard(x: jnp.ndarray): + # einops.rearrange(x, 'd b ... -> (d b) ...') + num_devices, batch_size = x.shape[:2] + rest = x.shape[2:] + return x.reshape(num_devices * batch_size, *rest) + + +def preprocess(image, dtype): + w, h = image.size + w, h = (x - x % 32 for x in (w, h)) # resize to integer multiple of 32 + image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]) + image = jnp.array(image).astype(dtype) / 255.0 + image = image[None].transpose(0, 3, 1, 2) + return 2.0 * image - 1.0 diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_inpaint.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..b7e17ba681a20b1382859b47b59fe815955adc6c --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_inpaint.py @@ -0,0 +1,593 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 warnings +from functools import partial +from typing import Dict, List, Optional, Union + +import jax +import jax.numpy as jnp +import numpy as np +from flax.core.frozen_dict import FrozenDict +from flax.jax_utils import unreplicate +from flax.training.common_utils import shard +from packaging import version +from PIL import Image +from transformers import CLIPImageProcessor, CLIPTokenizer, FlaxCLIPTextModel + +from ...models import FlaxAutoencoderKL, FlaxUNet2DConditionModel +from ...schedulers import ( + FlaxDDIMScheduler, + FlaxDPMSolverMultistepScheduler, + FlaxLMSDiscreteScheduler, + FlaxPNDMScheduler, +) +from ...utils import PIL_INTERPOLATION, deprecate, logging, replace_example_docstring +from ..pipeline_flax_utils import FlaxDiffusionPipeline +from .pipeline_output import FlaxStableDiffusionPipelineOutput +from .safety_checker_flax import FlaxStableDiffusionSafetyChecker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +# Set to True to use python for loop instead of jax.fori_loop for easier debugging +DEBUG = False + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import jax + >>> import numpy as np + >>> from flax.jax_utils import replicate + >>> from flax.training.common_utils import shard + >>> import PIL + >>> import requests + >>> from io import BytesIO + >>> from diffusers import FlaxStableDiffusionInpaintPipeline + + + >>> def download_image(url): + ... response = requests.get(url) + ... return PIL.Image.open(BytesIO(response.content)).convert("RGB") + + + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + + >>> init_image = download_image(img_url).resize((512, 512)) + >>> mask_image = download_image(mask_url).resize((512, 512)) + + >>> pipeline, params = FlaxStableDiffusionInpaintPipeline.from_pretrained( + ... "xvjiarui/stable-diffusion-2-inpainting" + ... ) + + >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench" + >>> prng_seed = jax.random.PRNGKey(0) + >>> num_inference_steps = 50 + + >>> num_samples = jax.device_count() + >>> prompt = num_samples * [prompt] + >>> init_image = num_samples * [init_image] + >>> mask_image = num_samples * [mask_image] + >>> prompt_ids, processed_masked_images, processed_masks = pipeline.prepare_inputs( + ... prompt, init_image, mask_image + ... ) + # shard inputs and rng + + >>> params = replicate(params) + >>> prng_seed = jax.random.split(prng_seed, jax.device_count()) + >>> prompt_ids = shard(prompt_ids) + >>> processed_masked_images = shard(processed_masked_images) + >>> processed_masks = shard(processed_masks) + + >>> images = pipeline( + ... prompt_ids, processed_masks, processed_masked_images, params, prng_seed, num_inference_steps, jit=True + ... ).images + >>> images = pipeline.numpy_to_pil(np.asarray(images.reshape((num_samples,) + images.shape[-3:]))) + ``` +""" + + +class FlaxStableDiffusionInpaintPipeline(FlaxDiffusionPipeline): + r""" + Flax-based pipeline for text-guided image inpainting using Stable Diffusion. + + + + 🧪 This is an experimental feature! + + + + This model inherits from [`FlaxDiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`FlaxAutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.FlaxCLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`FlaxUNet2DConditionModel`]): + A `FlaxUNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`FlaxDDIMScheduler`], [`FlaxLMSDiscreteScheduler`], [`FlaxPNDMScheduler`], or + [`FlaxDPMSolverMultistepScheduler`]. + safety_checker ([`FlaxStableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + def __init__( + self, + vae: FlaxAutoencoderKL, + text_encoder: FlaxCLIPTextModel, + tokenizer: CLIPTokenizer, + unet: FlaxUNet2DConditionModel, + scheduler: Union[ + FlaxDDIMScheduler, FlaxPNDMScheduler, FlaxLMSDiscreteScheduler, FlaxDPMSolverMultistepScheduler + ], + safety_checker: FlaxStableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + dtype: jnp.dtype = jnp.float32, + ): + super().__init__() + self.dtype = dtype + + if safety_checker is None: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + + def prepare_inputs( + self, + prompt: Union[str, List[str]], + image: Union[Image.Image, List[Image.Image]], + mask: Union[Image.Image, List[Image.Image]], + ): + if not isinstance(prompt, (str, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if not isinstance(image, (Image.Image, list)): + raise ValueError(f"image has to be of type `PIL.Image.Image` or list but is {type(image)}") + + if isinstance(image, Image.Image): + image = [image] + + if not isinstance(mask, (Image.Image, list)): + raise ValueError(f"image has to be of type `PIL.Image.Image` or list but is {type(image)}") + + if isinstance(mask, Image.Image): + mask = [mask] + + processed_images = jnp.concatenate([preprocess_image(img, jnp.float32) for img in image]) + processed_masks = jnp.concatenate([preprocess_mask(m, jnp.float32) for m in mask]) + # processed_masks[processed_masks < 0.5] = 0 + processed_masks = processed_masks.at[processed_masks < 0.5].set(0) + # processed_masks[processed_masks >= 0.5] = 1 + processed_masks = processed_masks.at[processed_masks >= 0.5].set(1) + + processed_masked_images = processed_images * (processed_masks < 0.5) + + text_input = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="np", + ) + return text_input.input_ids, processed_masked_images, processed_masks + + def _get_has_nsfw_concepts(self, features, params): + has_nsfw_concepts = self.safety_checker(features, params) + return has_nsfw_concepts + + def _run_safety_checker(self, images, safety_model_params, jit=False): + # safety_model_params should already be replicated when jit is True + pil_images = [Image.fromarray(image) for image in images] + features = self.feature_extractor(pil_images, return_tensors="np").pixel_values + + if jit: + features = shard(features) + has_nsfw_concepts = _p_get_has_nsfw_concepts(self, features, safety_model_params) + has_nsfw_concepts = unshard(has_nsfw_concepts) + safety_model_params = unreplicate(safety_model_params) + else: + has_nsfw_concepts = self._get_has_nsfw_concepts(features, safety_model_params) + + images_was_copied = False + for idx, has_nsfw_concept in enumerate(has_nsfw_concepts): + if has_nsfw_concept: + if not images_was_copied: + images_was_copied = True + images = images.copy() + + images[idx] = np.zeros(images[idx].shape, dtype=np.uint8) # black image + + if any(has_nsfw_concepts): + warnings.warn( + "Potential NSFW content was detected in one or more images. A black image will be returned" + " instead. Try again with a different prompt and/or seed." + ) + + return images, has_nsfw_concepts + + def _generate( + self, + prompt_ids: jnp.ndarray, + mask: jnp.ndarray, + masked_image: jnp.ndarray, + params: Union[Dict, FrozenDict], + prng_seed: jax.Array, + num_inference_steps: int, + height: int, + width: int, + guidance_scale: float, + latents: Optional[jnp.ndarray] = None, + neg_prompt_ids: Optional[jnp.ndarray] = None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + # get prompt text embeddings + prompt_embeds = self.text_encoder(prompt_ids, params=params["text_encoder"])[0] + + # TODO: currently it is assumed `do_classifier_free_guidance = guidance_scale > 1.0` + # implement this conditional `do_classifier_free_guidance = guidance_scale > 1.0` + batch_size = prompt_ids.shape[0] + + max_length = prompt_ids.shape[-1] + + if neg_prompt_ids is None: + uncond_input = self.tokenizer( + [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="np" + ).input_ids + else: + uncond_input = neg_prompt_ids + negative_prompt_embeds = self.text_encoder(uncond_input, params=params["text_encoder"])[0] + context = jnp.concatenate([negative_prompt_embeds, prompt_embeds]) + + latents_shape = ( + batch_size, + self.vae.config.latent_channels, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + if latents is None: + latents = jax.random.normal(prng_seed, shape=latents_shape, dtype=self.dtype) + else: + if latents.shape != latents_shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}") + + prng_seed, mask_prng_seed = jax.random.split(prng_seed) + + masked_image_latent_dist = self.vae.apply( + {"params": params["vae"]}, masked_image, method=self.vae.encode + ).latent_dist + masked_image_latents = masked_image_latent_dist.sample(key=mask_prng_seed).transpose((0, 3, 1, 2)) + masked_image_latents = self.vae.config.scaling_factor * masked_image_latents + del mask_prng_seed + + mask = jax.image.resize(mask, (*mask.shape[:-2], *masked_image_latents.shape[-2:]), method="nearest") + + # 8. Check that sizes of mask, masked image and latents match + num_channels_latents = self.vae.config.latent_channels + num_channels_mask = mask.shape[1] + num_channels_masked_image = masked_image_latents.shape[1] + if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of" + " `pipeline.unet` or your `mask_image` or `image` input." + ) + + def loop_body(step, args): + latents, mask, masked_image_latents, scheduler_state = args + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + latents_input = jnp.concatenate([latents] * 2) + mask_input = jnp.concatenate([mask] * 2) + masked_image_latents_input = jnp.concatenate([masked_image_latents] * 2) + + t = jnp.array(scheduler_state.timesteps, dtype=jnp.int32)[step] + timestep = jnp.broadcast_to(t, latents_input.shape[0]) + + latents_input = self.scheduler.scale_model_input(scheduler_state, latents_input, t) + # concat latents, mask, masked_image_latents in the channel dimension + latents_input = jnp.concatenate([latents_input, mask_input, masked_image_latents_input], axis=1) + + # predict the noise residual + noise_pred = self.unet.apply( + {"params": params["unet"]}, + jnp.array(latents_input), + jnp.array(timestep, dtype=jnp.int32), + encoder_hidden_states=context, + ).sample + # perform guidance + noise_pred_uncond, noise_prediction_text = jnp.split(noise_pred, 2, axis=0) + noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents, scheduler_state = self.scheduler.step(scheduler_state, noise_pred, t, latents).to_tuple() + return latents, mask, masked_image_latents, scheduler_state + + scheduler_state = self.scheduler.set_timesteps( + params["scheduler"], num_inference_steps=num_inference_steps, shape=latents.shape + ) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * params["scheduler"].init_noise_sigma + + if DEBUG: + # run with python for loop + for i in range(num_inference_steps): + latents, mask, masked_image_latents, scheduler_state = loop_body( + i, (latents, mask, masked_image_latents, scheduler_state) + ) + else: + latents, _, _, _ = jax.lax.fori_loop( + 0, num_inference_steps, loop_body, (latents, mask, masked_image_latents, scheduler_state) + ) + + # scale and decode the image latents with vae + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.apply({"params": params["vae"]}, latents, method=self.vae.decode).sample + + image = (image / 2 + 0.5).clip(0, 1).transpose(0, 2, 3, 1) + return image + + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt_ids: jnp.ndarray, + mask: jnp.ndarray, + masked_image: jnp.ndarray, + params: Union[Dict, FrozenDict], + prng_seed: jax.Array, + num_inference_steps: int = 50, + height: Optional[int] = None, + width: Optional[int] = None, + guidance_scale: Union[float, jnp.ndarray] = 7.5, + latents: jnp.ndarray = None, + neg_prompt_ids: jnp.ndarray = None, + return_dict: bool = True, + jit: bool = False, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide image generation. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter is modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + latents (`jnp.ndarray`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + array is generated by sampling using the supplied random `generator`. + jit (`bool`, defaults to `False`): + Whether to run `pmap` versions of the generation and safety scoring functions. + + + + This argument exists because `__call__` is not yet end-to-end pmap-able. It will be removed in a + future release. + + + + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] instead of + a plain tuple. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] is + returned, otherwise a `tuple` is returned where the first element is a list with the generated images + and the second element is a list of `bool`s indicating whether the corresponding generated image + contains "not-safe-for-work" (nsfw) content. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + masked_image = jax.image.resize(masked_image, (*masked_image.shape[:-2], height, width), method="bicubic") + mask = jax.image.resize(mask, (*mask.shape[:-2], height, width), method="nearest") + + if isinstance(guidance_scale, float): + # Convert to a tensor so each device gets a copy. Follow the prompt_ids for + # shape information, as they may be sharded (when `jit` is `True`), or not. + guidance_scale = jnp.array([guidance_scale] * prompt_ids.shape[0]) + if len(prompt_ids.shape) > 2: + # Assume sharded + guidance_scale = guidance_scale[:, None] + + if jit: + images = _p_generate( + self, + prompt_ids, + mask, + masked_image, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + neg_prompt_ids, + ) + else: + images = self._generate( + prompt_ids, + mask, + masked_image, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + neg_prompt_ids, + ) + + if self.safety_checker is not None: + safety_params = params["safety_checker"] + images_uint8_casted = (images * 255).round().astype("uint8") + num_devices, batch_size = images.shape[:2] + + images_uint8_casted = np.asarray(images_uint8_casted).reshape(num_devices * batch_size, height, width, 3) + images_uint8_casted, has_nsfw_concept = self._run_safety_checker(images_uint8_casted, safety_params, jit) + images = np.asarray(images) + + # block images + if any(has_nsfw_concept): + for i, is_nsfw in enumerate(has_nsfw_concept): + if is_nsfw: + images[i] = np.asarray(images_uint8_casted[i]) + + images = images.reshape(num_devices, batch_size, height, width, 3) + else: + images = np.asarray(images) + has_nsfw_concept = False + + if not return_dict: + return (images, has_nsfw_concept) + + return FlaxStableDiffusionPipelineOutput(images=images, nsfw_content_detected=has_nsfw_concept) + + +# Static argnums are pipe, num_inference_steps, height, width. A change would trigger recompilation. +# Non-static args are (sharded) input tensors mapped over their first dimension (hence, `0`). +@partial( + jax.pmap, + in_axes=(None, 0, 0, 0, 0, 0, None, None, None, 0, 0, 0), + static_broadcasted_argnums=(0, 6, 7, 8), +) +def _p_generate( + pipe, + prompt_ids, + mask, + masked_image, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + neg_prompt_ids, +): + return pipe._generate( + prompt_ids, + mask, + masked_image, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + neg_prompt_ids, + ) + + +@partial(jax.pmap, static_broadcasted_argnums=(0,)) +def _p_get_has_nsfw_concepts(pipe, features, params): + return pipe._get_has_nsfw_concepts(features, params) + + +def unshard(x: jnp.ndarray): + # einops.rearrange(x, 'd b ... -> (d b) ...') + num_devices, batch_size = x.shape[:2] + rest = x.shape[2:] + return x.reshape(num_devices * batch_size, *rest) + + +def preprocess_image(image, dtype): + w, h = image.size + w, h = (x - x % 32 for x in (w, h)) # resize to integer multiple of 32 + image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]) + image = jnp.array(image).astype(dtype) / 255.0 + image = image[None].transpose(0, 3, 1, 2) + return 2.0 * image - 1.0 + + +def preprocess_mask(mask, dtype): + w, h = mask.size + w, h = (x - x % 32 for x in (w, h)) # resize to integer multiple of 32 + mask = mask.resize((w, h)) + mask = jnp.array(mask.convert("L")).astype(dtype) / 255.0 + mask = jnp.expand_dims(mask, axis=(0, 1)) + + return mask diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..bd8609a11a200f25abe65fc2e595974cd802181a --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion.py @@ -0,0 +1,487 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Callable, List, Optional, Union + +import numpy as np +import torch +from transformers import CLIPImageProcessor, CLIPTokenizer + +from ...configuration_utils import FrozenDict +from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler +from ...utils import deprecate, logging +from ..onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel +from ..pipeline_utils import DiffusionPipeline +from . import StableDiffusionPipelineOutput + + +logger = logging.get_logger(__name__) + + +class OnnxStableDiffusionPipeline(DiffusionPipeline): + vae_encoder: OnnxRuntimeModel + vae_decoder: OnnxRuntimeModel + text_encoder: OnnxRuntimeModel + tokenizer: CLIPTokenizer + unet: OnnxRuntimeModel + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler] + safety_checker: OnnxRuntimeModel + feature_extractor: CLIPImageProcessor + + _optional_components = ["safety_checker", "feature_extractor"] + _is_onnx = True + + def __init__( + self, + vae_encoder: OnnxRuntimeModel, + vae_decoder: OnnxRuntimeModel, + text_encoder: OnnxRuntimeModel, + tokenizer: CLIPTokenizer, + unet: OnnxRuntimeModel, + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler], + safety_checker: OnnxRuntimeModel, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae_encoder=vae_encoder, + vae_decoder=vae_decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + def _encode_prompt( + self, + prompt: Union[str, List[str]], + num_images_per_prompt: Optional[int], + do_classifier_free_guidance: bool, + negative_prompt: Optional[str], + prompt_embeds: Optional[np.ndarray] = None, + negative_prompt_embeds: Optional[np.ndarray] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`): + prompt to be encoded + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + """ + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="np", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids + + if not np.array_equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0] + + prompt_embeds = np.repeat(prompt_embeds, num_images_per_prompt, axis=0) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] * batch_size + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="np", + ) + negative_prompt_embeds = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0] + + if do_classifier_free_guidance: + negative_prompt_embeds = np.repeat(negative_prompt_embeds, num_images_per_prompt, axis=0) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = np.concatenate([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + def check_inputs( + self, + prompt: Union[str, List[str]], + height: Optional[int], + width: Optional[int], + callback_steps: int, + negative_prompt: Optional[str] = None, + prompt_embeds: Optional[np.ndarray] = None, + negative_prompt_embeds: Optional[np.ndarray] = None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def __call__( + self, + prompt: Union[str, List[str]] = None, + height: Optional[int] = 512, + width: Optional[int] = 512, + num_inference_steps: Optional[int] = 50, + guidance_scale: Optional[float] = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: Optional[float] = 0.0, + generator: Optional[np.random.RandomState] = None, + latents: Optional[np.ndarray] = None, + prompt_embeds: Optional[np.ndarray] = None, + negative_prompt_embeds: Optional[np.ndarray] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, np.ndarray], None]] = None, + callback_steps: int = 1, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + image (`PIL.Image.Image` or List[`PIL.Image.Image`] or `torch.Tensor`): + `Image`, or tensor representing an image batch which will be upscaled. * + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds`. instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` + is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`np.random.RandomState`, *optional*): + One or a list of [numpy generator(s)](TODO) to make generation deterministic. + latents (`np.ndarray`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + + # check inputs. Raise error if not correct + self.check_inputs( + prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds + ) + + # define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if generator is None: + generator = np.random + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + prompt_embeds = self._encode_prompt( + prompt, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + # get the initial random noise unless the user supplied it + latents_dtype = prompt_embeds.dtype + latents_shape = (batch_size * num_images_per_prompt, 4, height // 8, width // 8) + if latents is None: + latents = generator.randn(*latents_shape).astype(latents_dtype) + elif latents.shape != latents_shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}") + + # set timesteps + self.scheduler.set_timesteps(num_inference_steps) + + latents = latents * np.float64(self.scheduler.init_noise_sigma) + + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + timestep_dtype = next( + (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)" + ) + timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype] + + for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(torch.from_numpy(latent_model_input), t) + latent_model_input = latent_model_input.cpu().numpy() + + # predict the noise residual + timestep = np.array([t], dtype=timestep_dtype) + noise_pred = self.unet(sample=latent_model_input, timestep=timestep, encoder_hidden_states=prompt_embeds) + noise_pred = noise_pred[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + scheduler_output = self.scheduler.step( + torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs + ) + latents = scheduler_output.prev_sample.numpy() + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + latents = 1 / 0.18215 * latents + # image = self.vae_decoder(latent_sample=latents)[0] + # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1 + image = np.concatenate( + [self.vae_decoder(latent_sample=latents[i : i + 1])[0] for i in range(latents.shape[0])] + ) + + image = np.clip(image / 2 + 0.5, 0, 1) + image = image.transpose((0, 2, 3, 1)) + + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor( + self.numpy_to_pil(image), return_tensors="np" + ).pixel_values.astype(image.dtype) + + images, has_nsfw_concept = [], [] + for i in range(image.shape[0]): + image_i, has_nsfw_concept_i = self.safety_checker( + clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1] + ) + images.append(image_i) + has_nsfw_concept.append(has_nsfw_concept_i[0]) + image = np.concatenate(images) + else: + has_nsfw_concept = None + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + +class StableDiffusionOnnxPipeline(OnnxStableDiffusionPipeline): + def __init__( + self, + vae_encoder: OnnxRuntimeModel, + vae_decoder: OnnxRuntimeModel, + text_encoder: OnnxRuntimeModel, + tokenizer: CLIPTokenizer, + unet: OnnxRuntimeModel, + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler], + safety_checker: OnnxRuntimeModel, + feature_extractor: CLIPImageProcessor, + ): + deprecation_message = "Please use `OnnxStableDiffusionPipeline` instead of `StableDiffusionOnnxPipeline`." + deprecate("StableDiffusionOnnxPipeline", "1.0.0", deprecation_message) + super().__init__( + vae_encoder=vae_encoder, + vae_decoder=vae_decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_img2img.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..d63bf3bf456459317790668ca8a7a39eb8024d69 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_img2img.py @@ -0,0 +1,550 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Callable, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTokenizer + +from ...configuration_utils import FrozenDict +from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler +from ...utils import PIL_INTERPOLATION, deprecate, logging +from ..onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel +from ..pipeline_utils import DiffusionPipeline +from . import StableDiffusionPipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess with 8->64 +def preprocess(image): + deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead" + deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 64 for x in (w, h)) # resize to integer multiple of 64 + + image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +class OnnxStableDiffusionImg2ImgPipeline(DiffusionPipeline): + r""" + Pipeline for text-guided image to image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + details. + feature_extractor ([`CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + vae_encoder: OnnxRuntimeModel + vae_decoder: OnnxRuntimeModel + text_encoder: OnnxRuntimeModel + tokenizer: CLIPTokenizer + unet: OnnxRuntimeModel + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler] + safety_checker: OnnxRuntimeModel + feature_extractor: CLIPImageProcessor + + _optional_components = ["safety_checker", "feature_extractor"] + _is_onnx = True + + def __init__( + self, + vae_encoder: OnnxRuntimeModel, + vae_decoder: OnnxRuntimeModel, + text_encoder: OnnxRuntimeModel, + tokenizer: CLIPTokenizer, + unet: OnnxRuntimeModel, + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler], + safety_checker: OnnxRuntimeModel, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae_encoder=vae_encoder, + vae_decoder=vae_decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt: Union[str, List[str]], + num_images_per_prompt: Optional[int], + do_classifier_free_guidance: bool, + negative_prompt: Optional[str], + prompt_embeds: Optional[np.ndarray] = None, + negative_prompt_embeds: Optional[np.ndarray] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`): + prompt to be encoded + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + """ + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="np", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids + + if not np.array_equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0] + + prompt_embeds = np.repeat(prompt_embeds, num_images_per_prompt, axis=0) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] * batch_size + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="np", + ) + negative_prompt_embeds = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0] + + if do_classifier_free_guidance: + negative_prompt_embeds = np.repeat(negative_prompt_embeds, num_images_per_prompt, axis=0) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = np.concatenate([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + def check_inputs( + self, + prompt: Union[str, List[str]], + callback_steps: int, + negative_prompt: Optional[Union[str, List[str]]] = None, + prompt_embeds: Optional[np.ndarray] = None, + negative_prompt_embeds: Optional[np.ndarray] = None, + ): + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def __call__( + self, + prompt: Union[str, List[str]], + image: Union[np.ndarray, PIL.Image.Image] = None, + strength: float = 0.8, + num_inference_steps: Optional[int] = 50, + guidance_scale: Optional[float] = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: Optional[float] = 0.0, + generator: Optional[np.random.RandomState] = None, + prompt_embeds: Optional[np.ndarray] = None, + negative_prompt_embeds: Optional[np.ndarray] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, np.ndarray], None]] = None, + callback_steps: int = 1, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + image (`np.ndarray` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image` + will be used as a starting point, adding more noise to it the larger the `strength`. The number of + denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will + be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter will be modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`np.random.RandomState`, *optional*): + A np.random.RandomState to make generation deterministic. + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: np.ndarray)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + + # check inputs. Raise error if not correct + self.check_inputs(prompt, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds) + + # define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if generator is None: + generator = np.random + + # set timesteps + self.scheduler.set_timesteps(num_inference_steps) + + image = preprocess(image).cpu().numpy() + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + prompt_embeds = self._encode_prompt( + prompt, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + latents_dtype = prompt_embeds.dtype + image = image.astype(latents_dtype) + # encode the init image into latents and scale the latents + init_latents = self.vae_encoder(sample=image)[0] + init_latents = 0.18215 * init_latents + + if isinstance(prompt, str): + prompt = [prompt] + if len(prompt) > init_latents.shape[0] and len(prompt) % init_latents.shape[0] == 0: + # expand init_latents for batch_size + deprecation_message = ( + f"You have passed {len(prompt)} text prompts (`prompt`), but only {init_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = len(prompt) // init_latents.shape[0] + init_latents = np.concatenate([init_latents] * additional_image_per_prompt * num_images_per_prompt, axis=0) + elif len(prompt) > init_latents.shape[0] and len(prompt) % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {len(prompt)} text prompts." + ) + else: + init_latents = np.concatenate([init_latents] * num_images_per_prompt, axis=0) + + # get the original timestep using init_timestep + offset = self.scheduler.config.get("steps_offset", 0) + init_timestep = int(num_inference_steps * strength) + offset + init_timestep = min(init_timestep, num_inference_steps) + + timesteps = self.scheduler.timesteps.numpy()[-init_timestep] + timesteps = np.array([timesteps] * batch_size * num_images_per_prompt) + + # add noise to latents using the timesteps + noise = generator.randn(*init_latents.shape).astype(latents_dtype) + init_latents = self.scheduler.add_noise( + torch.from_numpy(init_latents), torch.from_numpy(noise), torch.from_numpy(timesteps) + ) + init_latents = init_latents.numpy() + + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + latents = init_latents + + t_start = max(num_inference_steps - init_timestep + offset, 0) + timesteps = self.scheduler.timesteps[t_start:].numpy() + + timestep_dtype = next( + (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)" + ) + timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype] + + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(torch.from_numpy(latent_model_input), t) + latent_model_input = latent_model_input.cpu().numpy() + + # predict the noise residual + timestep = np.array([t], dtype=timestep_dtype) + noise_pred = self.unet(sample=latent_model_input, timestep=timestep, encoder_hidden_states=prompt_embeds)[ + 0 + ] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + scheduler_output = self.scheduler.step( + torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs + ) + latents = scheduler_output.prev_sample.numpy() + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + latents = 1 / 0.18215 * latents + # image = self.vae_decoder(latent_sample=latents)[0] + # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1 + image = np.concatenate( + [self.vae_decoder(latent_sample=latents[i : i + 1])[0] for i in range(latents.shape[0])] + ) + + image = np.clip(image / 2 + 0.5, 0, 1) + image = image.transpose((0, 2, 3, 1)) + + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor( + self.numpy_to_pil(image), return_tensors="np" + ).pixel_values.astype(image.dtype) + # safety_checker does not support batched inputs yet + images, has_nsfw_concept = [], [] + for i in range(image.shape[0]): + image_i, has_nsfw_concept_i = self.safety_checker( + clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1] + ) + images.append(image_i) + has_nsfw_concept.append(has_nsfw_concept_i[0]) + image = np.concatenate(images) + else: + has_nsfw_concept = None + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..6a952a7ae650e08b4aab96ccd65b24f20b31663f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py @@ -0,0 +1,564 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Callable, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTokenizer + +from ...configuration_utils import FrozenDict +from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler +from ...utils import PIL_INTERPOLATION, deprecate, logging +from ..onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel +from ..pipeline_utils import DiffusionPipeline +from . import StableDiffusionPipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +NUM_UNET_INPUT_CHANNELS = 9 +NUM_LATENT_CHANNELS = 4 + + +def prepare_mask_and_masked_image(image, mask, latents_shape): + image = np.array(image.convert("RGB").resize((latents_shape[1] * 8, latents_shape[0] * 8))) + image = image[None].transpose(0, 3, 1, 2) + image = image.astype(np.float32) / 127.5 - 1.0 + + image_mask = np.array(mask.convert("L").resize((latents_shape[1] * 8, latents_shape[0] * 8))) + masked_image = image * (image_mask < 127.5) + + mask = mask.resize((latents_shape[1], latents_shape[0]), PIL_INTERPOLATION["nearest"]) + mask = np.array(mask.convert("L")) + mask = mask.astype(np.float32) / 255.0 + mask = mask[None, None] + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + + return mask, masked_image + + +class OnnxStableDiffusionInpaintPipeline(DiffusionPipeline): + r""" + Pipeline for text-guided image inpainting using Stable Diffusion. *This is an experimental feature*. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + details. + feature_extractor ([`CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + vae_encoder: OnnxRuntimeModel + vae_decoder: OnnxRuntimeModel + text_encoder: OnnxRuntimeModel + tokenizer: CLIPTokenizer + unet: OnnxRuntimeModel + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler] + safety_checker: OnnxRuntimeModel + feature_extractor: CLIPImageProcessor + + _optional_components = ["safety_checker", "feature_extractor"] + _is_onnx = True + + def __init__( + self, + vae_encoder: OnnxRuntimeModel, + vae_decoder: OnnxRuntimeModel, + text_encoder: OnnxRuntimeModel, + tokenizer: CLIPTokenizer, + unet: OnnxRuntimeModel, + scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler], + safety_checker: OnnxRuntimeModel, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + logger.info("`OnnxStableDiffusionInpaintPipeline` is experimental and will very likely change in the future.") + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae_encoder=vae_encoder, + vae_decoder=vae_decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt: Union[str, List[str]], + num_images_per_prompt: Optional[int], + do_classifier_free_guidance: bool, + negative_prompt: Optional[str], + prompt_embeds: Optional[np.ndarray] = None, + negative_prompt_embeds: Optional[np.ndarray] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`): + prompt to be encoded + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + """ + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="np", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids + + if not np.array_equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0] + + prompt_embeds = np.repeat(prompt_embeds, num_images_per_prompt, axis=0) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] * batch_size + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="np", + ) + negative_prompt_embeds = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0] + + if do_classifier_free_guidance: + negative_prompt_embeds = np.repeat(negative_prompt_embeds, num_images_per_prompt, axis=0) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = np.concatenate([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline.check_inputs + def check_inputs( + self, + prompt: Union[str, List[str]], + height: Optional[int], + width: Optional[int], + callback_steps: int, + negative_prompt: Optional[str] = None, + prompt_embeds: Optional[np.ndarray] = None, + negative_prompt_embeds: Optional[np.ndarray] = None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]], + image: PIL.Image.Image, + mask_image: PIL.Image.Image, + height: Optional[int] = 512, + width: Optional[int] = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[np.random.RandomState] = None, + latents: Optional[np.ndarray] = None, + prompt_embeds: Optional[np.ndarray] = None, + negative_prompt_embeds: Optional[np.ndarray] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, np.ndarray], None]] = None, + callback_steps: int = 1, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will + be masked out with `mask_image` and repainted according to `prompt`. + mask_image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted + to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L) + instead of 3, so the expected shape would be `(B, H, W, 1)`. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`np.random.RandomState`, *optional*): + A np.random.RandomState to make generation deterministic. + latents (`np.ndarray`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: np.ndarray)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + + # check inputs. Raise error if not correct + self.check_inputs( + prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds + ) + + # define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if generator is None: + generator = np.random + + # set timesteps + self.scheduler.set_timesteps(num_inference_steps) + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + prompt_embeds = self._encode_prompt( + prompt, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + num_channels_latents = NUM_LATENT_CHANNELS + latents_shape = (batch_size * num_images_per_prompt, num_channels_latents, height // 8, width // 8) + latents_dtype = prompt_embeds.dtype + if latents is None: + latents = generator.randn(*latents_shape).astype(latents_dtype) + else: + if latents.shape != latents_shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}") + + # prepare mask and masked_image + mask, masked_image = prepare_mask_and_masked_image(image, mask_image, latents_shape[-2:]) + mask = mask.astype(latents.dtype) + masked_image = masked_image.astype(latents.dtype) + + masked_image_latents = self.vae_encoder(sample=masked_image)[0] + masked_image_latents = 0.18215 * masked_image_latents + + # duplicate mask and masked_image_latents for each generation per prompt + mask = mask.repeat(batch_size * num_images_per_prompt, 0) + masked_image_latents = masked_image_latents.repeat(batch_size * num_images_per_prompt, 0) + + mask = np.concatenate([mask] * 2) if do_classifier_free_guidance else mask + masked_image_latents = ( + np.concatenate([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + ) + + num_channels_mask = mask.shape[1] + num_channels_masked_image = masked_image_latents.shape[1] + + unet_input_channels = NUM_UNET_INPUT_CHANNELS + if num_channels_latents + num_channels_mask + num_channels_masked_image != unet_input_channels: + raise ValueError( + "Incorrect configuration settings! The config of `pipeline.unet` expects" + f" {unet_input_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of" + " `pipeline.unet` or your `mask_image` or `image` input." + ) + + # set timesteps + self.scheduler.set_timesteps(num_inference_steps) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * np.float64(self.scheduler.init_noise_sigma) + + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + timestep_dtype = next( + (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)" + ) + timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype] + + for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents + # concat latents, mask, masked_image_latnets in the channel dimension + latent_model_input = self.scheduler.scale_model_input(torch.from_numpy(latent_model_input), t) + latent_model_input = latent_model_input.cpu().numpy() + latent_model_input = np.concatenate([latent_model_input, mask, masked_image_latents], axis=1) + + # predict the noise residual + timestep = np.array([t], dtype=timestep_dtype) + noise_pred = self.unet(sample=latent_model_input, timestep=timestep, encoder_hidden_states=prompt_embeds)[ + 0 + ] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + scheduler_output = self.scheduler.step( + torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs + ) + latents = scheduler_output.prev_sample.numpy() + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + latents = 1 / 0.18215 * latents + # image = self.vae_decoder(latent_sample=latents)[0] + # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1 + image = np.concatenate( + [self.vae_decoder(latent_sample=latents[i : i + 1])[0] for i in range(latents.shape[0])] + ) + + image = np.clip(image / 2 + 0.5, 0, 1) + image = image.transpose((0, 2, 3, 1)) + + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor( + self.numpy_to_pil(image), return_tensors="np" + ).pixel_values.astype(image.dtype) + # safety_checker does not support batched inputs yet + images, has_nsfw_concept = [], [] + for i in range(image.shape[0]): + image_i, has_nsfw_concept_i = self.safety_checker( + clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1] + ) + images.append(image_i) + has_nsfw_concept.append(has_nsfw_concept_i[0]) + image = np.concatenate(images) + else: + has_nsfw_concept = None + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_upscale.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_upscale.py new file mode 100644 index 0000000000000000000000000000000000000000..3f10764dc7b436e1aacd200ecd442183d3755d23 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_upscale.py @@ -0,0 +1,586 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTokenizer + +from ...configuration_utils import FrozenDict +from ...schedulers import DDPMScheduler, KarrasDiffusionSchedulers +from ...utils import deprecate, logging +from ..onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel +from ..pipeline_utils import DiffusionPipeline +from . import StableDiffusionPipelineOutput + + +logger = logging.get_logger(__name__) + + +def preprocess(image): + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 64 for x in (w, h)) # resize to integer multiple of 32 + + image = [np.array(i.resize((w, h)))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + + return image + + +class OnnxStableDiffusionUpscalePipeline(DiffusionPipeline): + vae: OnnxRuntimeModel + text_encoder: OnnxRuntimeModel + tokenizer: CLIPTokenizer + unet: OnnxRuntimeModel + low_res_scheduler: DDPMScheduler + scheduler: KarrasDiffusionSchedulers + safety_checker: OnnxRuntimeModel + feature_extractor: CLIPImageProcessor + + _optional_components = ["safety_checker", "feature_extractor"] + _is_onnx = True + + def __init__( + self, + vae: OnnxRuntimeModel, + text_encoder: OnnxRuntimeModel, + tokenizer: Any, + unet: OnnxRuntimeModel, + low_res_scheduler: DDPMScheduler, + scheduler: KarrasDiffusionSchedulers, + safety_checker: Optional[OnnxRuntimeModel] = None, + feature_extractor: Optional[CLIPImageProcessor] = None, + max_noise_level: int = 350, + num_latent_channels=4, + num_unet_input_channels=7, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + low_res_scheduler=low_res_scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.register_to_config( + max_noise_level=max_noise_level, + num_latent_channels=num_latent_channels, + num_unet_input_channels=num_unet_input_channels, + ) + + def check_inputs( + self, + prompt: Union[str, List[str]], + image, + noise_level, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ( + not isinstance(image, torch.Tensor) + and not isinstance(image, PIL.Image.Image) + and not isinstance(image, np.ndarray) + and not isinstance(image, list) + ): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `np.ndarray`, `PIL.Image.Image` or `list` but is {type(image)}" + ) + + # verify batch size of prompt and image are same if image is a list or tensor or numpy array + if isinstance(image, (list, np.ndarray)): + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if isinstance(image, list): + image_batch_size = len(image) + else: + image_batch_size = image.shape[0] + if batch_size != image_batch_size: + raise ValueError( + f"`prompt` has batch size {batch_size} and `image` has batch size {image_batch_size}." + " Please make sure that passed `prompt` matches the batch size of `image`." + ) + + # check noise level + if noise_level > self.config.max_noise_level: + raise ValueError(f"`noise_level` has to be <= {self.config.max_noise_level} but is {noise_level}") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, generator, latents=None): + shape = (batch_size, num_channels_latents, height, width) + if latents is None: + latents = generator.randn(*shape).astype(dtype) + elif latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + + return latents + + def decode_latents(self, latents): + latents = 1 / 0.08333 * latents + image = self.vae(latent_sample=latents)[0] + image = np.clip(image / 2 + 0.5, 0, 1) + image = image.transpose((0, 2, 3, 1)) + return image + + def _encode_prompt( + self, + prompt: Union[str, List[str]], + num_images_per_prompt: Optional[int], + do_classifier_free_guidance: bool, + negative_prompt: Optional[str], + prompt_embeds: Optional[np.ndarray] = None, + negative_prompt_embeds: Optional[np.ndarray] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`): + prompt to be encoded + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + """ + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="np", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids + + if not np.array_equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0] + + prompt_embeds = np.repeat(prompt_embeds, num_images_per_prompt, axis=0) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] * batch_size + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="np", + ) + negative_prompt_embeds = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0] + + if do_classifier_free_guidance: + negative_prompt_embeds = np.repeat(negative_prompt_embeds, num_images_per_prompt, axis=0) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = np.concatenate([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + def __call__( + self, + prompt: Union[str, List[str]], + image: Union[np.ndarray, PIL.Image.Image, List[PIL.Image.Image]], + num_inference_steps: int = 75, + guidance_scale: float = 9.0, + noise_level: int = 20, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[np.random.RandomState, List[np.random.RandomState]]] = None, + latents: Optional[np.ndarray] = None, + prompt_embeds: Optional[np.ndarray] = None, + negative_prompt_embeds: Optional[np.ndarray] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, np.ndarray], None]] = None, + callback_steps: Optional[int] = 1, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + image (`np.ndarray` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter will be modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + noise_level (`float`, defaults to 0.2): + Deteremines the amount of noise to add to the initial image before performing upscaling. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`np.random.RandomState`, *optional*): + A np.random.RandomState to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: np.ndarray)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + + # 1. Check inputs + self.check_inputs( + prompt, + image, + noise_level, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if generator is None: + generator = np.random + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + prompt_embeds = self._encode_prompt( + prompt, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + latents_dtype = prompt_embeds.dtype + image = preprocess(image).cpu().numpy() + height, width = image.shape[2:] + + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + self.config.num_latent_channels, + height, + width, + latents_dtype, + generator, + ) + image = image.astype(latents_dtype) + + self.scheduler.set_timesteps(num_inference_steps) + timesteps = self.scheduler.timesteps + + # Scale the initial noise by the standard deviation required by the scheduler + latents = latents * np.float64(self.scheduler.init_noise_sigma) + + # 5. Add noise to image + noise_level = np.array([noise_level]).astype(np.int64) + noise = generator.randn(*image.shape).astype(latents_dtype) + + image = self.low_res_scheduler.add_noise( + torch.from_numpy(image), torch.from_numpy(noise), torch.from_numpy(noise_level) + ) + image = image.numpy() + + batch_multiplier = 2 if do_classifier_free_guidance else 1 + image = np.concatenate([image] * batch_multiplier * num_images_per_prompt) + noise_level = np.concatenate([noise_level] * image.shape[0]) + + # 7. Check that sizes of image and latents match + num_channels_image = image.shape[1] + if self.config.num_latent_channels + num_channels_image != self.config.num_unet_input_channels: + raise ValueError( + "Incorrect configuration settings! The config of `pipeline.unet` expects" + f" {self.config.num_unet_input_channels} but received `num_channels_latents`: {self.config.num_latent_channels} +" + f" `num_channels_image`: {num_channels_image} " + f" = {self.config.num_latent_channels + num_channels_image}. Please verify the config of" + " `pipeline.unet` or your `image` input." + ) + + # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + timestep_dtype = next( + (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)" + ) + timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype] + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents + + # concat latents, mask, masked_image_latents in the channel dimension + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + latent_model_input = np.concatenate([latent_model_input, image], axis=1) + + # timestep to tensor + timestep = np.array([t], dtype=timestep_dtype) + + # predict the noise residual + noise_pred = self.unet( + sample=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + class_labels=noise_level, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step( + torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs + ).prev_sample + latents = latents.numpy() + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + # 10. Post-processing + image = self.decode_latents(latents) + + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor( + self.numpy_to_pil(image), return_tensors="np" + ).pixel_values.astype(image.dtype) + + images, has_nsfw_concept = [], [] + for i in range(image.shape[0]): + image_i, has_nsfw_concept_i = self.safety_checker( + clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1] + ) + images.append(image_i) + has_nsfw_concept.append(has_nsfw_concept_i[0]) + image = np.concatenate(images) + else: + has_nsfw_concept = None + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_output.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..5fb9b1a1412d96b69144a4c2e960dcc8b75a615c --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_output.py @@ -0,0 +1,45 @@ +from dataclasses import dataclass +from typing import List, Optional, Union + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput, is_flax_available + + +@dataclass +class StableDiffusionPipelineOutput(BaseOutput): + """ + Output class for Stable Diffusion pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + nsfw_content_detected (`List[bool]`) + List indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content or + `None` if safety checking could not be performed. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] + nsfw_content_detected: Optional[List[bool]] + + +if is_flax_available(): + import flax + + @flax.struct.dataclass + class FlaxStableDiffusionPipelineOutput(BaseOutput): + """ + Output class for Flax-based Stable Diffusion pipelines. + + Args: + images (`np.ndarray`): + Denoised images of array shape of `(batch_size, height, width, num_channels)`. + nsfw_content_detected (`List[bool]`): + List indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content + or `None` if safety checking could not be performed. + """ + + images: np.ndarray + nsfw_content_detected: List[bool] diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..acf685784eb13e171fea0bce8dc5b9afdf5be71d --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py @@ -0,0 +1,1103 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...configuration_utils import FrozenDict +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from .pipeline_output import StableDiffusionPipelineOutput +from .safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionPipeline + + >>> pipe = StableDiffusionPipeline.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt).images[0] + ``` +""" + + +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + IPAdapterMixin, + FromSingleFileMixin, +): + """ + Pipeline for text-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + self._is_unet_config_sample_size_int = unet is not None and isinstance(unet.config.sample_size, int) + is_unet_sample_size_less_64 = ( + unet is not None + and hasattr(unet.config, "sample_size") + and self._is_unet_config_sample_size_int + and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + timesteps: List[int] = None, + sigmas: List[float] = None, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default height and width to unet + if not height or not width: + height = ( + self.unet.config.sample_size + if self._is_unet_config_sample_size_int + else self.unet.config.sample_size[0] + ) + width = ( + self.unet.config.sample_size + if self._is_unet_config_sample_size_int + else self.unet.config.sample_size[1] + ) + height, width = height * self.vae_scale_factor, width * self.vae_scale_factor + # to deal with lora scaling and other possible forward hooks + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": image_embeds} + if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) + else None + ) + + # 6.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py new file mode 100644 index 0000000000000000000000000000000000000000..e957c6661f873f35dd81488dd40b1fd1256a3a91 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py @@ -0,0 +1,897 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 contextlib +import inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from packaging import version +from transformers import CLIPTextModel, CLIPTokenizer, DPTForDepthEstimation, DPTImageProcessor + +from ...configuration_utils import FrozenDict +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + PIL_INTERPOLATION, + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess +def preprocess(image): + deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead" + deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + + image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +class StableDiffusionDepth2ImgPipeline(DiffusionPipeline, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin): + r""" + Pipeline for text-guided depth-based image-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "depth_mask"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + depth_estimator: DPTForDepthEstimation, + feature_extractor: DPTImageProcessor, + ): + super().__init__() + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + depth_estimator=depth_estimator, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + strength, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.prepare_latents + def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + if image.shape[0] < batch_size and batch_size % image.shape[0] == 0: + image = torch.cat([image] * (batch_size // image.shape[0]), dim=0) + elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} " + ) + + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + latents = init_latents + + return latents + + def prepare_depth_map(self, image, depth_map, batch_size, do_classifier_free_guidance, dtype, device): + if isinstance(image, PIL.Image.Image): + image = [image] + else: + image = list(image) + + if isinstance(image[0], PIL.Image.Image): + width, height = image[0].size + elif isinstance(image[0], np.ndarray): + width, height = image[0].shape[:-1] + else: + height, width = image[0].shape[-2:] + + if depth_map is None: + pixel_values = self.feature_extractor(images=image, return_tensors="pt").pixel_values + pixel_values = pixel_values.to(device=device, dtype=dtype) + # The DPT-Hybrid model uses batch-norm layers which are not compatible with fp16. + # So we use `torch.autocast` here for half precision inference. + if torch.backends.mps.is_available(): + autocast_ctx = contextlib.nullcontext() + logger.warning( + "The DPT-Hybrid model uses batch-norm layers which are not compatible with fp16, but autocast is not yet supported on MPS." + ) + else: + autocast_ctx = torch.autocast(device.type, dtype=dtype) + + with autocast_ctx: + depth_map = self.depth_estimator(pixel_values).predicted_depth + else: + depth_map = depth_map.to(device=device, dtype=dtype) + + depth_map = torch.nn.functional.interpolate( + depth_map.unsqueeze(1), + size=(height // self.vae_scale_factor, width // self.vae_scale_factor), + mode="bicubic", + align_corners=False, + ) + + depth_min = torch.amin(depth_map, dim=[1, 2, 3], keepdim=True) + depth_max = torch.amax(depth_map, dim=[1, 2, 3], keepdim=True) + depth_map = 2.0 * (depth_map - depth_min) / (depth_max - depth_min) - 1.0 + depth_map = depth_map.to(dtype) + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if depth_map.shape[0] < batch_size: + repeat_by = batch_size // depth_map.shape[0] + depth_map = depth_map.repeat(repeat_by, 1, 1, 1) + + depth_map = torch.cat([depth_map] * 2) if do_classifier_free_guidance else depth_map + return depth_map + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]] = None, + image: PipelineImageInput = None, + depth_map: Optional[torch.Tensor] = None, + strength: float = 0.8, + num_inference_steps: Optional[int] = 50, + guidance_scale: Optional[float] = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: Optional[float] = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image` or tensor representing an image batch to be used as the starting point. Can accept image + latents as `image` only if `depth_map` is not `None`. + depth_map (`torch.Tensor`, *optional*): + Depth prediction to be used as additional conditioning for the image generation process. If not + defined, it automatically predicts the depth with `self.depth_estimator`. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter is modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + Examples: + + ```py + >>> import torch + >>> import requests + >>> from PIL import Image + + >>> from diffusers import StableDiffusionDepth2ImgPipeline + + >>> pipe = StableDiffusionDepth2ImgPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-2-depth", + ... torch_dtype=torch.float16, + ... ) + >>> pipe.to("cuda") + + + >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" + >>> init_image = Image.open(requests.get(url, stream=True).raw) + >>> prompt = "two tigers" + >>> n_prompt = "bad, deformed, ugly, bad anotomy" + >>> image = pipe(prompt=prompt, image=init_image, negative_prompt=n_prompt, strength=0.7).images[0] + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + # 1. Check inputs + self.check_inputs( + prompt, + strength, + callback_steps, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + + if image is None: + raise ValueError("`image` input cannot be undefined.") + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare depth mask + depth_mask = self.prepare_depth_map( + image, + depth_map, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + prompt_embeds.dtype, + device, + ) + + # 5. Preprocess image + image = self.image_processor.preprocess(image) + + # 6. Set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 7. Prepare latent variables + latents = self.prepare_latents( + image, latent_timestep, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator + ) + + # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + latent_model_input = torch.cat([latent_model_input, depth_mask], dim=1) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=self.cross_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + depth_mask = callback_outputs.pop("depth_mask", depth_mask) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = latents + + image = self.image_processor.postprocess(image, output_type=output_type) + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py new file mode 100644 index 0000000000000000000000000000000000000000..d47e2f0593dd1e2f86bcfc4332eaa78a6c60563d --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py @@ -0,0 +1,439 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Callable, List, Optional, Union + +import PIL.Image +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection + +from ...configuration_utils import FrozenDict +from ...image_processor import VaeImageProcessor +from ...models import AutoencoderKL, UNet2DConditionModel +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import deprecate, is_torch_xla_available, logging +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from . import StableDiffusionPipelineOutput +from .safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class StableDiffusionImageVariationPipeline(DiffusionPipeline, StableDiffusionMixin): + r""" + Pipeline to generate image variations from an input image using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + image_encoder ([`~transformers.CLIPVisionModelWithProjection`]): + Frozen CLIP image-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + # TODO: feature_extractor is required to encode images (if they are in PIL format), + # we should give a descriptive message if the pipeline doesn't have one. + _optional_components = ["safety_checker"] + model_cpu_offload_seq = "image_encoder->unet->vae" + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + image_encoder: CLIPVisionModelWithProjection, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + image_encoder=image_encoder, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + def _encode_image(self, image, device, num_images_per_prompt, do_classifier_free_guidance): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(images=image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeddings = self.image_encoder(image).image_embeds + image_embeddings = image_embeddings.unsqueeze(1) + + # duplicate image embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = image_embeddings.shape + image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1) + image_embeddings = image_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + negative_prompt_embeds = torch.zeros_like(image_embeddings) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + image_embeddings = torch.cat([negative_prompt_embeds, image_embeddings]) + + return image_embeddings + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs(self, image, height, width, callback_steps): + if ( + not isinstance(image, torch.Tensor) + and not isinstance(image, PIL.Image.Image) + and not isinstance(image, list) + ): + raise ValueError( + "`image` has to be of type `torch.Tensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is" + f" {type(image)}" + ) + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def __call__( + self, + image: Union[PIL.Image.Image, List[PIL.Image.Image], torch.Tensor], + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.Tensor`): + Image or images to guide image generation. If you provide a tensor, it needs to be compatible with + [`CLIPImageProcessor`](https://huggingface.co/lambdalabs/sd-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json). + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter is modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + + Examples: + + ```py + from diffusers import StableDiffusionImageVariationPipeline + from PIL import Image + from io import BytesIO + import requests + + pipe = StableDiffusionImageVariationPipeline.from_pretrained( + "lambdalabs/sd-image-variations-diffusers", revision="v2.0" + ) + pipe = pipe.to("cuda") + + url = "https://lh3.googleusercontent.com/y-iFOHfLTwkuQSUegpwDdgKmOjRSTvPxat63dQLB25xkTs4lhIbRUFeNBWZzYf370g=s1200" + + response = requests.get(url) + image = Image.open(BytesIO(response.content)).convert("RGB") + + out = pipe(image, num_images_per_prompt=3, guidance_scale=15) + out["images"][0].save("result.jpg") + ``` + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs(image, height, width, callback_steps) + + # 2. Define call parameters + if isinstance(image, PIL.Image.Image): + batch_size = 1 + elif isinstance(image, list): + batch_size = len(image) + else: + batch_size = image.shape[0] + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input image + image_embeddings = self._encode_image(image, device, num_images_per_prompt, do_classifier_free_guidance) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + image_embeddings.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=image_embeddings).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + self.maybe_free_model_hooks() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, image_embeddings.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..95d3ab06f02a2f34a5d71f5796a79f578f276a73 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py @@ -0,0 +1,1161 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...configuration_utils import FrozenDict +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + PIL_INTERPOLATION, + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from . import StableDiffusionPipelineOutput +from .safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import requests + >>> import torch + >>> from PIL import Image + >>> from io import BytesIO + + >>> from diffusers import StableDiffusionImg2ImgPipeline + + >>> device = "cuda" + >>> model_id_or_path = "stable-diffusion-v1-5/stable-diffusion-v1-5" + >>> pipe = StableDiffusionImg2ImgPipeline.from_pretrained(model_id_or_path, torch_dtype=torch.float16) + >>> pipe = pipe.to(device) + + >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + + >>> response = requests.get(url) + >>> init_image = Image.open(BytesIO(response.content)).convert("RGB") + >>> init_image = init_image.resize((768, 512)) + + >>> prompt = "A fantasy landscape, trending on artstation" + + >>> images = pipe(prompt=prompt, image=init_image, strength=0.75, guidance_scale=7.5).images + >>> images[0].save("fantasy_landscape.png") + ``` +""" + + +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +def preprocess(image): + deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead" + deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + + image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionImg2ImgPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, + StableDiffusionLoraLoaderMixin, + FromSingleFileMixin, +): + r""" + Pipeline for text-guided image-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + strength, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + if image.shape[0] < batch_size and batch_size % image.shape[0] == 0: + image = torch.cat([image] * (batch_size // image.shape[0]), dim=0) + elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} " + ) + + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + latents = init_latents + + return latents + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + image: PipelineImageInput = None, + strength: float = 0.8, + num_inference_steps: Optional[int] = 50, + timesteps: List[int] = None, + sigmas: List[float] = None, + guidance_scale: Optional[float] = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: Optional[float] = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + clip_skip: int = None, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter is modulated by `strength`. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + strength, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 4. Preprocess image + image = self.image_processor.preprocess(image) + + # 5. set timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 6. Prepare latent variables + latents = self.prepare_latents( + image, + latent_timestep, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + ) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + # 7.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..148d7386a732dc5aea824d6dd4a2dda919551c2a --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py @@ -0,0 +1,1359 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import PIL.Image +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...configuration_utils import FrozenDict +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AsymmetricAutoencoderKL, AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from . import StableDiffusionPipelineOutput +from .safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionInpaintPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, + StableDiffusionLoraLoaderMixin, + FromSingleFileMixin, +): + r""" + Pipeline for text-guided image inpainting using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + + Args: + vae ([`AutoencoderKL`, `AsymmetricAutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "mask", "masked_image_latents"] + + def __init__( + self, + vae: Union[AutoencoderKL, AsymmetricAutoencoderKL], + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "skip_prk_steps", True) is False: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration" + " `skip_prk_steps`. `skip_prk_steps` should be set to True in the configuration file. Please make" + " sure to update the config accordingly as not setting `skip_prk_steps` in the config might lead to" + " incorrect results in future versions. If you have downloaded this checkpoint from the Hugging Face" + " Hub, it would be very nice if you could open a Pull request for the" + " `scheduler/scheduler_config.json` file" + ) + deprecate("skip_prk_steps not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["skip_prk_steps"] = True + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + # Check shapes, assume num_channels_latents == 4, num_channels_mask == 1, num_channels_masked == 4 + if unet is not None and unet.config.in_channels != 9: + logger.info(f"You have loaded a UNet with {unet.config.in_channels} input channels which.") + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True + ) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + image, + mask_image, + height, + width, + strength, + callback_steps, + output_type, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % self.vae_scale_factor != 0 or width % self.vae_scale_factor != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + image=None, + timestep=None, + is_strength_max=True, + return_noise=False, + return_image_latents=False, + ): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if (image is None or timestep is None) and not is_strength_max: + raise ValueError( + "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise." + "However, either the image or the noise timestep has not been provided." + ) + + if return_image_latents or (latents is None and not is_strength_max): + image = image.to(device=device, dtype=dtype) + + if image.shape[1] == 4: + image_latents = image + else: + image_latents = self._encode_vae_image(image=image, generator=generator) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + + if latents is None: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # if strength is 1. then initialise the latents to noise, else initial to image + noise + latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep) + # if pure noise then scale the initial latents by the Scheduler's init sigma + latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents + else: + noise = latents.to(device) + latents = noise * self.scheduler.init_noise_sigma + + outputs = (latents,) + + if return_noise: + outputs += (noise,) + + if return_image_latents: + outputs += (image_latents,) + + return outputs + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = self.vae.config.scaling_factor * image_latents + + return image_latents + + def prepare_mask_latents( + self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance + ): + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate( + mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor) + ) + mask = mask.to(device=device, dtype=dtype) + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == 4: + masked_image_latents = masked_image + else: + masked_image_latents = self._encode_vae_image(masked_image, generator=generator) + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + + mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask + masked_image_latents = ( + torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + ) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + return mask, masked_image_latents + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]] = None, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: torch.Tensor = None, + height: Optional[int] = None, + width: Optional[int] = None, + padding_mask_crop: Optional[int] = None, + strength: float = 1.0, + num_inference_steps: int = 50, + timesteps: List[int] = None, + sigmas: List[float] = None, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + clip_skip: int = None, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be inpainted (which parts of the image to + be masked out with `mask_image` and repainted according to `prompt`). For both numpy array and pytorch + tensor, the expected value range is between `[0, 1]` If it's a tensor or a list or tensors, the + expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a list of arrays, the + expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image latents as `image`, but + if passing latents directly it is not encoded again. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter is modulated by `strength`. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + Examples: + + ```py + >>> import PIL + >>> import requests + >>> import torch + >>> from io import BytesIO + + >>> from diffusers import StableDiffusionInpaintPipeline + + + >>> def download_image(url): + ... response = requests.get(url) + ... return PIL.Image.open(BytesIO(response.content)).convert("RGB") + + + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + + >>> init_image = download_image(img_url).resize((512, 512)) + >>> mask_image = download_image(mask_url).resize((512, 512)) + + >>> pipe = StableDiffusionInpaintPipeline.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-inpainting", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench" + >>> image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0] + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs + self.check_inputs( + prompt, + image, + mask_image, + height, + width, + strength, + callback_steps, + output_type, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + padding_mask_crop, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 4. set timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps=num_inference_steps, strength=strength, device=device + ) + # check that number of inference steps is not < 1 - as this doesn't make sense + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + # at which timestep to set the initial noise (n.b. 50% if strength is 0.5) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise + is_strength_max = strength == 1.0 + + # 5. Preprocess mask and image + + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + # 6. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + num_channels_unet = self.unet.config.in_channels + return_image_latents = num_channels_unet == 4 + + latents_outputs = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + image=init_image, + timestep=latent_timestep, + is_strength_max=is_strength_max, + return_noise=True, + return_image_latents=return_image_latents, + ) + + if return_image_latents: + latents, noise, image_latents = latents_outputs + else: + latents, noise = latents_outputs + + # 7. Prepare mask latent variables + mask_condition = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + if masked_image_latents is None: + masked_image = init_image * (mask_condition < 0.5) + else: + masked_image = masked_image_latents + + mask, masked_image_latents = self.prepare_mask_latents( + mask_condition, + masked_image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + self.do_classifier_free_guidance, + ) + + # 8. Check that sizes of mask, masked image and latents match + if num_channels_unet == 9: + # default case for stable-diffusion-v1-5/stable-diffusion-inpainting + num_channels_mask = mask.shape[1] + num_channels_masked_image = masked_image_latents.shape[1] + if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of" + " `pipeline.unet` or your `mask_image` or `image` input." + ) + elif num_channels_unet != 4: + raise ValueError( + f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}." + ) + + # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 9.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + # 9.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 10. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + # concat latents, mask, masked_image_latents in the channel dimension + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + if num_channels_unet == 9: + latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if num_channels_unet == 4: + init_latents_proper = image_latents + if self.do_classifier_free_guidance: + init_mask, _ = mask.chunk(2) + else: + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.add_noise( + init_latents_proper, noise, torch.tensor([noise_timestep]) + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + mask = callback_outputs.pop("mask", mask) + masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + condition_kwargs = {} + if isinstance(self.vae, AsymmetricAutoencoderKL): + init_image = init_image.to(device=device, dtype=masked_image_latents.dtype) + init_image_condition = init_image.clone() + init_image = self._encode_vae_image(init_image, generator=generator) + mask_condition = mask_condition.to(device=device, dtype=masked_image_latents.dtype) + condition_kwargs = {"image": init_image_condition, "mask": mask_condition} + image = self.vae.decode( + latents / self.vae.config.scaling_factor, return_dict=False, generator=generator, **condition_kwargs + )[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + if padding_mask_crop is not None: + image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image] + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py new file mode 100644 index 0000000000000000000000000000000000000000..843d25d67c1040d5f3a4906daa400e1afae644c6 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py @@ -0,0 +1,917 @@ +# Copyright 2025 The InstructPix2Pix Authors and The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import PIL_INTERPOLATION, deprecate, is_torch_xla_available, logging +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from . import StableDiffusionPipelineOutput +from .safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess +def preprocess(image): + deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead" + deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + + image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class StableDiffusionInstructPix2PixPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + IPAdapterMixin, + FromSingleFileMixin, +): + r""" + Pipeline for pixel-level image editing by following text instructions (based on Stable Diffusion). + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "image_latents"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: Optional[CLIPVisionModelWithProjection] = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]] = None, + image: PipelineImageInput = None, + num_inference_steps: int = 100, + guidance_scale: float = 7.5, + image_guidance_scale: float = 1.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor` `np.ndarray`, `PIL.Image.Image`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image` or tensor representing an image batch to be repainted according to `prompt`. Can also accept + image latents as `image`, but if passing latents directly it is not encoded again. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + image_guidance_scale (`float`, *optional*, defaults to 1.5): + Push the generated image towards the initial `image`. Image guidance scale is enabled by setting + `image_guidance_scale > 1`. Higher image guidance scale encourages generated images that are closely + linked to the source `image`, usually at the expense of lower image quality. This pipeline requires a + value of at least `1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + + Examples: + + ```py + >>> import PIL + >>> import requests + >>> import torch + >>> from io import BytesIO + + >>> from diffusers import StableDiffusionInstructPix2PixPipeline + + + >>> def download_image(url): + ... response = requests.get(url) + ... return PIL.Image.open(BytesIO(response.content)).convert("RGB") + + + >>> img_url = "https://huggingface.co/datasets/diffusers/diffusers-images-docs/resolve/main/mountain.png" + + >>> image = download_image(img_url).resize((512, 512)) + + >>> pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained( + ... "timbrooks/instruct-pix2pix", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "make the mountains snowy" + >>> image = pipe(prompt=prompt, image=image).images[0] + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Check inputs + self.check_inputs( + prompt, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + self._guidance_scale = guidance_scale + self._image_guidance_scale = image_guidance_scale + + device = self._execution_device + + if image is None: + raise ValueError("`image` input cannot be undefined.") + + # 1. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 2. Encode input prompt + prompt_embeds = self._encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + # 3. Preprocess image + image = self.image_processor.preprocess(image) + + # 4. set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare Image latents + image_latents = self.prepare_image_latents( + image, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + self.do_classifier_free_guidance, + ) + + height, width = image_latents.shape[-2:] + height = height * self.vae_scale_factor + width = width * self.vae_scale_factor + + # 6. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 7. Check that shapes of latents and image match the UNet channels + num_channels_image = image_latents.shape[1] + if num_channels_latents + num_channels_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_image`: {num_channels_image} " + f" = {num_channels_latents + num_channels_image}. Please verify the config of" + " `pipeline.unet` or your `image` input." + ) + + # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 8.1 Add image embeds for IP-Adapter + added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # Expand the latents if we are doing classifier free guidance. + # The latents are expanded 3 times because for pix2pix the guidance\ + # is applied for both the text and the input image. + latent_model_input = torch.cat([latents] * 3) if self.do_classifier_free_guidance else latents + + # concat latents, image_latents in the channel dimension + scaled_latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + scaled_latent_model_input = torch.cat([scaled_latent_model_input, image_latents], dim=1) + + # predict the noise residual + noise_pred = self.unet( + scaled_latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + added_cond_kwargs=added_cond_kwargs, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_text, noise_pred_image, noise_pred_uncond = noise_pred.chunk(3) + noise_pred = ( + noise_pred_uncond + + self.guidance_scale * (noise_pred_text - noise_pred_image) + + self.image_guidance_scale * (noise_pred_image - noise_pred_uncond) + ) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + image_latents = callback_outputs.pop("image_latents", image_latents) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_ prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + """ + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + else: + prompt_embeds_dtype = self.unet.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + # pix2pix has two negative embeddings, and unlike in other pipelines latents are ordered [prompt_embeds, negative_prompt_embeds, negative_prompt_embeds] + prompt_embeds = torch.cat([prompt_embeds, negative_prompt_embeds, negative_prompt_embeds]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + image_embeds = [] + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0) + single_negative_image_embeds = torch.stack( + [single_negative_image_embeds] * num_images_per_prompt, dim=0 + ) + + if do_classifier_free_guidance: + single_image_embeds = torch.cat( + [single_image_embeds, single_negative_image_embeds, single_negative_image_embeds] + ) + single_image_embeds = single_image_embeds.to(device) + + image_embeds.append(single_image_embeds) + else: + repeat_dims = [1] + image_embeds = [] + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + ( + single_image_embeds, + single_negative_image_embeds, + single_negative_image_embeds, + ) = single_image_embeds.chunk(3) + single_image_embeds = single_image_embeds.repeat( + num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:])) + ) + single_negative_image_embeds = single_negative_image_embeds.repeat( + num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:])) + ) + single_image_embeds = torch.cat( + [single_image_embeds, single_negative_image_embeds, single_negative_image_embeds] + ) + else: + single_image_embeds = single_image_embeds.repeat( + num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:])) + ) + image_embeds.append(single_image_embeds) + + return image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def check_inputs( + self, + prompt, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def prepare_image_latents( + self, image, batch_size, num_images_per_prompt, dtype, device, do_classifier_free_guidance, generator=None + ): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + image_latents = image + else: + image_latents = retrieve_latents(self.vae.encode(image), sample_mode="argmax") + + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand image_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {image_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + if do_classifier_free_guidance: + uncond_image_latents = torch.zeros_like(image_latents) + image_latents = torch.cat([image_latents, image_latents, uncond_image_latents], dim=0) + + return image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def image_guidance_scale(self): + return self._image_guidance_scale + + @property + def num_timesteps(self): + return self._num_timesteps + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self.guidance_scale > 1.0 and self.image_guidance_scale >= 1.0 diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py new file mode 100644 index 0000000000000000000000000000000000000000..66d5ffa6b8498ac9f9f9392b1c92f84fd52963a6 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py @@ -0,0 +1,665 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 warnings +from typing import Callable, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from transformers import CLIPTextModel, CLIPTokenizer + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...schedulers import EulerDiscreteScheduler +from ...utils import deprecate, is_torch_xla_available, logging +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput, StableDiffusionMixin + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.preprocess +def preprocess(image): + warnings.warn( + "The preprocess method is deprecated and will be removed in a future version. Please" + " use VaeImageProcessor.preprocess instead", + FutureWarning, + ) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 64 for x in (w, h)) # resize to integer multiple of 64 + + image = [np.array(i.resize((w, h)))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +class StableDiffusionLatentUpscalePipeline(DiffusionPipeline, StableDiffusionMixin, FromSingleFileMixin): + r""" + Pipeline for upscaling Stable Diffusion output image resolution by a factor of 2. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A [`EulerDiscreteScheduler`] to be used in combination with `unet` to denoise the encoded image latents. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: EulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, resample="bicubic") + + def _encode_prompt( + self, + prompt, + device, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + device=device, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + **kwargs, + ) + + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds]) + + return prompt_embeds, pooled_prompt_embeds + + def encode_prompt( + self, + prompt, + device, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list(int)`): + prompt to be encoded + device: (`torch.device`): + torch device + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + """ + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None or pooled_prompt_embeds is None: + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_length=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + text_encoder_out = self.text_encoder( + text_input_ids.to(device), + output_hidden_states=True, + ) + prompt_embeds = text_encoder_out.hidden_states[-1] + pooled_prompt_embeds = text_encoder_out.pooler_output + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + if negative_prompt_embeds is None or negative_pooled_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_length=True, + return_tensors="pt", + ) + + uncond_encoder_out = self.text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + negative_prompt_embeds = uncond_encoder_out.hidden_states[-1] + negative_pooled_prompt_embeds = uncond_encoder_out.pooler_output + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def check_inputs( + self, + prompt, + image, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ): + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and not isinstance(prompt, str) and not isinstance(prompt, list): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if ( + not isinstance(image, torch.Tensor) + and not isinstance(image, np.ndarray) + and not isinstance(image, PIL.Image.Image) + and not isinstance(image, list) + ): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or `list` but is {type(image)}" + ) + + # verify batch size of prompt and image are same if image is a list or tensor + if isinstance(image, (list, torch.Tensor)): + if prompt is not None: + if isinstance(prompt, str): + batch_size = 1 + else: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if isinstance(image, list): + image_batch_size = len(image) + else: + image_batch_size = image.shape[0] if image.ndim == 4 else 1 + if batch_size != image_batch_size: + raise ValueError( + f"`prompt` has batch size {batch_size} and `image` has batch size {image_batch_size}." + " Please make sure that passed `prompt` matches the batch size of `image`." + ) + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height, width) + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]] = None, + image: PipelineImageInput = None, + num_inference_steps: int = 75, + guidance_scale: float = 9.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide image upscaling. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image` or tensor representing an image batch to be upscaled. If it's a tensor, it can be either a + latent output from a Stable Diffusion model or an image tensor in the range `[-1, 1]`. It is considered + a `latent` if `image.shape[1]` is `4`; otherwise, it is considered to be an image representation and + encoded using this pipeline's `vae` encoder. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + + Examples: + ```py + >>> from diffusers import StableDiffusionLatentUpscalePipeline, StableDiffusionPipeline + >>> import torch + + + >>> pipeline = StableDiffusionPipeline.from_pretrained( + ... "CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16 + ... ) + >>> pipeline.to("cuda") + + >>> model_id = "stabilityai/sd-x2-latent-upscaler" + >>> upscaler = StableDiffusionLatentUpscalePipeline.from_pretrained(model_id, torch_dtype=torch.float16) + >>> upscaler.to("cuda") + + >>> prompt = "a photo of an astronaut high resolution, unreal engine, ultra realistic" + >>> generator = torch.manual_seed(33) + + >>> low_res_latents = pipeline(prompt, generator=generator, output_type="latent").images + + >>> with torch.no_grad(): + ... image = pipeline.decode_latents(low_res_latents) + >>> image = pipeline.numpy_to_pil(image)[0] + + >>> image.save("../images/a1.png") + + >>> upscaled_image = upscaler( + ... prompt=prompt, + ... image=low_res_latents, + ... num_inference_steps=20, + ... guidance_scale=0, + ... generator=generator, + ... ).images[0] + + >>> upscaled_image.save("../images/a2.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images. + """ + + # 1. Check inputs + self.check_inputs( + prompt, + image, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) + + # 2. Define call parameters + if prompt is not None: + batch_size = 1 if isinstance(prompt, str) else len(prompt) + else: + batch_size = prompt_embeds.shape[0] + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + if guidance_scale == 0: + prompt = [""] * batch_size + + # 3. Encode input prompt + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt, + device, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) + + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds]) + + # 4. Preprocess image + image = self.image_processor.preprocess(image) + image = image.to(dtype=prompt_embeds.dtype, device=device) + if image.shape[1] == 3: + # encode image if not in latent-space yet + image = retrieve_latents(self.vae.encode(image), generator=generator) * self.vae.config.scaling_factor + + # 5. set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + batch_multiplier = 2 if do_classifier_free_guidance else 1 + image = image[None, :] if image.ndim == 3 else image + image = torch.cat([image] * batch_multiplier) + + # 5. Add noise to image (set to be 0): + # (see below notes from the author): + # "the This step theoretically can make the model work better on out-of-distribution inputs, but mostly just seems to make it match the input less, so it's turned off by default." + noise_level = torch.tensor([0.0], dtype=torch.float32, device=device) + noise_level = torch.cat([noise_level] * image.shape[0]) + inv_noise_level = (noise_level**2 + 1) ** (-0.5) + + image_cond = F.interpolate(image, scale_factor=2, mode="nearest") * inv_noise_level[:, None, None, None] + image_cond = image_cond.to(prompt_embeds.dtype) + + noise_level_embed = torch.cat( + [ + torch.ones(pooled_prompt_embeds.shape[0], 64, dtype=pooled_prompt_embeds.dtype, device=device), + torch.zeros(pooled_prompt_embeds.shape[0], 64, dtype=pooled_prompt_embeds.dtype, device=device), + ], + dim=1, + ) + + timestep_condition = torch.cat([noise_level_embed, pooled_prompt_embeds], dim=1) + + # 6. Prepare latent variables + height, width = image.shape[2:] + num_channels_latents = self.vae.config.latent_channels + latents = self.prepare_latents( + batch_size, + num_channels_latents, + height * 2, # 2x upscale + width * 2, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 7. Check that sizes of image and latents match + num_channels_image = image.shape[1] + if num_channels_latents + num_channels_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_image`: {num_channels_image} " + f" = {num_channels_latents + num_channels_image}. Please verify the config of" + " `pipeline.unet` or your `image` input." + ) + + # 9. Denoising loop + num_warmup_steps = 0 + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + sigma = self.scheduler.sigmas[i] + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + scaled_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + scaled_model_input = torch.cat([scaled_model_input, image_cond], dim=1) + # preconditioning parameter based on Karras et al. (2022) (table 1) + timestep = torch.log(sigma) * 0.25 + + noise_pred = self.unet( + scaled_model_input, + timestep, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_condition, + ).sample + + # in original repo, the output contains a variance channel that's not used + noise_pred = noise_pred[:, :-1] + + # apply preconditioning, based on table 1 in Karras et al. (2022) + inv_sigma = 1 / (sigma**2 + 1) + noise_pred = inv_sigma * latent_model_input + self.scheduler.scale_model_input(sigma, t) * noise_pred + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = latents + + image = self.image_processor.postprocess(image, output_type=output_type) + + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py new file mode 100644 index 0000000000000000000000000000000000000000..02f7526b71028b2a1eb7aa69e5cf62c380850c94 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py @@ -0,0 +1,826 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +import warnings +from typing import Any, Callable, Dict, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.attention_processor import ( + AttnProcessor2_0, + XFormersAttnProcessor, +) +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import DDPMScheduler, KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from . import StableDiffusionPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def preprocess(image): + warnings.warn( + "The preprocess method is deprecated and will be removed in a future version. Please" + " use VaeImageProcessor.preprocess instead", + FutureWarning, + ) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 64 for x in (w, h)) # resize to integer multiple of 64 + + image = [np.array(i.resize((w, h)))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +class StableDiffusionUpscalePipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + FromSingleFileMixin, +): + r""" + Pipeline for text-guided image super-resolution using Stable Diffusion 2. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + low_res_scheduler ([`SchedulerMixin`]): + A scheduler used to add initial noise to the low resolution conditioning image. It must be an instance of + [`DDPMScheduler`]. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["watermarker", "safety_checker", "feature_extractor"] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + low_res_scheduler: DDPMScheduler, + scheduler: KarrasDiffusionSchedulers, + safety_checker: Optional[Any] = None, + feature_extractor: Optional[CLIPImageProcessor] = None, + watermarker: Optional[Any] = None, + max_noise_level: int = 350, + ): + super().__init__() + + if hasattr( + vae, "config" + ): # check if vae has a config attribute `scaling_factor` and if it is set to 0.08333, else set it to 0.08333 and deprecate + is_vae_scaling_factor_set_to_0_08333 = ( + hasattr(vae.config, "scaling_factor") and vae.config.scaling_factor == 0.08333 + ) + if not is_vae_scaling_factor_set_to_0_08333: + deprecation_message = ( + "The configuration file of the vae does not contain `scaling_factor` or it is set to" + f" {vae.config.scaling_factor}, which seems highly unlikely. If your checkpoint is a fine-tuned" + " version of `stabilityai/stable-diffusion-x4-upscaler` you should change 'scaling_factor' to" + " 0.08333 Please make sure to update the config accordingly, as not doing so might lead to" + " incorrect results in future versions. If you have downloaded this checkpoint from the Hugging" + " Face Hub, it would be very nice if you could open a Pull Request for the `vae/config.json` file" + ) + deprecate("wrong scaling_factor", "1.0.0", deprecation_message, standard_warn=False) + vae.register_to_config(scaling_factor=0.08333) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + low_res_scheduler=low_res_scheduler, + scheduler=scheduler, + safety_checker=safety_checker, + watermarker=watermarker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, resample="bicubic") + self.register_to_config(max_noise_level=max_noise_level) + + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is not None: + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, nsfw_detected, watermark_detected = self.safety_checker( + images=image, + clip_input=safety_checker_input.pixel_values.to(dtype=dtype), + ) + else: + nsfw_detected = None + watermark_detected = None + + if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None: + self.unet_offload_hook.offload() + + return image, nsfw_detected, watermark_detected + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def check_inputs( + self, + prompt, + image, + noise_level, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ( + not isinstance(image, torch.Tensor) + and not isinstance(image, PIL.Image.Image) + and not isinstance(image, np.ndarray) + and not isinstance(image, list) + ): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `np.ndarray`, `PIL.Image.Image` or `list` but is {type(image)}" + ) + + # verify batch size of prompt and image are same if image is a list or tensor or numpy array + if isinstance(image, (list, np.ndarray, torch.Tensor)): + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if isinstance(image, list): + image_batch_size = len(image) + else: + image_batch_size = image.shape[0] + if batch_size != image_batch_size: + raise ValueError( + f"`prompt` has batch size {batch_size} and `image` has batch size {image_batch_size}." + " Please make sure that passed `prompt` matches the batch size of `image`." + ) + + # check noise level + if noise_level > self.config.max_noise_level: + raise ValueError(f"`noise_level` has to be <= {self.config.max_noise_level} but is {noise_level}") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height, width) + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]] = None, + image: PipelineImageInput = None, + num_inference_steps: int = 75, + guidance_scale: float = 9.0, + noise_level: int = 20, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + clip_skip: int = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image` or tensor representing an image batch to be upscaled. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + ```py + >>> import requests + >>> from PIL import Image + >>> from io import BytesIO + >>> from diffusers import StableDiffusionUpscalePipeline + >>> import torch + + >>> # load model and scheduler + >>> model_id = "stabilityai/stable-diffusion-x4-upscaler" + >>> pipeline = StableDiffusionUpscalePipeline.from_pretrained( + ... model_id, variant="fp16", torch_dtype=torch.float16 + ... ) + >>> pipeline = pipeline.to("cuda") + + >>> # let's download an image + >>> url = "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-upscale/low_res_cat.png" + >>> response = requests.get(url) + >>> low_res_img = Image.open(BytesIO(response.content)).convert("RGB") + >>> low_res_img = low_res_img.resize((128, 128)) + >>> prompt = "a white cat" + + >>> upscaled_image = pipeline(prompt=prompt, image=low_res_img).images[0] + >>> upscaled_image.save("upsampled_cat.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + # 1. Check inputs + self.check_inputs( + prompt, + image, + noise_level, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ) + + if image is None: + raise ValueError("`image` input cannot be undefined.") + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Preprocess image + image = self.image_processor.preprocess(image) + image = image.to(dtype=prompt_embeds.dtype, device=device) + + # 5. set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Add noise to image + noise_level = torch.tensor([noise_level], dtype=torch.long, device=device) + noise = randn_tensor(image.shape, generator=generator, device=device, dtype=prompt_embeds.dtype) + image = self.low_res_scheduler.add_noise(image, noise, noise_level) + + batch_multiplier = 2 if do_classifier_free_guidance else 1 + image = torch.cat([image] * batch_multiplier * num_images_per_prompt) + noise_level = torch.cat([noise_level] * image.shape[0]) + + # 6. Prepare latent variables + height, width = image.shape[2:] + num_channels_latents = self.vae.config.latent_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 7. Check that sizes of image and latents match + num_channels_image = image.shape[1] + if num_channels_latents + num_channels_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_image`: {num_channels_image} " + f" = {num_channels_latents + num_channels_image}. Please verify the config of" + " `pipeline.unet` or your `image` input." + ) + + # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + + # concat latents, mask, masked_image_latents in the channel dimension + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + latent_model_input = torch.cat([latent_model_input, image], dim=1) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + class_labels=noise_level, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + + # Ensure latents are always the same type as the VAE + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + + image, has_nsfw_concept, _ = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # 11. Apply watermark + if output_type == "pil" and self.watermarker is not None: + image = self.watermarker.apply_watermark(image) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip.py new file mode 100644 index 0000000000000000000000000000000000000000..a134244e3ee4ecfbc881fe0c4b2f917a8fb153a4 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip.py @@ -0,0 +1,952 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import torch +from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer +from transformers.models.clip.modeling_clip import CLIPTextModelOutput + +from ...image_processor import VaeImageProcessor +from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, PriorTransformer, UNet2DConditionModel +from ...models.embeddings import get_timestep_embedding +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput, StableDiffusionMixin +from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableUnCLIPPipeline + + >>> pipe = StableUnCLIPPipeline.from_pretrained( + ... "fusing/stable-unclip-2-1-l", torch_dtype=torch.float16 + ... ) # TODO update model path + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> images = pipe(prompt).images + >>> images[0].save("astronaut_horse.png") + ``` +""" + + +class StableUnCLIPPipeline( + DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin +): + """ + Pipeline for text-to-image generation using stable unCLIP. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + prior_tokenizer ([`CLIPTokenizer`]): + A [`CLIPTokenizer`]. + prior_text_encoder ([`CLIPTextModelWithProjection`]): + Frozen [`CLIPTextModelWithProjection`] text-encoder. + prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + prior_scheduler ([`KarrasDiffusionSchedulers`]): + Scheduler used in the prior denoising process. + image_normalizer ([`StableUnCLIPImageNormalizer`]): + Used to normalize the predicted image embeddings before the noise is applied and un-normalize the image + embeddings after the noise has been applied. + image_noising_scheduler ([`KarrasDiffusionSchedulers`]): + Noise schedule for adding noise to the predicted image embeddings. The amount of noise to add is determined + by the `noise_level`. + tokenizer ([`CLIPTokenizer`]): + A [`CLIPTokenizer`]. + text_encoder ([`CLIPTextModel`]): + Frozen [`CLIPTextModel`] text-encoder. + unet ([`UNet2DConditionModel`]): + A [`UNet2DConditionModel`] to denoise the encoded image latents. + scheduler ([`KarrasDiffusionSchedulers`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + """ + + _exclude_from_cpu_offload = ["prior", "image_normalizer"] + model_cpu_offload_seq = "text_encoder->prior_text_encoder->unet->vae" + + # prior components + prior_tokenizer: CLIPTokenizer + prior_text_encoder: CLIPTextModelWithProjection + prior: PriorTransformer + prior_scheduler: KarrasDiffusionSchedulers + + # image noising components + image_normalizer: StableUnCLIPImageNormalizer + image_noising_scheduler: KarrasDiffusionSchedulers + + # regular denoising components + tokenizer: CLIPTokenizer + text_encoder: CLIPTextModel + unet: UNet2DConditionModel + scheduler: KarrasDiffusionSchedulers + + vae: AutoencoderKL + + def __init__( + self, + # prior components + prior_tokenizer: CLIPTokenizer, + prior_text_encoder: CLIPTextModelWithProjection, + prior: PriorTransformer, + prior_scheduler: KarrasDiffusionSchedulers, + # image noising components + image_normalizer: StableUnCLIPImageNormalizer, + image_noising_scheduler: KarrasDiffusionSchedulers, + # regular denoising components + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModel, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + # vae + vae: AutoencoderKL, + ): + super().__init__() + + self.register_modules( + prior_tokenizer=prior_tokenizer, + prior_text_encoder=prior_text_encoder, + prior=prior, + prior_scheduler=prior_scheduler, + image_normalizer=image_normalizer, + image_noising_scheduler=image_noising_scheduler, + tokenizer=tokenizer, + text_encoder=text_encoder, + unet=unet, + scheduler=scheduler, + vae=vae, + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline._encode_prompt with _encode_prompt->_encode_prior_prompt, tokenizer->prior_tokenizer, text_encoder->prior_text_encoder + def _encode_prior_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + text_model_output: Optional[Union[CLIPTextModelOutput, Tuple]] = None, + text_attention_mask: Optional[torch.Tensor] = None, + ): + if text_model_output is None: + batch_size = len(prompt) if isinstance(prompt, list) else 1 + # get prompt text embeddings + text_inputs = self.prior_tokenizer( + prompt, + padding="max_length", + max_length=self.prior_tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + text_mask = text_inputs.attention_mask.bool().to(device) + + untruncated_ids = self.prior_tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.prior_tokenizer.batch_decode( + untruncated_ids[:, self.prior_tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.prior_tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.prior_tokenizer.model_max_length] + + prior_text_encoder_output = self.prior_text_encoder(text_input_ids.to(device)) + + prompt_embeds = prior_text_encoder_output.text_embeds + text_enc_hid_states = prior_text_encoder_output.last_hidden_state + + else: + batch_size = text_model_output[0].shape[0] + prompt_embeds, text_enc_hid_states = text_model_output[0], text_model_output[1] + text_mask = text_attention_mask + + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_enc_hid_states = text_enc_hid_states.repeat_interleave(num_images_per_prompt, dim=0) + text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + uncond_tokens = [""] * batch_size + + uncond_input = self.prior_tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.prior_tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + uncond_text_mask = uncond_input.attention_mask.bool().to(device) + negative_prompt_embeds_prior_text_encoder_output = self.prior_text_encoder( + uncond_input.input_ids.to(device) + ) + + negative_prompt_embeds = negative_prompt_embeds_prior_text_encoder_output.text_embeds + uncond_text_enc_hid_states = negative_prompt_embeds_prior_text_encoder_output.last_hidden_state + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len) + + seq_len = uncond_text_enc_hid_states.shape[1] + uncond_text_enc_hid_states = uncond_text_enc_hid_states.repeat(1, num_images_per_prompt, 1) + uncond_text_enc_hid_states = uncond_text_enc_hid_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + text_enc_hid_states = torch.cat([uncond_text_enc_hid_states, text_enc_hid_states]) + + text_mask = torch.cat([uncond_text_mask, text_mask]) + + return prompt_embeds, text_enc_hid_states, text_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs with prepare_extra_step_kwargs->prepare_prior_extra_step_kwargs, scheduler->prior_scheduler + def prepare_prior_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the prior_scheduler step, since not all prior_schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other prior_schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.prior_scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the prior_scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.prior_scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + noise_level, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Please make sure to define only one of the two." + ) + + if prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + + if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + "Provide either `negative_prompt` or `negative_prompt_embeds`. Cannot leave both `negative_prompt` and `negative_prompt_embeds` undefined." + ) + + if prompt is not None and negative_prompt is not None: + if type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if noise_level < 0 or noise_level >= self.image_noising_scheduler.config.num_train_timesteps: + raise ValueError( + f"`noise_level` must be between 0 and {self.image_noising_scheduler.config.num_train_timesteps - 1}, inclusive." + ) + + # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def noise_image_embeddings( + self, + image_embeds: torch.Tensor, + noise_level: int, + noise: Optional[torch.Tensor] = None, + generator: Optional[torch.Generator] = None, + ): + """ + Add noise to the image embeddings. The amount of noise is controlled by a `noise_level` input. A higher + `noise_level` increases the variance in the final un-noised images. + + The noise is applied in two ways: + 1. A noise schedule is applied directly to the embeddings. + 2. A vector of sinusoidal time embeddings are appended to the output. + + In both cases, the amount of noise is controlled by the same `noise_level`. + + The embeddings are normalized before the noise is applied and un-normalized after the noise is applied. + """ + if noise is None: + noise = randn_tensor( + image_embeds.shape, generator=generator, device=image_embeds.device, dtype=image_embeds.dtype + ) + + noise_level = torch.tensor([noise_level] * image_embeds.shape[0], device=image_embeds.device) + + self.image_normalizer.to(image_embeds.device) + image_embeds = self.image_normalizer.scale(image_embeds) + + image_embeds = self.image_noising_scheduler.add_noise(image_embeds, timesteps=noise_level, noise=noise) + + image_embeds = self.image_normalizer.unscale(image_embeds) + + noise_level = get_timestep_embedding( + timesteps=noise_level, embedding_dim=image_embeds.shape[-1], flip_sin_to_cos=True, downscale_freq_shift=0 + ) + + # `get_timestep_embeddings` does not contain any weights and will always return f32 tensors, + # but we might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + noise_level = noise_level.to(image_embeds.dtype) + + image_embeds = torch.cat((image_embeds, noise_level), 1) + + return image_embeds + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + # regular denoising process args + prompt: Optional[Union[str, List[str]]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 20, + guidance_scale: float = 10.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + noise_level: int = 0, + # prior args + prior_num_inference_steps: int = 25, + prior_guidance_scale: float = 4.0, + prior_latents: Optional[torch.Tensor] = None, + clip_skip: Optional[int] = None, + ): + """ + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 20): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 10.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + noise_level (`int`, *optional*, defaults to `0`): + The amount of noise to add to the image embeddings. A higher `noise_level` increases the variance in + the final un-noised images. See [`StableUnCLIPPipeline.noise_image_embeddings`] for more details. + prior_num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps in the prior denoising process. More denoising steps usually lead to a + higher quality image at the expense of slower inference. + prior_guidance_scale (`float`, *optional*, defaults to 4.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + prior_latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + embedding generation in the prior denoising process. Can be used to tweak the same generation with + different prompts. If not provided, a latents tensor is generated by sampling using the supplied random + `generator`. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + [`~ pipeline_utils.ImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When returning + a tuple, the first element is a list with the generated images. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + height=height, + width=width, + callback_steps=callback_steps, + noise_level=noise_level, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + batch_size = batch_size * num_images_per_prompt + + device = self._execution_device + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + prior_do_classifier_free_guidance = prior_guidance_scale > 1.0 + + # 3. Encode input prompt + prior_prompt_embeds, prior_text_encoder_hidden_states, prior_text_mask = self._encode_prior_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=prior_do_classifier_free_guidance, + ) + + # 4. Prepare prior timesteps + self.prior_scheduler.set_timesteps(prior_num_inference_steps, device=device) + prior_timesteps_tensor = self.prior_scheduler.timesteps + + # 5. Prepare prior latent variables + embedding_dim = self.prior.config.embedding_dim + prior_latents = self.prepare_latents( + (batch_size, embedding_dim), + prior_prompt_embeds.dtype, + device, + generator, + prior_latents, + self.prior_scheduler, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + prior_extra_step_kwargs = self.prepare_prior_extra_step_kwargs(generator, eta) + + # 7. Prior denoising loop + for i, t in enumerate(self.progress_bar(prior_timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([prior_latents] * 2) if prior_do_classifier_free_guidance else prior_latents + latent_model_input = self.prior_scheduler.scale_model_input(latent_model_input, t) + + predicted_image_embedding = self.prior( + latent_model_input, + timestep=t, + proj_embedding=prior_prompt_embeds, + encoder_hidden_states=prior_text_encoder_hidden_states, + attention_mask=prior_text_mask, + ).predicted_image_embedding + + if prior_do_classifier_free_guidance: + predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2) + predicted_image_embedding = predicted_image_embedding_uncond + prior_guidance_scale * ( + predicted_image_embedding_text - predicted_image_embedding_uncond + ) + + prior_latents = self.prior_scheduler.step( + predicted_image_embedding, + timestep=t, + sample=prior_latents, + **prior_extra_step_kwargs, + return_dict=False, + )[0] + + if callback is not None and i % callback_steps == 0: + callback(i, t, prior_latents) + + prior_latents = self.prior.post_process_latents(prior_latents) + + image_embeds = prior_latents + + # done prior + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 8. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 9. Prepare image embeddings + image_embeds = self.noise_image_embeddings( + image_embeds=image_embeds, + noise_level=noise_level, + generator=generator, + ) + + if do_classifier_free_guidance: + negative_prompt_embeds = torch.zeros_like(image_embeds) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + image_embeds = torch.cat([negative_prompt_embeds, image_embeds]) + + # 10. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 11. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + latents = self.prepare_latents( + shape=shape, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=latents, + scheduler=self.scheduler, + ) + + # 12. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 13. Denoising loop + for i, t in enumerate(self.progress_bar(timesteps)): + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + class_labels=image_embeds, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = latents + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..abb4cc3a05d51ed0164120abb688267ec5e2af58 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py @@ -0,0 +1,858 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...image_processor import VaeImageProcessor +from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.embeddings import get_timestep_embedding +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput, StableDiffusionMixin +from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import requests + >>> import torch + >>> from PIL import Image + >>> from io import BytesIO + + >>> from diffusers import StableUnCLIPImg2ImgPipeline + + >>> pipe = StableUnCLIPImg2ImgPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-2-1-unclip-small", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + + >>> response = requests.get(url) + >>> init_image = Image.open(BytesIO(response.content)).convert("RGB") + >>> init_image = init_image.resize((768, 512)) + + >>> prompt = "A fantasy landscape, trending on artstation" + + >>> images = pipe(init_image, prompt).images + >>> images[0].save("fantasy_landscape.png") + ``` +""" + + +class StableUnCLIPImg2ImgPipeline( + DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin +): + """ + Pipeline for text-guided image-to-image generation using stable unCLIP. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + feature_extractor ([`CLIPImageProcessor`]): + Feature extractor for image pre-processing before being encoded. + image_encoder ([`CLIPVisionModelWithProjection`]): + CLIP vision model for encoding images. + image_normalizer ([`StableUnCLIPImageNormalizer`]): + Used to normalize the predicted image embeddings before the noise is applied and un-normalize the image + embeddings after the noise has been applied. + image_noising_scheduler ([`KarrasDiffusionSchedulers`]): + Noise schedule for adding noise to the predicted image embeddings. The amount of noise to add is determined + by the `noise_level`. + tokenizer (`~transformers.CLIPTokenizer`): + A [`~transformers.CLIPTokenizer`)]. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen [`~transformers.CLIPTextModel`] text-encoder. + unet ([`UNet2DConditionModel`]): + A [`UNet2DConditionModel`] to denoise the encoded image latents. + scheduler ([`KarrasDiffusionSchedulers`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _exclude_from_cpu_offload = ["image_normalizer"] + + # image encoding components + feature_extractor: CLIPImageProcessor + image_encoder: CLIPVisionModelWithProjection + + # image noising components + image_normalizer: StableUnCLIPImageNormalizer + image_noising_scheduler: KarrasDiffusionSchedulers + + # regular denoising components + tokenizer: CLIPTokenizer + text_encoder: CLIPTextModel + unet: UNet2DConditionModel + scheduler: KarrasDiffusionSchedulers + + vae: AutoencoderKL + + def __init__( + self, + # image encoding components + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection, + # image noising components + image_normalizer: StableUnCLIPImageNormalizer, + image_noising_scheduler: KarrasDiffusionSchedulers, + # regular denoising components + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModel, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + # vae + vae: AutoencoderKL, + ): + super().__init__() + + self.register_modules( + feature_extractor=feature_extractor, + image_encoder=image_encoder, + image_normalizer=image_normalizer, + image_noising_scheduler=image_noising_scheduler, + tokenizer=tokenizer, + text_encoder=text_encoder, + unet=unet, + scheduler=scheduler, + vae=vae, + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + def _encode_image( + self, + image, + device, + batch_size, + num_images_per_prompt, + do_classifier_free_guidance, + noise_level, + generator, + image_embeds, + ): + dtype = next(self.image_encoder.parameters()).dtype + + if isinstance(image, PIL.Image.Image): + # the image embedding should repeated so it matches the total batch size of the prompt + repeat_by = batch_size + else: + # assume the image input is already properly batched and just needs to be repeated so + # it matches the num_images_per_prompt. + # + # NOTE(will) this is probably missing a few number of side cases. I.e. batched/non-batched + # `image_embeds`. If those happen to be common use cases, let's think harder about + # what the expected dimensions of inputs should be and how we handle the encoding. + repeat_by = num_images_per_prompt + + if image_embeds is None: + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(images=image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeds = self.image_encoder(image).image_embeds + + image_embeds = self.noise_image_embeddings( + image_embeds=image_embeds, + noise_level=noise_level, + generator=generator, + ) + + # duplicate image embeddings for each generation per prompt, using mps friendly method + image_embeds = image_embeds.unsqueeze(1) + bs_embed, seq_len, _ = image_embeds.shape + image_embeds = image_embeds.repeat(1, repeat_by, 1) + image_embeds = image_embeds.view(bs_embed * repeat_by, seq_len, -1) + image_embeds = image_embeds.squeeze(1) + + if do_classifier_free_guidance: + negative_prompt_embeds = torch.zeros_like(image_embeds) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + image_embeds = torch.cat([negative_prompt_embeds, image_embeds]) + + return image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + image, + height, + width, + callback_steps, + noise_level, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + image_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Please make sure to define only one of the two." + ) + + if prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + + if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + "Provide either `negative_prompt` or `negative_prompt_embeds`. Cannot leave both `negative_prompt` and `negative_prompt_embeds` undefined." + ) + + if prompt is not None and negative_prompt is not None: + if type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if noise_level < 0 or noise_level >= self.image_noising_scheduler.config.num_train_timesteps: + raise ValueError( + f"`noise_level` must be between 0 and {self.image_noising_scheduler.config.num_train_timesteps - 1}, inclusive." + ) + + if image is not None and image_embeds is not None: + raise ValueError( + "Provide either `image` or `image_embeds`. Please make sure to define only one of the two." + ) + + if image is None and image_embeds is None: + raise ValueError( + "Provide either `image` or `image_embeds`. Cannot leave both `image` and `image_embeds` undefined." + ) + + if image is not None: + if ( + not isinstance(image, torch.Tensor) + and not isinstance(image, PIL.Image.Image) + and not isinstance(image, list) + ): + raise ValueError( + "`image` has to be of type `torch.Tensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is" + f" {type(image)}" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_unclip.StableUnCLIPPipeline.noise_image_embeddings + def noise_image_embeddings( + self, + image_embeds: torch.Tensor, + noise_level: int, + noise: Optional[torch.Tensor] = None, + generator: Optional[torch.Generator] = None, + ): + """ + Add noise to the image embeddings. The amount of noise is controlled by a `noise_level` input. A higher + `noise_level` increases the variance in the final un-noised images. + + The noise is applied in two ways: + 1. A noise schedule is applied directly to the embeddings. + 2. A vector of sinusoidal time embeddings are appended to the output. + + In both cases, the amount of noise is controlled by the same `noise_level`. + + The embeddings are normalized before the noise is applied and un-normalized after the noise is applied. + """ + if noise is None: + noise = randn_tensor( + image_embeds.shape, generator=generator, device=image_embeds.device, dtype=image_embeds.dtype + ) + + noise_level = torch.tensor([noise_level] * image_embeds.shape[0], device=image_embeds.device) + + self.image_normalizer.to(image_embeds.device) + image_embeds = self.image_normalizer.scale(image_embeds) + + image_embeds = self.image_noising_scheduler.add_noise(image_embeds, timesteps=noise_level, noise=noise) + + image_embeds = self.image_normalizer.unscale(image_embeds) + + noise_level = get_timestep_embedding( + timesteps=noise_level, embedding_dim=image_embeds.shape[-1], flip_sin_to_cos=True, downscale_freq_shift=0 + ) + + # `get_timestep_embeddings` does not contain any weights and will always return f32 tensors, + # but we might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + noise_level = noise_level.to(image_embeds.dtype) + + image_embeds = torch.cat((image_embeds, noise_level), 1) + + return image_embeds + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: Union[torch.Tensor, PIL.Image.Image] = None, + prompt: Union[str, List[str]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 20, + guidance_scale: float = 10, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + noise_level: int = 0, + image_embeds: Optional[torch.Tensor] = None, + clip_skip: Optional[int] = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, either `prompt_embeds` will be + used or prompt is initialized to `""`. + image (`torch.Tensor` or `PIL.Image.Image`): + `Image` or tensor representing an image batch. The image is encoded to its CLIP embedding which the + `unet` is conditioned on. The image is _not_ encoded by the `vae` and then used as the latents in the + denoising process like it is in the standard Stable Diffusion text-guided image variation process. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 20): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 10.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + noise_level (`int`, *optional*, defaults to `0`): + The amount of noise to add to the image embeddings. A higher `noise_level` increases the variance in + the final un-noised images. See [`StableUnCLIPPipeline.noise_image_embeddings`] for more details. + image_embeds (`torch.Tensor`, *optional*): + Pre-generated CLIP embeddings to condition the `unet` on. These latents are not used in the denoising + process. If you want to provide pre-generated latents, pass them to `__call__` as `latents`. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + [`~ pipeline_utils.ImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When returning + a tuple, the first element is a list with the generated images. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + if prompt is None and prompt_embeds is None: + prompt = len(image) * [""] if isinstance(image, list) else "" + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + image=image, + height=height, + width=width, + callback_steps=callback_steps, + noise_level=noise_level, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + image_embeds=image_embeds, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + batch_size = batch_size * num_images_per_prompt + + device = self._execution_device + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Encoder input image + noise_level = torch.tensor([noise_level], device=device) + image_embeds = self._encode_image( + image=image, + device=device, + batch_size=batch_size, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + noise_level=noise_level, + generator=generator, + image_embeds=image_embeds, + ) + + # 5. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 6. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + if latents is None: + latents = self.prepare_latents( + batch_size=batch_size, + num_channels_latents=num_channels_latents, + height=height, + width=width, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=latents, + ) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 8. Denoising loop + for i, t in enumerate(self.progress_bar(timesteps)): + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + class_labels=image_embeds, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 9. Post-processing + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = latents + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/safety_checker.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/safety_checker.py new file mode 100644 index 0000000000000000000000000000000000000000..16aff102599ca7fe3399dff9de54aafb9e5333df --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/safety_checker.py @@ -0,0 +1,126 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 numpy as np +import torch +import torch.nn as nn +from transformers import CLIPConfig, CLIPVisionModel, PreTrainedModel + +from ...utils import logging + + +logger = logging.get_logger(__name__) + + +def cosine_distance(image_embeds, text_embeds): + normalized_image_embeds = nn.functional.normalize(image_embeds) + normalized_text_embeds = nn.functional.normalize(text_embeds) + return torch.mm(normalized_image_embeds, normalized_text_embeds.t()) + + +class StableDiffusionSafetyChecker(PreTrainedModel): + config_class = CLIPConfig + main_input_name = "clip_input" + + _no_split_modules = ["CLIPEncoderLayer"] + + def __init__(self, config: CLIPConfig): + super().__init__(config) + + self.vision_model = CLIPVisionModel(config.vision_config) + self.visual_projection = nn.Linear(config.vision_config.hidden_size, config.projection_dim, bias=False) + + self.concept_embeds = nn.Parameter(torch.ones(17, config.projection_dim), requires_grad=False) + self.special_care_embeds = nn.Parameter(torch.ones(3, config.projection_dim), requires_grad=False) + + self.concept_embeds_weights = nn.Parameter(torch.ones(17), requires_grad=False) + self.special_care_embeds_weights = nn.Parameter(torch.ones(3), requires_grad=False) + + @torch.no_grad() + def forward(self, clip_input, images): + pooled_output = self.vision_model(clip_input)[1] # pooled_output + image_embeds = self.visual_projection(pooled_output) + + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds).cpu().float().numpy() + cos_dist = cosine_distance(image_embeds, self.concept_embeds).cpu().float().numpy() + + result = [] + batch_size = image_embeds.shape[0] + for i in range(batch_size): + result_img = {"special_scores": {}, "special_care": [], "concept_scores": {}, "bad_concepts": []} + + # increase this value to create a stronger `nfsw` filter + # at the cost of increasing the possibility of filtering benign images + adjustment = 0.0 + + for concept_idx in range(len(special_cos_dist[0])): + concept_cos = special_cos_dist[i][concept_idx] + concept_threshold = self.special_care_embeds_weights[concept_idx].item() + result_img["special_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3) + if result_img["special_scores"][concept_idx] > 0: + result_img["special_care"].append({concept_idx, result_img["special_scores"][concept_idx]}) + adjustment = 0.01 + + for concept_idx in range(len(cos_dist[0])): + concept_cos = cos_dist[i][concept_idx] + concept_threshold = self.concept_embeds_weights[concept_idx].item() + result_img["concept_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3) + if result_img["concept_scores"][concept_idx] > 0: + result_img["bad_concepts"].append(concept_idx) + + result.append(result_img) + + has_nsfw_concepts = [len(res["bad_concepts"]) > 0 for res in result] + + for idx, has_nsfw_concept in enumerate(has_nsfw_concepts): + if has_nsfw_concept: + if torch.is_tensor(images) or torch.is_tensor(images[0]): + images[idx] = torch.zeros_like(images[idx]) # black image + else: + images[idx] = np.zeros(images[idx].shape) # black image + + if any(has_nsfw_concepts): + logger.warning( + "Potential NSFW content was detected in one or more images. A black image will be returned instead." + " Try again with a different prompt and/or seed." + ) + + return images, has_nsfw_concepts + + @torch.no_grad() + def forward_onnx(self, clip_input: torch.Tensor, images: torch.Tensor): + pooled_output = self.vision_model(clip_input)[1] # pooled_output + image_embeds = self.visual_projection(pooled_output) + + special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds) + cos_dist = cosine_distance(image_embeds, self.concept_embeds) + + # increase this value to create a stronger `nsfw` filter + # at the cost of increasing the possibility of filtering benign images + adjustment = 0.0 + + special_scores = special_cos_dist - self.special_care_embeds_weights + adjustment + # special_scores = special_scores.round(decimals=3) + special_care = torch.any(special_scores > 0, dim=1) + special_adjustment = special_care * 0.01 + special_adjustment = special_adjustment.unsqueeze(1).expand(-1, cos_dist.shape[1]) + + concept_scores = (cos_dist - self.concept_embeds_weights) + special_adjustment + # concept_scores = concept_scores.round(decimals=3) + has_nsfw_concepts = torch.any(concept_scores > 0, dim=1) + + images[has_nsfw_concepts] = 0.0 # black image + + return images, has_nsfw_concepts diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/safety_checker_flax.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/safety_checker_flax.py new file mode 100644 index 0000000000000000000000000000000000000000..55d5023b686994a731041a83bd716862812e3fac --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/safety_checker_flax.py @@ -0,0 +1,112 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 + +import jax +import jax.numpy as jnp +from flax import linen as nn +from flax.core.frozen_dict import FrozenDict +from transformers import CLIPConfig, FlaxPreTrainedModel +from transformers.models.clip.modeling_flax_clip import FlaxCLIPVisionModule + + +def jax_cosine_distance(emb_1, emb_2, eps=1e-12): + norm_emb_1 = jnp.divide(emb_1.T, jnp.clip(jnp.linalg.norm(emb_1, axis=1), a_min=eps)).T + norm_emb_2 = jnp.divide(emb_2.T, jnp.clip(jnp.linalg.norm(emb_2, axis=1), a_min=eps)).T + return jnp.matmul(norm_emb_1, norm_emb_2.T) + + +class FlaxStableDiffusionSafetyCheckerModule(nn.Module): + config: CLIPConfig + dtype: jnp.dtype = jnp.float32 + + def setup(self): + self.vision_model = FlaxCLIPVisionModule(self.config.vision_config) + self.visual_projection = nn.Dense(self.config.projection_dim, use_bias=False, dtype=self.dtype) + + self.concept_embeds = self.param("concept_embeds", jax.nn.initializers.ones, (17, self.config.projection_dim)) + self.special_care_embeds = self.param( + "special_care_embeds", jax.nn.initializers.ones, (3, self.config.projection_dim) + ) + + self.concept_embeds_weights = self.param("concept_embeds_weights", jax.nn.initializers.ones, (17,)) + self.special_care_embeds_weights = self.param("special_care_embeds_weights", jax.nn.initializers.ones, (3,)) + + def __call__(self, clip_input): + pooled_output = self.vision_model(clip_input)[1] + image_embeds = self.visual_projection(pooled_output) + + special_cos_dist = jax_cosine_distance(image_embeds, self.special_care_embeds) + cos_dist = jax_cosine_distance(image_embeds, self.concept_embeds) + + # increase this value to create a stronger `nfsw` filter + # at the cost of increasing the possibility of filtering benign image inputs + adjustment = 0.0 + + special_scores = special_cos_dist - self.special_care_embeds_weights[None, :] + adjustment + special_scores = jnp.round(special_scores, 3) + is_special_care = jnp.any(special_scores > 0, axis=1, keepdims=True) + # Use a lower threshold if an image has any special care concept + special_adjustment = is_special_care * 0.01 + + concept_scores = cos_dist - self.concept_embeds_weights[None, :] + special_adjustment + concept_scores = jnp.round(concept_scores, 3) + has_nsfw_concepts = jnp.any(concept_scores > 0, axis=1) + + return has_nsfw_concepts + + +class FlaxStableDiffusionSafetyChecker(FlaxPreTrainedModel): + config_class = CLIPConfig + main_input_name = "clip_input" + module_class = FlaxStableDiffusionSafetyCheckerModule + + def __init__( + self, + config: CLIPConfig, + input_shape: Optional[Tuple] = None, + seed: int = 0, + dtype: jnp.dtype = jnp.float32, + _do_init: bool = True, + **kwargs, + ): + if input_shape is None: + input_shape = (1, 224, 224, 3) + module = self.module_class(config=config, dtype=dtype, **kwargs) + super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init) + + def init_weights(self, rng: jax.Array, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict: + # init input tensor + clip_input = jax.random.normal(rng, input_shape) + + params_rng, dropout_rng = jax.random.split(rng) + rngs = {"params": params_rng, "dropout": dropout_rng} + + random_params = self.module.init(rngs, clip_input)["params"] + + return random_params + + def __call__( + self, + clip_input, + params: dict = None, + ): + clip_input = jnp.transpose(clip_input, (0, 2, 3, 1)) + + return self.module.apply( + {"params": params or self.params}, + jnp.array(clip_input, dtype=jnp.float32), + rngs={}, + ) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/stable_unclip_image_normalizer.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/stable_unclip_image_normalizer.py new file mode 100644 index 0000000000000000000000000000000000000000..ffd66792fe463c775027362b1b63fff250f3b52f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion/stable_unclip_image_normalizer.py @@ -0,0 +1,57 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 torch import nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models.modeling_utils import ModelMixin + + +class StableUnCLIPImageNormalizer(ModelMixin, ConfigMixin): + """ + This class is used to hold the mean and standard deviation of the CLIP embedder used in stable unCLIP. + + It is used to normalize the image embeddings before the noise is applied and un-normalize the noised image + embeddings. + """ + + @register_to_config + def __init__( + self, + embedding_dim: int = 768, + ): + super().__init__() + + self.mean = nn.Parameter(torch.zeros(1, embedding_dim)) + self.std = nn.Parameter(torch.ones(1, embedding_dim)) + + def to( + self, + torch_device: Optional[Union[str, torch.device]] = None, + torch_dtype: Optional[torch.dtype] = None, + ): + self.mean = nn.Parameter(self.mean.to(torch_device).to(torch_dtype)) + self.std = nn.Parameter(self.std.to(torch_device).to(torch_dtype)) + return self + + def scale(self, embeds): + embeds = (embeds - self.mean) * 1.0 / self.std + return embeds + + def unscale(self, embeds): + embeds = (embeds * self.std) + self.mean + return embeds diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b0604589a208d0c5c1270629dcad144395b750cb --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__init__.py @@ -0,0 +1,54 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_flax_available, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_additional_imports = {} +_import_structure = {"pipeline_output": ["StableDiffusion3PipelineOutput"]} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_3"] = ["StableDiffusion3Pipeline"] + _import_structure["pipeline_stable_diffusion_3_img2img"] = ["StableDiffusion3Img2ImgPipeline"] + _import_structure["pipeline_stable_diffusion_3_inpaint"] = ["StableDiffusion3InpaintPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_stable_diffusion_3 import StableDiffusion3Pipeline + from .pipeline_stable_diffusion_3_img2img import StableDiffusion3Img2ImgPipeline + from .pipeline_stable_diffusion_3_inpaint import StableDiffusion3InpaintPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) + for name, value in _additional_imports.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a6e1e1d1aec20eabf1b3936c54c802f20b4e743d Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_output.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_output.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e6dd46d56117831d2deb81585da22cdbb628a214 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_output.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_stable_diffusion_3.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_stable_diffusion_3.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3c50861590ef21ab107333e928cebde5d2f7ca5a Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_stable_diffusion_3.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_stable_diffusion_3_img2img.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_stable_diffusion_3_img2img.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..635ec0213d9252f90e821ae256cdce6b89a4ad67 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_stable_diffusion_3_img2img.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_stable_diffusion_3_inpaint.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_stable_diffusion_3_inpaint.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0392b221d6d23e1b34b0d1d7a62cda8e7232db64 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/__pycache__/pipeline_stable_diffusion_3_inpaint.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/pipeline_output.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..4655f446102a7f2db93743abdba9acc038270862 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/pipeline_output.py @@ -0,0 +1,21 @@ +from dataclasses import dataclass +from typing import List, Union + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class StableDiffusion3PipelineOutput(BaseOutput): + """ + Output class for Stable Diffusion pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3.py new file mode 100644 index 0000000000000000000000000000000000000000..ad4c4b09172ee78b9f5eec5397b5afaa449a7c5b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3.py @@ -0,0 +1,1139 @@ +# Copyright 2025 Stability AI, The HuggingFace Team and The InstantX Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import torch +from transformers import ( + CLIPTextModelWithProjection, + CLIPTokenizer, + SiglipImageProcessor, + SiglipVisionModel, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, SD3IPAdapterMixin, SD3LoraLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import SD3Transformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import StableDiffusion3PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusion3Pipeline + + >>> pipe = StableDiffusion3Pipeline.from_pretrained( + ... "stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16 + ... ) + >>> pipe.to("cuda") + >>> prompt = "A cat holding a sign that says hello world" + >>> image = pipe(prompt).images[0] + >>> image.save("sd3.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingleFileMixin, SD3IPAdapterMixin): + r""" + Args: + transformer ([`SD3Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant, + with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size` + as its dimension. + text_encoder_2 ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + text_encoder_3 ([`T5EncoderModel`]): + Frozen text-encoder. Stable Diffusion 3 uses + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_3 (`T5TokenizerFast`): + Tokenizer of class + [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). + image_encoder (`SiglipVisionModel`, *optional*): + Pre-trained Vision Model for IP Adapter. + feature_extractor (`SiglipImageProcessor`, *optional*): + Image processor for IP Adapter. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->image_encoder->transformer->vae" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"] + + def __init__( + self, + transformer: SD3Transformer2DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer_2: CLIPTokenizer, + text_encoder_3: T5EncoderModel, + tokenizer_3: T5TokenizerFast, + image_encoder: SiglipVisionModel = None, + feature_extractor: SiglipImageProcessor = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + text_encoder_3=text_encoder_3, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + tokenizer_3=tokenizer_3, + transformer=transformer, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = ( + self.transformer.config.sample_size + if hasattr(self, "transformer") and self.transformer is not None + else 128 + ) + self.patch_size = ( + self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2 + ) + + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 256, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if self.text_encoder_3 is None: + return torch.zeros( + ( + batch_size * num_images_per_prompt, + self.tokenizer_max_length, + self.transformer.config.joint_attention_dim, + ), + device=device, + dtype=dtype, + ) + + text_inputs = self.tokenizer_3( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0] + + dtype = self.text_encoder_3.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + def _get_clip_prompt_embeds( + self, + prompt: Union[str, List[str]], + num_images_per_prompt: int = 1, + device: Optional[torch.device] = None, + clip_skip: Optional[int] = None, + clip_model_index: int = 0, + ): + device = device or self._execution_device + + clip_tokenizers = [self.tokenizer, self.tokenizer_2] + clip_text_encoders = [self.text_encoder, self.text_encoder_2] + + tokenizer = clip_tokenizers[clip_model_index] + text_encoder = clip_text_encoders[clip_model_index] + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt, 1) + pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds, pooled_prompt_embeds + + def encode_prompt( + self, + prompt: Union[str, List[str]], + prompt_2: Union[str, List[str]], + prompt_3: Union[str, List[str]], + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt_3: Optional[Union[str, List[str]]] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + clip_skip: Optional[int] = None, + max_sequence_length: int = 256, + lora_scale: Optional[float] = None, + ): + r""" + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + negative_prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used in all the text-encoders. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + prompt_3 = prompt_3 or prompt + prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3 + + prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=0, + ) + prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds( + prompt=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=1, + ) + clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1) + + t5_prompt_embed = self._get_t5_prompt_embeds( + prompt=prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + clip_prompt_embeds = torch.nn.functional.pad( + clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1]) + ) + + prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2) + pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + negative_prompt_3 = negative_prompt_3 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + negative_prompt_3 = ( + batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3 + ) + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds( + negative_prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=0, + ) + negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds( + negative_prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=1, + ) + negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1) + + t5_negative_prompt_embed = self._get_t5_prompt_embeds( + prompt=negative_prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + negative_clip_prompt_embeds = torch.nn.functional.pad( + negative_clip_prompt_embeds, + (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]), + ) + + negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2) + negative_pooled_prompt_embeds = torch.cat( + [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1 + ) + + if self.text_encoder is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + def check_inputs( + self, + prompt, + prompt_2, + prompt_3, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + negative_prompt_3=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if ( + height % (self.vae_scale_factor * self.patch_size) != 0 + or width % (self.vae_scale_factor * self.patch_size) != 0 + ): + raise ValueError( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}." + f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_3 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_3`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)): + raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_3 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def skip_guidance_layers(self): + return self._skip_guidance_layers + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + # Adapted from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_image + def encode_image(self, image: PipelineImageInput, device: torch.device) -> torch.Tensor: + """Encodes the given image into a feature representation using a pre-trained image encoder. + + Args: + image (`PipelineImageInput`): + Input image to be encoded. + device: (`torch.device`): + Torch device. + + Returns: + `torch.Tensor`: The encoded image feature representation. + """ + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=self.dtype) + + return self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + + # Adapted from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[torch.Tensor] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + ) -> torch.Tensor: + """Prepares image embeddings for use in the IP-Adapter. + + Either `ip_adapter_image` or `ip_adapter_image_embeds` must be passed. + + Args: + ip_adapter_image (`PipelineImageInput`, *optional*): + The input image to extract features from for IP-Adapter. + ip_adapter_image_embeds (`torch.Tensor`, *optional*): + Precomputed image embeddings. + device: (`torch.device`, *optional*): + Torch device. + num_images_per_prompt (`int`, defaults to 1): + Number of images that should be generated per prompt. + do_classifier_free_guidance (`bool`, defaults to True): + Whether to use classifier free guidance or not. + """ + device = device or self._execution_device + + if ip_adapter_image_embeds is not None: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = ip_adapter_image_embeds.chunk(2) + else: + single_image_embeds = ip_adapter_image_embeds + elif ip_adapter_image is not None: + single_image_embeds = self.encode_image(ip_adapter_image, device) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.zeros_like(single_image_embeds) + else: + raise ValueError("Neither `ip_adapter_image_embeds` or `ip_adapter_image_embeds` were provided.") + + image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + negative_image_embeds = torch.cat([single_negative_image_embeds] * num_images_per_prompt, dim=0) + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + + return image_embeds.to(device=device) + + def enable_sequential_cpu_offload(self, *args, **kwargs): + if self.image_encoder is not None and "image_encoder" not in self._exclude_from_cpu_offload: + logger.warning( + "`pipe.enable_sequential_cpu_offload()` might fail for `image_encoder` if it uses " + "`torch.nn.MultiheadAttention`. You can exclude `image_encoder` from CPU offloading by calling " + "`pipe._exclude_from_cpu_offload.append('image_encoder')` before `pipe.enable_sequential_cpu_offload()`." + ) + + super().enable_sequential_cpu_offload(*args, **kwargs) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + prompt_3: Optional[Union[str, List[str]]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 28, + sigmas: Optional[List[float]] = None, + guidance_scale: float = 7.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt_3: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 256, + skip_guidance_layers: List[int] = None, + skip_layer_guidance_scale: float = 2.8, + skip_layer_guidance_stop: float = 0.2, + skip_layer_guidance_start: float = 0.01, + mu: Optional[float] = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + will be used instead + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used instead + negative_prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used instead + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`torch.Tensor`, *optional*): + Pre-generated image embeddings for IP-Adapter. Should be a tensor of shape `(batch_size, num_images, + emb_dim)`. It should contain the negative image embedding if `do_classifier_free_guidance` is set to + `True`. If not provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] instead of + a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + skip_guidance_layers (`List[int]`, *optional*): + A list of integers that specify layers to skip during guidance. If not provided, all layers will be + used for guidance. If provided, the guidance will only be applied to the layers specified in the list. + Recommended value by StabiltyAI for Stable Diffusion 3.5 Medium is [7, 8, 9]. + skip_layer_guidance_scale (`int`, *optional*): The scale of the guidance for the layers specified in + `skip_guidance_layers`. The guidance will be applied to the layers specified in `skip_guidance_layers` + with a scale of `skip_layer_guidance_scale`. The guidance will be applied to the rest of the layers + with a scale of `1`. + skip_layer_guidance_stop (`int`, *optional*): The step at which the guidance for the layers specified in + `skip_guidance_layers` will stop. The guidance will be applied to the layers specified in + `skip_guidance_layers` until the fraction specified in `skip_layer_guidance_stop`. Recommended value by + StabiltyAI for Stable Diffusion 3.5 Medium is 0.2. + skip_layer_guidance_start (`int`, *optional*): The step at which the guidance for the layers specified in + `skip_guidance_layers` will start. The guidance will be applied to the layers specified in + `skip_guidance_layers` from the fraction specified in `skip_layer_guidance_start`. Recommended value by + StabiltyAI for Stable Diffusion 3.5 Medium is 0.01. + mu (`float`, *optional*): `mu` value used for `dynamic_shifting`. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + prompt_3, + height, + width, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._skip_layer_guidance_scale = skip_layer_guidance_scale + self._clip_skip = clip_skip + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_3=prompt_3, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + clip_skip=self.clip_skip, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + if self.do_classifier_free_guidance: + if skip_guidance_layers is not None: + original_prompt_embeds = prompt_embeds + original_pooled_prompt_embeds = pooled_prompt_embeds + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5. Prepare timesteps + scheduler_kwargs = {} + if self.scheduler.config.get("use_dynamic_shifting", None) and mu is None: + _, _, height, width = latents.shape + image_seq_len = (height // self.transformer.config.patch_size) * ( + width // self.transformer.config.patch_size + ) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.16), + ) + scheduler_kwargs["mu"] = mu + elif mu is not None: + scheduler_kwargs["mu"] = mu + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + **scheduler_kwargs, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 6. Prepare image embeddings + if (ip_adapter_image is not None and self.is_ip_adapter_active) or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {"ip_adapter_image_embeds": ip_adapter_image_embeds} + else: + self._joint_attention_kwargs.update(ip_adapter_image_embeds=ip_adapter_image_embeds) + + # 7. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + pooled_projections=pooled_prompt_embeds, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + should_skip_layers = ( + True + if i > num_inference_steps * skip_layer_guidance_start + and i < num_inference_steps * skip_layer_guidance_stop + else False + ) + if skip_guidance_layers is not None and should_skip_layers: + timestep = t.expand(latents.shape[0]) + latent_model_input = latents + noise_pred_skip_layers = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=original_prompt_embeds, + pooled_projections=original_pooled_prompt_embeds, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + skip_layers=skip_guidance_layers, + )[0] + noise_pred = ( + noise_pred + (noise_pred_text - noise_pred_skip_layers) * self._skip_layer_guidance_scale + ) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + + else: + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusion3PipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_img2img.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..fa1e0a4f3270c5c5c84d5c546962ba4030f9cf1c --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_img2img.py @@ -0,0 +1,1154 @@ +# Copyright 2025 Stability AI and The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import PIL.Image +import torch +from transformers import ( + CLIPTextModelWithProjection, + CLIPTokenizer, + SiglipImageProcessor, + SiglipVisionModel, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, SD3IPAdapterMixin, SD3LoraLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import SD3Transformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import StableDiffusion3PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + + >>> from diffusers import AutoPipelineForImage2Image + >>> from diffusers.utils import load_image + + >>> device = "cuda" + >>> model_id_or_path = "stabilityai/stable-diffusion-3-medium-diffusers" + >>> pipe = AutoPipelineForImage2Image.from_pretrained(model_id_or_path, torch_dtype=torch.float16) + >>> pipe = pipe.to(device) + + >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + >>> init_image = load_image(url).resize((1024, 1024)) + + >>> prompt = "cat wizard, gandalf, lord of the rings, detailed, fantasy, cute, adorable, Pixar, Disney, 8k" + + >>> images = pipe(prompt=prompt, image=init_image, strength=0.95, guidance_scale=7.5).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusion3Img2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingleFileMixin, SD3IPAdapterMixin): + r""" + Args: + transformer ([`SD3Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant, + with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size` + as its dimension. + text_encoder_2 ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + text_encoder_3 ([`T5EncoderModel`]): + Frozen text-encoder. Stable Diffusion 3 uses + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_3 (`T5TokenizerFast`): + Tokenizer of class + [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). + image_encoder (`SiglipVisionModel`, *optional*): + Pre-trained Vision Model for IP Adapter. + feature_extractor (`SiglipImageProcessor`, *optional*): + Image processor for IP Adapter. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->image_encoder->transformer->vae" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"] + + def __init__( + self, + transformer: SD3Transformer2DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer_2: CLIPTokenizer, + text_encoder_3: T5EncoderModel, + tokenizer_3: T5TokenizerFast, + image_encoder: Optional[SiglipVisionModel] = None, + feature_extractor: Optional[SiglipImageProcessor] = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + text_encoder_3=text_encoder_3, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + tokenizer_3=tokenizer_3, + transformer=transformer, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, vae_latent_channels=latent_channels + ) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = ( + self.transformer.config.sample_size + if hasattr(self, "transformer") and self.transformer is not None + else 128 + ) + self.patch_size = ( + self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2 + ) + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 256, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if self.text_encoder_3 is None: + return torch.zeros( + ( + batch_size * num_images_per_prompt, + self.tokenizer_max_length, + self.transformer.config.joint_attention_dim, + ), + device=device, + dtype=dtype, + ) + + text_inputs = self.tokenizer_3( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0] + + dtype = self.text_encoder_3.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: Union[str, List[str]], + num_images_per_prompt: int = 1, + device: Optional[torch.device] = None, + clip_skip: Optional[int] = None, + clip_model_index: int = 0, + ): + device = device or self._execution_device + + clip_tokenizers = [self.tokenizer, self.tokenizer_2] + clip_text_encoders = [self.text_encoder, self.text_encoder_2] + + tokenizer = clip_tokenizers[clip_model_index] + text_encoder = clip_text_encoders[clip_model_index] + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt, 1) + pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds, pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_prompt + def encode_prompt( + self, + prompt: Union[str, List[str]], + prompt_2: Union[str, List[str]], + prompt_3: Union[str, List[str]], + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt_3: Optional[Union[str, List[str]]] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + clip_skip: Optional[int] = None, + max_sequence_length: int = 256, + lora_scale: Optional[float] = None, + ): + r""" + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + negative_prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used in all the text-encoders. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + prompt_3 = prompt_3 or prompt + prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3 + + prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=0, + ) + prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds( + prompt=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=1, + ) + clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1) + + t5_prompt_embed = self._get_t5_prompt_embeds( + prompt=prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + clip_prompt_embeds = torch.nn.functional.pad( + clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1]) + ) + + prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2) + pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + negative_prompt_3 = negative_prompt_3 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + negative_prompt_3 = ( + batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3 + ) + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds( + negative_prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=0, + ) + negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds( + negative_prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=1, + ) + negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1) + + t5_negative_prompt_embed = self._get_t5_prompt_embeds( + prompt=negative_prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + negative_clip_prompt_embeds = torch.nn.functional.pad( + negative_clip_prompt_embeds, + (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]), + ) + + negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2) + negative_pooled_prompt_embeds = torch.cat( + [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1 + ) + + if self.text_encoder is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + def check_inputs( + self, + prompt, + prompt_2, + prompt_3, + height, + width, + strength, + negative_prompt=None, + negative_prompt_2=None, + negative_prompt_3=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if ( + height % (self.vae_scale_factor * self.patch_size) != 0 + or width % (self.vae_scale_factor * self.patch_size) != 0 + ): + raise ValueError( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}." + f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}." + ) + + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_3 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_3`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)): + raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_3 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + if image.shape[1] == self.vae.config.latent_channels: + init_latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + init_latents = (init_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.scale_noise(init_latents, timestep, noise) + latents = init_latents.to(device=device, dtype=dtype) + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_image + def encode_image(self, image: PipelineImageInput, device: torch.device) -> torch.Tensor: + """Encodes the given image into a feature representation using a pre-trained image encoder. + + Args: + image (`PipelineImageInput`): + Input image to be encoded. + device: (`torch.device`): + Torch device. + + Returns: + `torch.Tensor`: The encoded image feature representation. + """ + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=self.dtype) + + return self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[torch.Tensor] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + ) -> torch.Tensor: + """Prepares image embeddings for use in the IP-Adapter. + + Either `ip_adapter_image` or `ip_adapter_image_embeds` must be passed. + + Args: + ip_adapter_image (`PipelineImageInput`, *optional*): + The input image to extract features from for IP-Adapter. + ip_adapter_image_embeds (`torch.Tensor`, *optional*): + Precomputed image embeddings. + device: (`torch.device`, *optional*): + Torch device. + num_images_per_prompt (`int`, defaults to 1): + Number of images that should be generated per prompt. + do_classifier_free_guidance (`bool`, defaults to True): + Whether to use classifier free guidance or not. + """ + device = device or self._execution_device + + if ip_adapter_image_embeds is not None: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = ip_adapter_image_embeds.chunk(2) + else: + single_image_embeds = ip_adapter_image_embeds + elif ip_adapter_image is not None: + single_image_embeds = self.encode_image(ip_adapter_image, device) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.zeros_like(single_image_embeds) + else: + raise ValueError("Neither `ip_adapter_image_embeds` or `ip_adapter_image_embeds` were provided.") + + image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + negative_image_embeds = torch.cat([single_negative_image_embeds] * num_images_per_prompt, dim=0) + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + + return image_embeds.to(device=device) + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.enable_sequential_cpu_offload + def enable_sequential_cpu_offload(self, *args, **kwargs): + if self.image_encoder is not None and "image_encoder" not in self._exclude_from_cpu_offload: + logger.warning( + "`pipe.enable_sequential_cpu_offload()` might fail for `image_encoder` if it uses " + "`torch.nn.MultiheadAttention`. You can exclude `image_encoder` from CPU offloading by calling " + "`pipe._exclude_from_cpu_offload.append('image_encoder')` before `pipe.enable_sequential_cpu_offload()`." + ) + + super().enable_sequential_cpu_offload(*args, **kwargs) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + prompt_3: Optional[Union[str, List[str]]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + image: PipelineImageInput = None, + strength: float = 0.6, + num_inference_steps: int = 50, + sigmas: Optional[List[float]] = None, + guidance_scale: float = 7.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt_3: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[torch.Tensor] = None, + return_dict: bool = True, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 256, + mu: Optional[float] = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + will be used instead + height (`int`, *optional*, defaults to self.transformer.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.transformer.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used instead + negative_prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used instead + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`torch.Tensor`, *optional*): + Pre-generated image embeddings for IP-Adapter. Should be a tensor of shape `(batch_size, num_images, + emb_dim)`. It should contain the negative image embedding if `do_classifier_free_guidance` is set to + `True`. If not provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] instead of + a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + mu (`float`, *optional*): `mu` value used for `dynamic_shifting`. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + prompt_3, + height, + width, + strength, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_3=prompt_3, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + clip_skip=self.clip_skip, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + + # 3. Preprocess image + image = self.image_processor.preprocess(image, height=height, width=width) + + # 4. Prepare timesteps + scheduler_kwargs = {} + if self.scheduler.config.get("use_dynamic_shifting", None) and mu is None: + image_seq_len = (int(height) // self.vae_scale_factor // self.transformer.config.patch_size) * ( + int(width) // self.vae_scale_factor // self.transformer.config.patch_size + ) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.16), + ) + scheduler_kwargs["mu"] = mu + elif mu is not None: + scheduler_kwargs["mu"] = mu + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, sigmas=sigmas, **scheduler_kwargs + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 5. Prepare latent variables + if latents is None: + latents = self.prepare_latents( + image, + latent_timestep, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + ) + + # 6. Prepare image embeddings + if (ip_adapter_image is not None and self.is_ip_adapter_active) or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {"ip_adapter_image_embeds": ip_adapter_image_embeds} + else: + self._joint_attention_kwargs.update(ip_adapter_image_embeds=ip_adapter_image_embeds) + + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + pooled_projections=pooled_prompt_embeds, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + + else: + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusion3PipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_inpaint.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..937f7195b21d94b9103f3b130ec6b1951a5bda40 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_inpaint.py @@ -0,0 +1,1379 @@ +# Copyright 2025 Stability AI and The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import torch +from transformers import ( + CLIPTextModelWithProjection, + CLIPTokenizer, + SiglipImageProcessor, + SiglipVisionModel, + T5EncoderModel, + T5TokenizerFast, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, SD3IPAdapterMixin, SD3LoraLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import SD3Transformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import StableDiffusion3PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusion3InpaintPipeline + >>> from diffusers.utils import load_image + + >>> pipe = StableDiffusion3InpaintPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16 + ... ) + >>> pipe.to("cuda") + >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench" + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + >>> source = load_image(img_url) + >>> mask = load_image(mask_url) + >>> image = pipe(prompt=prompt, image=source, mask_image=mask).images[0] + >>> image.save("sd3_inpainting.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusion3InpaintPipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingleFileMixin, SD3IPAdapterMixin): + r""" + Args: + transformer ([`SD3Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant, + with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size` + as its dimension. + text_encoder_2 ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + text_encoder_3 ([`T5EncoderModel`]): + Frozen text-encoder. Stable Diffusion 3 uses + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_3 (`T5TokenizerFast`): + Tokenizer of class + [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). + image_encoder (`SiglipVisionModel`, *optional*): + Pre-trained Vision Model for IP Adapter. + feature_extractor (`SiglipImageProcessor`, *optional*): + Image processor for IP Adapter. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->image_encoder->transformer->vae" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"] + + def __init__( + self, + transformer: SD3Transformer2DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer_2: CLIPTokenizer, + text_encoder_3: T5EncoderModel, + tokenizer_3: T5TokenizerFast, + image_encoder: Optional[SiglipVisionModel] = None, + feature_extractor: Optional[SiglipImageProcessor] = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + text_encoder_3=text_encoder_3, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + tokenizer_3=tokenizer_3, + transformer=transformer, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, vae_latent_channels=latent_channels + ) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, + vae_latent_channels=latent_channels, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = ( + self.transformer.config.sample_size + if hasattr(self, "transformer") and self.transformer is not None + else 128 + ) + self.patch_size = ( + self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2 + ) + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 256, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if self.text_encoder_3 is None: + return torch.zeros( + ( + batch_size * num_images_per_prompt, + self.tokenizer_max_length, + self.transformer.config.joint_attention_dim, + ), + device=device, + dtype=dtype, + ) + + text_inputs = self.tokenizer_3( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0] + + dtype = self.text_encoder_3.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: Union[str, List[str]], + num_images_per_prompt: int = 1, + device: Optional[torch.device] = None, + clip_skip: Optional[int] = None, + clip_model_index: int = 0, + ): + device = device or self._execution_device + + clip_tokenizers = [self.tokenizer, self.tokenizer_2] + clip_text_encoders = [self.text_encoder, self.text_encoder_2] + + tokenizer = clip_tokenizers[clip_model_index] + text_encoder = clip_text_encoders[clip_model_index] + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt, 1) + pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds, pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_prompt + def encode_prompt( + self, + prompt: Union[str, List[str]], + prompt_2: Union[str, List[str]], + prompt_3: Union[str, List[str]], + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt_3: Optional[Union[str, List[str]]] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + clip_skip: Optional[int] = None, + max_sequence_length: int = 256, + lora_scale: Optional[float] = None, + ): + r""" + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + negative_prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used in all the text-encoders. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + prompt_3 = prompt_3 or prompt + prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3 + + prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=0, + ) + prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds( + prompt=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=1, + ) + clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1) + + t5_prompt_embed = self._get_t5_prompt_embeds( + prompt=prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + clip_prompt_embeds = torch.nn.functional.pad( + clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1]) + ) + + prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2) + pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + negative_prompt_3 = negative_prompt_3 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + negative_prompt_3 = ( + batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3 + ) + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds( + negative_prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=0, + ) + negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds( + negative_prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=1, + ) + negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1) + + t5_negative_prompt_embed = self._get_t5_prompt_embeds( + prompt=negative_prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + negative_clip_prompt_embeds = torch.nn.functional.pad( + negative_clip_prompt_embeds, + (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]), + ) + + negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2) + negative_pooled_prompt_embeds = torch.cat( + [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1 + ) + + if self.text_encoder is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + prompt_3, + height, + width, + strength, + negative_prompt=None, + negative_prompt_2=None, + negative_prompt_3=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if ( + height % (self.vae_scale_factor * self.patch_size) != 0 + or width % (self.vae_scale_factor * self.patch_size) != 0 + ): + raise ValueError( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}." + f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}." + ) + + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_3 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_3`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)): + raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_3 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + image=None, + timestep=None, + is_strength_max=True, + return_noise=False, + return_image_latents=False, + ): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if (image is None or timestep is None) and not is_strength_max: + raise ValueError( + "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise." + "However, either the image or the noise timestep has not been provided." + ) + + if return_image_latents or (latents is None and not is_strength_max): + image = image.to(device=device, dtype=dtype) + + if image.shape[1] == 16: + image_latents = image + else: + image_latents = self._encode_vae_image(image=image, generator=generator) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + + if latents is None: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # if strength is 1. then initialise the latents to noise, else initial to image + noise + latents = noise if is_strength_max else self.scheduler.scale_noise(image_latents, timestep, noise) + else: + noise = latents.to(device) + latents = noise + + outputs = (latents,) + + if return_noise: + outputs += (noise,) + + if return_image_latents: + outputs += (image_latents,) + + return outputs + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + def prepare_mask_latents( + self, + mask, + masked_image, + batch_size, + num_images_per_prompt, + height, + width, + dtype, + device, + generator, + do_classifier_free_guidance, + ): + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate( + mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor) + ) + mask = mask.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == 16: + masked_image_latents = masked_image + else: + masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator) + + masked_image_latents = (masked_image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + + mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask + masked_image_latents = ( + torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + ) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + return mask, masked_image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_image + def encode_image(self, image: PipelineImageInput, device: torch.device) -> torch.Tensor: + """Encodes the given image into a feature representation using a pre-trained image encoder. + + Args: + image (`PipelineImageInput`): + Input image to be encoded. + device: (`torch.device`): + Torch device. + + Returns: + `torch.Tensor`: The encoded image feature representation. + """ + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=self.dtype) + + return self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[torch.Tensor] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + ) -> torch.Tensor: + """Prepares image embeddings for use in the IP-Adapter. + + Either `ip_adapter_image` or `ip_adapter_image_embeds` must be passed. + + Args: + ip_adapter_image (`PipelineImageInput`, *optional*): + The input image to extract features from for IP-Adapter. + ip_adapter_image_embeds (`torch.Tensor`, *optional*): + Precomputed image embeddings. + device: (`torch.device`, *optional*): + Torch device. + num_images_per_prompt (`int`, defaults to 1): + Number of images that should be generated per prompt. + do_classifier_free_guidance (`bool`, defaults to True): + Whether to use classifier free guidance or not. + """ + device = device or self._execution_device + + if ip_adapter_image_embeds is not None: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = ip_adapter_image_embeds.chunk(2) + else: + single_image_embeds = ip_adapter_image_embeds + elif ip_adapter_image is not None: + single_image_embeds = self.encode_image(ip_adapter_image, device) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.zeros_like(single_image_embeds) + else: + raise ValueError("Neither `ip_adapter_image_embeds` or `ip_adapter_image_embeds` were provided.") + + image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + negative_image_embeds = torch.cat([single_negative_image_embeds] * num_images_per_prompt, dim=0) + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + + return image_embeds.to(device=device) + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.enable_sequential_cpu_offload + def enable_sequential_cpu_offload(self, *args, **kwargs): + if self.image_encoder is not None and "image_encoder" not in self._exclude_from_cpu_offload: + logger.warning( + "`pipe.enable_sequential_cpu_offload()` might fail for `image_encoder` if it uses " + "`torch.nn.MultiheadAttention`. You can exclude `image_encoder` from CPU offloading by calling " + "`pipe._exclude_from_cpu_offload.append('image_encoder')` before `pipe.enable_sequential_cpu_offload()`." + ) + + super().enable_sequential_cpu_offload(*args, **kwargs) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + prompt_3: Optional[Union[str, List[str]]] = None, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: PipelineImageInput = None, + height: int = None, + width: int = None, + padding_mask_crop: Optional[int] = None, + strength: float = 0.6, + num_inference_steps: int = 50, + sigmas: Optional[List[float]] = None, + guidance_scale: float = 7.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt_3: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 256, + mu: Optional[float] = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + will be used instead + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + mask_image_latent (`torch.Tensor`, `List[torch.Tensor]`): + `Tensor` representing an image batch to mask `image` generated by VAE. If not provided, the mask + latents tensor will ge generated by `mask_image`. + height (`int`, *optional*, defaults to self.transformer.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.transformer.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used instead + negative_prompt_3 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used instead + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`torch.Tensor`, *optional*): + Pre-generated image embeddings for IP-Adapter. Should be a tensor of shape `(batch_size, num_images, + emb_dim)`. It should contain the negative image embedding if `do_classifier_free_guidance` is set to + `True`. If not provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] instead of + a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + mu (`float`, *optional*): `mu` value used for `dynamic_shifting`. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + height = height or self.transformer.config.sample_size * self.vae_scale_factor + width = width or self.transformer.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + prompt_3, + height, + width, + strength, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_3=prompt_3, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + clip_skip=self.clip_skip, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + + # 3. Prepare timesteps + scheduler_kwargs = {} + if self.scheduler.config.get("use_dynamic_shifting", None) and mu is None: + image_seq_len = (int(height) // self.vae_scale_factor // self.transformer.config.patch_size) * ( + int(width) // self.vae_scale_factor // self.transformer.config.patch_size + ) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.16), + ) + scheduler_kwargs["mu"] = mu + elif mu is not None: + scheduler_kwargs["mu"] = mu + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, sigmas=sigmas, **scheduler_kwargs + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + # check that number of inference steps is not < 1 - as this doesn't make sense + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise + is_strength_max = strength == 1.0 + + # 4. Preprocess mask and image + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + # 5. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + num_channels_transformer = self.transformer.config.in_channels + return_image_latents = num_channels_transformer == 16 + + latents_outputs = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + image=init_image, + timestep=latent_timestep, + is_strength_max=is_strength_max, + return_noise=True, + return_image_latents=return_image_latents, + ) + + if return_image_latents: + latents, noise, image_latents = latents_outputs + else: + latents, noise = latents_outputs + + # 6. Prepare mask latent variables + mask_condition = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + if masked_image_latents is None: + masked_image = init_image * (mask_condition < 0.5) + else: + masked_image = masked_image_latents + + mask, masked_image_latents = self.prepare_mask_latents( + mask_condition, + masked_image, + batch_size, + num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + self.do_classifier_free_guidance, + ) + + # match the inpainting pipeline and will be updated with input + mask inpainting model later + if num_channels_transformer == 33: + # default case for runwayml/stable-diffusion-inpainting + num_channels_mask = mask.shape[1] + num_channels_masked_image = masked_image_latents.shape[1] + if ( + num_channels_latents + num_channels_mask + num_channels_masked_image + != self.transformer.config.in_channels + ): + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.transformer`: {self.transformer.config} expects" + f" {self.transformer.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of" + " `pipeline.transformer` or your `mask_image` or `image` input." + ) + elif num_channels_transformer != 16: + raise ValueError( + f"The transformer {self.transformer.__class__} should have 16 input channels or 33 input channels, not {self.transformer.config.in_channels}." + ) + + # 7. Prepare image embeddings + if (ip_adapter_image is not None and self.is_ip_adapter_active) or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {"ip_adapter_image_embeds": ip_adapter_image_embeds} + else: + self._joint_attention_kwargs.update(ip_adapter_image_embeds=ip_adapter_image_embeds) + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + if num_channels_transformer == 33: + latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + pooled_projections=pooled_prompt_embeds, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + if num_channels_transformer == 16: + init_latents_proper = image_latents + if self.do_classifier_free_guidance: + init_mask, _ = mask.chunk(2) + else: + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.scale_noise( + init_latents_proper, torch.tensor([noise_timestep]), noise + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + mask = callback_outputs.pop("mask", mask) + masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + else: + image = latents + + do_denormalize = [True] * image.shape[0] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + if padding_mask_crop is not None: + image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image] + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusion3PipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_attend_and_excite/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_attend_and_excite/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..cce556fceb2379be482d383e380a09836d25ce3b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_attend_and_excite/__init__.py @@ -0,0 +1,48 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_attend_and_excite"] = ["StableDiffusionAttendAndExcitePipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_stable_diffusion_attend_and_excite import StableDiffusionAttendAndExcitePipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_attend_and_excite/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_attend_and_excite/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1a4f8e8d8780411f0e6f551142032c2ea0b85223 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_attend_and_excite/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_attend_and_excite/__pycache__/pipeline_stable_diffusion_attend_and_excite.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_attend_and_excite/__pycache__/pipeline_stable_diffusion_attend_and_excite.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9cc1895cc65e2e90fe31897e208d25fee6fb7974 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_attend_and_excite/__pycache__/pipeline_stable_diffusion_attend_and_excite.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_attend_and_excite/pipeline_stable_diffusion_attend_and_excite.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_attend_and_excite/pipeline_stable_diffusion_attend_and_excite.py new file mode 100644 index 0000000000000000000000000000000000000000..a1ff99b6aa3492ec44dacc75b8fc52705635731a --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_attend_and_excite/pipeline_stable_diffusion_attend_and_excite.py @@ -0,0 +1,1113 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +import math +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import numpy as np +import torch +from torch.nn import functional as F +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ...image_processor import VaeImageProcessor +from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.attention_processor import Attention +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion import StableDiffusionPipelineOutput +from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionAttendAndExcitePipeline + + >>> pipe = StableDiffusionAttendAndExcitePipeline.from_pretrained( + ... "CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16 + ... ).to("cuda") + + + >>> prompt = "a cat and a frog" + + >>> # use get_indices function to find out indices of the tokens you want to alter + >>> pipe.get_indices(prompt) + {0: '<|startoftext|>', 1: 'a', 2: 'cat', 3: 'and', 4: 'a', 5: 'frog', 6: '<|endoftext|>'} + + >>> token_indices = [2, 5] + >>> seed = 6141 + >>> generator = torch.Generator("cuda").manual_seed(seed) + + >>> images = pipe( + ... prompt=prompt, + ... token_indices=token_indices, + ... guidance_scale=7.5, + ... generator=generator, + ... num_inference_steps=50, + ... max_iter_to_alter=25, + ... ).images + + >>> image = images[0] + >>> image.save(f"../images/{prompt}_{seed}.png") + ``` +""" + + +class AttentionStore: + @staticmethod + def get_empty_store(): + return {"down": [], "mid": [], "up": []} + + def __call__(self, attn, is_cross: bool, place_in_unet: str): + if self.cur_att_layer >= 0 and is_cross: + if attn.shape[1] == np.prod(self.attn_res): + self.step_store[place_in_unet].append(attn) + + self.cur_att_layer += 1 + if self.cur_att_layer == self.num_att_layers: + self.cur_att_layer = 0 + self.between_steps() + + def between_steps(self): + self.attention_store = self.step_store + self.step_store = self.get_empty_store() + + def get_average_attention(self): + average_attention = self.attention_store + return average_attention + + def aggregate_attention(self, from_where: List[str]) -> torch.Tensor: + """Aggregates the attention across the different layers and heads at the specified resolution.""" + out = [] + attention_maps = self.get_average_attention() + for location in from_where: + for item in attention_maps[location]: + cross_maps = item.reshape(-1, self.attn_res[0], self.attn_res[1], item.shape[-1]) + out.append(cross_maps) + out = torch.cat(out, dim=0) + out = out.sum(0) / out.shape[0] + return out + + def reset(self): + self.cur_att_layer = 0 + self.step_store = self.get_empty_store() + self.attention_store = {} + + def __init__(self, attn_res): + """ + Initialize an empty AttentionStore :param step_index: used to visualize only a specific step in the diffusion + process + """ + self.num_att_layers = -1 + self.cur_att_layer = 0 + self.step_store = self.get_empty_store() + self.attention_store = {} + self.curr_step_index = 0 + self.attn_res = attn_res + + +class AttendExciteAttnProcessor: + def __init__(self, attnstore, place_in_unet): + super().__init__() + self.attnstore = attnstore + self.place_in_unet = place_in_unet + + def __call__(self, attn: Attention, hidden_states, encoder_hidden_states=None, attention_mask=None): + batch_size, sequence_length, _ = hidden_states.shape + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + + query = attn.to_q(hidden_states) + + is_cross = encoder_hidden_states is not None + encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + query = attn.head_to_batch_dim(query) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + + # only need to store attention maps during the Attend and Excite process + if attention_probs.requires_grad: + self.attnstore(attention_probs, is_cross, self.place_in_unet) + + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + return hidden_states + + +class StableDiffusionAttendAndExcitePipeline( + DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin +): + r""" + Pipeline for text-to-image generation using Stable Diffusion and Attend-and-Excite. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + _last_supported_version = "0.33.1" + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor"] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + indices, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + indices_is_list_ints = isinstance(indices, list) and isinstance(indices[0], int) + indices_is_list_list_ints = ( + isinstance(indices, list) and isinstance(indices[0], list) and isinstance(indices[0][0], int) + ) + + if not indices_is_list_ints and not indices_is_list_list_ints: + raise TypeError("`indices` must be a list of ints or a list of a list of ints") + + if indices_is_list_ints: + indices_batch_size = 1 + elif indices_is_list_list_ints: + indices_batch_size = len(indices) + + if prompt is not None and isinstance(prompt, str): + prompt_batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + prompt_batch_size = len(prompt) + elif prompt_embeds is not None: + prompt_batch_size = prompt_embeds.shape[0] + + if indices_batch_size != prompt_batch_size: + raise ValueError( + f"indices batch size must be same as prompt batch size. indices batch size: {indices_batch_size}, prompt batch size: {prompt_batch_size}" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @staticmethod + def _compute_max_attention_per_index( + attention_maps: torch.Tensor, + indices: List[int], + ) -> List[torch.Tensor]: + """Computes the maximum attention value for each of the tokens we wish to alter.""" + attention_for_text = attention_maps[:, :, 1:-1] + attention_for_text *= 100 + attention_for_text = torch.nn.functional.softmax(attention_for_text, dim=-1) + + # Shift indices since we removed the first token + indices = [index - 1 for index in indices] + + # Extract the maximum values + max_indices_list = [] + for i in indices: + image = attention_for_text[:, :, i] + smoothing = GaussianSmoothing().to(attention_maps.device) + input = F.pad(image.unsqueeze(0).unsqueeze(0), (1, 1, 1, 1), mode="reflect") + image = smoothing(input).squeeze(0).squeeze(0) + max_indices_list.append(image.max()) + return max_indices_list + + def _aggregate_and_get_max_attention_per_token( + self, + indices: List[int], + ): + """Aggregates the attention for each token and computes the max activation value for each token to alter.""" + attention_maps = self.attention_store.aggregate_attention( + from_where=("up", "down", "mid"), + ) + max_attention_per_index = self._compute_max_attention_per_index( + attention_maps=attention_maps, + indices=indices, + ) + return max_attention_per_index + + @staticmethod + def _compute_loss(max_attention_per_index: List[torch.Tensor]) -> torch.Tensor: + """Computes the attend-and-excite loss using the maximum attention value for each token.""" + losses = [max(0, 1.0 - curr_max) for curr_max in max_attention_per_index] + loss = max(losses) + return loss + + @staticmethod + def _update_latent(latents: torch.Tensor, loss: torch.Tensor, step_size: float) -> torch.Tensor: + """Update the latent according to the computed loss.""" + grad_cond = torch.autograd.grad(loss.requires_grad_(True), [latents], retain_graph=True)[0] + latents = latents - step_size * grad_cond + return latents + + def _perform_iterative_refinement_step( + self, + latents: torch.Tensor, + indices: List[int], + loss: torch.Tensor, + threshold: float, + text_embeddings: torch.Tensor, + step_size: float, + t: int, + max_refinement_steps: int = 20, + ): + """ + Performs the iterative latent refinement introduced in the paper. Here, we continuously update the latent code + according to our loss objective until the given threshold is reached for all tokens. + """ + iteration = 0 + target_loss = max(0, 1.0 - threshold) + while loss > target_loss: + iteration += 1 + + latents = latents.clone().detach().requires_grad_(True) + self.unet(latents, t, encoder_hidden_states=text_embeddings).sample + self.unet.zero_grad() + + # Get max activation value for each subject token + max_attention_per_index = self._aggregate_and_get_max_attention_per_token( + indices=indices, + ) + + loss = self._compute_loss(max_attention_per_index) + + if loss != 0: + latents = self._update_latent(latents, loss, step_size) + + logger.info(f"\t Try {iteration}. loss: {loss}") + + if iteration >= max_refinement_steps: + logger.info(f"\t Exceeded max number of iterations ({max_refinement_steps})! ") + break + + # Run one more time but don't compute gradients and update the latents. + # We just need to compute the new loss - the grad update will occur below + latents = latents.clone().detach().requires_grad_(True) + _ = self.unet(latents, t, encoder_hidden_states=text_embeddings).sample + self.unet.zero_grad() + + # Get max activation value for each subject token + max_attention_per_index = self._aggregate_and_get_max_attention_per_token( + indices=indices, + ) + loss = self._compute_loss(max_attention_per_index) + logger.info(f"\t Finished with loss of: {loss}") + return loss, latents, max_attention_per_index + + def register_attention_control(self): + attn_procs = {} + cross_att_count = 0 + for name in self.unet.attn_processors.keys(): + if name.startswith("mid_block"): + place_in_unet = "mid" + elif name.startswith("up_blocks"): + place_in_unet = "up" + elif name.startswith("down_blocks"): + place_in_unet = "down" + else: + continue + + cross_att_count += 1 + attn_procs[name] = AttendExciteAttnProcessor(attnstore=self.attention_store, place_in_unet=place_in_unet) + + self.unet.set_attn_processor(attn_procs) + self.attention_store.num_att_layers = cross_att_count + + def get_indices(self, prompt: str) -> Dict[str, int]: + """Utility function to list the indices of the tokens you wish to alte""" + ids = self.tokenizer(prompt).input_ids + indices = {i: tok for tok, i in zip(self.tokenizer.convert_ids_to_tokens(ids), range(len(ids)))} + return indices + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]], + token_indices: Union[List[int], List[List[int]]], + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: int = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + max_iter_to_alter: int = 25, + thresholds: dict = {0: 0.05, 10: 0.5, 20: 0.8}, + scale_factor: int = 20, + attn_res: Optional[Tuple[int]] = (16, 16), + clip_skip: Optional[int] = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + token_indices (`List[int]`): + The token indices to alter with attend-and-excite. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + max_iter_to_alter (`int`, *optional*, defaults to `25`): + Number of denoising steps to apply attend-and-excite. The `max_iter_to_alter` denoising steps are when + attend-and-excite is applied. For example, if `max_iter_to_alter` is `25` and there are a total of `30` + denoising steps, the first `25` denoising steps applies attend-and-excite and the last `5` will not. + thresholds (`dict`, *optional*, defaults to `{0: 0.05, 10: 0.5, 20: 0.8}`): + Dictionary defining the iterations and desired thresholds to apply iterative latent refinement in. + scale_factor (`int`, *optional*, default to 20): + Scale factor to control the step size of each attend-and-excite update. + attn_res (`tuple`, *optional*, default computed from width and height): + The 2D resolution of the semantic attention map. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + token_indices, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + if attn_res is None: + attn_res = int(np.ceil(width / 32)), int(np.ceil(height / 32)) + self.attention_store = AttentionStore(attn_res) + original_attn_proc = self.unet.attn_processors + self.register_attention_control() + + # default config for step size from original repo + scale_range = np.linspace(1.0, 0.5, len(self.scheduler.timesteps)) + step_size = scale_factor * np.sqrt(scale_range) + + text_embeddings = ( + prompt_embeds[batch_size * num_images_per_prompt :] if do_classifier_free_guidance else prompt_embeds + ) + + if isinstance(token_indices[0], int): + token_indices = [token_indices] + + indices = [] + + for ind in token_indices: + indices = indices + [ind] * num_images_per_prompt + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # Attend and excite process + with torch.enable_grad(): + latents = latents.clone().detach().requires_grad_(True) + updated_latents = [] + for latent, index, text_embedding in zip(latents, indices, text_embeddings): + # Forward pass of denoising with text conditioning + latent = latent.unsqueeze(0) + text_embedding = text_embedding.unsqueeze(0) + + self.unet( + latent, + t, + encoder_hidden_states=text_embedding, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + self.unet.zero_grad() + + # Get max activation value for each subject token + max_attention_per_index = self._aggregate_and_get_max_attention_per_token( + indices=index, + ) + + loss = self._compute_loss(max_attention_per_index=max_attention_per_index) + + # If this is an iterative refinement step, verify we have reached the desired threshold for all + if i in thresholds.keys() and loss > 1.0 - thresholds[i]: + loss, latent, max_attention_per_index = self._perform_iterative_refinement_step( + latents=latent, + indices=index, + loss=loss, + threshold=thresholds[i], + text_embeddings=text_embedding, + step_size=step_size[i], + t=t, + ) + + # Perform gradient update + if i < max_iter_to_alter: + if loss != 0: + latent = self._update_latent( + latents=latent, + loss=loss, + step_size=step_size[i], + ) + logger.info(f"Iteration {i} | Loss: {loss:0.4f}") + + updated_latents.append(latent) + + latents = torch.cat(updated_latents, dim=0) + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 8. Post-processing + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + self.maybe_free_model_hooks() + # make sure to set the original attention processors back + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + +class GaussianSmoothing(torch.nn.Module): + """ + Arguments: + Apply gaussian smoothing on a 1d, 2d or 3d tensor. Filtering is performed separately for each channel in the input + using a depthwise convolution. + channels (int, sequence): Number of channels of the input tensors. Output will + have this number of channels as well. + kernel_size (int, sequence): Size of the gaussian kernel. sigma (float, sequence): Standard deviation of the + gaussian kernel. dim (int, optional): The number of dimensions of the data. + Default value is 2 (spatial). + """ + + # channels=1, kernel_size=kernel_size, sigma=sigma, dim=2 + def __init__( + self, + channels: int = 1, + kernel_size: int = 3, + sigma: float = 0.5, + dim: int = 2, + ): + super().__init__() + + if isinstance(kernel_size, int): + kernel_size = [kernel_size] * dim + if isinstance(sigma, float): + sigma = [sigma] * dim + + # The gaussian kernel is the product of the + # gaussian function of each dimension. + kernel = 1 + meshgrids = torch.meshgrid([torch.arange(size, dtype=torch.float32) for size in kernel_size]) + for size, std, mgrid in zip(kernel_size, sigma, meshgrids): + mean = (size - 1) / 2 + kernel *= 1 / (std * math.sqrt(2 * math.pi)) * torch.exp(-(((mgrid - mean) / (2 * std)) ** 2)) + + # Make sure sum of values in gaussian kernel equals 1. + kernel = kernel / torch.sum(kernel) + + # Reshape to depthwise convolutional weight + kernel = kernel.view(1, 1, *kernel.size()) + kernel = kernel.repeat(channels, *[1] * (kernel.dim() - 1)) + + self.register_buffer("weight", kernel) + self.groups = channels + + if dim == 1: + self.conv = F.conv1d + elif dim == 2: + self.conv = F.conv2d + elif dim == 3: + self.conv = F.conv3d + else: + raise RuntimeError("Only 1, 2 and 3 dimensions are supported. Received {}.".format(dim)) + + def forward(self, input): + """ + Arguments: + Apply gaussian filter to input. + input (torch.Tensor): Input to apply gaussian filter on. + Returns: + filtered (torch.Tensor): Filtered output. + """ + return self.conv(input, weight=self.weight.to(input.dtype), groups=self.groups) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_diffedit/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_diffedit/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e2145edb96c6be124abf9e9a21b9a5e8a3f3d641 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_diffedit/__init__.py @@ -0,0 +1,48 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_diffedit"] = ["StableDiffusionDiffEditPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_stable_diffusion_diffedit import StableDiffusionDiffEditPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_diffedit/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_diffedit/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..61866386c3a7fa83a87ea47d814aa5f72d322f43 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_diffedit/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_diffedit/__pycache__/pipeline_stable_diffusion_diffedit.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_diffedit/__pycache__/pipeline_stable_diffusion_diffedit.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a9a64fd539aa6ed9577da209ad1e0f6559c68bac Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_diffedit/__pycache__/pipeline_stable_diffusion_diffedit.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_diffedit/pipeline_stable_diffusion_diffedit.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_diffedit/pipeline_stable_diffusion_diffedit.py new file mode 100644 index 0000000000000000000000000000000000000000..87bd9f4444ac954da3bdd067cbace2dcb6f2df44 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_diffedit/pipeline_stable_diffusion_diffedit.py @@ -0,0 +1,1552 @@ +# Copyright 2025 DiffEdit Authors and Pix2Pix Zero Authors and The HuggingFace Team. All rights reserved. +# +# 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 inspect +from dataclasses import dataclass +from typing import Any, Callable, Dict, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ...configuration_utils import FrozenDict +from ...image_processor import VaeImageProcessor +from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import DDIMInverseScheduler, KarrasDiffusionSchedulers +from ...utils import ( + PIL_INTERPOLATION, + USE_PEFT_BACKEND, + BaseOutput, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion import StableDiffusionPipelineOutput +from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class DiffEditInversionPipelineOutput(BaseOutput): + """ + Output class for Stable Diffusion pipelines. + + Args: + latents (`torch.Tensor`) + inverted latents tensor + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `num_timesteps * batch_size` or numpy array of shape `(num_timesteps, + batch_size, height, width, num_channels)`. PIL images or numpy array present the denoised images of the + diffusion pipeline. + """ + + latents: torch.Tensor + images: Union[List[PIL.Image.Image], np.ndarray] + + +EXAMPLE_DOC_STRING = """ + + ```py + >>> import PIL + >>> import requests + >>> import torch + >>> from io import BytesIO + + >>> from diffusers import StableDiffusionDiffEditPipeline + + + >>> def download_image(url): + ... response = requests.get(url) + ... return PIL.Image.open(BytesIO(response.content)).convert("RGB") + + + >>> img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png" + + >>> init_image = download_image(img_url).resize((768, 768)) + + >>> pipeline = StableDiffusionDiffEditPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16 + ... ) + + >>> pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config) + >>> pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config) + >>> pipeline.enable_model_cpu_offload() + + >>> mask_prompt = "A bowl of fruits" + >>> prompt = "A bowl of pears" + + >>> mask_image = pipeline.generate_mask(image=init_image, source_prompt=prompt, target_prompt=mask_prompt) + >>> image_latents = pipeline.invert(image=init_image, prompt=mask_prompt).latents + >>> image = pipeline(prompt=prompt, mask_image=mask_image, image_latents=image_latents).images[0] + ``` +""" + +EXAMPLE_INVERT_DOC_STRING = """ + ```py + >>> import PIL + >>> import requests + >>> import torch + >>> from io import BytesIO + + >>> from diffusers import StableDiffusionDiffEditPipeline + + + >>> def download_image(url): + ... response = requests.get(url) + ... return PIL.Image.open(BytesIO(response.content)).convert("RGB") + + + >>> img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png" + + >>> init_image = download_image(img_url).resize((768, 768)) + + >>> pipeline = StableDiffusionDiffEditPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16 + ... ) + + >>> pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config) + >>> pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config) + >>> pipeline.enable_model_cpu_offload() + + >>> prompt = "A bowl of fruits" + + >>> inverted_latents = pipeline.invert(image=init_image, prompt=prompt).latents + ``` +""" + + +def auto_corr_loss(hidden_states, generator=None): + reg_loss = 0.0 + for i in range(hidden_states.shape[0]): + for j in range(hidden_states.shape[1]): + noise = hidden_states[i : i + 1, j : j + 1, :, :] + while True: + roll_amount = torch.randint(noise.shape[2] // 2, (1,), generator=generator).item() + reg_loss += (noise * torch.roll(noise, shifts=roll_amount, dims=2)).mean() ** 2 + reg_loss += (noise * torch.roll(noise, shifts=roll_amount, dims=3)).mean() ** 2 + + if noise.shape[2] <= 8: + break + noise = torch.nn.functional.avg_pool2d(noise, kernel_size=2) + return reg_loss + + +def kl_divergence(hidden_states): + return hidden_states.var() + hidden_states.mean() ** 2 - 1 - torch.log(hidden_states.var() + 1e-7) + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess +def preprocess(image): + deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead" + deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + + image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +def preprocess_mask(mask, batch_size: int = 1): + if not isinstance(mask, torch.Tensor): + # preprocess mask + if isinstance(mask, (PIL.Image.Image, np.ndarray)): + mask = [mask] + + if isinstance(mask, list): + if isinstance(mask[0], PIL.Image.Image): + mask = [np.array(m.convert("L")).astype(np.float32) / 255.0 for m in mask] + if isinstance(mask[0], np.ndarray): + mask = np.stack(mask, axis=0) if mask[0].ndim < 3 else np.concatenate(mask, axis=0) + mask = torch.from_numpy(mask) + elif isinstance(mask[0], torch.Tensor): + mask = torch.stack(mask, dim=0) if mask[0].ndim < 3 else torch.cat(mask, dim=0) + + # Batch and add channel dim for single mask + if mask.ndim == 2: + mask = mask.unsqueeze(0).unsqueeze(0) + + # Batch single mask or add channel dim + if mask.ndim == 3: + # Single batched mask, no channel dim or single mask not batched but channel dim + if mask.shape[0] == 1: + mask = mask.unsqueeze(0) + + # Batched masks no channel dim + else: + mask = mask.unsqueeze(1) + + # Check mask shape + if batch_size > 1: + if mask.shape[0] == 1: + mask = torch.cat([mask] * batch_size) + elif mask.shape[0] > 1 and mask.shape[0] != batch_size: + raise ValueError( + f"`mask_image` with batch size {mask.shape[0]} cannot be broadcasted to batch size {batch_size} " + f"inferred by prompt inputs" + ) + + if mask.shape[1] != 1: + raise ValueError(f"`mask_image` must have 1 channel, but has {mask.shape[1]} channels") + + # Check mask is in [0, 1] + if mask.min() < 0 or mask.max() > 1: + raise ValueError("`mask_image` should be in [0, 1] range") + + # Binarize mask + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + + return mask + + +class StableDiffusionDiffEditPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, +): + r""" + + + This is an experimental feature! + + + + Pipeline for text-guided image inpainting using Stable Diffusion and DiffEdit. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading and saving methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. + inverse_scheduler ([`DDIMInverseScheduler`]): + A scheduler to be used in combination with `unet` to fill in the unmasked part of the input latents. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + _last_supported_version = "0.33.1" + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "inverse_scheduler"] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + inverse_scheduler: DDIMInverseScheduler, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "skip_prk_steps", True) is False: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration" + " `skip_prk_steps`. `skip_prk_steps` should be set to True in the configuration file. Please make" + " sure to update the config accordingly as not setting `skip_prk_steps` in the config might lead to" + " incorrect results in future versions. If you have downloaded this checkpoint from the Hugging Face" + " Hub, it would be very nice if you could open a Pull request for the" + " `scheduler/scheduler_config.json` file" + ) + deprecate("skip_prk_steps not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["skip_prk_steps"] = True + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + inverse_scheduler=inverse_scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def check_inputs( + self, + prompt, + strength, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if (strength is None) or (strength is not None and (strength < 0 or strength > 1)): + raise ValueError( + f"The value of `strength` should in [0.0, 1.0] but is, but is {strength} of type {type(strength)}." + ) + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def check_source_inputs( + self, + source_prompt=None, + source_negative_prompt=None, + source_prompt_embeds=None, + source_negative_prompt_embeds=None, + ): + if source_prompt is not None and source_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `source_prompt`: {source_prompt} and `source_prompt_embeds`: {source_prompt_embeds}." + " Please make sure to only forward one of the two." + ) + elif source_prompt is None and source_prompt_embeds is None: + raise ValueError( + "Provide either `source_image` or `source_prompt_embeds`. Cannot leave all both of the arguments undefined." + ) + elif source_prompt is not None and ( + not isinstance(source_prompt, str) and not isinstance(source_prompt, list) + ): + raise ValueError(f"`source_prompt` has to be of type `str` or `list` but is {type(source_prompt)}") + + if source_negative_prompt is not None and source_negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `source_negative_prompt`: {source_negative_prompt} and `source_negative_prompt_embeds`:" + f" {source_negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if source_prompt_embeds is not None and source_negative_prompt_embeds is not None: + if source_prompt_embeds.shape != source_negative_prompt_embeds.shape: + raise ValueError( + "`source_prompt_embeds` and `source_negative_prompt_embeds` must have the same shape when passed" + f" directly, but got: `source_prompt_embeds` {source_prompt_embeds.shape} !=" + f" `source_negative_prompt_embeds` {source_negative_prompt_embeds.shape}." + ) + + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + + return timesteps, num_inference_steps - t_start + + def get_inverse_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + + # safety for t_start overflow to prevent empty timsteps slice + if t_start == 0: + return self.inverse_scheduler.timesteps, num_inference_steps + timesteps = self.inverse_scheduler.timesteps[:-t_start] + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def prepare_image_latents(self, image, batch_size, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + if image.shape[1] == 4: + latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if isinstance(generator, list): + latents = [ + self.vae.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size) + ] + latents = torch.cat(latents, dim=0) + else: + latents = self.vae.encode(image).latent_dist.sample(generator) + + latents = self.vae.config.scaling_factor * latents + + if batch_size != latents.shape[0]: + if batch_size % latents.shape[0] == 0: + # expand image_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_latents_per_image = batch_size // latents.shape[0] + latents = torch.cat([latents] * additional_latents_per_image, dim=0) + else: + raise ValueError( + f"Cannot duplicate `image` of batch size {latents.shape[0]} to {batch_size} text prompts." + ) + else: + latents = torch.cat([latents], dim=0) + + return latents + + def get_epsilon(self, model_output: torch.Tensor, sample: torch.Tensor, timestep: int): + pred_type = self.inverse_scheduler.config.prediction_type + alpha_prod_t = self.inverse_scheduler.alphas_cumprod[timestep] + + beta_prod_t = 1 - alpha_prod_t + + if pred_type == "epsilon": + return model_output + elif pred_type == "sample": + return (sample - alpha_prod_t ** (0.5) * model_output) / beta_prod_t ** (0.5) + elif pred_type == "v_prediction": + return (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {pred_type} must be one of `epsilon`, `sample`, or `v_prediction`" + ) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def generate_mask( + self, + image: Union[torch.Tensor, PIL.Image.Image] = None, + target_prompt: Optional[Union[str, List[str]]] = None, + target_negative_prompt: Optional[Union[str, List[str]]] = None, + target_prompt_embeds: Optional[torch.Tensor] = None, + target_negative_prompt_embeds: Optional[torch.Tensor] = None, + source_prompt: Optional[Union[str, List[str]]] = None, + source_negative_prompt: Optional[Union[str, List[str]]] = None, + source_prompt_embeds: Optional[torch.Tensor] = None, + source_negative_prompt_embeds: Optional[torch.Tensor] = None, + num_maps_per_mask: Optional[int] = 10, + mask_encode_strength: Optional[float] = 0.5, + mask_thresholding_ratio: Optional[float] = 3.0, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + output_type: Optional[str] = "np", + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + ): + r""" + Generate a latent mask given a mask prompt, a target prompt, and an image. + + Args: + image (`PIL.Image.Image`): + `Image` or tensor representing an image batch to be used for computing the mask. + target_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide semantic mask generation. If not defined, you need to pass + `prompt_embeds`. + target_negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + target_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + target_negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + source_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide semantic mask generation using DiffEdit. If not defined, you need to + pass `source_prompt_embeds` or `source_image` instead. + source_negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide semantic mask generation away from using DiffEdit. If not defined, you + need to pass `source_negative_prompt_embeds` or `source_image` instead. + source_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings to guide the semantic mask generation. Can be used to easily tweak text + inputs (prompt weighting). If not provided, text embeddings are generated from `source_prompt` input + argument. + source_negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings to negatively guide the semantic mask generation. Can be used to easily + tweak text inputs (prompt weighting). If not provided, text embeddings are generated from + `source_negative_prompt` input argument. + num_maps_per_mask (`int`, *optional*, defaults to 10): + The number of noise maps sampled to generate the semantic mask using DiffEdit. + mask_encode_strength (`float`, *optional*, defaults to 0.5): + The strength of the noise maps sampled to generate the semantic mask using DiffEdit. Must be between 0 + and 1. + mask_thresholding_ratio (`float`, *optional*, defaults to 3.0): + The maximum multiple of the mean absolute difference used to clamp the semantic guidance map before + mask binarization. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the + [`~models.attention_processor.AttnProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + + Examples: + + Returns: + `List[PIL.Image.Image]` or `np.array`: + When returning a `List[PIL.Image.Image]`, the list consists of a batch of single-channel binary images + with dimensions `(height // self.vae_scale_factor, width // self.vae_scale_factor)`. If it's + `np.array`, the shape is `(batch_size, height // self.vae_scale_factor, width // + self.vae_scale_factor)`. + """ + + # 1. Check inputs (Provide dummy argument for callback_steps) + self.check_inputs( + target_prompt, + mask_encode_strength, + 1, + target_negative_prompt, + target_prompt_embeds, + target_negative_prompt_embeds, + ) + + self.check_source_inputs( + source_prompt, + source_negative_prompt, + source_prompt_embeds, + source_negative_prompt_embeds, + ) + + if (num_maps_per_mask is None) or ( + num_maps_per_mask is not None and (not isinstance(num_maps_per_mask, int) or num_maps_per_mask <= 0) + ): + raise ValueError( + f"`num_maps_per_mask` has to be a positive integer but is {num_maps_per_mask} of type" + f" {type(num_maps_per_mask)}." + ) + + if mask_thresholding_ratio is None or mask_thresholding_ratio <= 0: + raise ValueError( + f"`mask_thresholding_ratio` has to be positive but is {mask_thresholding_ratio} of type" + f" {type(mask_thresholding_ratio)}." + ) + + # 2. Define call parameters + if target_prompt is not None and isinstance(target_prompt, str): + batch_size = 1 + elif target_prompt is not None and isinstance(target_prompt, list): + batch_size = len(target_prompt) + else: + batch_size = target_prompt_embeds.shape[0] + if cross_attention_kwargs is None: + cross_attention_kwargs = {} + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompts + (cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None) + target_negative_prompt_embeds, target_prompt_embeds = self.encode_prompt( + target_prompt, + device, + num_maps_per_mask, + do_classifier_free_guidance, + target_negative_prompt, + prompt_embeds=target_prompt_embeds, + negative_prompt_embeds=target_negative_prompt_embeds, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + target_prompt_embeds = torch.cat([target_negative_prompt_embeds, target_prompt_embeds]) + + source_negative_prompt_embeds, source_prompt_embeds = self.encode_prompt( + source_prompt, + device, + num_maps_per_mask, + do_classifier_free_guidance, + source_negative_prompt, + prompt_embeds=source_prompt_embeds, + negative_prompt_embeds=source_negative_prompt_embeds, + ) + if do_classifier_free_guidance: + source_prompt_embeds = torch.cat([source_negative_prompt_embeds, source_prompt_embeds]) + + # 4. Preprocess image + image = self.image_processor.preprocess(image).repeat_interleave(num_maps_per_mask, dim=0) + + # 5. Set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, _ = self.get_timesteps(num_inference_steps, mask_encode_strength, device) + encode_timestep = timesteps[0] + + # 6. Prepare image latents and add noise with specified strength + image_latents = self.prepare_image_latents( + image, batch_size * num_maps_per_mask, self.vae.dtype, device, generator + ) + noise = randn_tensor(image_latents.shape, generator=generator, device=device, dtype=self.vae.dtype) + image_latents = self.scheduler.add_noise(image_latents, noise, encode_timestep) + + latent_model_input = torch.cat([image_latents] * (4 if do_classifier_free_guidance else 2)) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, encode_timestep) + + # 7. Predict the noise residual + prompt_embeds = torch.cat([source_prompt_embeds, target_prompt_embeds]) + noise_pred = self.unet( + latent_model_input, + encode_timestep, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + if do_classifier_free_guidance: + noise_pred_neg_src, noise_pred_source, noise_pred_uncond, noise_pred_target = noise_pred.chunk(4) + noise_pred_source = noise_pred_neg_src + guidance_scale * (noise_pred_source - noise_pred_neg_src) + noise_pred_target = noise_pred_uncond + guidance_scale * (noise_pred_target - noise_pred_uncond) + else: + noise_pred_source, noise_pred_target = noise_pred.chunk(2) + + # 8. Compute the mask from the absolute difference of predicted noise residuals + # TODO: Consider smoothing mask guidance map + mask_guidance_map = ( + torch.abs(noise_pred_target - noise_pred_source) + .reshape(batch_size, num_maps_per_mask, *noise_pred_target.shape[-3:]) + .mean([1, 2]) + ) + clamp_magnitude = mask_guidance_map.mean() * mask_thresholding_ratio + semantic_mask_image = mask_guidance_map.clamp(0, clamp_magnitude) / clamp_magnitude + semantic_mask_image = torch.where(semantic_mask_image <= 0.5, 0, 1) + mask_image = semantic_mask_image.cpu().numpy() + + # 9. Convert to Numpy array or PIL. + if output_type == "pil": + mask_image = self.image_processor.numpy_to_pil(mask_image) + + # Offload all models + self.maybe_free_model_hooks() + + return mask_image + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_INVERT_DOC_STRING) + def invert( + self, + prompt: Optional[Union[str, List[str]]] = None, + image: Union[torch.Tensor, PIL.Image.Image] = None, + num_inference_steps: int = 50, + inpaint_strength: float = 0.8, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + decode_latents: bool = False, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: Optional[int] = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + lambda_auto_corr: float = 20.0, + lambda_kl: float = 20.0, + num_reg_steps: int = 0, + num_auto_corr_rolls: int = 5, + ): + r""" + Generate inverted latents given a prompt and image. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`PIL.Image.Image`): + `Image` or tensor representing an image batch to produce the inverted latents guided by `prompt`. + inpaint_strength (`float`, *optional*, defaults to 0.8): + Indicates extent of the noising process to run latent inversion. Must be between 0 and 1. When + `inpaint_strength` is 1, the inversion process is run for the full number of iterations specified in + `num_inference_steps`. `image` is used as a reference for the inversion process, and adding more noise + increases `inpaint_strength`. If `inpaint_strength` is 0, no inpainting occurs. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + decode_latents (`bool`, *optional*, defaults to `False`): + Whether or not to decode the inverted latents into a generated image. Setting this argument to `True` + decodes all inverted latents for each timestep into a list of generated images. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.DiffEditInversionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the + [`~models.attention_processor.AttnProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + lambda_auto_corr (`float`, *optional*, defaults to 20.0): + Lambda parameter to control auto correction. + lambda_kl (`float`, *optional*, defaults to 20.0): + Lambda parameter to control Kullback-Leibler divergence output. + num_reg_steps (`int`, *optional*, defaults to 0): + Number of regularization loss steps. + num_auto_corr_rolls (`int`, *optional*, defaults to 5): + Number of auto correction roll steps. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.pipeline_stable_diffusion_diffedit.DiffEditInversionPipelineOutput`] or + `tuple`: + If `return_dict` is `True`, + [`~pipelines.stable_diffusion.pipeline_stable_diffusion_diffedit.DiffEditInversionPipelineOutput`] is + returned, otherwise a `tuple` is returned where the first element is the inverted latents tensors + ordered by increasing noise, and the second is the corresponding decoded images if `decode_latents` is + `True`, otherwise `None`. + """ + + # 1. Check inputs + self.check_inputs( + prompt, + inpaint_strength, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ) + + if image is None: + raise ValueError("`image` input cannot be undefined.") + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + if cross_attention_kwargs is None: + cross_attention_kwargs = {} + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Preprocess image + image = self.image_processor.preprocess(image) + + # 4. Prepare latent variables + num_images_per_prompt = 1 + latents = self.prepare_image_latents( + image, batch_size * num_images_per_prompt, self.vae.dtype, device, generator + ) + + # 5. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 6. Prepare timesteps + self.inverse_scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_inverse_timesteps(num_inference_steps, inpaint_strength, device) + + # 7. Noising loop where we obtain the intermediate noised latent image for each timestep. + num_warmup_steps = len(timesteps) - num_inference_steps * self.inverse_scheduler.order + inverted_latents = [] + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.inverse_scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # regularization of the noise prediction (not in original code or paper but borrowed from Pix2PixZero) + if num_reg_steps > 0: + with torch.enable_grad(): + for _ in range(num_reg_steps): + if lambda_auto_corr > 0: + for _ in range(num_auto_corr_rolls): + var = torch.autograd.Variable(noise_pred.detach().clone(), requires_grad=True) + + # Derive epsilon from model output before regularizing to IID standard normal + var_epsilon = self.get_epsilon(var, latent_model_input.detach(), t) + + l_ac = auto_corr_loss(var_epsilon, generator=generator) + l_ac.backward() + + grad = var.grad.detach() / num_auto_corr_rolls + noise_pred = noise_pred - lambda_auto_corr * grad + + if lambda_kl > 0: + var = torch.autograd.Variable(noise_pred.detach().clone(), requires_grad=True) + + # Derive epsilon from model output before regularizing to IID standard normal + var_epsilon = self.get_epsilon(var, latent_model_input.detach(), t) + + l_kld = kl_divergence(var_epsilon) + l_kld.backward() + + grad = var.grad.detach() + noise_pred = noise_pred - lambda_kl * grad + + noise_pred = noise_pred.detach() + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.inverse_scheduler.step(noise_pred, t, latents).prev_sample + inverted_latents.append(latents.detach().clone()) + + # call the callback, if provided + if i == len(timesteps) - 1 or ( + (i + 1) > num_warmup_steps and (i + 1) % self.inverse_scheduler.order == 0 + ): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + assert len(inverted_latents) == len(timesteps) + latents = torch.stack(list(reversed(inverted_latents)), 1) + + # 8. Post-processing + image = None + if decode_latents: + image = self.decode_latents(latents.flatten(0, 1)) + + # 9. Convert to PIL. + if decode_latents and output_type == "pil": + image = self.image_processor.numpy_to_pil(image) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (latents, image) + + return DiffEditInversionPipelineOutput(latents=latents, images=image) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Optional[Union[str, List[str]]] = None, + mask_image: Union[torch.Tensor, PIL.Image.Image] = None, + image_latents: Union[torch.Tensor, PIL.Image.Image] = None, + inpaint_strength: Optional[float] = 0.8, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + clip_skip: int = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + mask_image (`PIL.Image.Image`): + `Image` or tensor representing an image batch to mask the generated image. White pixels in the mask are + repainted, while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a tensor, it should contain one color channel (L) + instead of 3, so the expected shape would be `(B, 1, H, W)`. + image_latents (`PIL.Image.Image` or `torch.Tensor`): + Partially noised image latents from the inversion process to be used as inputs for image generation. + inpaint_strength (`float`, *optional*, defaults to 0.8): + Indicates extent to inpaint the masked area. Must be between 0 and 1. When `inpaint_strength` is 1, the + denoising process is run on the masked area for the full number of iterations specified in + `num_inference_steps`. `image_latents` is used as a reference for the masked area, and adding more + noise to a region increases `inpaint_strength`. If `inpaint_strength` is 0, no inpainting occurs. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + # 1. Check inputs + self.check_inputs( + prompt, + inpaint_strength, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ) + + if mask_image is None: + raise ValueError( + "`mask_image` input cannot be undefined. Use `generate_mask()` to compute `mask_image` from text prompts." + ) + if image_latents is None: + raise ValueError( + "`image_latents` input cannot be undefined. Use `invert()` to compute `image_latents` from input images." + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + if cross_attention_kwargs is None: + cross_attention_kwargs = {} + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Preprocess mask + mask_image = preprocess_mask(mask_image, batch_size) + latent_height, latent_width = mask_image.shape[-2:] + mask_image = torch.cat([mask_image] * num_images_per_prompt) + mask_image = mask_image.to(device=device, dtype=prompt_embeds.dtype) + + # 5. Set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, inpaint_strength, device) + + # 6. Preprocess image latents + if isinstance(image_latents, list) and any(isinstance(l, torch.Tensor) and l.ndim == 5 for l in image_latents): + image_latents = torch.cat(image_latents).detach() + elif isinstance(image_latents, torch.Tensor) and image_latents.ndim == 5: + image_latents = image_latents.detach() + else: + image_latents = self.image_processor.preprocess(image_latents).detach() + + latent_shape = (self.vae.config.latent_channels, latent_height, latent_width) + if image_latents.shape[-3:] != latent_shape: + raise ValueError( + f"Each latent image in `image_latents` must have shape {latent_shape}, " + f"but has shape {image_latents.shape[-3:]}" + ) + if image_latents.ndim == 4: + image_latents = image_latents.reshape(batch_size, len(timesteps), *latent_shape) + if image_latents.shape[:2] != (batch_size, len(timesteps)): + raise ValueError( + f"`image_latents` must have batch size {batch_size} with latent images from {len(timesteps)}" + f" timesteps, but has batch size {image_latents.shape[0]} with latent images from" + f" {image_latents.shape[1]} timesteps." + ) + image_latents = image_latents.transpose(0, 1).repeat_interleave(num_images_per_prompt, dim=1) + image_latents = image_latents.to(device=device, dtype=prompt_embeds.dtype) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 8. Denoising loop + latents = image_latents[0].clone() + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # mask with inverted latents from appropriate timestep - use original image latent for last step + latents = latents * mask_image + image_latents[i] * (1 - mask_image) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..147980cbf9e5c3418fc1854787ae37b25e4fed56 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/__init__.py @@ -0,0 +1,50 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_gligen"] = ["StableDiffusionGLIGENPipeline"] + _import_structure["pipeline_stable_diffusion_gligen_text_image"] = ["StableDiffusionGLIGENTextImagePipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_stable_diffusion_gligen import StableDiffusionGLIGENPipeline + from .pipeline_stable_diffusion_gligen_text_image import StableDiffusionGLIGENTextImagePipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c753d6f6890599f4eaac42c342260007e671fb78 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/__pycache__/pipeline_stable_diffusion_gligen.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/__pycache__/pipeline_stable_diffusion_gligen.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..03e2784dcbb4f68f174601dd74bb928c81c37b07 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/__pycache__/pipeline_stable_diffusion_gligen.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/__pycache__/pipeline_stable_diffusion_gligen_text_image.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/__pycache__/pipeline_stable_diffusion_gligen_text_image.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4dd97f39547d6a619758c851db7f461492af602e Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/__pycache__/pipeline_stable_diffusion_gligen_text_image.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen.py new file mode 100644 index 0000000000000000000000000000000000000000..78b026684cfaa8aef49bf4360bb472ade9fa24e9 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen.py @@ -0,0 +1,865 @@ +# Copyright 2025 The GLIGEN Authors and HuggingFace Team. All rights reserved. +# +# 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 inspect +import warnings +from typing import Any, Callable, Dict, List, Optional, Union + +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ...image_processor import VaeImageProcessor +from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.attention import GatedSelfAttentionDense +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion import StableDiffusionPipelineOutput +from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionGLIGENPipeline + >>> from diffusers.utils import load_image + + >>> # Insert objects described by text at the region defined by bounding boxes + >>> pipe = StableDiffusionGLIGENPipeline.from_pretrained( + ... "masterful/gligen-1-4-inpainting-text-box", variant="fp16", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> input_image = load_image( + ... "https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/gligen/livingroom_modern.png" + ... ) + >>> prompt = "a birthday cake" + >>> boxes = [[0.2676, 0.6088, 0.4773, 0.7183]] + >>> phrases = ["a birthday cake"] + + >>> images = pipe( + ... prompt=prompt, + ... gligen_phrases=phrases, + ... gligen_inpaint_image=input_image, + ... gligen_boxes=boxes, + ... gligen_scheduled_sampling_beta=1, + ... output_type="pil", + ... num_inference_steps=50, + ... ).images + + >>> images[0].save("./gligen-1-4-inpainting-text-box.jpg") + + >>> # Generate an image described by the prompt and + >>> # insert objects described by text at the region defined by bounding boxes + >>> pipe = StableDiffusionGLIGENPipeline.from_pretrained( + ... "masterful/gligen-1-4-generation-text-box", variant="fp16", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a waterfall and a modern high speed train running through the tunnel in a beautiful forest with fall foliage" + >>> boxes = [[0.1387, 0.2051, 0.4277, 0.7090], [0.4980, 0.4355, 0.8516, 0.7266]] + >>> phrases = ["a waterfall", "a modern high speed train running through the tunnel"] + + >>> images = pipe( + ... prompt=prompt, + ... gligen_phrases=phrases, + ... gligen_boxes=boxes, + ... gligen_scheduled_sampling_beta=1, + ... output_type="pil", + ... num_inference_steps=50, + ... ).images + + >>> images[0].save("./gligen-1-4-generation-text-box.jpg") + ``` +""" + + +class StableDiffusionGLIGENPipeline(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin): + r""" + Pipeline for text-to-image generation using Stable Diffusion with Grounded-Language-to-Image Generation (GLIGEN). + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + _last_supported_version = "0.33.1" + + _optional_components = ["safety_checker", "feature_extractor"] + model_cpu_offload_seq = "text_encoder->unet->vae" + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + gligen_phrases, + gligen_boxes, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if len(gligen_phrases) != len(gligen_boxes): + raise ValueError( + "length of `gligen_phrases` and `gligen_boxes` has to be same, but" + f" got: `gligen_phrases` {len(gligen_phrases)} != `gligen_boxes` {len(gligen_boxes)}" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def enable_fuser(self, enabled=True): + for module in self.unet.modules(): + if type(module) is GatedSelfAttentionDense: + module.enabled = enabled + + def draw_inpaint_mask_from_boxes(self, boxes, size): + inpaint_mask = torch.ones(size[0], size[1]) + for box in boxes: + x0, x1 = box[0] * size[0], box[2] * size[0] + y0, y1 = box[1] * size[1], box[3] * size[1] + inpaint_mask[int(y0) : int(y1), int(x0) : int(x1)] = 0 + return inpaint_mask + + def crop(self, im, new_width, new_height): + width, height = im.size + left = (width - new_width) / 2 + top = (height - new_height) / 2 + right = (width + new_width) / 2 + bottom = (height + new_height) / 2 + return im.crop((left, top, right, bottom)) + + def target_size_center_crop(self, im, new_hw): + width, height = im.size + if width != height: + im = self.crop(im, min(height, width), min(height, width)) + return im.resize((new_hw, new_hw), PIL.Image.LANCZOS) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + gligen_scheduled_sampling_beta: float = 0.3, + gligen_phrases: List[str] = None, + gligen_boxes: List[List[float]] = None, + gligen_inpaint_image: Optional[PIL.Image.Image] = None, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + clip_skip: Optional[int] = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + gligen_phrases (`List[str]`): + The phrases to guide what to include in each of the regions defined by the corresponding + `gligen_boxes`. There should only be one phrase per bounding box. + gligen_boxes (`List[List[float]]`): + The bounding boxes that identify rectangular regions of the image that are going to be filled with the + content described by the corresponding `gligen_phrases`. Each rectangular box is defined as a + `List[float]` of 4 elements `[xmin, ymin, xmax, ymax]` where each value is between [0,1]. + gligen_inpaint_image (`PIL.Image.Image`, *optional*): + The input image, if provided, is inpainted with objects described by the `gligen_boxes` and + `gligen_phrases`. Otherwise, it is treated as a generation task on a blank input image. + gligen_scheduled_sampling_beta (`float`, defaults to 0.3): + Scheduled Sampling factor from [GLIGEN: Open-Set Grounded Text-to-Image + Generation](https://huggingface.co/papers/2301.07093). Scheduled Sampling factor is only varied for + scheduled sampling during inference for improved quality and controllability. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + callback_steps, + gligen_phrases, + gligen_boxes, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5.1 Prepare GLIGEN variables + max_objs = 30 + if len(gligen_boxes) > max_objs: + warnings.warn( + f"More that {max_objs} objects found. Only first {max_objs} objects will be processed.", + FutureWarning, + ) + gligen_phrases = gligen_phrases[:max_objs] + gligen_boxes = gligen_boxes[:max_objs] + # prepare batched input to the GLIGENTextBoundingboxProjection (boxes, phrases, mask) + # Get tokens for phrases from pre-trained CLIPTokenizer + tokenizer_inputs = self.tokenizer(gligen_phrases, padding=True, return_tensors="pt").to(device) + # For the token, we use the same pre-trained text encoder + # to obtain its text feature + _text_embeddings = self.text_encoder(**tokenizer_inputs).pooler_output + n_objs = len(gligen_boxes) + # For each entity, described in phrases, is denoted with a bounding box, + # we represent the location information as (xmin,ymin,xmax,ymax) + boxes = torch.zeros(max_objs, 4, device=device, dtype=self.text_encoder.dtype) + boxes[:n_objs] = torch.tensor(gligen_boxes) + text_embeddings = torch.zeros( + max_objs, self.unet.config.cross_attention_dim, device=device, dtype=self.text_encoder.dtype + ) + text_embeddings[:n_objs] = _text_embeddings + # Generate a mask for each object that is entity described by phrases + masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype) + masks[:n_objs] = 1 + + repeat_batch = batch_size * num_images_per_prompt + boxes = boxes.unsqueeze(0).expand(repeat_batch, -1, -1).clone() + text_embeddings = text_embeddings.unsqueeze(0).expand(repeat_batch, -1, -1).clone() + masks = masks.unsqueeze(0).expand(repeat_batch, -1).clone() + if do_classifier_free_guidance: + repeat_batch = repeat_batch * 2 + boxes = torch.cat([boxes] * 2) + text_embeddings = torch.cat([text_embeddings] * 2) + masks = torch.cat([masks] * 2) + masks[: repeat_batch // 2] = 0 + if cross_attention_kwargs is None: + cross_attention_kwargs = {} + cross_attention_kwargs["gligen"] = {"boxes": boxes, "positive_embeddings": text_embeddings, "masks": masks} + + # Prepare latent variables for GLIGEN inpainting + if gligen_inpaint_image is not None: + # if the given input image is not of the same size as expected by VAE + # center crop and resize the input image to expected shape + if gligen_inpaint_image.size != (self.vae.sample_size, self.vae.sample_size): + gligen_inpaint_image = self.target_size_center_crop(gligen_inpaint_image, self.vae.sample_size) + # Convert a single image into a batch of images with a batch size of 1 + # The resulting shape becomes (1, C, H, W), where C is the number of channels, + # and H and W are the height and width of the image. + # scales the pixel values to a range [-1, 1] + gligen_inpaint_image = self.image_processor.preprocess(gligen_inpaint_image) + gligen_inpaint_image = gligen_inpaint_image.to(dtype=self.vae.dtype, device=self.vae.device) + # Run AutoEncoder to get corresponding latents + gligen_inpaint_latent = self.vae.encode(gligen_inpaint_image).latent_dist.sample() + gligen_inpaint_latent = self.vae.config.scaling_factor * gligen_inpaint_latent + # Generate an inpainting mask + # pixel value = 0, where the object is present (defined by bounding boxes above) + # 1, everywhere else + gligen_inpaint_mask = self.draw_inpaint_mask_from_boxes(gligen_boxes, gligen_inpaint_latent.shape[2:]) + gligen_inpaint_mask = gligen_inpaint_mask.to( + dtype=gligen_inpaint_latent.dtype, device=gligen_inpaint_latent.device + ) + gligen_inpaint_mask = gligen_inpaint_mask[None, None] + gligen_inpaint_mask_addition = torch.cat( + (gligen_inpaint_latent * gligen_inpaint_mask, gligen_inpaint_mask), dim=1 + ) + # Convert a single mask into a batch of masks with a batch size of 1 + gligen_inpaint_mask_addition = gligen_inpaint_mask_addition.expand(repeat_batch, -1, -1, -1).clone() + + num_grounding_steps = int(gligen_scheduled_sampling_beta * len(timesteps)) + self.enable_fuser(True) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # Scheduled sampling + if i == num_grounding_steps: + self.enable_fuser(False) + + if latents.shape[1] != 4: + latents = torch.randn_like(latents[:, :4]) + + if gligen_inpaint_image is not None: + gligen_inpaint_latent_with_noise = ( + self.scheduler.add_noise( + gligen_inpaint_latent, torch.randn_like(gligen_inpaint_latent), torch.tensor([t]) + ) + .expand(latents.shape[0], -1, -1, -1) + .clone() + ) + latents = gligen_inpaint_latent_with_noise * gligen_inpaint_mask + latents * ( + 1 - gligen_inpaint_mask + ) + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + if gligen_inpaint_image is not None: + latent_model_input = torch.cat((latent_model_input, gligen_inpaint_mask_addition), dim=1) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen_text_image.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen_text_image.py new file mode 100644 index 0000000000000000000000000000000000000000..05cbad139d923bdcb6c7e5a10ac14fb531bfcaa8 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen_text_image.py @@ -0,0 +1,1053 @@ +# Copyright 2025 The GLIGEN Authors and HuggingFace Team. All rights reserved. +# +# 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 inspect +import warnings +from typing import Any, Callable, Dict, List, Optional, Union + +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPProcessor, + CLIPTextModel, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...image_processor import VaeImageProcessor +from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.attention import GatedSelfAttentionDense +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion import StableDiffusionPipelineOutput +from ..stable_diffusion.clip_image_project_model import CLIPImageProjection +from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionGLIGENTextImagePipeline + >>> from diffusers.utils import load_image + + >>> # Insert objects described by image at the region defined by bounding boxes + >>> pipe = StableDiffusionGLIGENTextImagePipeline.from_pretrained( + ... "anhnct/Gligen_Inpainting_Text_Image", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> input_image = load_image( + ... "https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/gligen/livingroom_modern.png" + ... ) + >>> prompt = "a backpack" + >>> boxes = [[0.2676, 0.4088, 0.4773, 0.7183]] + >>> phrases = None + >>> gligen_image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/gligen/backpack.jpeg" + ... ) + + >>> images = pipe( + ... prompt=prompt, + ... gligen_phrases=phrases, + ... gligen_inpaint_image=input_image, + ... gligen_boxes=boxes, + ... gligen_images=[gligen_image], + ... gligen_scheduled_sampling_beta=1, + ... output_type="pil", + ... num_inference_steps=50, + ... ).images + + >>> images[0].save("./gligen-inpainting-text-image-box.jpg") + + >>> # Generate an image described by the prompt and + >>> # insert objects described by text and image at the region defined by bounding boxes + >>> pipe = StableDiffusionGLIGENTextImagePipeline.from_pretrained( + ... "anhnct/Gligen_Text_Image", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a flower sitting on the beach" + >>> boxes = [[0.0, 0.09, 0.53, 0.76]] + >>> phrases = ["flower"] + >>> gligen_image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/gligen/pexels-pixabay-60597.jpg" + ... ) + + >>> images = pipe( + ... prompt=prompt, + ... gligen_phrases=phrases, + ... gligen_images=[gligen_image], + ... gligen_boxes=boxes, + ... gligen_scheduled_sampling_beta=1, + ... output_type="pil", + ... num_inference_steps=50, + ... ).images + + >>> images[0].save("./gligen-generation-text-image-box.jpg") + + >>> # Generate an image described by the prompt and + >>> # transfer style described by image at the region defined by bounding boxes + >>> pipe = StableDiffusionGLIGENTextImagePipeline.from_pretrained( + ... "anhnct/Gligen_Text_Image", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a dragon flying on the sky" + >>> boxes = [[0.4, 0.2, 1.0, 0.8], [0.0, 1.0, 0.0, 1.0]] # Set `[0.0, 1.0, 0.0, 1.0]` for the style + + >>> gligen_image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/landscape.png" + ... ) + + >>> gligen_placeholder = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/landscape.png" + ... ) + + >>> images = pipe( + ... prompt=prompt, + ... gligen_phrases=[ + ... "dragon", + ... "placeholder", + ... ], # Can use any text instead of `placeholder` token, because we will use mask here + ... gligen_images=[ + ... gligen_placeholder, + ... gligen_image, + ... ], # Can use any image in gligen_placeholder, because we will use mask here + ... input_phrases_mask=[1, 0], # Set 0 for the placeholder token + ... input_images_mask=[0, 1], # Set 0 for the placeholder image + ... gligen_boxes=boxes, + ... gligen_scheduled_sampling_beta=1, + ... output_type="pil", + ... num_inference_steps=50, + ... ).images + + >>> images[0].save("./gligen-generation-text-image-box-style-transfer.jpg") + ``` +""" + + +class StableDiffusionGLIGENTextImagePipeline(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin): + r""" + Pipeline for text-to-image generation using Stable Diffusion with Grounded-Language-to-Image Generation (GLIGEN). + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + processor ([`~transformers.CLIPProcessor`]): + A `CLIPProcessor` to process reference image. + image_encoder ([`~transformers.CLIPVisionModelWithProjection`]): + Frozen image-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + image_project ([`CLIPImageProjection`]): + A `CLIPImageProjection` to project image embedding into phrases embedding space. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + _last_supported_version = "0.33.1" + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor"] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + processor: CLIPProcessor, + image_encoder: CLIPVisionModelWithProjection, + image_project: CLIPImageProjection, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + image_encoder=image_encoder, + processor=processor, + image_project=image_project, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + gligen_images, + gligen_phrases, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if gligen_images is not None and gligen_phrases is not None: + if len(gligen_images) != len(gligen_phrases): + raise ValueError( + "`gligen_images` and `gligen_phrases` must have the same length when both are provided, but" + f" got: `gligen_images` with length {len(gligen_images)} != `gligen_phrases` with length {len(gligen_phrases)}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def enable_fuser(self, enabled=True): + for module in self.unet.modules(): + if type(module) is GatedSelfAttentionDense: + module.enabled = enabled + + def draw_inpaint_mask_from_boxes(self, boxes, size): + """ + Create an inpainting mask based on given boxes. This function generates an inpainting mask using the provided + boxes to mark regions that need to be inpainted. + """ + inpaint_mask = torch.ones(size[0], size[1]) + for box in boxes: + x0, x1 = box[0] * size[0], box[2] * size[0] + y0, y1 = box[1] * size[1], box[3] * size[1] + inpaint_mask[int(y0) : int(y1), int(x0) : int(x1)] = 0 + return inpaint_mask + + def crop(self, im, new_width, new_height): + """ + Crop the input image to the specified dimensions. + """ + width, height = im.size + left = (width - new_width) / 2 + top = (height - new_height) / 2 + right = (width + new_width) / 2 + bottom = (height + new_height) / 2 + return im.crop((left, top, right, bottom)) + + def target_size_center_crop(self, im, new_hw): + """ + Crop and resize the image to the target size while keeping the center. + """ + width, height = im.size + if width != height: + im = self.crop(im, min(height, width), min(height, width)) + return im.resize((new_hw, new_hw), PIL.Image.LANCZOS) + + def complete_mask(self, has_mask, max_objs, device): + """ + Based on the input mask corresponding value `0 or 1` for each phrases and image, mask the features + corresponding to phrases and images. + """ + mask = torch.ones(1, max_objs).type(self.text_encoder.dtype).to(device) + if has_mask is None: + return mask + + if isinstance(has_mask, int): + return mask * has_mask + else: + for idx, value in enumerate(has_mask): + mask[0, idx] = value + return mask + + def get_clip_feature(self, input, normalize_constant, device, is_image=False): + """ + Get image and phrases embedding by using CLIP pretrain model. The image embedding is transformed into the + phrases embedding space through a projection. + """ + if is_image: + if input is None: + return None + inputs = self.processor(images=[input], return_tensors="pt").to(device) + inputs["pixel_values"] = inputs["pixel_values"].to(self.image_encoder.dtype) + + outputs = self.image_encoder(**inputs) + feature = outputs.image_embeds + feature = self.image_project(feature).squeeze(0) + feature = (feature / feature.norm()) * normalize_constant + feature = feature.unsqueeze(0) + else: + if input is None: + return None + inputs = self.tokenizer(input, return_tensors="pt", padding=True).to(device) + outputs = self.text_encoder(**inputs) + feature = outputs.pooler_output + return feature + + def get_cross_attention_kwargs_with_grounded( + self, + hidden_size, + gligen_phrases, + gligen_images, + gligen_boxes, + input_phrases_mask, + input_images_mask, + repeat_batch, + normalize_constant, + max_objs, + device, + ): + """ + Prepare the cross-attention kwargs containing information about the grounded input (boxes, mask, image + embedding, phrases embedding). + """ + phrases, images = gligen_phrases, gligen_images + images = [None] * len(phrases) if images is None else images + phrases = [None] * len(images) if phrases is None else phrases + + boxes = torch.zeros(max_objs, 4, device=device, dtype=self.text_encoder.dtype) + masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype) + phrases_masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype) + image_masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype) + phrases_embeddings = torch.zeros(max_objs, hidden_size, device=device, dtype=self.text_encoder.dtype) + image_embeddings = torch.zeros(max_objs, hidden_size, device=device, dtype=self.text_encoder.dtype) + + text_features = [] + image_features = [] + for phrase, image in zip(phrases, images): + text_features.append(self.get_clip_feature(phrase, normalize_constant, device, is_image=False)) + image_features.append(self.get_clip_feature(image, normalize_constant, device, is_image=True)) + + for idx, (box, text_feature, image_feature) in enumerate(zip(gligen_boxes, text_features, image_features)): + boxes[idx] = torch.tensor(box) + masks[idx] = 1 + if text_feature is not None: + phrases_embeddings[idx] = text_feature + phrases_masks[idx] = 1 + if image_feature is not None: + image_embeddings[idx] = image_feature + image_masks[idx] = 1 + + input_phrases_mask = self.complete_mask(input_phrases_mask, max_objs, device) + phrases_masks = phrases_masks.unsqueeze(0).repeat(repeat_batch, 1) * input_phrases_mask + input_images_mask = self.complete_mask(input_images_mask, max_objs, device) + image_masks = image_masks.unsqueeze(0).repeat(repeat_batch, 1) * input_images_mask + boxes = boxes.unsqueeze(0).repeat(repeat_batch, 1, 1) + masks = masks.unsqueeze(0).repeat(repeat_batch, 1) + phrases_embeddings = phrases_embeddings.unsqueeze(0).repeat(repeat_batch, 1, 1) + image_embeddings = image_embeddings.unsqueeze(0).repeat(repeat_batch, 1, 1) + + out = { + "boxes": boxes, + "masks": masks, + "phrases_masks": phrases_masks, + "image_masks": image_masks, + "phrases_embeddings": phrases_embeddings, + "image_embeddings": image_embeddings, + } + + return out + + def get_cross_attention_kwargs_without_grounded(self, hidden_size, repeat_batch, max_objs, device): + """ + Prepare the cross-attention kwargs without information about the grounded input (boxes, mask, image embedding, + phrases embedding) (All are zero tensor). + """ + boxes = torch.zeros(max_objs, 4, device=device, dtype=self.text_encoder.dtype) + masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype) + phrases_masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype) + image_masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype) + phrases_embeddings = torch.zeros(max_objs, hidden_size, device=device, dtype=self.text_encoder.dtype) + image_embeddings = torch.zeros(max_objs, hidden_size, device=device, dtype=self.text_encoder.dtype) + + out = { + "boxes": boxes.unsqueeze(0).repeat(repeat_batch, 1, 1), + "masks": masks.unsqueeze(0).repeat(repeat_batch, 1), + "phrases_masks": phrases_masks.unsqueeze(0).repeat(repeat_batch, 1), + "image_masks": image_masks.unsqueeze(0).repeat(repeat_batch, 1), + "phrases_embeddings": phrases_embeddings.unsqueeze(0).repeat(repeat_batch, 1, 1), + "image_embeddings": image_embeddings.unsqueeze(0).repeat(repeat_batch, 1, 1), + } + + return out + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + gligen_scheduled_sampling_beta: float = 0.3, + gligen_phrases: List[str] = None, + gligen_images: List[PIL.Image.Image] = None, + input_phrases_mask: Union[int, List[int]] = None, + input_images_mask: Union[int, List[int]] = None, + gligen_boxes: List[List[float]] = None, + gligen_inpaint_image: Optional[PIL.Image.Image] = None, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + gligen_normalize_constant: float = 28.7, + clip_skip: int = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + gligen_phrases (`List[str]`): + The phrases to guide what to include in each of the regions defined by the corresponding + `gligen_boxes`. There should only be one phrase per bounding box. + gligen_images (`List[PIL.Image.Image]`): + The images to guide what to include in each of the regions defined by the corresponding `gligen_boxes`. + There should only be one image per bounding box + input_phrases_mask (`int` or `List[int]`): + pre phrases mask input defined by the correspongding `input_phrases_mask` + input_images_mask (`int` or `List[int]`): + pre images mask input defined by the correspongding `input_images_mask` + gligen_boxes (`List[List[float]]`): + The bounding boxes that identify rectangular regions of the image that are going to be filled with the + content described by the corresponding `gligen_phrases`. Each rectangular box is defined as a + `List[float]` of 4 elements `[xmin, ymin, xmax, ymax]` where each value is between [0,1]. + gligen_inpaint_image (`PIL.Image.Image`, *optional*): + The input image, if provided, is inpainted with objects described by the `gligen_boxes` and + `gligen_phrases`. Otherwise, it is treated as a generation task on a blank input image. + gligen_scheduled_sampling_beta (`float`, defaults to 0.3): + Scheduled Sampling factor from [GLIGEN: Open-Set Grounded Text-to-Image + Generation](https://huggingface.co/papers/2301.07093). Scheduled Sampling factor is only varied for + scheduled sampling during inference for improved quality and controllability. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + gligen_normalize_constant (`float`, *optional*, defaults to 28.7): + The normalize value of the image embedding. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + callback_steps, + gligen_images, + gligen_phrases, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5.1 Prepare GLIGEN variables + max_objs = 30 + if len(gligen_boxes) > max_objs: + warnings.warn( + f"More that {max_objs} objects found. Only first {max_objs} objects will be processed.", + FutureWarning, + ) + gligen_phrases = gligen_phrases[:max_objs] + gligen_boxes = gligen_boxes[:max_objs] + gligen_images = gligen_images[:max_objs] + + repeat_batch = batch_size * num_images_per_prompt + + if do_classifier_free_guidance: + repeat_batch = repeat_batch * 2 + + if cross_attention_kwargs is None: + cross_attention_kwargs = {} + + hidden_size = prompt_embeds.shape[2] + + cross_attention_kwargs["gligen"] = self.get_cross_attention_kwargs_with_grounded( + hidden_size=hidden_size, + gligen_phrases=gligen_phrases, + gligen_images=gligen_images, + gligen_boxes=gligen_boxes, + input_phrases_mask=input_phrases_mask, + input_images_mask=input_images_mask, + repeat_batch=repeat_batch, + normalize_constant=gligen_normalize_constant, + max_objs=max_objs, + device=device, + ) + + cross_attention_kwargs_without_grounded = {} + cross_attention_kwargs_without_grounded["gligen"] = self.get_cross_attention_kwargs_without_grounded( + hidden_size=hidden_size, repeat_batch=repeat_batch, max_objs=max_objs, device=device + ) + + # Prepare latent variables for GLIGEN inpainting + if gligen_inpaint_image is not None: + # if the given input image is not of the same size as expected by VAE + # center crop and resize the input image to expected shape + if gligen_inpaint_image.size != (self.vae.sample_size, self.vae.sample_size): + gligen_inpaint_image = self.target_size_center_crop(gligen_inpaint_image, self.vae.sample_size) + # Convert a single image into a batch of images with a batch size of 1 + # The resulting shape becomes (1, C, H, W), where C is the number of channels, + # and H and W are the height and width of the image. + # scales the pixel values to a range [-1, 1] + gligen_inpaint_image = self.image_processor.preprocess(gligen_inpaint_image) + gligen_inpaint_image = gligen_inpaint_image.to(dtype=self.vae.dtype, device=self.vae.device) + # Run AutoEncoder to get corresponding latents + gligen_inpaint_latent = self.vae.encode(gligen_inpaint_image).latent_dist.sample() + gligen_inpaint_latent = self.vae.config.scaling_factor * gligen_inpaint_latent + # Generate an inpainting mask + # pixel value = 0, where the object is present (defined by bounding boxes above) + # 1, everywhere else + gligen_inpaint_mask = self.draw_inpaint_mask_from_boxes(gligen_boxes, gligen_inpaint_latent.shape[2:]) + gligen_inpaint_mask = gligen_inpaint_mask.to( + dtype=gligen_inpaint_latent.dtype, device=gligen_inpaint_latent.device + ) + gligen_inpaint_mask = gligen_inpaint_mask[None, None] + gligen_inpaint_mask_addition = torch.cat( + (gligen_inpaint_latent * gligen_inpaint_mask, gligen_inpaint_mask), dim=1 + ) + # Convert a single mask into a batch of masks with a batch size of 1 + gligen_inpaint_mask_addition = gligen_inpaint_mask_addition.expand(repeat_batch, -1, -1, -1).clone() + + int(gligen_scheduled_sampling_beta * len(timesteps)) + self.enable_fuser(True) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if latents.shape[1] != 4: + latents = torch.randn_like(latents[:, :4]) + + if gligen_inpaint_image is not None: + gligen_inpaint_latent_with_noise = ( + self.scheduler.add_noise( + gligen_inpaint_latent, torch.randn_like(gligen_inpaint_latent), torch.tensor([t]) + ) + .expand(latents.shape[0], -1, -1, -1) + .clone() + ) + latents = gligen_inpaint_latent_with_noise * gligen_inpaint_mask + latents * ( + 1 - gligen_inpaint_mask + ) + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + if gligen_inpaint_image is not None: + latent_model_input = torch.cat((latent_model_input, gligen_inpaint_mask_addition), dim=1) + + # predict the noise residual with grounded information + noise_pred_with_grounding = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + # predict the noise residual without grounded information + noise_pred_without_grounding = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs_without_grounded, + ).sample + + # perform guidance + if do_classifier_free_guidance: + # Using noise_pred_text from noise residual with grounded information and noise_pred_uncond from noise residual without grounded information + _, noise_pred_text = noise_pred_with_grounding.chunk(2) + noise_pred_uncond, _ = noise_pred_without_grounding.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + else: + noise_pred = noise_pred_with_grounding + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..7eb5bf8c229b584c934e33806efa73a3727dbb5b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/__init__.py @@ -0,0 +1,62 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_k_diffusion_available, + is_k_diffusion_version, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not ( + is_transformers_available() + and is_torch_available() + and is_k_diffusion_available() + and is_k_diffusion_version(">=", "0.0.12") + ): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_and_k_diffusion_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_k_diffusion_objects)) +else: + _import_structure["pipeline_stable_diffusion_k_diffusion"] = ["StableDiffusionKDiffusionPipeline"] + _import_structure["pipeline_stable_diffusion_xl_k_diffusion"] = ["StableDiffusionXLKDiffusionPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not ( + is_transformers_available() + and is_torch_available() + and is_k_diffusion_available() + and is_k_diffusion_version(">=", "0.0.12") + ): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_and_k_diffusion_objects import * + else: + from .pipeline_stable_diffusion_k_diffusion import StableDiffusionKDiffusionPipeline + from .pipeline_stable_diffusion_xl_k_diffusion import StableDiffusionXLKDiffusionPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6833e43130ffe1f894361d829fab564441aba03a Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/__pycache__/pipeline_stable_diffusion_k_diffusion.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/__pycache__/pipeline_stable_diffusion_k_diffusion.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7e331b2a032b69c412715ac5166a4672b321032c Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/__pycache__/pipeline_stable_diffusion_k_diffusion.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/__pycache__/pipeline_stable_diffusion_xl_k_diffusion.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/__pycache__/pipeline_stable_diffusion_xl_k_diffusion.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7940b1169c54a1228a800151e9d3cdf398495807 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/__pycache__/pipeline_stable_diffusion_xl_k_diffusion.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_k_diffusion.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_k_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..350a49282693868ca93fa9a8b6a4829f1a919825 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_k_diffusion.py @@ -0,0 +1,693 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 importlib +import inspect +from typing import Callable, List, Optional, Union + +import torch +from k_diffusion.external import CompVisDenoiser, CompVisVDenoiser +from k_diffusion.sampling import BrownianTreeNoiseSampler, get_sigmas_karras +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTokenizer, + CLIPTokenizerFast, +) + +from ...image_processor import VaeImageProcessor +from ...loaders import ( + StableDiffusionLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers, LMSDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + logging, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class ModelWrapper: + def __init__(self, model, alphas_cumprod): + self.model = model + self.alphas_cumprod = alphas_cumprod + + def apply_model(self, *args, **kwargs): + if len(args) == 3: + encoder_hidden_states = args[-1] + args = args[:2] + if kwargs.get("cond", None) is not None: + encoder_hidden_states = kwargs.pop("cond") + return self.model(*args, encoder_hidden_states=encoder_hidden_states, **kwargs).sample + + +class StableDiffusionKDiffusionPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + + + This is an experimental pipeline and is likely to change in the future. + + + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + details. + feature_extractor ([`CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + _last_supported_version = "0.33.1" + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor"] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: Union[CLIPTokenizer, CLIPTokenizerFast], + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + logger.info( + f"{self.__class__} is an experimental pipeline and is likely to change in the future. We recommend to use" + " this pipeline for fast experimentation / iteration if needed, but advice to rely on existing pipelines" + " as defined in https://huggingface.co/docs/diffusers/api/schedulers#implemented-schedulers for" + " production settings." + ) + + # get correct sigmas from LMS + scheduler = LMSDiscreteScheduler.from_config(scheduler.config) + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.register_to_config(requires_safety_checker=requires_safety_checker) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + model = ModelWrapper(unet, scheduler.alphas_cumprod) + if scheduler.config.prediction_type == "v_prediction": + self.k_diffusion_model = CompVisVDenoiser(model) + else: + self.k_diffusion_model = CompVisDenoiser(model) + + def set_scheduler(self, scheduler_type: str): + library = importlib.import_module("k_diffusion") + sampling = getattr(library, "sampling") + try: + self.sampler = getattr(sampling, scheduler_type) + except Exception: + valid_samplers = [] + for s in dir(sampling): + if "sample_" in s: + valid_samplers.append(s) + + raise ValueError(f"Invalid scheduler type {scheduler_type}. Please choose one of {valid_samplers}.") + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + return latents + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + use_karras_sigmas: Optional[bool] = False, + noise_sampler_seed: Optional[int] = None, + clip_skip: int = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds`. instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` + is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Use karras sigmas. For example, specifying `sample_dpmpp_2m` to `set_scheduler` will be equivalent to + `DPM++2M` in stable-diffusion-webui. On top of that, setting this option to True will make it `DPM++2M + Karras`. + noise_sampler_seed (`int`, *optional*, defaults to `None`): + The random seed to use for the noise sampler. If `None`, a random seed will be generated. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = True + if guidance_scale <= 1.0: + raise ValueError("has to use guidance_scale") + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=prompt_embeds.device) + + # 5. Prepare sigmas + if use_karras_sigmas: + sigma_min: float = self.k_diffusion_model.sigmas[0].item() + sigma_max: float = self.k_diffusion_model.sigmas[-1].item() + sigmas = get_sigmas_karras(n=num_inference_steps, sigma_min=sigma_min, sigma_max=sigma_max) + else: + sigmas = self.scheduler.sigmas + sigmas = sigmas.to(device) + sigmas = sigmas.to(prompt_embeds.dtype) + + # 6. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + latents = latents * sigmas[0] + self.k_diffusion_model.sigmas = self.k_diffusion_model.sigmas.to(latents.device) + self.k_diffusion_model.log_sigmas = self.k_diffusion_model.log_sigmas.to(latents.device) + + # 7. Define model function + def model_fn(x, t): + latent_model_input = torch.cat([x] * 2) + t = torch.cat([t] * 2) + + noise_pred = self.k_diffusion_model(latent_model_input, t, cond=prompt_embeds) + + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + return noise_pred + + # 8. Run k-diffusion solver + sampler_kwargs = {} + + if "noise_sampler" in inspect.signature(self.sampler).parameters: + min_sigma, max_sigma = sigmas[sigmas > 0].min(), sigmas.max() + noise_sampler = BrownianTreeNoiseSampler(latents, min_sigma, max_sigma, noise_sampler_seed) + sampler_kwargs["noise_sampler"] = noise_sampler + + if "generator" in inspect.signature(self.sampler).parameters: + sampler_kwargs["generator"] = generator + + latents = self.sampler(model_fn, latents, sigmas, **sampler_kwargs) + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_xl_k_diffusion.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_xl_k_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..3b57555071f350339568fb1f99e15a2e3fecc449 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_xl_k_diffusion.py @@ -0,0 +1,903 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 importlib +import inspect +from typing import List, Optional, Tuple, Union + +import torch +from k_diffusion.external import CompVisDenoiser, CompVisVDenoiser +from k_diffusion.sampling import BrownianTreeNoiseSampler, get_sigmas_karras +from transformers import ( + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, +) + +from ...image_processor import VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.attention_processor import ( + AttnProcessor2_0, + FusedAttnProcessor2_0, + XFormersAttnProcessor, +) +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers, LMSDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionXLKDiffusionPipeline + + >>> pipe = StableDiffusionXLKDiffusionPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + >>> pipe.set_scheduler("sample_dpmpp_2m_sde") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.ModelWrapper +class ModelWrapper: + def __init__(self, model, alphas_cumprod): + self.model = model + self.alphas_cumprod = alphas_cumprod + + def apply_model(self, *args, **kwargs): + if len(args) == 3: + encoder_hidden_states = args[-1] + args = args[:2] + if kwargs.get("cond", None) is not None: + encoder_hidden_states = kwargs.pop("cond") + return self.model(*args, encoder_hidden_states=encoder_hidden_states, **kwargs).sample + + +class StableDiffusionXLKDiffusionPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + FromSingleFileMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, +): + _last_supported_version = "0.33.1" + + r""" + Pipeline for text-to-image generation using Stable Diffusion XL and k-diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "feature_extractor", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + force_zeros_for_empty_prompt: bool = True, + ): + super().__init__() + + # get correct sigmas from LMS + scheduler = LMSDiscreteScheduler.from_config(scheduler.config) + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + scheduler=scheduler, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.default_sample_size = ( + self.unet.config.sample_size + if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size") + else 128 + ) + + model = ModelWrapper(unet, scheduler.alphas_cumprod) + if scheduler.config.prediction_type == "v_prediction": + self.k_diffusion_model = CompVisVDenoiser(model) + else: + self.k_diffusion_model = CompVisDenoiser(model) + + # Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.StableDiffusionKDiffusionPipeline.set_scheduler + def set_scheduler(self, scheduler_type: str): + library = importlib.import_module("k_diffusion") + sampling = getattr(library, "sampling") + try: + self.sampler = getattr(sampling, scheduler_type) + except Exception: + valid_samplers = [] + for s in dir(sampling): + if "sample_" in s: + valid_samplers.append(s) + + raise ValueError(f"Invalid scheduler type {scheduler_type}. Please choose one of {valid_samplers}.") + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + return latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.upcast_vae + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + FusedAttnProcessor2_0, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + original_size: Optional[Tuple[int, int]] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Optional[Tuple[int, int]] = None, + negative_original_size: Optional[Tuple[int, int]] = None, + negative_crops_coords_top_left: Tuple[int, int] = (0, 0), + negative_target_size: Optional[Tuple[int, int]] = None, + use_karras_sigmas: Optional[bool] = False, + noise_sampler_seed: Optional[int] = None, + clip_skip: Optional[int] = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + + # 0. Default height and width to unet + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) + + if guidance_scale <= 1.0: + raise ValueError("has to use guidance_scale") + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + lora_scale = None + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=prompt_embeds.device) + + # 5. Prepare sigmas + if use_karras_sigmas: + sigma_min: float = self.k_diffusion_model.sigmas[0].item() + sigma_max: float = self.k_diffusion_model.sigmas[-1].item() + sigmas = get_sigmas_karras(n=num_inference_steps, sigma_min=sigma_min, sigma_max=sigma_max) + else: + sigmas = self.scheduler.sigmas + sigmas = sigmas.to(dtype=prompt_embeds.dtype, device=device) + + # 6. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + latents = latents * sigmas[0] + + self.k_diffusion_model.sigmas = self.k_diffusion_model.sigmas.to(latents.device) + self.k_diffusion_model.log_sigmas = self.k_diffusion_model.log_sigmas.to(latents.device) + + # 7. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + if negative_original_size is not None and negative_target_size is not None: + negative_add_time_ids = self._get_add_time_ids( + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + else: + negative_add_time_ids = add_time_ids + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + + # 8. Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 9. Define model function + def model_fn(x, t): + latent_model_input = torch.cat([x] * 2) + t = torch.cat([t] * 2) + + noise_pred = self.k_diffusion_model( + latent_model_input, + t, + cond=prompt_embeds, + timestep_cond=timestep_cond, + added_cond_kwargs=added_cond_kwargs, + ) + + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + return noise_pred + + # 10. Run k-diffusion solver + sampler_kwargs = {} + + if "noise_sampler" in inspect.signature(self.sampler).parameters: + min_sigma, max_sigma = sigmas[sigmas > 0].min(), sigmas.max() + noise_sampler = BrownianTreeNoiseSampler(latents, min_sigma, max_sigma, noise_sampler_seed) + sampler_kwargs["noise_sampler"] = noise_sampler + + if "generator" in inspect.signature(self.sampler).parameters: + sampler_kwargs["generator"] = generator + + latents = self.sampler(model_fn, latents, sigmas, **sampler_kwargs) + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + + if not output_type == "latent": + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_ldm3d/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_ldm3d/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..dae2affddd1fd5952f454ed9cee906277dcceb16 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_ldm3d/__init__.py @@ -0,0 +1,48 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_ldm3d"] = ["StableDiffusionLDM3DPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_stable_diffusion_ldm3d import StableDiffusionLDM3DPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_ldm3d/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_ldm3d/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5d220c5f3348b5b2c295788a0638d44e07925cdd Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_ldm3d/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_ldm3d/__pycache__/pipeline_stable_diffusion_ldm3d.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_ldm3d/__pycache__/pipeline_stable_diffusion_ldm3d.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e8412b4f54660acc9f5d37348e36b1a814bd913a Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_ldm3d/__pycache__/pipeline_stable_diffusion_ldm3d.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_ldm3d/pipeline_stable_diffusion_ldm3d.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_ldm3d/pipeline_stable_diffusion_ldm3d.py new file mode 100644 index 0000000000000000000000000000000000000000..c32121c88c9bb173f567311518db2ae05d3a1da6 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_ldm3d/pipeline_stable_diffusion_ldm3d.py @@ -0,0 +1,1040 @@ +# Copyright 2025 The Intel Labs Team Authors and the HuggingFace Team. All rights reserved. +# +# 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 inspect +from dataclasses import dataclass +from typing import Any, Callable, Dict, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...image_processor import PipelineImageInput, VaeImageProcessorLDM3D +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + BaseOutput, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> from diffusers import StableDiffusionLDM3DPipeline + + >>> pipe = StableDiffusionLDM3DPipeline.from_pretrained("Intel/ldm3d-4c") + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> output = pipe(prompt) + >>> rgb_image, depth_image = output.rgb, output.depth + >>> rgb_image[0].save("astronaut_ldm3d_rgb.jpg") + >>> depth_image[0].save("astronaut_ldm3d_depth.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +@dataclass +class LDM3DPipelineOutput(BaseOutput): + """ + Output class for Stable Diffusion pipelines. + + Args: + rgb (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + depth (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + nsfw_content_detected (`List[bool]`) + List indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content or + `None` if safety checking could not be performed. + """ + + rgb: Union[List[PIL.Image.Image], np.ndarray] + depth: Union[List[PIL.Image.Image], np.ndarray] + nsfw_content_detected: Optional[List[bool]] + + +class StableDiffusionLDM3DPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, + StableDiffusionLoraLoaderMixin, + FromSingleFileMixin, +): + _last_supported_version = "0.33.1" + + r""" + Pipeline for text-to-image and 3D generation using LDM3D. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: Optional[CLIPVisionModelWithProjection], + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessorLDM3D(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + rgb_feature_extractor_input = feature_extractor_input[0] + safety_checker_input = self.feature_extractor(rgb_feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 49, + timesteps: List[int] = None, + sigmas: List[float] = None, + guidance_scale: float = 5.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 5.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Add image embeds for IP-Adapter + added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None + + # 6.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + rgb, depth = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return ((rgb, depth), has_nsfw_concept) + + return LDM3DPipelineOutput(rgb=rgb, depth=depth, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_panorama/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_panorama/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..f7572db7236cd6bcfd7dd032abcb29fd5f67cf1c --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_panorama/__init__.py @@ -0,0 +1,48 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_panorama"] = ["StableDiffusionPanoramaPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_stable_diffusion_panorama import StableDiffusionPanoramaPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_panorama/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_panorama/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..23839fd61058d62f0de8bc5a9179bc44f22b2c95 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_panorama/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_panorama/__pycache__/pipeline_stable_diffusion_panorama.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_panorama/__pycache__/pipeline_stable_diffusion_panorama.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..45be563f812fc162aa2fe5a9ef63ee918e7dcfae Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_panorama/__pycache__/pipeline_stable_diffusion_panorama.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_panorama/pipeline_stable_diffusion_panorama.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_panorama/pipeline_stable_diffusion_panorama.py new file mode 100644 index 0000000000000000000000000000000000000000..295095947a121243d5b415c1fbaa3d4b2d324ea2 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_panorama/pipeline_stable_diffusion_panorama.py @@ -0,0 +1,1195 @@ +# Copyright 2025 MultiDiffusion Authors and The HuggingFace Team. All rights reserved." +# 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 copy +import inspect +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import DDIMScheduler +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion import StableDiffusionPipelineOutput +from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionPanoramaPipeline, DDIMScheduler + + >>> model_ckpt = "stabilityai/stable-diffusion-2-base" + >>> scheduler = DDIMScheduler.from_pretrained(model_ckpt, subfolder="scheduler") + >>> pipe = StableDiffusionPanoramaPipeline.from_pretrained( + ... model_ckpt, scheduler=scheduler, torch_dtype=torch.float16 + ... ) + + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of the dolomites" + >>> image = pipe(prompt).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionPanoramaPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + IPAdapterMixin, +): + _last_supported_version = "0.33.1" + + r""" + Pipeline for text-to-image generation using MultiDiffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: DDIMScheduler, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: Optional[CLIPVisionModelWithProjection] = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def decode_latents_with_padding(self, latents: torch.Tensor, padding: int = 8) -> torch.Tensor: + """ + Decode the given latents with padding for circular inference. + + Args: + latents (torch.Tensor): The input latents to decode. + padding (int, optional): The number of latents to add on each side for padding. Defaults to 8. + + Returns: + torch.Tensor: The decoded image with padding removed. + + Notes: + - The padding is added to remove boundary artifacts and improve the output quality. + - This would slightly increase the memory usage. + - The padding pixels are then removed from the decoded image. + + """ + latents = 1 / self.vae.config.scaling_factor * latents + latents_left = latents[..., :padding] + latents_right = latents[..., -padding:] + latents = torch.cat((latents_right, latents, latents_left), axis=-1) + image = self.vae.decode(latents, return_dict=False)[0] + padding_pix = self.vae_scale_factor * padding + image = image[..., padding_pix:-padding_pix] + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + def get_views( + self, + panorama_height: int, + panorama_width: int, + window_size: int = 64, + stride: int = 8, + circular_padding: bool = False, + ) -> List[Tuple[int, int, int, int]]: + """ + Generates a list of views based on the given parameters. Here, we define the mappings F_i (see Eq. 7 in the + MultiDiffusion paper https://huggingface.co/papers/2302.08113). If panorama's height/width < window_size, + num_blocks of height/width should return 1. + + Args: + panorama_height (int): The height of the panorama. + panorama_width (int): The width of the panorama. + window_size (int, optional): The size of the window. Defaults to 64. + stride (int, optional): The stride value. Defaults to 8. + circular_padding (bool, optional): Whether to apply circular padding. Defaults to False. + + Returns: + List[Tuple[int, int, int, int]]: A list of tuples representing the views. Each tuple contains four integers + representing the start and end coordinates of the window in the panorama. + + """ + panorama_height /= 8 + panorama_width /= 8 + num_blocks_height = (panorama_height - window_size) // stride + 1 if panorama_height > window_size else 1 + if circular_padding: + num_blocks_width = panorama_width // stride if panorama_width > window_size else 1 + else: + num_blocks_width = (panorama_width - window_size) // stride + 1 if panorama_width > window_size else 1 + total_num_blocks = int(num_blocks_height * num_blocks_width) + views = [] + for i in range(total_num_blocks): + h_start = int((i // num_blocks_width) * stride) + h_end = h_start + window_size + w_start = int((i % num_blocks_width) * stride) + w_end = w_start + window_size + views.append((h_start, h_end, w_start, w_end)) + return views + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def clip_skip(self): + return self._clip_skip + + @property + def do_classifier_free_guidance(self): + return False + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + height: Optional[int] = 512, + width: Optional[int] = 2048, + num_inference_steps: int = 50, + timesteps: List[int] = None, + guidance_scale: float = 7.5, + view_batch_size: int = 1, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + circular_padding: bool = False, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs: Any, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 2048): + The width in pixels of the generated image. The width is kept high because the pipeline is supposed + generate panorama-like images. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + The timesteps at which to generate the images. If not specified, then the default timestep spacing + strategy of the scheduler is used. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + view_batch_size (`int`, *optional*, defaults to 1): + The batch size to denoise split views. For some GPUs with high performance, higher view batch size can + speedup the generation and increase the VRAM usage. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescaling factor for the guidance embeddings. A value of 0.0 means no rescaling is applied. + circular_padding (`bool`, *optional*, defaults to `False`): + If set to `True`, circular padding is applied to ensure there are no stitching artifacts. Circular + padding allows the model to seamlessly generate a transition from the rightmost part of the image to + the leftmost part, maintaining consistency in a 360-degree sense. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List[str]`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Define panorama grid and initialize views for synthesis. + # prepare batch grid + views = self.get_views(height, width, circular_padding=circular_padding) + views_batch = [views[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)] + views_scheduler_status = [copy.deepcopy(self.scheduler.__dict__)] * len(views_batch) + count = torch.zeros_like(latents) + value = torch.zeros_like(latents) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + # 7.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 8. Denoising loop + # Each denoising step also includes refinement of the latents with respect to the + # views. + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + count.zero_() + value.zero_() + + # generate views + # Here, we iterate through different spatial crops of the latents and denoise them. These + # denoised (latent) crops are then averaged to produce the final latent + # for the current timestep via MultiDiffusion. Please see Sec. 4.1 in the + # MultiDiffusion paper for more details: https://huggingface.co/papers/2302.08113 + # Batch views denoise + for j, batch_view in enumerate(views_batch): + vb_size = len(batch_view) + # get the latents corresponding to the current view coordinates + if circular_padding: + latents_for_view = [] + for h_start, h_end, w_start, w_end in batch_view: + if w_end > latents.shape[3]: + # Add circular horizontal padding + latent_view = torch.cat( + ( + latents[:, :, h_start:h_end, w_start:], + latents[:, :, h_start:h_end, : w_end - latents.shape[3]], + ), + axis=-1, + ) + else: + latent_view = latents[:, :, h_start:h_end, w_start:w_end] + latents_for_view.append(latent_view) + latents_for_view = torch.cat(latents_for_view) + else: + latents_for_view = torch.cat( + [ + latents[:, :, h_start:h_end, w_start:w_end] + for h_start, h_end, w_start, w_end in batch_view + ] + ) + + # rematch block's scheduler status + self.scheduler.__dict__.update(views_scheduler_status[j]) + + # expand the latents if we are doing classifier free guidance + latent_model_input = ( + latents_for_view.repeat_interleave(2, dim=0) + if do_classifier_free_guidance + else latents_for_view + ) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # repeat prompt_embeds for batch + prompt_embeds_input = torch.cat([prompt_embeds] * vb_size) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds_input, + timestep_cond=timestep_cond, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2] + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg( + noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale + ) + + # compute the previous noisy sample x_t -> x_t-1 + latents_denoised_batch = self.scheduler.step( + noise_pred, t, latents_for_view, **extra_step_kwargs + ).prev_sample + + # save views scheduler status after sample + views_scheduler_status[j] = copy.deepcopy(self.scheduler.__dict__) + + # extract value from batch + for latents_view_denoised, (h_start, h_end, w_start, w_end) in zip( + latents_denoised_batch.chunk(vb_size), batch_view + ): + if circular_padding and w_end > latents.shape[3]: + # Case for circular padding + value[:, :, h_start:h_end, w_start:] += latents_view_denoised[ + :, :, h_start:h_end, : latents.shape[3] - w_start + ] + value[:, :, h_start:h_end, : w_end - latents.shape[3]] += latents_view_denoised[ + :, :, h_start:h_end, latents.shape[3] - w_start : + ] + count[:, :, h_start:h_end, w_start:] += 1 + count[:, :, h_start:h_end, : w_end - latents.shape[3]] += 1 + else: + value[:, :, h_start:h_end, w_start:w_end] += latents_view_denoised + count[:, :, h_start:h_end, w_start:w_end] += 1 + + # take the MultiDiffusion step. Eq. 5 in MultiDiffusion paper: https://huggingface.co/papers/2302.08113 + latents = torch.where(count > 0, value / count, value) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type != "latent": + if circular_padding: + image = self.decode_latents_with_padding(latents) + else: + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b432b9418c46257913d81c5bf56edc0f1fa74ed1 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__init__.py @@ -0,0 +1,99 @@ +from dataclasses import dataclass +from enum import Enum +from typing import TYPE_CHECKING, List, Optional, Union + +import numpy as np +import PIL +from PIL import Image + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + BaseOutput, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +@dataclass +class SafetyConfig(object): + WEAK = { + "sld_warmup_steps": 15, + "sld_guidance_scale": 20, + "sld_threshold": 0.0, + "sld_momentum_scale": 0.0, + "sld_mom_beta": 0.0, + } + MEDIUM = { + "sld_warmup_steps": 10, + "sld_guidance_scale": 1000, + "sld_threshold": 0.01, + "sld_momentum_scale": 0.3, + "sld_mom_beta": 0.4, + } + STRONG = { + "sld_warmup_steps": 7, + "sld_guidance_scale": 2000, + "sld_threshold": 0.025, + "sld_momentum_scale": 0.5, + "sld_mom_beta": 0.7, + } + MAX = { + "sld_warmup_steps": 0, + "sld_guidance_scale": 5000, + "sld_threshold": 1.0, + "sld_momentum_scale": 0.5, + "sld_mom_beta": 0.7, + } + + +_dummy_objects = {} +_additional_imports = {} +_import_structure = {} + +_additional_imports.update({"SafetyConfig": SafetyConfig}) + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure.update( + { + "pipeline_output": ["StableDiffusionSafePipelineOutput"], + "pipeline_stable_diffusion_safe": ["StableDiffusionPipelineSafe"], + "safety_checker": ["StableDiffusionSafetyChecker"], + } + ) + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_output import StableDiffusionSafePipelineOutput + from .pipeline_stable_diffusion_safe import StableDiffusionPipelineSafe + from .safety_checker import SafeStableDiffusionSafetyChecker + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) + for name, value in _additional_imports.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..834fb9a945c4f08236b9f101cf099eeb0db6e8e1 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__pycache__/pipeline_output.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__pycache__/pipeline_output.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d1e98b19acf9ba0d781005e60f9b17300b0654d0 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__pycache__/pipeline_output.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__pycache__/pipeline_stable_diffusion_safe.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__pycache__/pipeline_stable_diffusion_safe.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9ecf423fba723c69844682357c4a53d2bdeb7213 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__pycache__/pipeline_stable_diffusion_safe.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__pycache__/safety_checker.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__pycache__/safety_checker.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1951800804dabcaf476afc4fd9439f3acea29323 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/__pycache__/safety_checker.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/pipeline_output.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..69a064d6638df556d3007f59daf7e767ec7c298b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/pipeline_output.py @@ -0,0 +1,34 @@ +from dataclasses import dataclass +from typing import List, Optional, Union + +import numpy as np +import PIL.Image + +from ...utils import ( + BaseOutput, +) + + +@dataclass +class StableDiffusionSafePipelineOutput(BaseOutput): + """ + Output class for Safe Stable Diffusion pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + nsfw_content_detected (`List[bool]`) + List of flags denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, or `None` if safety checking could not be performed. + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images that were flagged by the safety checker any may contain "not-safe-for-work" + (nsfw) content, or `None` if no safety check was performed or no images were flagged. + applied_safety_concept (`str`) + The safety concept that was applied for safety guidance, or `None` if safety guidance was disabled + """ + + images: Union[List[PIL.Image.Image], np.ndarray] + nsfw_content_detected: Optional[List[bool]] + unsafe_images: Optional[Union[List[PIL.Image.Image], np.ndarray]] + applied_safety_concept: Optional[str] diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py new file mode 100644 index 0000000000000000000000000000000000000000..d334107b07036aaaed2960cbaca887fa1584d780 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py @@ -0,0 +1,785 @@ +import inspect +import warnings +from typing import Callable, List, Optional, Union + +import numpy as np +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...configuration_utils import FrozenDict +from ...image_processor import PipelineImageInput +from ...loaders import IPAdapterMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import deprecate, is_torch_xla_available, logging +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from . import StableDiffusionSafePipelineOutput +from .safety_checker import SafeStableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class StableDiffusionPipelineSafe(DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin, IPAdapterMixin): + _last_supported_version = "0.33.1" + + r""" + Pipeline based on the [`StableDiffusionPipeline`] for text-to-image generation using Safe Latent Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: SafeStableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: Optional[CLIPVisionModelWithProjection] = None, + requires_safety_checker: bool = True, + ): + super().__init__() + safety_concept: Optional[str] = ( + "an image showing hate, harassment, violence, suffering, humiliation, harm, suicide, sexual, nudity," + " bodily fluids, blood, obscene gestures, illegal activity, drug use, theft, vandalism, weapons, child" + " abuse, brutality, cruelty" + ) + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self._safety_text_concept = safety_concept + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.register_to_config(requires_safety_checker=requires_safety_checker) + + @property + def safety_concept(self): + r""" + Getter method for the safety concept used with SLD + + Returns: + `str`: The text describing the safety concept + """ + return self._safety_text_concept + + @safety_concept.setter + def safety_concept(self, concept): + r""" + Setter method for the safety concept used with SLD + + Args: + concept (`str`): + The text of the new safety concept + """ + self._safety_text_concept = concept + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + enable_safety_guidance, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + """ + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="pt").input_ids + + if not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + prompt_embeds = self.text_encoder( + text_input_ids.to(device), + attention_mask=attention_mask, + ) + prompt_embeds = prompt_embeds[0] + + # duplicate text embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + # Encode the safety concept text + if enable_safety_guidance: + safety_concept_input = self.tokenizer( + [self._safety_text_concept], + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + safety_embeddings = self.text_encoder(safety_concept_input.input_ids.to(self.device))[0] + + # duplicate safety embeddings for each generation per prompt, using mps friendly method + seq_len = safety_embeddings.shape[1] + safety_embeddings = safety_embeddings.repeat(batch_size, num_images_per_prompt, 1) + safety_embeddings = safety_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance + sld, we need to do three forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing three forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, safety_embeddings]) + + else: + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + def run_safety_checker(self, image, device, dtype, enable_safety_guidance): + if self.safety_checker is not None: + images = image.copy() + safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + flagged_images = np.zeros((2, *image.shape[1:])) + if any(has_nsfw_concept): + logger.warning( + "Potential NSFW content was detected in one or more images. A black image will be returned" + " instead." + f"{'You may look at this images in the `unsafe_images` variable of the output at your own discretion.' if enable_safety_guidance else 'Try again with a different prompt and/or seed.'}" + ) + for idx, has_nsfw_concept in enumerate(has_nsfw_concept): + if has_nsfw_concept: + flagged_images[idx] = images[idx] + image[idx] = np.zeros(image[idx].shape) # black image + else: + has_nsfw_concept = None + flagged_images = None + return image, has_nsfw_concept, flagged_images + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.StableDiffusionKDiffusionPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def perform_safety_guidance( + self, + enable_safety_guidance, + safety_momentum, + noise_guidance, + noise_pred_out, + i, + sld_guidance_scale, + sld_warmup_steps, + sld_threshold, + sld_momentum_scale, + sld_mom_beta, + ): + # Perform SLD guidance + if enable_safety_guidance: + if safety_momentum is None: + safety_momentum = torch.zeros_like(noise_guidance) + noise_pred_text, noise_pred_uncond = noise_pred_out[0], noise_pred_out[1] + noise_pred_safety_concept = noise_pred_out[2] + + # Equation 6 + scale = torch.clamp(torch.abs((noise_pred_text - noise_pred_safety_concept)) * sld_guidance_scale, max=1.0) + + # Equation 6 + safety_concept_scale = torch.where( + (noise_pred_text - noise_pred_safety_concept) >= sld_threshold, torch.zeros_like(scale), scale + ) + + # Equation 4 + noise_guidance_safety = torch.mul((noise_pred_safety_concept - noise_pred_uncond), safety_concept_scale) + + # Equation 7 + noise_guidance_safety = noise_guidance_safety + sld_momentum_scale * safety_momentum + + # Equation 8 + safety_momentum = sld_mom_beta * safety_momentum + (1 - sld_mom_beta) * noise_guidance_safety + + if i >= sld_warmup_steps: # Warmup + # Equation 3 + noise_guidance = noise_guidance - noise_guidance_safety + return noise_guidance, safety_momentum + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]], + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + sld_guidance_scale: Optional[float] = 1000, + sld_warmup_steps: Optional[int] = 10, + sld_threshold: Optional[float] = 0.01, + sld_momentum_scale: Optional[float] = 0.3, + sld_mom_beta: Optional[float] = 0.4, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + sld_guidance_scale (`float`, *optional*, defaults to 1000): + If `sld_guidance_scale < 1`, safety guidance is disabled. + sld_warmup_steps (`int`, *optional*, defaults to 10): + Number of warmup steps for safety guidance. SLD is only be applied for diffusion steps greater than + `sld_warmup_steps`. + sld_threshold (`float`, *optional*, defaults to 0.01): + Threshold that separates the hyperplane between appropriate and inappropriate images. + sld_momentum_scale (`float`, *optional*, defaults to 0.3): + Scale of the SLD momentum to be added to the safety guidance at each diffusion step. If set to 0.0, + momentum is disabled. Momentum is built up during warmup for diffusion steps smaller than + `sld_warmup_steps`. + sld_mom_beta (`float`, *optional*, defaults to 0.4): + Defines how safety guidance momentum builds up. `sld_mom_beta` indicates how much of the previous + momentum is kept. Momentum is built up during warmup for diffusion steps smaller than + `sld_warmup_steps`. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + + Examples: + + ```py + import torch + from diffusers import StableDiffusionPipelineSafe + from diffusers.pipelines.stable_diffusion_safe import SafetyConfig + + pipeline = StableDiffusionPipelineSafe.from_pretrained( + "AIML-TUDA/stable-diffusion-safe", torch_dtype=torch.float16 + ).to("cuda") + prompt = "the four horsewomen of the apocalypse, painting by tom of finland, gaston bussiere, craig mullins, j. c. leyendecker" + image = pipeline(prompt=prompt, **SafetyConfig.MEDIUM).images[0] + ``` + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, height, width, callback_steps) + + # 2. Define call parameters + batch_size = 1 if isinstance(prompt, str) else len(prompt) + device = self._execution_device + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + enable_safety_guidance = sld_guidance_scale > 1.0 and do_classifier_free_guidance + if not enable_safety_guidance: + warnings.warn("Safety checker disabled!") + + if ip_adapter_image is not None: + output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True + image_embeds, negative_image_embeds = self.encode_image( + ip_adapter_image, device, num_images_per_prompt, output_hidden_state + ) + if do_classifier_free_guidance: + if enable_safety_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds, image_embeds]) + else: + image_embeds = torch.cat([negative_image_embeds, image_embeds]) + + # 3. Encode input prompt + prompt_embeds = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt, enable_safety_guidance + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Add image embeds for IP-Adapter + added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None + + safety_momentum = None + + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = ( + torch.cat([latents] * (3 if enable_safety_guidance else 2)) + if do_classifier_free_guidance + else latents + ) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, t, encoder_hidden_states=prompt_embeds, added_cond_kwargs=added_cond_kwargs + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_out = noise_pred.chunk((3 if enable_safety_guidance else 2)) + noise_pred_uncond, noise_pred_text = noise_pred_out[0], noise_pred_out[1] + + # default classifier free guidance + noise_guidance = noise_pred_text - noise_pred_uncond + + # Perform SLD guidance + if enable_safety_guidance: + if safety_momentum is None: + safety_momentum = torch.zeros_like(noise_guidance) + noise_pred_safety_concept = noise_pred_out[2] + + # Equation 6 + scale = torch.clamp( + torch.abs((noise_pred_text - noise_pred_safety_concept)) * sld_guidance_scale, max=1.0 + ) + + # Equation 6 + safety_concept_scale = torch.where( + (noise_pred_text - noise_pred_safety_concept) >= sld_threshold, + torch.zeros_like(scale), + scale, + ) + + # Equation 4 + noise_guidance_safety = torch.mul( + (noise_pred_safety_concept - noise_pred_uncond), safety_concept_scale + ) + + # Equation 7 + noise_guidance_safety = noise_guidance_safety + sld_momentum_scale * safety_momentum + + # Equation 8 + safety_momentum = sld_mom_beta * safety_momentum + (1 - sld_mom_beta) * noise_guidance_safety + + if i >= sld_warmup_steps: # Warmup + # Equation 3 + noise_guidance = noise_guidance - noise_guidance_safety + + noise_pred = noise_pred_uncond + guidance_scale * noise_guidance + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 8. Post-processing + image = self.decode_latents(latents) + + # 9. Run safety checker + image, has_nsfw_concept, flagged_images = self.run_safety_checker( + image, device, prompt_embeds.dtype, enable_safety_guidance + ) + + # 10. Convert to PIL + if output_type == "pil": + image = self.numpy_to_pil(image) + if flagged_images is not None: + flagged_images = self.numpy_to_pil(flagged_images) + + if not return_dict: + return ( + image, + has_nsfw_concept, + self._safety_text_concept if enable_safety_guidance else None, + flagged_images, + ) + + return StableDiffusionSafePipelineOutput( + images=image, + nsfw_content_detected=has_nsfw_concept, + applied_safety_concept=self._safety_text_concept if enable_safety_guidance else None, + unsafe_images=flagged_images, + ) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/safety_checker.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/safety_checker.py new file mode 100644 index 0000000000000000000000000000000000000000..1f6ad5f2a348c3a0aed208011e40e507fe0195e5 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_safe/safety_checker.py @@ -0,0 +1,109 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 as nn +from transformers import CLIPConfig, CLIPVisionModel, PreTrainedModel + +from ...utils import logging + + +logger = logging.get_logger(__name__) + + +def cosine_distance(image_embeds, text_embeds): + normalized_image_embeds = nn.functional.normalize(image_embeds) + normalized_text_embeds = nn.functional.normalize(text_embeds) + return torch.mm(normalized_image_embeds, normalized_text_embeds.t()) + + +class SafeStableDiffusionSafetyChecker(PreTrainedModel): + config_class = CLIPConfig + + _no_split_modules = ["CLIPEncoderLayer"] + + def __init__(self, config: CLIPConfig): + super().__init__(config) + + self.vision_model = CLIPVisionModel(config.vision_config) + self.visual_projection = nn.Linear(config.vision_config.hidden_size, config.projection_dim, bias=False) + + self.concept_embeds = nn.Parameter(torch.ones(17, config.projection_dim), requires_grad=False) + self.special_care_embeds = nn.Parameter(torch.ones(3, config.projection_dim), requires_grad=False) + + self.concept_embeds_weights = nn.Parameter(torch.ones(17), requires_grad=False) + self.special_care_embeds_weights = nn.Parameter(torch.ones(3), requires_grad=False) + + @torch.no_grad() + def forward(self, clip_input, images): + pooled_output = self.vision_model(clip_input)[1] # pooled_output + image_embeds = self.visual_projection(pooled_output) + + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds).cpu().float().numpy() + cos_dist = cosine_distance(image_embeds, self.concept_embeds).cpu().float().numpy() + + result = [] + batch_size = image_embeds.shape[0] + for i in range(batch_size): + result_img = {"special_scores": {}, "special_care": [], "concept_scores": {}, "bad_concepts": []} + + # increase this value to create a stronger `nfsw` filter + # at the cost of increasing the possibility of filtering benign images + adjustment = 0.0 + + for concept_idx in range(len(special_cos_dist[0])): + concept_cos = special_cos_dist[i][concept_idx] + concept_threshold = self.special_care_embeds_weights[concept_idx].item() + result_img["special_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3) + if result_img["special_scores"][concept_idx] > 0: + result_img["special_care"].append({concept_idx, result_img["special_scores"][concept_idx]}) + adjustment = 0.01 + + for concept_idx in range(len(cos_dist[0])): + concept_cos = cos_dist[i][concept_idx] + concept_threshold = self.concept_embeds_weights[concept_idx].item() + result_img["concept_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3) + if result_img["concept_scores"][concept_idx] > 0: + result_img["bad_concepts"].append(concept_idx) + + result.append(result_img) + + has_nsfw_concepts = [len(res["bad_concepts"]) > 0 for res in result] + + return images, has_nsfw_concepts + + @torch.no_grad() + def forward_onnx(self, clip_input: torch.Tensor, images: torch.Tensor): + pooled_output = self.vision_model(clip_input)[1] # pooled_output + image_embeds = self.visual_projection(pooled_output) + + special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds) + cos_dist = cosine_distance(image_embeds, self.concept_embeds) + + # increase this value to create a stronger `nsfw` filter + # at the cost of increasing the possibility of filtering benign images + adjustment = 0.0 + + special_scores = special_cos_dist - self.special_care_embeds_weights + adjustment + # special_scores = special_scores.round(decimals=3) + special_care = torch.any(special_scores > 0, dim=1) + special_adjustment = special_care * 0.01 + special_adjustment = special_adjustment.unsqueeze(1).expand(-1, cos_dist.shape[1]) + + concept_scores = (cos_dist - self.concept_embeds_weights) + special_adjustment + # concept_scores = concept_scores.round(decimals=3) + has_nsfw_concepts = torch.any(concept_scores > 0, dim=1) + + return images, has_nsfw_concepts diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_sag/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_sag/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..378e0e57817f58a0a28afed5d6110f6ee3effb3a --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_sag/__init__.py @@ -0,0 +1,48 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_sag"] = ["StableDiffusionSAGPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_stable_diffusion_sag import StableDiffusionSAGPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_sag/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_sag/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c89d88646427dd4b9f80ee5015d1aef2bc0f397a Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_sag/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_sag/__pycache__/pipeline_stable_diffusion_sag.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_sag/__pycache__/pipeline_stable_diffusion_sag.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bbfd1dbce53ff467f8a27cc2ae291bdb3c72b3a6 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_sag/__pycache__/pipeline_stable_diffusion_sag.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_sag/pipeline_stable_diffusion_sag.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_sag/pipeline_stable_diffusion_sag.py new file mode 100644 index 0000000000000000000000000000000000000000..48add535a81da083f65b2fb9843ee72bf5577819 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_sag/pipeline_stable_diffusion_sag.py @@ -0,0 +1,970 @@ +# Copyright 2025 Susung Hong and The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import torch +import torch.nn.functional as F +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion import StableDiffusionPipelineOutput +from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionSAGPipeline + + >>> pipe = StableDiffusionSAGPipeline.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt, sag_scale=0.75).images[0] + ``` +""" + + +# processes and stores attention probabilities +class CrossAttnStoreProcessor: + def __init__(self): + self.attention_probs = None + + def __call__( + self, + attn, + hidden_states, + encoder_hidden_states=None, + attention_mask=None, + ): + batch_size, sequence_length, _ = hidden_states.shape + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + query = attn.to_q(hidden_states) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + query = attn.head_to_batch_dim(query) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + self.attention_probs = attn.get_attention_scores(query, key, attention_mask) + hidden_states = torch.bmm(self.attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + return hidden_states + + +# Modified to get self-attention guidance scale in this paper (https://huggingface.co/papers/2210.00939) as an input +class StableDiffusionSAGPipeline( + DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, IPAdapterMixin +): + _last_supported_version = "0.33.1" + + r""" + Pipeline for text-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: Optional[CLIPVisionModelWithProjection] = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + image_embeds = [] + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0) + single_negative_image_embeds = torch.stack( + [single_negative_image_embeds] * num_images_per_prompt, dim=0 + ) + + if do_classifier_free_guidance: + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds]) + single_image_embeds = single_image_embeds.to(device) + + image_embeds.append(single_image_embeds) + else: + image_embeds = ip_adapter_image_embeds + return image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.StableDiffusionKDiffusionPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + sag_scale: float = 0.75, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: Optional[int] = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + clip_skip: Optional[int] = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + sag_scale (`float`, *optional*, defaults to 0.75): + Chosen between [0, 1.0] for better quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. If not provided, embeddings are computed from the + `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + # and `sag_scale` is` `s` of equation (16) + # of the self-attention guidance paper: https://huggingface.co/papers/2210.00939 + # `sag_scale = 0` means no self-attention guidance + do_self_attention_guidance = sag_scale > 0.0 + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + do_classifier_free_guidance, + ) + + if do_classifier_free_guidance: + image_embeds = [] + negative_image_embeds = [] + for tmp_image_embeds in ip_adapter_image_embeds: + single_negative_image_embeds, single_image_embeds = tmp_image_embeds.chunk(2) + image_embeds.append(single_image_embeds) + negative_image_embeds.append(single_negative_image_embeds) + else: + image_embeds = ip_adapter_image_embeds + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + if timesteps.dtype not in [torch.int16, torch.int32, torch.int64]: + raise ValueError( + f"{self.__class__.__name__} does not support using a scheduler of type {self.scheduler.__class__.__name__}. Please make sure to use one of 'DDIMScheduler, PNDMScheduler, DDPMScheduler, DEISMultistepScheduler, UniPCMultistepScheduler, DPMSolverMultistepScheduler, DPMSolverSinglestepScheduler'." + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + if do_classifier_free_guidance: + added_uncond_kwargs = ( + {"image_embeds": negative_image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + # 7. Denoising loop + original_attn_proc = self.unet.attn_processors + store_processor = CrossAttnStoreProcessor() + self.unet.mid_block.attentions[0].transformer_blocks[0].attn1.processor = store_processor + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + + map_size = None + + def get_map_size(module, input, output): + nonlocal map_size + map_size = output[0].shape[-2:] + + with self.unet.mid_block.attentions[0].register_forward_hook(get_map_size): + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # perform self-attention guidance with the stored self-attention map + if do_self_attention_guidance: + # classifier-free guidance produces two chunks of attention map + # and we only use unconditional one according to equation (25) + # in https://huggingface.co/papers/2210.00939 + if do_classifier_free_guidance: + # DDIM-like prediction of x0 + pred_x0 = self.pred_x0(latents, noise_pred_uncond, t) + # get the stored attention maps + uncond_attn, cond_attn = store_processor.attention_probs.chunk(2) + # self-attention-based degrading of latents + degraded_latents = self.sag_masking( + pred_x0, uncond_attn, map_size, t, self.pred_epsilon(latents, noise_pred_uncond, t) + ) + uncond_emb, _ = prompt_embeds.chunk(2) + # forward and give guidance + degraded_pred = self.unet( + degraded_latents, + t, + encoder_hidden_states=uncond_emb, + added_cond_kwargs=added_uncond_kwargs, + ).sample + noise_pred += sag_scale * (noise_pred_uncond - degraded_pred) + else: + # DDIM-like prediction of x0 + pred_x0 = self.pred_x0(latents, noise_pred, t) + # get the stored attention maps + cond_attn = store_processor.attention_probs + # self-attention-based degrading of latents + degraded_latents = self.sag_masking( + pred_x0, cond_attn, map_size, t, self.pred_epsilon(latents, noise_pred, t) + ) + # forward and give guidance + degraded_pred = self.unet( + degraded_latents, + t, + encoder_hidden_states=prompt_embeds, + added_cond_kwargs=added_cond_kwargs, + ).sample + noise_pred += sag_scale * (noise_pred - degraded_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + self.maybe_free_model_hooks() + # make sure to set the original attention processors back + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + def sag_masking(self, original_latents, attn_map, map_size, t, eps): + # Same masking process as in SAG paper: https://huggingface.co/papers/2210.00939 + bh, hw1, hw2 = attn_map.shape + b, latent_channel, latent_h, latent_w = original_latents.shape + h = self.unet.config.attention_head_dim + if isinstance(h, list): + h = h[-1] + + # Produce attention mask + attn_map = attn_map.reshape(b, h, hw1, hw2) + attn_mask = attn_map.mean(1, keepdim=False).sum(1, keepdim=False) > 1.0 + attn_mask = ( + attn_mask.reshape(b, map_size[0], map_size[1]) + .unsqueeze(1) + .repeat(1, latent_channel, 1, 1) + .type(attn_map.dtype) + ) + attn_mask = F.interpolate(attn_mask, (latent_h, latent_w)) + + # Blur according to the self-attention mask + degraded_latents = gaussian_blur_2d(original_latents, kernel_size=9, sigma=1.0) + degraded_latents = degraded_latents * attn_mask + original_latents * (1 - attn_mask) + + # Noise it again to match the noise level + degraded_latents = self.scheduler.add_noise(degraded_latents, noise=eps, timesteps=t[None]) + + return degraded_latents + + # Modified from diffusers.schedulers.scheduling_ddim.DDIMScheduler.step + # Note: there are some schedulers that clip or do not return x_0 (PNDMScheduler, DDIMScheduler, etc.) + def pred_x0(self, sample, model_output, timestep): + alpha_prod_t = self.scheduler.alphas_cumprod[timestep].to(sample.device) + + beta_prod_t = 1 - alpha_prod_t + if self.scheduler.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + elif self.scheduler.config.prediction_type == "sample": + pred_original_sample = model_output + elif self.scheduler.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + # predict V + model_output = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.scheduler.config.prediction_type} must be one of `epsilon`, `sample`," + " or `v_prediction`" + ) + + return pred_original_sample + + def pred_epsilon(self, sample, model_output, timestep): + alpha_prod_t = self.scheduler.alphas_cumprod[timestep] + + beta_prod_t = 1 - alpha_prod_t + if self.scheduler.config.prediction_type == "epsilon": + pred_eps = model_output + elif self.scheduler.config.prediction_type == "sample": + pred_eps = (sample - (alpha_prod_t**0.5) * model_output) / (beta_prod_t**0.5) + elif self.scheduler.config.prediction_type == "v_prediction": + pred_eps = (beta_prod_t**0.5) * sample + (alpha_prod_t**0.5) * model_output + else: + raise ValueError( + f"prediction_type given as {self.scheduler.config.prediction_type} must be one of `epsilon`, `sample`," + " or `v_prediction`" + ) + + return pred_eps + + +# Gaussian blur +def gaussian_blur_2d(img, kernel_size, sigma): + ksize_half = (kernel_size - 1) * 0.5 + + x = torch.linspace(-ksize_half, ksize_half, steps=kernel_size) + + pdf = torch.exp(-0.5 * (x / sigma).pow(2)) + + x_kernel = pdf / pdf.sum() + x_kernel = x_kernel.to(device=img.device, dtype=img.dtype) + + kernel2d = torch.mm(x_kernel[:, None], x_kernel[None, :]) + kernel2d = kernel2d.expand(img.shape[-3], 1, kernel2d.shape[0], kernel2d.shape[1]) + + padding = [kernel_size // 2, kernel_size // 2, kernel_size // 2, kernel_size // 2] + + img = F.pad(img, padding, mode="reflect") + img = F.conv2d(img, kernel2d, groups=img.shape[-3]) + + return img diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..8088fbcfceba205b9b908613f4ca3fdc579120e8 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__init__.py @@ -0,0 +1,76 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_flax_available, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_additional_imports = {} +_import_structure = {"pipeline_output": ["StableDiffusionXLPipelineOutput"]} + +if is_transformers_available() and is_flax_available(): + _import_structure["pipeline_output"].extend(["FlaxStableDiffusionXLPipelineOutput"]) +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_xl"] = ["StableDiffusionXLPipeline"] + _import_structure["pipeline_stable_diffusion_xl_img2img"] = ["StableDiffusionXLImg2ImgPipeline"] + _import_structure["pipeline_stable_diffusion_xl_inpaint"] = ["StableDiffusionXLInpaintPipeline"] + _import_structure["pipeline_stable_diffusion_xl_instruct_pix2pix"] = ["StableDiffusionXLInstructPix2PixPipeline"] + +if is_transformers_available() and is_flax_available(): + from ...schedulers.scheduling_pndm_flax import PNDMSchedulerState + + _additional_imports.update({"PNDMSchedulerState": PNDMSchedulerState}) + _import_structure["pipeline_flax_stable_diffusion_xl"] = ["FlaxStableDiffusionXLPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_stable_diffusion_xl import StableDiffusionXLPipeline + from .pipeline_stable_diffusion_xl_img2img import StableDiffusionXLImg2ImgPipeline + from .pipeline_stable_diffusion_xl_inpaint import StableDiffusionXLInpaintPipeline + from .pipeline_stable_diffusion_xl_instruct_pix2pix import StableDiffusionXLInstructPix2PixPipeline + + try: + if not (is_transformers_available() and is_flax_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_flax_objects import * + else: + from .pipeline_flax_stable_diffusion_xl import ( + FlaxStableDiffusionXLPipeline, + ) + from .pipeline_output import FlaxStableDiffusionXLPipelineOutput + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) + for name, value in _additional_imports.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b6fcf27f9e718da6f1a5b1bfbaf130cb8357dca3 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_flax_stable_diffusion_xl.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_flax_stable_diffusion_xl.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f455f6f4f2d10a11d48c4dc9c2c22bda9825502b Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_flax_stable_diffusion_xl.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_output.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_output.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e5dc8857d8eb0c9ea9884a3e121d52a64a7065e0 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_output.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6d105cc8677fbc82f01dc0999da852aad223e944 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_img2img.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_img2img.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a4877c1d526a993ac01b36e6015637577cd9e56d Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_img2img.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_inpaint.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_inpaint.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f9da4ec6e4db2516677f5ddc97ff06f8cf5e0ea2 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_inpaint.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_instruct_pix2pix.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_instruct_pix2pix.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c9d5074192e2f129802e04718f4f8f0a80d89256 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_instruct_pix2pix.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/watermark.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/watermark.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5dd201b2ed520d291c10bb8226cc59925d442720 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/__pycache__/watermark.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_flax_stable_diffusion_xl.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_flax_stable_diffusion_xl.py new file mode 100644 index 0000000000000000000000000000000000000000..3227fd9a08a47306eeb0804f301d16cdd1f4ce5f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_flax_stable_diffusion_xl.py @@ -0,0 +1,308 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 partial +from typing import Dict, List, Optional, Union + +import jax +import jax.numpy as jnp +from flax.core.frozen_dict import FrozenDict +from transformers import CLIPTokenizer, FlaxCLIPTextModel + +from diffusers.utils import logging + +from ...models import FlaxAutoencoderKL, FlaxUNet2DConditionModel +from ...schedulers import ( + FlaxDDIMScheduler, + FlaxDPMSolverMultistepScheduler, + FlaxLMSDiscreteScheduler, + FlaxPNDMScheduler, +) +from ..pipeline_flax_utils import FlaxDiffusionPipeline +from .pipeline_output import FlaxStableDiffusionXLPipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +# Set to True to use python for loop instead of jax.fori_loop for easier debugging +DEBUG = False + + +class FlaxStableDiffusionXLPipeline(FlaxDiffusionPipeline): + def __init__( + self, + text_encoder: FlaxCLIPTextModel, + text_encoder_2: FlaxCLIPTextModel, + vae: FlaxAutoencoderKL, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: FlaxUNet2DConditionModel, + scheduler: Union[ + FlaxDDIMScheduler, FlaxPNDMScheduler, FlaxLMSDiscreteScheduler, FlaxDPMSolverMultistepScheduler + ], + dtype: jnp.dtype = jnp.float32, + ): + super().__init__() + self.dtype = dtype + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + + def prepare_inputs(self, prompt: Union[str, List[str]]): + if not isinstance(prompt, (str, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + # Assume we have the two encoders + inputs = [] + for tokenizer in [self.tokenizer, self.tokenizer_2]: + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="np", + ) + inputs.append(text_inputs.input_ids) + inputs = jnp.stack(inputs, axis=1) + return inputs + + def __call__( + self, + prompt_ids: jax.Array, + params: Union[Dict, FrozenDict], + prng_seed: jax.Array, + num_inference_steps: int = 50, + guidance_scale: Union[float, jax.Array] = 7.5, + height: Optional[int] = None, + width: Optional[int] = None, + latents: jnp.array = None, + neg_prompt_ids: jnp.array = None, + return_dict: bool = True, + output_type: str = None, + jit: bool = False, + ): + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + if isinstance(guidance_scale, float) and jit: + # Convert to a tensor so each device gets a copy. + guidance_scale = jnp.array([guidance_scale] * prompt_ids.shape[0]) + guidance_scale = guidance_scale[:, None] + + return_latents = output_type == "latent" + + if jit: + images = _p_generate( + self, + prompt_ids, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + neg_prompt_ids, + return_latents, + ) + else: + images = self._generate( + prompt_ids, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + neg_prompt_ids, + return_latents, + ) + + if not return_dict: + return (images,) + + return FlaxStableDiffusionXLPipelineOutput(images=images) + + def get_embeddings(self, prompt_ids: jnp.array, params): + # We assume we have the two encoders + + # bs, encoder_input, seq_length + te_1_inputs = prompt_ids[:, 0, :] + te_2_inputs = prompt_ids[:, 1, :] + + prompt_embeds = self.text_encoder(te_1_inputs, params=params["text_encoder"], output_hidden_states=True) + prompt_embeds = prompt_embeds["hidden_states"][-2] + prompt_embeds_2_out = self.text_encoder_2( + te_2_inputs, params=params["text_encoder_2"], output_hidden_states=True + ) + prompt_embeds_2 = prompt_embeds_2_out["hidden_states"][-2] + text_embeds = prompt_embeds_2_out["text_embeds"] + prompt_embeds = jnp.concatenate([prompt_embeds, prompt_embeds_2], axis=-1) + return prompt_embeds, text_embeds + + def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, bs, dtype): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + add_time_ids = jnp.array([add_time_ids] * bs, dtype=dtype) + return add_time_ids + + def _generate( + self, + prompt_ids: jnp.array, + params: Union[Dict, FrozenDict], + prng_seed: jax.Array, + num_inference_steps: int, + height: int, + width: int, + guidance_scale: float, + latents: Optional[jnp.array] = None, + neg_prompt_ids: Optional[jnp.array] = None, + return_latents=False, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + # Encode input prompt + prompt_embeds, pooled_embeds = self.get_embeddings(prompt_ids, params) + + # Get unconditional embeddings + batch_size = prompt_embeds.shape[0] + if neg_prompt_ids is None: + neg_prompt_embeds = jnp.zeros_like(prompt_embeds) + negative_pooled_embeds = jnp.zeros_like(pooled_embeds) + else: + neg_prompt_embeds, negative_pooled_embeds = self.get_embeddings(neg_prompt_ids, params) + + add_time_ids = self._get_add_time_ids( + (height, width), (0, 0), (height, width), prompt_embeds.shape[0], dtype=prompt_embeds.dtype + ) + + prompt_embeds = jnp.concatenate([neg_prompt_embeds, prompt_embeds], axis=0) # (2, 77, 2048) + add_text_embeds = jnp.concatenate([negative_pooled_embeds, pooled_embeds], axis=0) + add_time_ids = jnp.concatenate([add_time_ids, add_time_ids], axis=0) + + # Ensure model output will be `float32` before going into the scheduler + guidance_scale = jnp.array([guidance_scale], dtype=jnp.float32) + + # Create random latents + latents_shape = ( + batch_size, + self.unet.config.in_channels, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + if latents is None: + latents = jax.random.normal(prng_seed, shape=latents_shape, dtype=jnp.float32) + else: + if latents.shape != latents_shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}") + + # Prepare scheduler state + scheduler_state = self.scheduler.set_timesteps( + params["scheduler"], num_inference_steps=num_inference_steps, shape=latents.shape + ) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * scheduler_state.init_noise_sigma + + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + + # Denoising loop + def loop_body(step, args): + latents, scheduler_state = args + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + latents_input = jnp.concatenate([latents] * 2) + + t = jnp.array(scheduler_state.timesteps, dtype=jnp.int32)[step] + timestep = jnp.broadcast_to(t, latents_input.shape[0]) + + latents_input = self.scheduler.scale_model_input(scheduler_state, latents_input, t) + + # predict the noise residual + noise_pred = self.unet.apply( + {"params": params["unet"]}, + jnp.array(latents_input), + jnp.array(timestep, dtype=jnp.int32), + encoder_hidden_states=prompt_embeds, + added_cond_kwargs=added_cond_kwargs, + ).sample + # perform guidance + noise_pred_uncond, noise_prediction_text = jnp.split(noise_pred, 2, axis=0) + noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents, scheduler_state = self.scheduler.step(scheduler_state, noise_pred, t, latents).to_tuple() + return latents, scheduler_state + + if DEBUG: + # run with python for loop + for i in range(num_inference_steps): + latents, scheduler_state = loop_body(i, (latents, scheduler_state)) + else: + latents, _ = jax.lax.fori_loop(0, num_inference_steps, loop_body, (latents, scheduler_state)) + + if return_latents: + return latents + + # Decode latents + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.apply({"params": params["vae"]}, latents, method=self.vae.decode).sample + + image = (image / 2 + 0.5).clip(0, 1).transpose(0, 2, 3, 1) + return image + + +# Static argnums are pipe, num_inference_steps, height, width, return_latents. A change would trigger recompilation. +# Non-static args are (sharded) input tensors mapped over their first dimension (hence, `0`). +@partial( + jax.pmap, + in_axes=(None, 0, 0, 0, None, None, None, 0, 0, 0, None), + static_broadcasted_argnums=(0, 4, 5, 6, 10), +) +def _p_generate( + pipe, + prompt_ids, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + neg_prompt_ids, + return_latents, +): + return pipe._generate( + prompt_ids, + params, + prng_seed, + num_inference_steps, + height, + width, + guidance_scale, + latents, + neg_prompt_ids, + return_latents, + ) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_output.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..0783f44486ee1448bd15529f745af381ee7fa69f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_output.py @@ -0,0 +1,37 @@ +from dataclasses import dataclass +from typing import List, Union + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput, is_flax_available + + +@dataclass +class StableDiffusionXLPipelineOutput(BaseOutput): + """ + Output class for Stable Diffusion pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] + + +if is_flax_available(): + import flax + + @flax.struct.dataclass + class FlaxStableDiffusionXLPipelineOutput(BaseOutput): + """ + Output class for Flax Stable Diffusion XL pipelines. + + Args: + images (`np.ndarray`) + Array of shape `(batch_size, height, width, num_channels)` with images from the diffusion pipeline. + """ + + images: np.ndarray diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py new file mode 100644 index 0000000000000000000000000000000000000000..9ac64a0d842023a4b9b600039b9d27f6b2fec583 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py @@ -0,0 +1,1313 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.attention_processor import ( + AttnProcessor2_0, + FusedAttnProcessor2_0, + XFormersAttnProcessor, +) +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from .pipeline_output import StableDiffusionXLPipelineOutput + + +if is_invisible_watermark_available(): + from .watermark import StableDiffusionXLWatermarker + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionXLPipeline + + >>> pipe = StableDiffusionXLPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLPipeline( + DiffusionPipeline, + StableDiffusionMixin, + FromSingleFileMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "add_text_embeds", + "add_time_ids", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: Optional[bool] = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.default_sample_size = ( + self.unet.config.sample_size + if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size") + else 128 + ) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + FusedAttnProcessor2_0, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + timesteps: List[int] = None, + sigmas: List[float] = None, + denoising_end: Optional[float] = None, + guidance_scale: float = 5.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + original_size: Optional[Tuple[int, int]] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Optional[Tuple[int, int]] = None, + negative_original_size: Optional[Tuple[int, int]] = None, + negative_crops_coords_top_left: Tuple[int, int] = (0, 0), + negative_target_size: Optional[Tuple[int, int]] = None, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default height and width to unet + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + if negative_original_size is not None and negative_target_size is not None: + negative_add_time_ids = self._get_add_time_ids( + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + else: + negative_add_time_ids = add_time_ids + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 8.1 Apply denoising_end + if ( + self.denoising_end is not None + and isinstance(self.denoising_end, float) + and self.denoising_end > 0 + and self.denoising_end < 1 + ): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 9. Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + + if not output_type == "latent": + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..e63c7a55ce7b76ca2f5b3cffcc254f3e8feee829 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py @@ -0,0 +1,1503 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.attention_processor import ( + AttnProcessor2_0, + XFormersAttnProcessor, +) +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import empty_device_cache, randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from .pipeline_output import StableDiffusionXLPipelineOutput + + +if is_invisible_watermark_available(): + from .watermark import StableDiffusionXLWatermarker + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionXLImg2ImgPipeline + >>> from diffusers.utils import load_image + + >>> pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-xl-refiner-1.0", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + >>> url = "https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/aa_xl/000000009.png" + + >>> init_image = load_image(url).convert("RGB") + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt, image=init_image).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLImg2ImgPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + FromSingleFileMixin, + StableDiffusionXLLoraLoaderMixin, + IPAdapterMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`): + Whether the `unet` requires an `aesthetic_score` condition to be passed during inference. Also see the + config of `stabilityai/stable-diffusion-xl-refiner-1-0`. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "add_text_embeds", + "add_time_ids", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + requires_aesthetics_score: bool = False, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: Optional[bool] = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + scheduler=scheduler, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.register_to_config(requires_aesthetics_score=requires_aesthetics_score) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + prompt_2, + strength, + num_inference_steps, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + if num_inference_steps is None: + raise ValueError("`num_inference_steps` cannot be None.") + elif not isinstance(num_inference_steps, int) or num_inference_steps <= 0: + raise ValueError( + f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type" + f" {type(num_inference_steps)}." + ) + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None): + # get the original timestep using init_timestep + if denoising_start is None: + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + t_start = max(num_inference_steps - init_timestep, 0) + + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + else: + # Strength is irrelevant if we directly request a timestep to start at; + # that is, strength is determined by the denoising_start instead. + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_start * self.scheduler.config.num_train_timesteps) + ) + ) + + num_inference_steps = (self.scheduler.timesteps < discrete_timestep_cutoff).sum().item() + if self.scheduler.order == 2 and num_inference_steps % 2 == 0: + # if the scheduler is a 2nd order scheduler we might have to do +1 + # because `num_inference_steps` might be even given that every timestep + # (except the highest one) is duplicated. If `num_inference_steps` is even it would + # mean that we cut the timesteps in the middle of the denoising step + # (between 1st and 2nd derivative) which leads to incorrect results. By adding 1 + # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler + num_inference_steps = num_inference_steps + 1 + + # because t_n+1 >= t_n, we slice the timesteps starting from the end + t_start = len(self.scheduler.timesteps) - num_inference_steps + timesteps = self.scheduler.timesteps[t_start:] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start) + return timesteps, num_inference_steps + + def prepare_latents( + self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None, add_noise=True + ): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + latents_mean = latents_std = None + if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None: + latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1) + if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None: + latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1) + + # Offload text encoder if `enable_model_cpu_offload` was enabled + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.text_encoder_2.to("cpu") + empty_device_cache() + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + # make sure the VAE is in float32 mode, as it overflows in float16 + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + if image.shape[0] < batch_size and batch_size % image.shape[0] == 0: + image = torch.cat([image] * (batch_size // image.shape[0]), dim=0) + elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} " + ) + + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + if self.vae.config.force_upcast: + self.vae.to(dtype) + + init_latents = init_latents.to(dtype) + if latents_mean is not None and latents_std is not None: + latents_mean = latents_mean.to(device=device, dtype=dtype) + latents_std = latents_std.to(device=device, dtype=dtype) + init_latents = (init_latents - latents_mean) * self.vae.config.scaling_factor / latents_std + else: + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + if add_noise: + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + latents = init_latents + + return latents + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + def _get_add_time_ids( + self, + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype, + text_encoder_projection_dim=None, + ): + if self.config.requires_aesthetics_score: + add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,)) + add_neg_time_ids = list( + negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,) + ) + else: + add_time_ids = list(original_size + crops_coords_top_left + target_size) + add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if ( + expected_add_embed_dim > passed_add_embed_dim + and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model." + ) + elif ( + expected_add_embed_dim < passed_add_embed_dim + and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model." + ) + elif expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype) + + return add_time_ids, add_neg_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def denoising_start(self): + return self._denoising_start + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + image: PipelineImageInput = None, + strength: float = 0.3, + num_inference_steps: int = 50, + timesteps: List[int] = None, + sigmas: List[float] = None, + denoising_start: Optional[float] = None, + denoising_end: Optional[float] = None, + guidance_scale: float = 5.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + original_size: Tuple[int, int] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Tuple[int, int] = None, + negative_original_size: Optional[Tuple[int, int]] = None, + negative_crops_coords_top_left: Tuple[int, int] = (0, 0), + negative_target_size: Optional[Tuple[int, int]] = None, + aesthetic_score: float = 6.0, + negative_aesthetic_score: float = 2.5, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + image (`torch.Tensor` or `PIL.Image.Image` or `np.ndarray` or `List[torch.Tensor]` or `List[PIL.Image.Image]` or `List[np.ndarray]`): + The image(s) to modify with the pipeline. + strength (`float`, *optional*, defaults to 0.3): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image` + will be used as a starting point, adding more noise to it the larger the `strength`. The number of + denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will + be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. Note that in the case of + `denoising_start` being declared as an integer, the value of `strength` will be ignored. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_start (`float`, *optional*): + When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be + bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and + it is assumed that the passed `image` is a partly denoised image. Note that when this is specified, + strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline + is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refine Image + Quality**](https://huggingface.co/docs/diffusers/using-diffusers/sdxl#refine-image-quality). + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be + denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the + final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline + forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refine Image + Quality**](https://huggingface.co/docs/diffusers/using-diffusers/sdxl#refine-image-quality). + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + aesthetic_score (`float`, *optional*, defaults to 6.0): + Used to simulate an aesthetic score of the generated image by influencing the positive text condition. + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_aesthetic_score (`float`, *optional*, defaults to 2.5): + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to + simulate an aesthetic score of the generated image by influencing the negative text condition. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple. When returning a tuple, the first element is a list with the generated images. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + strength, + num_inference_steps, + callback_steps, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._denoising_start = denoising_start + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. Preprocess image + image = self.image_processor.preprocess(image) + + # 5. Prepare timesteps + def denoising_value_valid(dnv): + return isinstance(dnv, float) and 0 < dnv < 1 + + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps, + strength, + device, + denoising_start=self.denoising_start if denoising_value_valid(self.denoising_start) else None, + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + add_noise = True if self.denoising_start is None else False + + # 6. Prepare latent variables + if latents is None: + latents = self.prepare_latents( + image, + latent_timestep, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + add_noise, + ) + # 7. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + height, width = latents.shape[-2:] + height = height * self.vae_scale_factor + width = width * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 8. Prepare added time ids & embeddings + if negative_original_size is None: + negative_original_size = original_size + if negative_target_size is None: + negative_target_size = target_size + + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids, add_neg_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1) + add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 9. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 9.1 Apply denoising_end + if ( + self.denoising_end is not None + and self.denoising_start is not None + and denoising_value_valid(self.denoising_end) + and denoising_value_valid(self.denoising_start) + and self.denoising_start >= self.denoising_end + ): + raise ValueError( + f"`denoising_start`: {self.denoising_start} cannot be larger than or equal to `denoising_end`: " + + f" {self.denoising_end} when using type float." + ) + elif self.denoising_end is not None and denoising_value_valid(self.denoising_end): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 9.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..f0bc9b9bb3e2f623c5d45dc22798c3d685d9e347 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py @@ -0,0 +1,1740 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import numpy as np +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.attention_processor import ( + AttnProcessor2_0, + XFormersAttnProcessor, +) +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from .pipeline_output import StableDiffusionXLPipelineOutput + + +if is_invisible_watermark_available(): + from .watermark import StableDiffusionXLWatermarker + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionXLInpaintPipeline + >>> from diffusers.utils import load_image + + >>> pipe = StableDiffusionXLInpaintPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", + ... torch_dtype=torch.float16, + ... variant="fp16", + ... use_safetensors=True, + ... ) + >>> pipe.to("cuda") + + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + + >>> init_image = load_image(img_url).convert("RGB") + >>> mask_image = load_image(mask_url).convert("RGB") + + >>> prompt = "A majestic tiger sitting on a bench" + >>> image = pipe( + ... prompt=prompt, image=init_image, mask_image=mask_image, num_inference_steps=50, strength=0.80 + ... ).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +def mask_pil_to_torch(mask, height, width): + # preprocess mask + if isinstance(mask, (PIL.Image.Image, np.ndarray)): + mask = [mask] + + if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image): + mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask] + mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0) + mask = mask.astype(np.float32) / 255.0 + elif isinstance(mask, list) and isinstance(mask[0], np.ndarray): + mask = np.concatenate([m[None, None, :] for m in mask], axis=0) + + mask = torch.from_numpy(mask) + return mask + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLInpaintPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionXLLoraLoaderMixin, + FromSingleFileMixin, + IPAdapterMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`): + Whether the `unet` requires a aesthetic_score condition to be passed during inference. Also see the config + of `stabilityai/stable-diffusion-xl-refiner-1-0`. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "add_text_embeds", + "add_time_ids", + "mask", + "masked_image_latents", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + requires_aesthetics_score: bool = False, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: Optional[bool] = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + scheduler=scheduler, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.register_to_config(requires_aesthetics_score=requires_aesthetics_score) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True + ) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + prompt_2, + image, + mask_image, + height, + width, + strength, + callback_steps, + output_type, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + image=None, + timestep=None, + is_strength_max=True, + add_noise=True, + return_noise=False, + return_image_latents=False, + ): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if (image is None or timestep is None) and not is_strength_max: + raise ValueError( + "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise." + "However, either the image or the noise timestep has not been provided." + ) + + if image.shape[1] == 4: + image_latents = image.to(device=device, dtype=dtype) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + elif return_image_latents or (latents is None and not is_strength_max): + image = image.to(device=device, dtype=dtype) + image_latents = self._encode_vae_image(image=image, generator=generator) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + + if latents is None and add_noise: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # if strength is 1. then initialise the latents to noise, else initial to image + noise + latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep) + # if pure noise then scale the initial latents by the Scheduler's init sigma + latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents + elif add_noise: + noise = latents.to(device) + latents = noise * self.scheduler.init_noise_sigma + else: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = image_latents.to(device) + + outputs = (latents,) + + if return_noise: + outputs += (noise,) + + if return_image_latents: + outputs += (image_latents,) + + return outputs + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + dtype = image.dtype + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + if self.vae.config.force_upcast: + self.vae.to(dtype) + + image_latents = image_latents.to(dtype) + image_latents = self.vae.config.scaling_factor * image_latents + + return image_latents + + def prepare_mask_latents( + self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance + ): + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate( + mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor) + ) + mask = mask.to(device=device, dtype=dtype) + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + + mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask + + if masked_image is not None and masked_image.shape[1] == 4: + masked_image_latents = masked_image + else: + masked_image_latents = None + + if masked_image is not None: + if masked_image_latents is None: + masked_image = masked_image.to(device=device, dtype=dtype) + masked_image_latents = self._encode_vae_image(masked_image, generator=generator) + + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat( + batch_size // masked_image_latents.shape[0], 1, 1, 1 + ) + + masked_image_latents = ( + torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + ) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + + return mask, masked_image_latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None): + # get the original timestep using init_timestep + if denoising_start is None: + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + t_start = max(num_inference_steps - init_timestep, 0) + + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + else: + # Strength is irrelevant if we directly request a timestep to start at; + # that is, strength is determined by the denoising_start instead. + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_start * self.scheduler.config.num_train_timesteps) + ) + ) + + num_inference_steps = (self.scheduler.timesteps < discrete_timestep_cutoff).sum().item() + if self.scheduler.order == 2 and num_inference_steps % 2 == 0: + # if the scheduler is a 2nd order scheduler we might have to do +1 + # because `num_inference_steps` might be even given that every timestep + # (except the highest one) is duplicated. If `num_inference_steps` is even it would + # mean that we cut the timesteps in the middle of the denoising step + # (between 1st and 2nd derivative) which leads to incorrect results. By adding 1 + # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler + num_inference_steps = num_inference_steps + 1 + + # because t_n+1 >= t_n, we slice the timesteps starting from the end + t_start = len(self.scheduler.timesteps) - num_inference_steps + timesteps = self.scheduler.timesteps[t_start:] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start) + return timesteps, num_inference_steps + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline._get_add_time_ids + def _get_add_time_ids( + self, + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype, + text_encoder_projection_dim=None, + ): + if self.config.requires_aesthetics_score: + add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,)) + add_neg_time_ids = list( + negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,) + ) + else: + add_time_ids = list(original_size + crops_coords_top_left + target_size) + add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if ( + expected_add_embed_dim > passed_add_embed_dim + and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model." + ) + elif ( + expected_add_embed_dim < passed_add_embed_dim + and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model." + ) + elif expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype) + + return add_time_ids, add_neg_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def denoising_start(self): + return self._denoising_start + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: torch.Tensor = None, + height: Optional[int] = None, + width: Optional[int] = None, + padding_mask_crop: Optional[int] = None, + strength: float = 0.9999, + num_inference_steps: int = 50, + timesteps: List[int] = None, + sigmas: List[float] = None, + denoising_start: Optional[float] = None, + denoising_end: Optional[float] = None, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + original_size: Tuple[int, int] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Tuple[int, int] = None, + negative_original_size: Optional[Tuple[int, int]] = None, + negative_crops_coords_top_left: Tuple[int, int] = (0, 0), + negative_target_size: Optional[Tuple[int, int]] = None, + aesthetic_score: float = 6.0, + negative_aesthetic_score: float = 2.5, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will + be masked out with `mask_image` and repainted according to `prompt`. + mask_image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted + to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L) + instead of 3, so the expected shape would be `(B, H, W, 1)`. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 0.9999): + Conceptually, indicates how much to transform the masked portion of the reference `image`. Must be + between 0 and 1. `image` will be used as a starting point, adding more noise to it the larger the + `strength`. The number of denoising steps depends on the amount of noise initially added. When + `strength` is 1, added noise will be maximum and the denoising process will run for the full number of + iterations specified in `num_inference_steps`. A value of 1, therefore, essentially ignores the masked + portion of the reference `image`. Note that in the case of `denoising_start` being declared as an + integer, the value of `strength` will be ignored. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_start (`float`, *optional*): + When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be + bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and + it is assumed that the passed `image` is a partly denoised image. Note that when this is specified, + strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline + is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output). + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be + denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the + final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline + forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output). + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + aesthetic_score (`float`, *optional*, defaults to 6.0): + Used to simulate an aesthetic score of the generated image by influencing the positive text condition. + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_aesthetic_score (`float`, *optional*, defaults to 2.5): + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to + simulate an aesthetic score of the generated image by influencing the negative text condition. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple. `tuple. When returning a tuple, the first element is a list with the generated images. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs + self.check_inputs( + prompt, + prompt_2, + image, + mask_image, + height, + width, + strength, + callback_steps, + output_type, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + padding_mask_crop, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._denoising_start = denoising_start + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. set timesteps + def denoising_value_valid(dnv): + return isinstance(dnv, float) and 0 < dnv < 1 + + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps, + strength, + device, + denoising_start=self.denoising_start if denoising_value_valid(self.denoising_start) else None, + ) + # check that number of inference steps is not < 1 - as this doesn't make sense + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + # at which timestep to set the initial noise (n.b. 50% if strength is 0.5) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise + is_strength_max = strength == 1.0 + + # 5. Preprocess mask and image + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + mask = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + if masked_image_latents is not None: + masked_image = masked_image_latents + elif init_image.shape[1] == 4: + # if images are in latent space, we can't mask it + masked_image = None + else: + masked_image = init_image * (mask < 0.5) + + # 6. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + num_channels_unet = self.unet.config.in_channels + return_image_latents = num_channels_unet == 4 + + add_noise = True if self.denoising_start is None else False + latents_outputs = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + image=init_image, + timestep=latent_timestep, + is_strength_max=is_strength_max, + add_noise=add_noise, + return_noise=True, + return_image_latents=return_image_latents, + ) + + if return_image_latents: + latents, noise, image_latents = latents_outputs + else: + latents, noise = latents_outputs + + # 7. Prepare mask latent variables + mask, masked_image_latents = self.prepare_mask_latents( + mask, + masked_image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + self.do_classifier_free_guidance, + ) + + # 8. Check that sizes of mask, masked image and latents match + if num_channels_unet == 9: + # default case for runwayml/stable-diffusion-inpainting + num_channels_mask = mask.shape[1] + num_channels_masked_image = masked_image_latents.shape[1] + if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of" + " `pipeline.unet` or your `mask_image` or `image` input." + ) + elif num_channels_unet != 4: + raise ValueError( + f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}." + ) + # 8.1 Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + height, width = latents.shape[-2:] + height = height * self.vae_scale_factor + width = width * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 10. Prepare added time ids & embeddings + if negative_original_size is None: + negative_original_size = original_size + if negative_target_size is None: + negative_target_size = target_size + + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids, add_neg_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1) + add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 11. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + if ( + self.denoising_end is not None + and self.denoising_start is not None + and denoising_value_valid(self.denoising_end) + and denoising_value_valid(self.denoising_start) + and self.denoising_start >= self.denoising_end + ): + raise ValueError( + f"`denoising_start`: {self.denoising_start} cannot be larger than or equal to `denoising_end`: " + + f" {self.denoising_end} when using type float." + ) + elif self.denoising_end is not None and denoising_value_valid(self.denoising_end): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 11.1 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + # concat latents, mask, masked_image_latents in the channel dimension + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + if num_channels_unet == 9: + latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if num_channels_unet == 4: + init_latents_proper = image_latents + if self.do_classifier_free_guidance: + init_mask, _ = mask.chunk(2) + else: + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.add_noise( + init_latents_proper, noise, torch.tensor([noise_timestep]) + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + mask = callback_outputs.pop("mask", mask) + masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + return StableDiffusionXLPipelineOutput(images=latents) + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + if padding_mask_crop is not None: + image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image] + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py new file mode 100644 index 0000000000000000000000000000000000000000..b1379d1b2955e1f2d67794dd3c3b6c8ec23848ff --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py @@ -0,0 +1,1000 @@ +# Copyright 2025 Harutatsu Akiyama and The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import PIL.Image +import torch +from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.attention_processor import ( + AttnProcessor2_0, + FusedAttnProcessor2_0, + XFormersAttnProcessor, +) +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from .pipeline_output import StableDiffusionXLPipelineOutput + + +if is_invisible_watermark_available(): + from .watermark import StableDiffusionXLWatermarker + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionXLInstructPix2PixPipeline + >>> from diffusers.utils import load_image + + >>> resolution = 768 + >>> image = load_image( + ... "https://hf.co/datasets/diffusers/diffusers-images-docs/resolve/main/mountain.png" + ... ).resize((resolution, resolution)) + >>> edit_instruction = "Turn sky into a cloudy one" + + >>> pipe = StableDiffusionXLInstructPix2PixPipeline.from_pretrained( + ... "diffusers/sdxl-instructpix2pix-768", torch_dtype=torch.float16 + ... ).to("cuda") + + >>> edited_image = pipe( + ... prompt=edit_instruction, + ... image=image, + ... height=resolution, + ... width=resolution, + ... guidance_scale=3.0, + ... image_guidance_scale=1.5, + ... num_inference_steps=30, + ... ).images[0] + >>> edited_image + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + """ + Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891). See Section 3.4 + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class StableDiffusionXLInstructPix2PixPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + FromSingleFileMixin, + StableDiffusionXLLoraLoaderMixin, +): + r""" + Pipeline for pixel-level image editing by following text instructions. Based on Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`): + Whether the `unet` requires a aesthetic_score condition to be passed during inference. Also see the config + of `stabilityai/stable-diffusion-xl-refiner-1-0`. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + is_cosxl_edit (`bool`, *optional*): + When set the image latents are scaled. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae" + _optional_components = ["tokenizer", "tokenizer_2", "text_encoder", "text_encoder_2"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: Optional[bool] = None, + is_cosxl_edit: Optional[bool] = False, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + scheduler=scheduler, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.default_sample_size = ( + self.unet.config.sample_size + if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size") + else 128 + ) + self.is_cosxl_edit = is_cosxl_edit + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder( + text_input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + prompt_embeds = prompt_embeds.hidden_states[-2] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt, negative_prompt_2] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + prompt_embeds_dtype = self.text_encoder_2.dtype if self.text_encoder_2 is not None else self.unet.dtype + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def prepare_image_latents( + self, image, batch_size, num_images_per_prompt, dtype, device, do_classifier_free_guidance, generator=None + ): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + image_latents = image + else: + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + if needs_upcasting: + image = image.float() + self.upcast_vae() + + image_latents = retrieve_latents(self.vae.encode(image), sample_mode="argmax") + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand image_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {image_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + if do_classifier_free_guidance: + uncond_image_latents = torch.zeros_like(image_latents) + image_latents = torch.cat([image_latents, image_latents, uncond_image_latents], dim=0) + + if image_latents.dtype != self.vae.dtype: + image_latents = image_latents.to(dtype=self.vae.dtype) + + if self.is_cosxl_edit: + image_latents = image_latents * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.upcast_vae + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + FusedAttnProcessor2_0, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + image: PipelineImageInput = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 100, + denoising_end: Optional[float] = None, + guidance_scale: float = 5.0, + image_guidance_scale: float = 1.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + original_size: Tuple[int, int] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Tuple[int, int] = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + image (`torch.Tensor` or `PIL.Image.Image` or `np.ndarray` or `List[torch.Tensor]` or `List[PIL.Image.Image]` or `List[np.ndarray]`): + The image(s) to modify with the pipeline. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + image_guidance_scale (`float`, *optional*, defaults to 1.5): + Image guidance scale is to push the generated image towards the initial image `image`. Image guidance + scale is enabled by setting `image_guidance_scale > 1`. Higher image guidance scale encourages to + generate images that are closely linked to the source image `image`, usually at the expense of lower + image quality. This pipeline requires a value of at least `1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + aesthetic_score (`float`, *optional*, defaults to 6.0): + Used to simulate an aesthetic score of the generated image by influencing the positive text condition. + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_aesthetic_score (`float`, *optional*, defaults to 2.5): + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to + simulate an aesthetic score of the generated image by influencing the negative text condition. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + # 0. Default height and width to unet + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds) + + if image is None: + raise ValueError("`image` input cannot be undefined.") + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 and image_guidance_scale >= 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + ) + + # 4. Preprocess image + image = self.image_processor.preprocess(image, height=height, width=width).to(device) + + # 5. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 6. Prepare Image latents + image_latents = self.prepare_image_latents( + image, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + do_classifier_free_guidance, + ) + + # 7. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 8. Check that shapes of latents and image match the UNet channels + num_channels_image = image_latents.shape[1] + if num_channels_latents + num_channels_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_image`: {num_channels_image} " + f" = {num_channels_latents + num_channels_image}. Please verify the config of" + " `pipeline.unet` or your `image` input." + ) + + # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 10. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + + if do_classifier_free_guidance: + # The extra concat similar to how it's done in SD InstructPix2Pix. + prompt_embeds = torch.cat([prompt_embeds, negative_prompt_embeds, negative_prompt_embeds], dim=0) + add_text_embeds = torch.cat( + [add_text_embeds, negative_pooled_prompt_embeds, negative_pooled_prompt_embeds], dim=0 + ) + add_time_ids = torch.cat([add_time_ids, add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + # 11. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1: + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # Expand the latents if we are doing classifier free guidance. + # The latents are expanded 3 times because for pix2pix the guidance + # is applied for both the text and the input image. + latent_model_input = torch.cat([latents] * 3) if do_classifier_free_guidance else latents + + # concat latents, image_latents in the channel dimension + scaled_latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + scaled_latent_model_input = torch.cat([scaled_latent_model_input, image_latents], dim=1) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + noise_pred = self.unet( + scaled_latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_text, noise_pred_image, noise_pred_uncond = noise_pred.chunk(3) + noise_pred = ( + noise_pred_uncond + + guidance_scale * (noise_pred_text - noise_pred_image) + + image_guidance_scale * (noise_pred_image - noise_pred_uncond) + ) + + if do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + return StableDiffusionXLPipelineOutput(images=latents) + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/watermark.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/watermark.py new file mode 100644 index 0000000000000000000000000000000000000000..70d06bb6320d890ea90542cc7551bc42d7c59451 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_diffusion_xl/watermark.py @@ -0,0 +1,42 @@ +import numpy as np +import torch + +from ...utils import is_invisible_watermark_available + + +if is_invisible_watermark_available(): + from imwatermark import WatermarkEncoder + + +# Copied from https://github.com/Stability-AI/generative-models/blob/613af104c6b85184091d42d374fef420eddb356d/scripts/demo/streamlit_helpers.py#L66 +WATERMARK_MESSAGE = 0b101100111110110010010000011110111011000110011110 +# bin(x)[2:] gives bits of x as str, use int to convert them to 0/1 +WATERMARK_BITS = [int(bit) for bit in bin(WATERMARK_MESSAGE)[2:]] + + +class StableDiffusionXLWatermarker: + def __init__(self): + self.watermark = WATERMARK_BITS + self.encoder = WatermarkEncoder() + + self.encoder.set_watermark("bits", self.watermark) + + def apply_watermark(self, images: torch.Tensor): + # can't encode images that are smaller than 256 + if images.shape[-1] < 256: + return images + + images = (255 * (images / 2 + 0.5)).cpu().permute(0, 2, 3, 1).float().numpy() + + # Convert RGB to BGR, which is the channel order expected by the watermark encoder. + images = images[:, :, :, ::-1] + + # Add watermark and convert BGR back to RGB + images = [self.encoder.encode(image, "dwtDct")[:, :, ::-1] for image in images] + + images = np.array(images) + + images = torch.from_numpy(images).permute(0, 3, 1, 2) + + images = torch.clamp(2 * (images / 255 - 0.5), min=-1.0, max=1.0) + return images diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_video_diffusion/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_video_diffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..3bd4dc78966e217d85769691b98ed8fb0b6ac05c --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_video_diffusion/__init__.py @@ -0,0 +1,58 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + BaseOutput, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure.update( + { + "pipeline_stable_video_diffusion": [ + "StableVideoDiffusionPipeline", + "StableVideoDiffusionPipelineOutput", + ], + } + ) + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_stable_video_diffusion import ( + StableVideoDiffusionPipeline, + StableVideoDiffusionPipelineOutput, + ) + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_video_diffusion/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_video_diffusion/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4922a2f3b8789c96cdca9dbe753177dd945e3ba3 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_video_diffusion/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_video_diffusion/__pycache__/pipeline_stable_video_diffusion.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_video_diffusion/__pycache__/pipeline_stable_video_diffusion.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6b6e0eaf5c1f660781ef596a61261f2db2fa7370 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_video_diffusion/__pycache__/pipeline_stable_video_diffusion.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_video_diffusion/pipeline_stable_video_diffusion.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_video_diffusion/pipeline_stable_video_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..6d9053faaec8a8af0fce4b941b25667473a5e4a8 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/stable_video_diffusion/pipeline_stable_video_diffusion.py @@ -0,0 +1,737 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from dataclasses import dataclass +from typing import Callable, Dict, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection + +from ...image_processor import PipelineImageInput +from ...models import AutoencoderKLTemporalDecoder, UNetSpatioTemporalConditionModel +from ...schedulers import EulerDiscreteScheduler +from ...utils import BaseOutput, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import is_compiled_module, randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import StableVideoDiffusionPipeline + >>> from diffusers.utils import load_image, export_to_video + + >>> pipe = StableVideoDiffusionPipeline.from_pretrained( + ... "stabilityai/stable-video-diffusion-img2vid-xt", torch_dtype=torch.float16, variant="fp16" + ... ) + >>> pipe.to("cuda") + + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd-docstring-example.jpeg" + ... ) + >>> image = image.resize((1024, 576)) + + >>> frames = pipe(image, num_frames=25, decode_chunk_size=8).frames[0] + >>> export_to_video(frames, "generated.mp4", fps=7) + ``` +""" + + +def _append_dims(x, target_dims): + """Appends dimensions to the end of a tensor until it has target_dims dimensions.""" + dims_to_append = target_dims - x.ndim + if dims_to_append < 0: + raise ValueError(f"input has {x.ndim} dims but target_dims is {target_dims}, which is less") + return x[(...,) + (None,) * dims_to_append] + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +@dataclass +class StableVideoDiffusionPipelineOutput(BaseOutput): + r""" + Output class for Stable Video Diffusion pipeline. + + Args: + frames (`[List[List[PIL.Image.Image]]`, `np.ndarray`, `torch.Tensor`]): + List of denoised PIL images of length `batch_size` or numpy array or torch tensor of shape `(batch_size, + num_frames, height, width, num_channels)`. + """ + + frames: Union[List[List[PIL.Image.Image]], np.ndarray, torch.Tensor] + + +class StableVideoDiffusionPipeline(DiffusionPipeline): + r""" + Pipeline to generate video from an input image using Stable Video Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKLTemporalDecoder`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + image_encoder ([`~transformers.CLIPVisionModelWithProjection`]): + Frozen CLIP image-encoder + ([laion/CLIP-ViT-H-14-laion2B-s32B-b79K](https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K)). + unet ([`UNetSpatioTemporalConditionModel`]): + A `UNetSpatioTemporalConditionModel` to denoise the encoded image latents. + scheduler ([`EulerDiscreteScheduler`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images. + """ + + model_cpu_offload_seq = "image_encoder->unet->vae" + _callback_tensor_inputs = ["latents"] + + def __init__( + self, + vae: AutoencoderKLTemporalDecoder, + image_encoder: CLIPVisionModelWithProjection, + unet: UNetSpatioTemporalConditionModel, + scheduler: EulerDiscreteScheduler, + feature_extractor: CLIPImageProcessor, + ): + super().__init__() + + self.register_modules( + vae=vae, + image_encoder=image_encoder, + unet=unet, + scheduler=scheduler, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(do_resize=True, vae_scale_factor=self.vae_scale_factor) + + def _encode_image( + self, + image: PipelineImageInput, + device: Union[str, torch.device], + num_videos_per_prompt: int, + do_classifier_free_guidance: bool, + ) -> torch.Tensor: + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.video_processor.pil_to_numpy(image) + image = self.video_processor.numpy_to_pt(image) + + # We normalize the image before resizing to match with the original implementation. + # Then we unnormalize it after resizing. + image = image * 2.0 - 1.0 + image = _resize_with_antialiasing(image, (224, 224)) + image = (image + 1.0) / 2.0 + + # Normalize the image with for CLIP input + image = self.feature_extractor( + images=image, + do_normalize=True, + do_center_crop=False, + do_resize=False, + do_rescale=False, + return_tensors="pt", + ).pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeddings = self.image_encoder(image).image_embeds + image_embeddings = image_embeddings.unsqueeze(1) + + # duplicate image embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = image_embeddings.shape + image_embeddings = image_embeddings.repeat(1, num_videos_per_prompt, 1) + image_embeddings = image_embeddings.view(bs_embed * num_videos_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + negative_image_embeddings = torch.zeros_like(image_embeddings) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + image_embeddings = torch.cat([negative_image_embeddings, image_embeddings]) + + return image_embeddings + + def _encode_vae_image( + self, + image: torch.Tensor, + device: Union[str, torch.device], + num_videos_per_prompt: int, + do_classifier_free_guidance: bool, + ): + image = image.to(device=device) + image_latents = self.vae.encode(image).latent_dist.mode() + + # duplicate image_latents for each generation per prompt, using mps friendly method + image_latents = image_latents.repeat(num_videos_per_prompt, 1, 1, 1) + + if do_classifier_free_guidance: + negative_image_latents = torch.zeros_like(image_latents) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + image_latents = torch.cat([negative_image_latents, image_latents]) + + return image_latents + + def _get_add_time_ids( + self, + fps: int, + motion_bucket_id: int, + noise_aug_strength: float, + dtype: torch.dtype, + batch_size: int, + num_videos_per_prompt: int, + do_classifier_free_guidance: bool, + ): + add_time_ids = [fps, motion_bucket_id, noise_aug_strength] + + passed_add_embed_dim = self.unet.config.addition_time_embed_dim * len(add_time_ids) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + add_time_ids = add_time_ids.repeat(batch_size * num_videos_per_prompt, 1) + + if do_classifier_free_guidance: + add_time_ids = torch.cat([add_time_ids, add_time_ids]) + + return add_time_ids + + def decode_latents(self, latents: torch.Tensor, num_frames: int, decode_chunk_size: int = 14): + # [batch, frames, channels, height, width] -> [batch*frames, channels, height, width] + latents = latents.flatten(0, 1) + + latents = 1 / self.vae.config.scaling_factor * latents + + forward_vae_fn = self.vae._orig_mod.forward if is_compiled_module(self.vae) else self.vae.forward + accepts_num_frames = "num_frames" in set(inspect.signature(forward_vae_fn).parameters.keys()) + + # decode decode_chunk_size frames at a time to avoid OOM + frames = [] + for i in range(0, latents.shape[0], decode_chunk_size): + num_frames_in = latents[i : i + decode_chunk_size].shape[0] + decode_kwargs = {} + if accepts_num_frames: + # we only pass num_frames_in if it's expected + decode_kwargs["num_frames"] = num_frames_in + + frame = self.vae.decode(latents[i : i + decode_chunk_size], **decode_kwargs).sample + frames.append(frame) + frames = torch.cat(frames, dim=0) + + # [batch*frames, channels, height, width] -> [batch, channels, frames, height, width] + frames = frames.reshape(-1, num_frames, *frames.shape[1:]).permute(0, 2, 1, 3, 4) + + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + frames = frames.float() + return frames + + def check_inputs(self, image, height, width): + if ( + not isinstance(image, torch.Tensor) + and not isinstance(image, PIL.Image.Image) + and not isinstance(image, list) + ): + raise ValueError( + "`image` has to be of type `torch.Tensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is" + f" {type(image)}" + ) + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + def prepare_latents( + self, + batch_size: int, + num_frames: int, + num_channels_latents: int, + height: int, + width: int, + dtype: torch.dtype, + device: Union[str, torch.device], + generator: torch.Generator, + latents: Optional[torch.Tensor] = None, + ): + shape = ( + batch_size, + num_frames, + num_channels_latents // 2, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + if isinstance(self.guidance_scale, (int, float)): + return self.guidance_scale > 1 + return self.guidance_scale.max() > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: Union[PIL.Image.Image, List[PIL.Image.Image], torch.Tensor], + height: int = 576, + width: int = 1024, + num_frames: Optional[int] = None, + num_inference_steps: int = 25, + sigmas: Optional[List[float]] = None, + min_guidance_scale: float = 1.0, + max_guidance_scale: float = 3.0, + fps: int = 7, + motion_bucket_id: int = 127, + noise_aug_strength: float = 0.02, + decode_chunk_size: Optional[int] = None, + num_videos_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + return_dict: bool = True, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.Tensor`): + Image(s) to guide image generation. If you provide a tensor, the expected value range is between `[0, + 1]`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_frames (`int`, *optional*): + The number of video frames to generate. Defaults to `self.unet.config.num_frames` (14 for + `stable-video-diffusion-img2vid` and to 25 for `stable-video-diffusion-img2vid-xt`). + num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps. More denoising steps usually lead to a higher quality video at the + expense of slower inference. This parameter is modulated by `strength`. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + min_guidance_scale (`float`, *optional*, defaults to 1.0): + The minimum guidance scale. Used for the classifier free guidance with first frame. + max_guidance_scale (`float`, *optional*, defaults to 3.0): + The maximum guidance scale. Used for the classifier free guidance with last frame. + fps (`int`, *optional*, defaults to 7): + Frames per second. The rate at which the generated images shall be exported to a video after + generation. Note that Stable Diffusion Video's UNet was micro-conditioned on fps-1 during training. + motion_bucket_id (`int`, *optional*, defaults to 127): + Used for conditioning the amount of motion for the generation. The higher the number the more motion + will be in the video. + noise_aug_strength (`float`, *optional*, defaults to 0.02): + The amount of noise added to the init image, the higher it is the less the video will look like the + init image. Increase it for more motion. + decode_chunk_size (`int`, *optional*): + The number of frames to decode at a time. Higher chunk size leads to better temporal consistency at the + expense of more memory usage. By default, the decoder decodes all frames at once for maximal quality. + For lower memory usage, reduce `decode_chunk_size`. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `pil`, `np` or `pt`. + callback_on_step_end (`Callable`, *optional*): + A function that is called at the end of each denoising step during inference. The function is called + with the following arguments: + `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. + `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableVideoDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableVideoDiffusionPipelineOutput`] is + returned, otherwise a `tuple` of (`List[List[PIL.Image.Image]]` or `np.ndarray` or `torch.Tensor`) is + returned. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + num_frames = num_frames if num_frames is not None else self.unet.config.num_frames + decode_chunk_size = decode_chunk_size if decode_chunk_size is not None else num_frames + + # 1. Check inputs. Raise error if not correct + self.check_inputs(image, height, width) + + # 2. Define call parameters + if isinstance(image, PIL.Image.Image): + batch_size = 1 + elif isinstance(image, list): + batch_size = len(image) + else: + batch_size = image.shape[0] + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + self._guidance_scale = max_guidance_scale + + # 3. Encode input image + image_embeddings = self._encode_image(image, device, num_videos_per_prompt, self.do_classifier_free_guidance) + + # NOTE: Stable Video Diffusion was conditioned on fps - 1, which is why it is reduced here. + # See: https://github.com/Stability-AI/generative-models/blob/ed0997173f98eaf8f4edf7ba5fe8f15c6b877fd3/scripts/sampling/simple_video_sample.py#L188 + fps = fps - 1 + + # 4. Encode input image using VAE + image = self.video_processor.preprocess(image, height=height, width=width).to(device) + noise = randn_tensor(image.shape, generator=generator, device=device, dtype=image.dtype) + image = image + noise_aug_strength * noise + + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + if needs_upcasting: + self.vae.to(dtype=torch.float32) + + image_latents = self._encode_vae_image( + image, + device=device, + num_videos_per_prompt=num_videos_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + image_latents = image_latents.to(image_embeddings.dtype) + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + + # Repeat the image latents for each frame so we can concatenate them with the noise + # image_latents [batch, channels, height, width] ->[batch, num_frames, channels, height, width] + image_latents = image_latents.unsqueeze(1).repeat(1, num_frames, 1, 1, 1) + + # 5. Get Added Time IDs + added_time_ids = self._get_add_time_ids( + fps, + motion_bucket_id, + noise_aug_strength, + image_embeddings.dtype, + batch_size, + num_videos_per_prompt, + self.do_classifier_free_guidance, + ) + added_time_ids = added_time_ids.to(device) + + # 6. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, None, sigmas) + + # 7. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_frames, + num_channels_latents, + height, + width, + image_embeddings.dtype, + device, + generator, + latents, + ) + + # 8. Prepare guidance scale + guidance_scale = torch.linspace(min_guidance_scale, max_guidance_scale, num_frames).unsqueeze(0) + guidance_scale = guidance_scale.to(device, latents.dtype) + guidance_scale = guidance_scale.repeat(batch_size * num_videos_per_prompt, 1) + guidance_scale = _append_dims(guidance_scale, latents.ndim) + + self._guidance_scale = guidance_scale + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # Concatenate image_latents over channels dimension + latent_model_input = torch.cat([latent_model_input, image_latents], dim=2) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=image_embeddings, + added_time_ids=added_time_ids, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + frames = self.decode_latents(latents, num_frames, decode_chunk_size) + frames = self.video_processor.postprocess_video(video=frames, output_type=output_type) + else: + frames = latents + + self.maybe_free_model_hooks() + + if not return_dict: + return frames + + return StableVideoDiffusionPipelineOutput(frames=frames) + + +# resizing utils +# TODO: clean up later +def _resize_with_antialiasing(input, size, interpolation="bicubic", align_corners=True): + h, w = input.shape[-2:] + factors = (h / size[0], w / size[1]) + + # First, we have to determine sigma + # Taken from skimage: https://github.com/scikit-image/scikit-image/blob/v0.19.2/skimage/transform/_warps.py#L171 + sigmas = ( + max((factors[0] - 1.0) / 2.0, 0.001), + max((factors[1] - 1.0) / 2.0, 0.001), + ) + + # Now kernel size. Good results are for 3 sigma, but that is kind of slow. Pillow uses 1 sigma + # https://github.com/python-pillow/Pillow/blob/master/src/libImaging/Resample.c#L206 + # But they do it in the 2 passes, which gives better results. Let's try 2 sigmas for now + ks = int(max(2.0 * 2 * sigmas[0], 3)), int(max(2.0 * 2 * sigmas[1], 3)) + + # Make sure it is odd + if (ks[0] % 2) == 0: + ks = ks[0] + 1, ks[1] + + if (ks[1] % 2) == 0: + ks = ks[0], ks[1] + 1 + + input = _gaussian_blur2d(input, ks, sigmas) + + output = torch.nn.functional.interpolate(input, size=size, mode=interpolation, align_corners=align_corners) + return output + + +def _compute_padding(kernel_size): + """Compute padding tuple.""" + # 4 or 6 ints: (padding_left, padding_right,padding_top,padding_bottom) + # https://pytorch.org/docs/stable/nn.html#torch.nn.functional.pad + if len(kernel_size) < 2: + raise AssertionError(kernel_size) + computed = [k - 1 for k in kernel_size] + + # for even kernels we need to do asymmetric padding :( + out_padding = 2 * len(kernel_size) * [0] + + for i in range(len(kernel_size)): + computed_tmp = computed[-(i + 1)] + + pad_front = computed_tmp // 2 + pad_rear = computed_tmp - pad_front + + out_padding[2 * i + 0] = pad_front + out_padding[2 * i + 1] = pad_rear + + return out_padding + + +def _filter2d(input, kernel): + # prepare kernel + b, c, h, w = input.shape + tmp_kernel = kernel[:, None, ...].to(device=input.device, dtype=input.dtype) + + tmp_kernel = tmp_kernel.expand(-1, c, -1, -1) + + height, width = tmp_kernel.shape[-2:] + + padding_shape: List[int] = _compute_padding([height, width]) + input = torch.nn.functional.pad(input, padding_shape, mode="reflect") + + # kernel and input tensor reshape to align element-wise or batch-wise params + tmp_kernel = tmp_kernel.reshape(-1, 1, height, width) + input = input.view(-1, tmp_kernel.size(0), input.size(-2), input.size(-1)) + + # convolve the tensor with the kernel. + output = torch.nn.functional.conv2d(input, tmp_kernel, groups=tmp_kernel.size(0), padding=0, stride=1) + + out = output.view(b, c, h, w) + return out + + +def _gaussian(window_size: int, sigma): + if isinstance(sigma, float): + sigma = torch.tensor([[sigma]]) + + batch_size = sigma.shape[0] + + x = (torch.arange(window_size, device=sigma.device, dtype=sigma.dtype) - window_size // 2).expand(batch_size, -1) + + if window_size % 2 == 0: + x = x + 0.5 + + gauss = torch.exp(-x.pow(2.0) / (2 * sigma.pow(2.0))) + + return gauss / gauss.sum(-1, keepdim=True) + + +def _gaussian_blur2d(input, kernel_size, sigma): + if isinstance(sigma, tuple): + sigma = torch.tensor([sigma], dtype=input.dtype) + else: + sigma = sigma.to(dtype=input.dtype) + + ky, kx = int(kernel_size[0]), int(kernel_size[1]) + bs = sigma.shape[0] + kernel_x = _gaussian(kx, sigma[:, 1].view(bs, 1)) + kernel_y = _gaussian(ky, sigma[:, 0].view(bs, 1)) + out_x = _filter2d(input, kernel_x[..., None, :]) + out = _filter2d(out_x, kernel_y[..., None]) + + return out diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..08c22a2707fe55770a519db481954881c1cad26e --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/__init__.py @@ -0,0 +1,47 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_adapter"] = ["StableDiffusionAdapterPipeline"] + _import_structure["pipeline_stable_diffusion_xl_adapter"] = ["StableDiffusionXLAdapterPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_stable_diffusion_adapter import StableDiffusionAdapterPipeline + from .pipeline_stable_diffusion_xl_adapter import StableDiffusionXLAdapterPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..75c6a0d06b12b69865d3b3e92f775095fb71644f Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/__pycache__/pipeline_stable_diffusion_adapter.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/__pycache__/pipeline_stable_diffusion_adapter.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f97402643f14b5fb324dae0db7f3518c3d0b97ef Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/__pycache__/pipeline_stable_diffusion_adapter.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/__pycache__/pipeline_stable_diffusion_xl_adapter.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/__pycache__/pipeline_stable_diffusion_xl_adapter.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..12941edcb8c390da8b236f907a2a4d91dd31e5f9 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/__pycache__/pipeline_stable_diffusion_xl_adapter.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_adapter.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_adapter.py new file mode 100644 index 0000000000000000000000000000000000000000..5c561721fcc76f860c762cb0347a37fa70c8a121 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_adapter.py @@ -0,0 +1,956 @@ +# Copyright 2025 TencentARC and The HuggingFace Team. All rights reserved. +# +# 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 inspect +from dataclasses import dataclass +from typing import Any, Callable, Dict, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ...image_processor import VaeImageProcessor +from ...loaders import FromSingleFileMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, MultiAdapter, T2IAdapter, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + PIL_INTERPOLATION, + USE_PEFT_BACKEND, + BaseOutput, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +@dataclass +class StableDiffusionAdapterPipelineOutput(BaseOutput): + """ + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + nsfw_content_detected (`List[bool]`) + List of flags denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, or `None` if safety checking could not be performed. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] + nsfw_content_detected: Optional[List[bool]] + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from PIL import Image + >>> from diffusers.utils import load_image + >>> import torch + >>> from diffusers import StableDiffusionAdapterPipeline, T2IAdapter + + >>> image = load_image( + ... "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/t2i-adapter/color_ref.png" + ... ) + + >>> color_palette = image.resize((8, 8)) + >>> color_palette = color_palette.resize((512, 512), resample=Image.Resampling.NEAREST) + + >>> adapter = T2IAdapter.from_pretrained("TencentARC/t2iadapter_color_sd14v1", torch_dtype=torch.float16) + >>> pipe = StableDiffusionAdapterPipeline.from_pretrained( + ... "CompVis/stable-diffusion-v1-4", + ... adapter=adapter, + ... torch_dtype=torch.float16, + ... ) + + >>> pipe.to("cuda") + + >>> out_image = pipe( + ... "At night, glowing cubes in front of the beach", + ... image=color_palette, + ... ).images[0] + ``` +""" + + +def _preprocess_adapter_image(image, height, width): + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + image = [np.array(i.resize((width, height), resample=PIL_INTERPOLATION["lanczos"])) for i in image] + image = [ + i[None, ..., None] if i.ndim == 2 else i[None, ...] for i in image + ] # expand [h, w] or [h, w, c] to [b, h, w, c] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + if image[0].ndim == 3: + image = torch.stack(image, dim=0) + elif image[0].ndim == 4: + image = torch.cat(image, dim=0) + else: + raise ValueError( + f"Invalid image tensor! Expecting image tensor with 3 or 4 dimension, but receive: {image[0].ndim}" + ) + return image + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionAdapterPipeline(DiffusionPipeline, StableDiffusionMixin, FromSingleFileMixin): + r""" + Pipeline for text-to-image generation using Stable Diffusion augmented with T2I-Adapter + https://huggingface.co/papers/2302.08453 + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + adapter ([`T2IAdapter`] or [`MultiAdapter`] or `List[T2IAdapter]`): + Provides additional conditioning to the unet during the denoising process. If you set multiple Adapter as a + list, the outputs from each Adapter are added together to create one combined additional conditioning. + adapter_weights (`List[float]`, *optional*, defaults to None): + List of floats representing the weight which will be multiply to each adapter's output before adding them + together. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + details. + feature_extractor ([`CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->adapter->unet->vae" + _optional_components = ["safety_checker", "feature_extractor"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + adapter: Union[T2IAdapter, MultiAdapter, List[T2IAdapter]], + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + if isinstance(adapter, (list, tuple)): + adapter = MultiAdapter(adapter) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + adapter=adapter, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + image, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if isinstance(self.adapter, MultiAdapter): + if not isinstance(image, list): + raise ValueError( + "MultiAdapter is enabled, but `image` is not a list. Please pass a list of images to `image`." + ) + + if len(image) != len(self.adapter.adapters): + raise ValueError( + f"MultiAdapter requires passing the same number of images as adapters. Given {len(image)} images and {len(self.adapter.adapters)} adapters." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def _default_height_width(self, height, width, image): + # NOTE: It is possible that a list of images have different + # dimensions for each image, so just checking the first image + # is not _exactly_ correct, but it is simple. + while isinstance(image, list): + image = image[0] + + if height is None: + if isinstance(image, PIL.Image.Image): + height = image.height + elif isinstance(image, torch.Tensor): + height = image.shape[-2] + + # round down to nearest multiple of `self.adapter.downscale_factor` + height = (height // self.adapter.downscale_factor) * self.adapter.downscale_factor + + if width is None: + if isinstance(image, PIL.Image.Image): + width = image.width + elif isinstance(image, torch.Tensor): + width = image.shape[-1] + + # round down to nearest multiple of `self.adapter.downscale_factor` + width = (width // self.adapter.downscale_factor) * self.adapter.downscale_factor + + return height, width + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + image: Union[torch.Tensor, PIL.Image.Image, List[PIL.Image.Image]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + timesteps: List[int] = None, + sigmas: List[float] = None, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + adapter_conditioning_scale: Union[float, List[float]] = 1.0, + clip_skip: Optional[int] = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + image (`torch.Tensor`, `PIL.Image.Image`, `List[torch.Tensor]` or `List[PIL.Image.Image]` or `List[List[PIL.Image.Image]]`): + The Adapter input condition. Adapter uses this input condition to generate guidance to Unet. If the + type is specified as `torch.Tensor`, it is passed to Adapter as is. PIL.Image.Image` can also be + accepted as an image. The control image is automatically resized to fit the output image. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead. + Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] instead + of a plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttnProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + adapter_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0): + The outputs of the adapter are multiplied by `adapter_conditioning_scale` before they are added to the + residual in the original unet. If multiple adapters are specified in init, you can set the + corresponding scale as a list. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] if `return_dict` is True, otherwise a + `tuple. When returning a tuple, the first element is a list with the generated images, and the second + element is a list of `bool`s denoting whether the corresponding generated image likely represents + "not-safe-for-work" (nsfw) content, according to the `safety_checker`. + """ + # 0. Default height and width to unet + height, width = self._default_height_width(height, width, image) + device = self._execution_device + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, height, width, callback_steps, image, negative_prompt, prompt_embeds, negative_prompt_embeds + ) + + self._guidance_scale = guidance_scale + + if isinstance(self.adapter, MultiAdapter): + adapter_input = [] + + for one_image in image: + one_image = _preprocess_adapter_image(one_image, height, width) + one_image = one_image.to(device=device, dtype=self.adapter.dtype) + adapter_input.append(one_image) + else: + adapter_input = _preprocess_adapter_image(image, height, width) + adapter_input = adapter_input.to(device=device, dtype=self.adapter.dtype) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.5 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Denoising loop + if isinstance(self.adapter, MultiAdapter): + adapter_state = self.adapter(adapter_input, adapter_conditioning_scale) + for k, v in enumerate(adapter_state): + adapter_state[k] = v + else: + adapter_state = self.adapter(adapter_input) + for k, v in enumerate(adapter_state): + adapter_state[k] = v * adapter_conditioning_scale + if num_images_per_prompt > 1: + for k, v in enumerate(adapter_state): + adapter_state[k] = v.repeat(num_images_per_prompt, 1, 1, 1) + if self.do_classifier_free_guidance: + for k, v in enumerate(adapter_state): + adapter_state[k] = torch.cat([v] * 2, dim=0) + + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=cross_attention_kwargs, + down_intrablock_additional_residuals=[state.clone() for state in adapter_state], + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + has_nsfw_concept = None + elif output_type == "pil": + # 8. Post-processing + image = self.decode_latents(latents) + + # 9. Run safety checker + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + + # 10. Convert to PIL + image = self.numpy_to_pil(image) + else: + # 8. Post-processing + image = self.decode_latents(latents) + + # 9. Run safety checker + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionAdapterPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py new file mode 100644 index 0000000000000000000000000000000000000000..13183df47d4baef54118de17ef08d82632feb558 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py @@ -0,0 +1,1311 @@ +# Copyright 2025 TencentARC and The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import numpy as np +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ImageProjection, MultiAdapter, T2IAdapter, UNet2DConditionModel +from ...models.attention_processor import ( + AttnProcessor2_0, + XFormersAttnProcessor, +) +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + PIL_INTERPOLATION, + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import T2IAdapter, StableDiffusionXLAdapterPipeline, DDPMScheduler + >>> from diffusers.utils import load_image + + >>> sketch_image = load_image("https://huggingface.co/Adapter/t2iadapter/resolve/main/sketch.png").convert("L") + + >>> model_id = "stabilityai/stable-diffusion-xl-base-1.0" + + >>> adapter = T2IAdapter.from_pretrained( + ... "Adapter/t2iadapter", + ... subfolder="sketch_sdxl_1.0", + ... torch_dtype=torch.float16, + ... adapter_type="full_adapter_xl", + ... ) + >>> scheduler = DDPMScheduler.from_pretrained(model_id, subfolder="scheduler") + + >>> pipe = StableDiffusionXLAdapterPipeline.from_pretrained( + ... model_id, adapter=adapter, torch_dtype=torch.float16, variant="fp16", scheduler=scheduler + ... ).to("cuda") + + >>> generator = torch.manual_seed(42) + >>> sketch_image_out = pipe( + ... prompt="a photo of a dog in real world, high quality", + ... negative_prompt="extra digit, fewer digits, cropped, worst quality, low quality", + ... image=sketch_image, + ... generator=generator, + ... guidance_scale=7.5, + ... ).images[0] + ``` +""" + + +def _preprocess_adapter_image(image, height, width): + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + image = [np.array(i.resize((width, height), resample=PIL_INTERPOLATION["lanczos"])) for i in image] + image = [ + i[None, ..., None] if i.ndim == 2 else i[None, ...] for i in image + ] # expand [h, w] or [h, w, c] to [b, h, w, c] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + if image[0].ndim == 3: + image = torch.stack(image, dim=0) + elif image[0].ndim == 4: + image = torch.cat(image, dim=0) + else: + raise ValueError( + f"Invalid image tensor! Expecting image tensor with 3 or 4 dimension, but receive: {image[0].ndim}" + ) + return image + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLAdapterPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionXLLoraLoaderMixin, + IPAdapterMixin, + FromSingleFileMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion augmented with T2I-Adapter + https://huggingface.co/papers/2302.08453 + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + adapter ([`T2IAdapter`] or [`MultiAdapter`] or `List[T2IAdapter]`): + Provides additional conditioning to the unet during the denoising process. If you set multiple Adapter as a + list, the outputs from each Adapter are added together to create one combined additional conditioning. + adapter_weights (`List[float]`, *optional*, defaults to None): + List of floats representing the weight which will be multiply to each adapter's output before adding them + together. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + details. + feature_extractor ([`CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "feature_extractor", + "image_encoder", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + adapter: Union[T2IAdapter, MultiAdapter, List[T2IAdapter]], + scheduler: KarrasDiffusionSchedulers, + force_zeros_for_empty_prompt: bool = True, + feature_extractor: CLIPImageProcessor = None, + image_encoder: CLIPVisionModelWithProjection = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + adapter=adapter, + scheduler=scheduler, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.default_sample_size = ( + self.unet.config.sample_size + if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size") + else 128 + ) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + # Copied from diffusers.pipelines.t2i_adapter.pipeline_stable_diffusion_adapter.StableDiffusionAdapterPipeline._default_height_width + def _default_height_width(self, height, width, image): + # NOTE: It is possible that a list of images have different + # dimensions for each image, so just checking the first image + # is not _exactly_ correct, but it is simple. + while isinstance(image, list): + image = image[0] + + if height is None: + if isinstance(image, PIL.Image.Image): + height = image.height + elif isinstance(image, torch.Tensor): + height = image.shape[-2] + + # round down to nearest multiple of `self.adapter.downscale_factor` + height = (height // self.adapter.downscale_factor) * self.adapter.downscale_factor + + if width is None: + if isinstance(image, PIL.Image.Image): + width = image.width + elif isinstance(image, torch.Tensor): + width = image.shape[-1] + + # round down to nearest multiple of `self.adapter.downscale_factor` + width = (width // self.adapter.downscale_factor) * self.adapter.downscale_factor + + return height, width + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + image: PipelineImageInput = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + timesteps: List[int] = None, + sigmas: List[float] = None, + denoising_end: Optional[float] = None, + guidance_scale: float = 5.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + original_size: Optional[Tuple[int, int]] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Optional[Tuple[int, int]] = None, + negative_original_size: Optional[Tuple[int, int]] = None, + negative_crops_coords_top_left: Tuple[int, int] = (0, 0), + negative_target_size: Optional[Tuple[int, int]] = None, + adapter_conditioning_scale: Union[float, List[float]] = 1.0, + adapter_conditioning_factor: float = 1.0, + clip_skip: Optional[int] = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + image (`torch.Tensor`, `PIL.Image.Image`, `List[torch.Tensor]` or `List[PIL.Image.Image]` or `List[List[PIL.Image.Image]]`): + The Adapter input condition. Adapter uses this input condition to generate guidance to Unet. If the + type is specified as `torch.Tensor`, it is passed to Adapter as is. PIL.Image.Image` can also be + accepted as an image. The control image is automatically resized to fit the output image. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionAdapterPipelineOutput`] + instead of a plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + adapter_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0): + The outputs of the adapter are multiplied by `adapter_conditioning_scale` before they are added to the + residual in the original unet. If multiple adapters are specified in init, you can set the + corresponding scale as a list. + adapter_conditioning_factor (`float`, *optional*, defaults to 1.0): + The fraction of timesteps for which adapter should be applied. If `adapter_conditioning_factor` is + `0.0`, adapter is not applied at all. If `adapter_conditioning_factor` is `1.0`, adapter is applied for + all timesteps. If `adapter_conditioning_factor` is `0.5`, adapter is applied for half of the timesteps. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + # 0. Default height and width to unet + + height, width = self._default_height_width(height, width, image) + device = self._execution_device + + if isinstance(self.adapter, MultiAdapter): + adapter_input = [] + + for one_image in image: + one_image = _preprocess_adapter_image(one_image, height, width) + one_image = one_image.to(device=device, dtype=self.adapter.dtype) + adapter_input.append(one_image) + else: + adapter_input = _preprocess_adapter_image(image, height, width) + adapter_input = adapter_input.to(device=device, dtype=self.adapter.dtype) + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + ) + + self._guidance_scale = guidance_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3.1 Encode input prompt + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + clip_skip=clip_skip, + ) + + # 3.2 Encode ip_adapter_image + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6.1 Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Prepare added time ids & embeddings & adapter features + if isinstance(self.adapter, MultiAdapter): + adapter_state = self.adapter(adapter_input, adapter_conditioning_scale) + for k, v in enumerate(adapter_state): + adapter_state[k] = v + else: + adapter_state = self.adapter(adapter_input) + for k, v in enumerate(adapter_state): + adapter_state[k] = v * adapter_conditioning_scale + if num_images_per_prompt > 1: + for k, v in enumerate(adapter_state): + adapter_state[k] = v.repeat(num_images_per_prompt, 1, 1, 1) + if self.do_classifier_free_guidance: + for k, v in enumerate(adapter_state): + adapter_state[k] = torch.cat([v] * 2, dim=0) + + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + if negative_original_size is not None and negative_target_size is not None: + negative_add_time_ids = self._get_add_time_ids( + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + else: + negative_add_time_ids = add_time_ids + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + # Apply denoising_end + if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1: + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + + # predict the noise residual + if i < int(num_inference_steps * adapter_conditioning_factor): + down_intrablock_additional_residuals = [state.clone() for state in adapter_state] + else: + down_intrablock_additional_residuals = None + + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=cross_attention_kwargs, + down_intrablock_additional_residuals=down_intrablock_additional_residuals, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + return StableDiffusionXLPipelineOutput(images=image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..8d8fdb92769bb3dcb2dd7696115b16f197062262 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__init__.py @@ -0,0 +1,54 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_output"] = ["TextToVideoSDPipelineOutput"] + _import_structure["pipeline_text_to_video_synth"] = ["TextToVideoSDPipeline"] + _import_structure["pipeline_text_to_video_synth_img2img"] = ["VideoToVideoSDPipeline"] + _import_structure["pipeline_text_to_video_zero"] = ["TextToVideoZeroPipeline"] + _import_structure["pipeline_text_to_video_zero_sdxl"] = ["TextToVideoZeroSDXLPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_output import TextToVideoSDPipelineOutput + from .pipeline_text_to_video_synth import TextToVideoSDPipeline + from .pipeline_text_to_video_synth_img2img import VideoToVideoSDPipeline + from .pipeline_text_to_video_zero import TextToVideoZeroPipeline + from .pipeline_text_to_video_zero_sdxl import TextToVideoZeroSDXLPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8087ce05d888efa31510f5bb4e5112b40002ba54 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_output.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_output.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9dea9dfb831a3e8c93a4bcb8217b452e6bb6a992 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_output.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_synth.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_synth.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8a1578c1bad4e59fa9d62f18ae2554958e8b985b Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_synth.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_synth_img2img.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_synth_img2img.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2bf2493e91c51dceb09f12c6539eaa048dc51f5c Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_synth_img2img.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_zero.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_zero.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4f019e02fceb1af5dcc81f6c6e242b364c603d47 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_zero.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_zero_sdxl.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_zero_sdxl.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..640486f7064e9e179e7ffff330931b214fb0d574 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_zero_sdxl.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_output.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..040bf0efba8465d830af5e6686bc3f544585f63f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_output.py @@ -0,0 +1,26 @@ +from dataclasses import dataclass +from typing import List, Union + +import numpy as np +import PIL +import torch + +from ...utils import ( + BaseOutput, +) + + +@dataclass +class TextToVideoSDPipelineOutput(BaseOutput): + """ + Output class for text-to-video pipelines. + + Args: + frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]): + List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised + PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)` + """ + + frames: Union[torch.Tensor, np.ndarray, List[List[PIL.Image.Image]]] diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth.py new file mode 100644 index 0000000000000000000000000000000000000000..3ce7b4d1990fb5fcf9a96f6beb2934d653e52e30 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth.py @@ -0,0 +1,660 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import torch +from transformers import CLIPTextModel, CLIPTokenizer + +from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet3DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from . import TextToVideoSDPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import TextToVideoSDPipeline + >>> from diffusers.utils import export_to_video + + >>> pipe = TextToVideoSDPipeline.from_pretrained( + ... "damo-vilab/text-to-video-ms-1.7b", torch_dtype=torch.float16, variant="fp16" + ... ) + >>> pipe.enable_model_cpu_offload() + + >>> prompt = "Spiderman is surfing" + >>> video_frames = pipe(prompt).frames[0] + >>> video_path = export_to_video(video_frames) + >>> video_path + ``` +""" + + +class TextToVideoSDPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, +): + _last_supported_version = "0.33.1" + r""" + Pipeline for text-to-video generation. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer (`CLIPTokenizer`): + A [`~transformers.CLIPTokenizer`] to tokenize text. + unet ([`UNet3DConditionModel`]): + A [`UNet3DConditionModel`] to denoise the encoded video latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet3DConditionModel, + scheduler: KarrasDiffusionSchedulers, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(do_resize=False, vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + def decode_latents(self, latents): + latents = 1 / self.vae.config.scaling_factor * latents + + batch_size, channels, num_frames, height, width = latents.shape + latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width) + + image = self.vae.decode(latents).sample + video = image[None, :].reshape((batch_size, num_frames, -1) + image.shape[2:]).permute(0, 2, 1, 3, 4) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + video = video.float() + return video + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.StableDiffusionKDiffusionPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def prepare_latents( + self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None + ): + shape = ( + batch_size, + num_channels_latents, + num_frames, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_frames: int = 16, + num_inference_steps: int = 50, + guidance_scale: float = 9.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "np", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + clip_skip: Optional[int] = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated video. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated video. + num_frames (`int`, *optional*, defaults to 16): + The number of video frames that are generated. Defaults to 16 frames which at 8 frames per seconds + amounts to 2 seconds of video. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality videos at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. Latents should be of shape + `(batch_size, num_channel, num_frames, height, width)`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated video. Choose between `torch.Tensor` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] instead + of a plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] is + returned, otherwise a `tuple` is returned where the first element is a list with the generated frames. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + num_images_per_prompt = 1 + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + num_frames, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # reshape latents + bsz, channel, frames, width, height = latents.shape + latents = latents.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height) + noise_pred = noise_pred.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # reshape latents back + latents = latents[None, :].reshape(bsz, frames, channel, width, height).permute(0, 2, 1, 3, 4) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 8. Post processing + if output_type == "latent": + video = latents + else: + video_tensor = self.decode_latents(latents) + video = self.video_processor.postprocess_video(video=video_tensor, output_type=output_type) + + # 9. Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return TextToVideoSDPipelineOutput(frames=video) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..9d0b7e3dbc32ac782fcaa683dde74c74d3f87b19 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py @@ -0,0 +1,716 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import numpy as np +import torch +from transformers import CLIPTextModel, CLIPTokenizer + +from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet3DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from . import TextToVideoSDPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler + >>> from diffusers.utils import export_to_video + + >>> pipe = DiffusionPipeline.from_pretrained("cerspense/zeroscope_v2_576w", torch_dtype=torch.float16) + >>> pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config) + >>> pipe.to("cuda") + + >>> prompt = "spiderman running in the desert" + >>> video_frames = pipe(prompt, num_inference_steps=40, height=320, width=576, num_frames=24).frames[0] + >>> # safe low-res video + >>> video_path = export_to_video(video_frames, output_video_path="./video_576_spiderman.mp4") + + >>> # let's offload the text-to-image model + >>> pipe.to("cpu") + + >>> # and load the image-to-image model + >>> pipe = DiffusionPipeline.from_pretrained( + ... "cerspense/zeroscope_v2_XL", torch_dtype=torch.float16, revision="refs/pr/15" + ... ) + >>> pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config) + >>> pipe.enable_model_cpu_offload() + + >>> # The VAE consumes A LOT of memory, let's make sure we run it in sliced mode + >>> pipe.vae.enable_slicing() + + >>> # now let's upscale it + >>> video = [Image.fromarray(frame).resize((1024, 576)) for frame in video_frames] + + >>> # and denoise it + >>> video_frames = pipe(prompt, video=video, strength=0.6).frames[0] + >>> video_path = export_to_video(video_frames, output_video_path="./video_1024_spiderman.mp4") + >>> video_path + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class VideoToVideoSDPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, +): + _last_supported_version = "0.33.1" + r""" + Pipeline for text-guided video-to-video generation. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer (`CLIPTokenizer`): + A [`~transformers.CLIPTokenizer`] to tokenize text. + unet ([`UNet3DConditionModel`]): + A [`UNet3DConditionModel`] to denoise the encoded video latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet3DConditionModel, + scheduler: KarrasDiffusionSchedulers, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(do_resize=False, vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.decode_latents + def decode_latents(self, latents): + latents = 1 / self.vae.config.scaling_factor * latents + + batch_size, channels, num_frames, height, width = latents.shape + latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width) + + image = self.vae.decode(latents).sample + video = image[None, :].reshape((batch_size, num_frames, -1) + image.shape[2:]).permute(0, 2, 1, 3, 4) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + video = video.float() + return video + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + strength, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def prepare_latents(self, video, timestep, batch_size, dtype, device, generator=None): + video = video.to(device=device, dtype=dtype) + + # change from (b, c, f, h, w) -> (b * f, c, w, h) + bsz, channel, frames, width, height = video.shape + video = video.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height) + + if video.shape[1] == 4: + init_latents = video + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + elif isinstance(generator, list): + init_latents = [ + retrieve_latents(self.vae.encode(video[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(video), generator=generator) + + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `video` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + latents = init_latents + + latents = latents[None, :].reshape((bsz, frames, latents.shape[1]) + latents.shape[2:]).permute(0, 2, 1, 3, 4) + + return latents + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + video: Union[List[np.ndarray], torch.Tensor] = None, + strength: float = 0.6, + num_inference_steps: int = 50, + guidance_scale: float = 15.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "np", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + clip_skip: Optional[int] = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + video (`List[np.ndarray]` or `torch.Tensor`): + `video` frames or tensor representing a video batch to be used as the starting point for the process. + Can also accept video latents as `image`, if passing latents directly, it will not be encoded again. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `video`. Must be between 0 and 1. `video` is used as a + starting point, adding more noise to it the larger the `strength`. The number of denoising steps + depends on the amount of noise initially added. When `strength` is 1, added noise is maximum and the + denoising process runs for the full number of iterations specified in `num_inference_steps`. A value of + 1 essentially ignores `video`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality videos at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in video generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. Latents should be of shape + `(batch_size, num_channel, num_frames, height, width)`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated video. Choose between `torch.Tensor` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] instead + of a plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] is + returned, otherwise a `tuple` is returned where the first element is a list with the generated frames. + """ + # 0. Default height and width to unet + num_images_per_prompt = 1 + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, strength, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Preprocess video + video = self.video_processor.preprocess_video(video) + + # 5. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 6. Prepare latent variables + latents = self.prepare_latents(video, latent_timestep, batch_size, prompt_embeds.dtype, device, generator) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # reshape latents + bsz, channel, frames, width, height = latents.shape + latents = latents.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height) + noise_pred = noise_pred.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # reshape latents back + latents = latents[None, :].reshape(bsz, frames, channel, width, height).permute(0, 2, 1, 3, 4) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # manually for max memory savings + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.unet.to("cpu") + + # 9. Post processing + if output_type == "latent": + video = latents + else: + video_tensor = self.decode_latents(latents) + video = self.video_processor.postprocess_video(video=video_tensor, output_type=output_type) + + # 10. Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return TextToVideoSDPipelineOutput(frames=video) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero.py new file mode 100644 index 0000000000000000000000000000000000000000..96316f8e91e59d3a8f33f778f494654fee63dbc4 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero.py @@ -0,0 +1,1005 @@ +import copy +import inspect +from dataclasses import dataclass +from typing import Callable, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from torch.nn.functional import grid_sample +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ...image_processor import VaeImageProcessor +from ...loaders import FromSingleFileMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + BaseOutput, + is_torch_xla_available, + logging, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import empty_device_cache, randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def rearrange_0(tensor, f): + F, C, H, W = tensor.size() + tensor = torch.permute(torch.reshape(tensor, (F // f, f, C, H, W)), (0, 2, 1, 3, 4)) + return tensor + + +def rearrange_1(tensor): + B, C, F, H, W = tensor.size() + return torch.reshape(torch.permute(tensor, (0, 2, 1, 3, 4)), (B * F, C, H, W)) + + +def rearrange_3(tensor, f): + F, D, C = tensor.size() + return torch.reshape(tensor, (F // f, f, D, C)) + + +def rearrange_4(tensor): + B, F, D, C = tensor.size() + return torch.reshape(tensor, (B * F, D, C)) + + +class CrossFrameAttnProcessor: + """ + Cross frame attention processor. Each frame attends the first frame. + + Args: + batch_size: The number that represents actual batch size, other than the frames. + For example, calling unet with a single prompt and num_images_per_prompt=1, batch_size should be equal to + 2, due to classifier-free guidance. + """ + + def __init__(self, batch_size=2): + self.batch_size = batch_size + + def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None): + batch_size, sequence_length, _ = hidden_states.shape + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + query = attn.to_q(hidden_states) + + is_cross_attention = encoder_hidden_states is not None + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + # Cross Frame Attention + if not is_cross_attention: + video_length = key.size()[0] // self.batch_size + first_frame_index = [0] * video_length + + # rearrange keys to have batch and frames in the 1st and 2nd dims respectively + key = rearrange_3(key, video_length) + key = key[:, first_frame_index] + # rearrange values to have batch and frames in the 1st and 2nd dims respectively + value = rearrange_3(value, video_length) + value = value[:, first_frame_index] + + # rearrange back to original shape + key = rearrange_4(key) + value = rearrange_4(value) + + query = attn.head_to_batch_dim(query) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + return hidden_states + + +class CrossFrameAttnProcessor2_0: + """ + Cross frame attention processor with scaled_dot_product attention of Pytorch 2.0. + + Args: + batch_size: The number that represents actual batch size, other than the frames. + For example, calling unet with a single prompt and num_images_per_prompt=1, batch_size should be equal to + 2, due to classifier-free guidance. + """ + + def __init__(self, batch_size=2): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") + self.batch_size = batch_size + + def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None): + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + inner_dim = hidden_states.shape[-1] + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + query = attn.to_q(hidden_states) + + is_cross_attention = encoder_hidden_states is not None + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + # Cross Frame Attention + if not is_cross_attention: + video_length = max(1, key.size()[0] // self.batch_size) + first_frame_index = [0] * video_length + + # rearrange keys to have batch and frames in the 1st and 2nd dims respectively + key = rearrange_3(key, video_length) + key = key[:, first_frame_index] + # rearrange values to have batch and frames in the 1st and 2nd dims respectively + value = rearrange_3(value, video_length) + value = value[:, first_frame_index] + + # rearrange back to original shape + key = rearrange_4(key) + value = rearrange_4(value) + + head_dim = inner_dim // attn.heads + query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + # the output of sdp = (batch, num_heads, seq_len, head_dim) + # TODO: add support for attn.scale when we move to Torch 2.1 + hidden_states = F.scaled_dot_product_attention( + query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + ) + + hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) + hidden_states = hidden_states.to(query.dtype) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + return hidden_states + + +@dataclass +class TextToVideoPipelineOutput(BaseOutput): + r""" + Output class for zero-shot text-to-video pipeline. + + Args: + images (`[List[PIL.Image.Image]`, `np.ndarray`]): + List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + nsfw_content_detected (`[List[bool]]`): + List indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content or + `None` if safety checking could not be performed. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] + nsfw_content_detected: Optional[List[bool]] + + +def coords_grid(batch, ht, wd, device): + # Adapted from https://github.com/princeton-vl/RAFT/blob/master/core/utils/utils.py + coords = torch.meshgrid(torch.arange(ht, device=device), torch.arange(wd, device=device)) + coords = torch.stack(coords[::-1], dim=0).float() + return coords[None].repeat(batch, 1, 1, 1) + + +def warp_single_latent(latent, reference_flow): + """ + Warp latent of a single frame with given flow + + Args: + latent: latent code of a single frame + reference_flow: flow which to warp the latent with + + Returns: + warped: warped latent + """ + _, _, H, W = reference_flow.size() + _, _, h, w = latent.size() + coords0 = coords_grid(1, H, W, device=latent.device).to(latent.dtype) + + coords_t0 = coords0 + reference_flow + coords_t0[:, 0] /= W + coords_t0[:, 1] /= H + + coords_t0 = coords_t0 * 2.0 - 1.0 + coords_t0 = F.interpolate(coords_t0, size=(h, w), mode="bilinear") + coords_t0 = torch.permute(coords_t0, (0, 2, 3, 1)) + + warped = grid_sample(latent, coords_t0, mode="nearest", padding_mode="reflection") + return warped + + +def create_motion_field(motion_field_strength_x, motion_field_strength_y, frame_ids, device, dtype): + """ + Create translation motion field + + Args: + motion_field_strength_x: motion strength along x-axis + motion_field_strength_y: motion strength along y-axis + frame_ids: indexes of the frames the latents of which are being processed. + This is needed when we perform chunk-by-chunk inference + device: device + dtype: dtype + + Returns: + + """ + seq_length = len(frame_ids) + reference_flow = torch.zeros((seq_length, 2, 512, 512), device=device, dtype=dtype) + for fr_idx in range(seq_length): + reference_flow[fr_idx, 0, :, :] = motion_field_strength_x * (frame_ids[fr_idx]) + reference_flow[fr_idx, 1, :, :] = motion_field_strength_y * (frame_ids[fr_idx]) + return reference_flow + + +def create_motion_field_and_warp_latents(motion_field_strength_x, motion_field_strength_y, frame_ids, latents): + """ + Creates translation motion and warps the latents accordingly + + Args: + motion_field_strength_x: motion strength along x-axis + motion_field_strength_y: motion strength along y-axis + frame_ids: indexes of the frames the latents of which are being processed. + This is needed when we perform chunk-by-chunk inference + latents: latent codes of frames + + Returns: + warped_latents: warped latents + """ + motion_field = create_motion_field( + motion_field_strength_x=motion_field_strength_x, + motion_field_strength_y=motion_field_strength_y, + frame_ids=frame_ids, + device=latents.device, + dtype=latents.dtype, + ) + warped_latents = latents.clone().detach() + for i in range(len(warped_latents)): + warped_latents[i] = warp_single_latent(latents[i][None], motion_field[i][None]) + return warped_latents + + +class TextToVideoZeroPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + FromSingleFileMixin, +): + _last_supported_version = "0.33.1" + r""" + Pipeline for zero-shot text-to-video generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer (`CLIPTokenizer`): + A [`~transformers.CLIPTokenizer`] to tokenize text. + unet ([`UNet2DConditionModel`]): + A [`UNet3DConditionModel`] to denoise the encoded video latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`CLIPImageProcessor`]): + A [`CLIPImageProcessor`] to extract features from generated images; used as inputs to the `safety_checker`. + """ + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + def forward_loop(self, x_t0, t0, t1, generator): + """ + Perform DDPM forward process from time t0 to t1. This is the same as adding noise with corresponding variance. + + Args: + x_t0: + Latent code at time t0. + t0: + Timestep at t0. + t1: + Timestamp at t1. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + + Returns: + x_t1: + Forward process applied to x_t0 from time t0 to t1. + """ + eps = randn_tensor(x_t0.size(), generator=generator, dtype=x_t0.dtype, device=x_t0.device) + alpha_vec = torch.prod(self.scheduler.alphas[t0:t1]) + x_t1 = torch.sqrt(alpha_vec) * x_t0 + torch.sqrt(1 - alpha_vec) * eps + return x_t1 + + def backward_loop( + self, + latents, + timesteps, + prompt_embeds, + guidance_scale, + callback, + callback_steps, + num_warmup_steps, + extra_step_kwargs, + cross_attention_kwargs=None, + ): + """ + Perform backward process given list of time steps. + + Args: + latents: + Latents at time timesteps[0]. + timesteps: + Time steps along which to perform backward process. + prompt_embeds: + Pre-generated text embeddings. + guidance_scale: + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + extra_step_kwargs: + Extra_step_kwargs. + cross_attention_kwargs: + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + num_warmup_steps: + number of warmup steps. + + Returns: + latents: + Latents of backward process output at time timesteps[-1]. + """ + do_classifier_free_guidance = guidance_scale > 1.0 + num_steps = (len(timesteps) - num_warmup_steps) // self.scheduler.order + with self.progress_bar(total=num_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + ).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + return latents.clone().detach() + + # Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.StableDiffusionKDiffusionPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]], + video_length: Optional[int] = 8, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_videos_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + motion_field_strength_x: float = 12, + motion_field_strength_y: float = 12, + output_type: Optional[str] = "tensor", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: Optional[int] = 1, + t0: int = 44, + t1: int = 47, + frame_ids: Optional[List[int]] = None, + ): + """ + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + video_length (`int`, *optional*, defaults to 8): + The number of generated video frames. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in video generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated video. Choose between `"latent"` and `"np"`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a + [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoPipelineOutput`] instead of + a plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + motion_field_strength_x (`float`, *optional*, defaults to 12): + Strength of motion in generated video along x-axis. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + motion_field_strength_y (`float`, *optional*, defaults to 12): + Strength of motion in generated video along y-axis. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + t0 (`int`, *optional*, defaults to 44): + Timestep t0. Should be in the range [0, num_inference_steps - 1]. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + t1 (`int`, *optional*, defaults to 47): + Timestep t0. Should be in the range [t0 + 1, num_inference_steps - 1]. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + frame_ids (`List[int]`, *optional*): + Indexes of the frames that are being generated. This is used when generating longer videos + chunk-by-chunk. + + Returns: + [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoPipelineOutput`]: + The output contains a `ndarray` of the generated video, when `output_type` != `"latent"`, otherwise a + latent code of generated videos and a list of `bool`s indicating whether the corresponding generated + video contains "not-safe-for-work" (nsfw) content.. + """ + assert video_length > 0 + if frame_ids is None: + frame_ids = list(range(video_length)) + assert len(frame_ids) == video_length + + assert num_videos_per_prompt == 1 + + # set the processor + original_attn_proc = self.unet.attn_processors + processor = ( + CrossFrameAttnProcessor2_0(batch_size=2) + if hasattr(F, "scaled_dot_product_attention") + else CrossFrameAttnProcessor(batch_size=2) + ) + self.unet.set_attn_processor(processor) + + if isinstance(prompt, str): + prompt = [prompt] + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] + + # Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # Check inputs. Raise error if not correct + self.check_inputs(prompt, height, width, callback_steps) + + # Define call parameters + batch_size = 1 if isinstance(prompt, str) else len(prompt) + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # Encode input prompt + prompt_embeds_tuple = self.encode_prompt( + prompt, device, num_videos_per_prompt, do_classifier_free_guidance, negative_prompt + ) + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + # Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + # Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + + # Perform the first backward process up to time T_1 + x_1_t1 = self.backward_loop( + timesteps=timesteps[: -t1 - 1], + prompt_embeds=prompt_embeds, + latents=latents, + guidance_scale=guidance_scale, + callback=callback, + callback_steps=callback_steps, + extra_step_kwargs=extra_step_kwargs, + num_warmup_steps=num_warmup_steps, + ) + scheduler_copy = copy.deepcopy(self.scheduler) + + # Perform the second backward process up to time T_0 + x_1_t0 = self.backward_loop( + timesteps=timesteps[-t1 - 1 : -t0 - 1], + prompt_embeds=prompt_embeds, + latents=x_1_t1, + guidance_scale=guidance_scale, + callback=callback, + callback_steps=callback_steps, + extra_step_kwargs=extra_step_kwargs, + num_warmup_steps=0, + ) + + # Propagate first frame latents at time T_0 to remaining frames + x_2k_t0 = x_1_t0.repeat(video_length - 1, 1, 1, 1) + + # Add motion in latents at time T_0 + x_2k_t0 = create_motion_field_and_warp_latents( + motion_field_strength_x=motion_field_strength_x, + motion_field_strength_y=motion_field_strength_y, + latents=x_2k_t0, + frame_ids=frame_ids[1:], + ) + + # Perform forward process up to time T_1 + x_2k_t1 = self.forward_loop( + x_t0=x_2k_t0, + t0=timesteps[-t0 - 1].item(), + t1=timesteps[-t1 - 1].item(), + generator=generator, + ) + + # Perform backward process from time T_1 to 0 + x_1k_t1 = torch.cat([x_1_t1, x_2k_t1]) + b, l, d = prompt_embeds.size() + prompt_embeds = prompt_embeds[:, None].repeat(1, video_length, 1, 1).reshape(b * video_length, l, d) + + self.scheduler = scheduler_copy + x_1k_0 = self.backward_loop( + timesteps=timesteps[-t1 - 1 :], + prompt_embeds=prompt_embeds, + latents=x_1k_t1, + guidance_scale=guidance_scale, + callback=callback, + callback_steps=callback_steps, + extra_step_kwargs=extra_step_kwargs, + num_warmup_steps=0, + ) + latents = x_1k_0 + + # manually for max memory savings + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.unet.to("cpu") + empty_device_cache() + + if output_type == "latent": + image = latents + has_nsfw_concept = None + else: + image = self.decode_latents(latents) + # Run safety checker + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + + # Offload all models + self.maybe_free_model_hooks() + # make sure to set the original attention processors back + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image, has_nsfw_concept) + + return TextToVideoPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + def decode_latents(self, latents): + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero_sdxl.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero_sdxl.py new file mode 100644 index 0000000000000000000000000000000000000000..a9fa43c1f5c528b2393098d1854fb46750637f95 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero_sdxl.py @@ -0,0 +1,1354 @@ +import copy +import inspect +from dataclasses import dataclass +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import numpy as np +import PIL +import torch +import torch.nn.functional as F +from torch.nn.functional import grid_sample +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...image_processor import VaeImageProcessor +from ...loaders import StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.attention_processor import ( + AttnProcessor2_0, + FusedAttnProcessor2_0, + XFormersAttnProcessor, +) +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + BaseOutput, + is_invisible_watermark_available, + logging, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin + + +if is_invisible_watermark_available(): + from ..stable_diffusion_xl.watermark import StableDiffusionXLWatermarker + + +from ...utils import is_torch_xla_available + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.rearrange_0 +def rearrange_0(tensor, f): + F, C, H, W = tensor.size() + tensor = torch.permute(torch.reshape(tensor, (F // f, f, C, H, W)), (0, 2, 1, 3, 4)) + return tensor + + +# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.rearrange_1 +def rearrange_1(tensor): + B, C, F, H, W = tensor.size() + return torch.reshape(torch.permute(tensor, (0, 2, 1, 3, 4)), (B * F, C, H, W)) + + +# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.rearrange_3 +def rearrange_3(tensor, f): + F, D, C = tensor.size() + return torch.reshape(tensor, (F // f, f, D, C)) + + +# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.rearrange_4 +def rearrange_4(tensor): + B, F, D, C = tensor.size() + return torch.reshape(tensor, (B * F, D, C)) + + +# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.CrossFrameAttnProcessor +class CrossFrameAttnProcessor: + """ + Cross frame attention processor. Each frame attends the first frame. + + Args: + batch_size: The number that represents actual batch size, other than the frames. + For example, calling unet with a single prompt and num_images_per_prompt=1, batch_size should be equal to + 2, due to classifier-free guidance. + """ + + def __init__(self, batch_size=2): + self.batch_size = batch_size + + def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None): + batch_size, sequence_length, _ = hidden_states.shape + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + query = attn.to_q(hidden_states) + + is_cross_attention = encoder_hidden_states is not None + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + # Cross Frame Attention + if not is_cross_attention: + video_length = key.size()[0] // self.batch_size + first_frame_index = [0] * video_length + + # rearrange keys to have batch and frames in the 1st and 2nd dims respectively + key = rearrange_3(key, video_length) + key = key[:, first_frame_index] + # rearrange values to have batch and frames in the 1st and 2nd dims respectively + value = rearrange_3(value, video_length) + value = value[:, first_frame_index] + + # rearrange back to original shape + key = rearrange_4(key) + value = rearrange_4(value) + + query = attn.head_to_batch_dim(query) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + return hidden_states + + +# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.CrossFrameAttnProcessor2_0 +class CrossFrameAttnProcessor2_0: + """ + Cross frame attention processor with scaled_dot_product attention of Pytorch 2.0. + + Args: + batch_size: The number that represents actual batch size, other than the frames. + For example, calling unet with a single prompt and num_images_per_prompt=1, batch_size should be equal to + 2, due to classifier-free guidance. + """ + + def __init__(self, batch_size=2): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") + self.batch_size = batch_size + + def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None): + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + inner_dim = hidden_states.shape[-1] + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + query = attn.to_q(hidden_states) + + is_cross_attention = encoder_hidden_states is not None + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + # Cross Frame Attention + if not is_cross_attention: + video_length = max(1, key.size()[0] // self.batch_size) + first_frame_index = [0] * video_length + + # rearrange keys to have batch and frames in the 1st and 2nd dims respectively + key = rearrange_3(key, video_length) + key = key[:, first_frame_index] + # rearrange values to have batch and frames in the 1st and 2nd dims respectively + value = rearrange_3(value, video_length) + value = value[:, first_frame_index] + + # rearrange back to original shape + key = rearrange_4(key) + value = rearrange_4(value) + + head_dim = inner_dim // attn.heads + query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + # the output of sdp = (batch, num_heads, seq_len, head_dim) + # TODO: add support for attn.scale when we move to Torch 2.1 + hidden_states = F.scaled_dot_product_attention( + query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + ) + + hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) + hidden_states = hidden_states.to(query.dtype) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + return hidden_states + + +@dataclass +class TextToVideoSDXLPipelineOutput(BaseOutput): + """ + Output class for zero-shot text-to-video pipeline. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] + + +# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.coords_grid +def coords_grid(batch, ht, wd, device): + # Adapted from https://github.com/princeton-vl/RAFT/blob/master/core/utils/utils.py + coords = torch.meshgrid(torch.arange(ht, device=device), torch.arange(wd, device=device)) + coords = torch.stack(coords[::-1], dim=0).float() + return coords[None].repeat(batch, 1, 1, 1) + + +# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.warp_single_latent +def warp_single_latent(latent, reference_flow): + """ + Warp latent of a single frame with given flow + + Args: + latent: latent code of a single frame + reference_flow: flow which to warp the latent with + + Returns: + warped: warped latent + """ + _, _, H, W = reference_flow.size() + _, _, h, w = latent.size() + coords0 = coords_grid(1, H, W, device=latent.device).to(latent.dtype) + + coords_t0 = coords0 + reference_flow + coords_t0[:, 0] /= W + coords_t0[:, 1] /= H + + coords_t0 = coords_t0 * 2.0 - 1.0 + coords_t0 = F.interpolate(coords_t0, size=(h, w), mode="bilinear") + coords_t0 = torch.permute(coords_t0, (0, 2, 3, 1)) + + warped = grid_sample(latent, coords_t0, mode="nearest", padding_mode="reflection") + return warped + + +# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.create_motion_field +def create_motion_field(motion_field_strength_x, motion_field_strength_y, frame_ids, device, dtype): + """ + Create translation motion field + + Args: + motion_field_strength_x: motion strength along x-axis + motion_field_strength_y: motion strength along y-axis + frame_ids: indexes of the frames the latents of which are being processed. + This is needed when we perform chunk-by-chunk inference + device: device + dtype: dtype + + Returns: + + """ + seq_length = len(frame_ids) + reference_flow = torch.zeros((seq_length, 2, 512, 512), device=device, dtype=dtype) + for fr_idx in range(seq_length): + reference_flow[fr_idx, 0, :, :] = motion_field_strength_x * (frame_ids[fr_idx]) + reference_flow[fr_idx, 1, :, :] = motion_field_strength_y * (frame_ids[fr_idx]) + return reference_flow + + +# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.create_motion_field_and_warp_latents +def create_motion_field_and_warp_latents(motion_field_strength_x, motion_field_strength_y, frame_ids, latents): + """ + Creates translation motion and warps the latents accordingly + + Args: + motion_field_strength_x: motion strength along x-axis + motion_field_strength_y: motion strength along y-axis + frame_ids: indexes of the frames the latents of which are being processed. + This is needed when we perform chunk-by-chunk inference + latents: latent codes of frames + + Returns: + warped_latents: warped latents + """ + motion_field = create_motion_field( + motion_field_strength_x=motion_field_strength_x, + motion_field_strength_y=motion_field_strength_y, + frame_ids=frame_ids, + device=latents.device, + dtype=latents.dtype, + ) + warped_latents = latents.clone().detach() + for i in range(len(warped_latents)): + warped_latents[i] = warp_single_latent(latents[i][None], motion_field[i][None]) + return warped_latents + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class TextToVideoZeroSDXLPipeline( + DeprecatedPipelineMixin, + DiffusionPipeline, + StableDiffusionMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +): + _last_supported_version = "0.33.1" + r""" + Pipeline for zero-shot text-to-video generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: Optional[bool] = None, + ): + super().__init__() + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.default_sample_size = ( + self.unet.config.sample_size + if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size") + else 128 + ) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.upcast_vae + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + FusedAttnProcessor2_0, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoZeroPipeline.forward_loop + def forward_loop(self, x_t0, t0, t1, generator): + """ + Perform DDPM forward process from time t0 to t1. This is the same as adding noise with corresponding variance. + + Args: + x_t0: + Latent code at time t0. + t0: + Timestep at t0. + t1: + Timestamp at t1. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + + Returns: + x_t1: + Forward process applied to x_t0 from time t0 to t1. + """ + eps = randn_tensor(x_t0.size(), generator=generator, dtype=x_t0.dtype, device=x_t0.device) + alpha_vec = torch.prod(self.scheduler.alphas[t0:t1]) + x_t1 = torch.sqrt(alpha_vec) * x_t0 + torch.sqrt(1 - alpha_vec) * eps + return x_t1 + + def backward_loop( + self, + latents, + timesteps, + prompt_embeds, + guidance_scale, + callback, + callback_steps, + num_warmup_steps, + extra_step_kwargs, + add_text_embeds, + add_time_ids, + cross_attention_kwargs=None, + guidance_rescale: float = 0.0, + ): + """ + Perform backward process given list of time steps + + Args: + latents: + Latents at time timesteps[0]. + timesteps: + Time steps along which to perform backward process. + prompt_embeds: + Pre-generated text embeddings. + guidance_scale: + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + extra_step_kwargs: + Extra_step_kwargs. + cross_attention_kwargs: + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + num_warmup_steps: + number of warmup steps. + + Returns: + latents: latents of backward process output at time timesteps[-1] + """ + + do_classifier_free_guidance = guidance_scale > 1.0 + num_steps = (len(timesteps) - num_warmup_steps) // self.scheduler.order + + with self.progress_bar(total=num_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + callback(i, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + return latents.clone().detach() + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]], + prompt_2: Optional[Union[str, List[str]]] = None, + video_length: Optional[int] = 8, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + denoising_end: Optional[float] = None, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_videos_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + frame_ids: Optional[List[int]] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + pooled_prompt_embeds: Optional[torch.Tensor] = None, + negative_pooled_prompt_embeds: Optional[torch.Tensor] = None, + latents: Optional[torch.Tensor] = None, + motion_field_strength_x: float = 12, + motion_field_strength_y: float = 12, + output_type: Optional[str] = "tensor", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + original_size: Optional[Tuple[int, int]] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Optional[Tuple[int, int]] = None, + t0: int = 44, + t1: int = 47, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + video_length (`int`, *optional*, defaults to 8): + The number of generated video frames. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + frame_ids (`List[int]`, *optional*): + Indexes of the frames that are being generated. This is used when generating longer videos + chunk-by-chunk. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + motion_field_strength_x (`float`, *optional*, defaults to 12): + Strength of motion in generated video along x-axis. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + motion_field_strength_y (`float`, *optional*, defaults to 12): + Strength of motion in generated video along y-axis. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py). + guidance_rescale (`float`, *optional*, defaults to 0.7): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + t0 (`int`, *optional*, defaults to 44): + Timestep t0. Should be in the range [0, num_inference_steps - 1]. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + t1 (`int`, *optional*, defaults to 47): + Timestep t0. Should be in the range [t0 + 1, num_inference_steps - 1]. See the + [paper](https://huggingface.co/papers/2303.13439), Sect. 3.3.1. + + Returns: + [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoSDXLPipelineOutput`] or + `tuple`: [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoSDXLPipelineOutput`] + if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + assert video_length > 0 + if frame_ids is None: + frame_ids = list(range(video_length)) + assert len(frame_ids) == video_length + + assert num_videos_per_prompt == 1 + + # set the processor + original_attn_proc = self.unet.attn_processors + processor = ( + CrossFrameAttnProcessor2_0(batch_size=2) + if hasattr(F, "scaled_dot_product_attention") + else CrossFrameAttnProcessor(batch_size=2) + ) + self.unet.set_attn_processor(processor) + + if isinstance(prompt, str): + prompt = [prompt] + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] + + # 0. Default height and width to unet + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) + + # 2. Define call parameters + batch_size = ( + 1 if isinstance(prompt, str) else len(prompt) if isinstance(prompt, list) else prompt_embeds.shape[0] + ) + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_videos_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_videos_per_prompt, 1) + + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + + # Perform the first backward process up to time T_1 + x_1_t1 = self.backward_loop( + timesteps=timesteps[: -t1 - 1], + prompt_embeds=prompt_embeds, + latents=latents, + guidance_scale=guidance_scale, + callback=callback, + callback_steps=callback_steps, + extra_step_kwargs=extra_step_kwargs, + num_warmup_steps=num_warmup_steps, + add_text_embeds=add_text_embeds, + add_time_ids=add_time_ids, + ) + + scheduler_copy = copy.deepcopy(self.scheduler) + + # Perform the second backward process up to time T_0 + x_1_t0 = self.backward_loop( + timesteps=timesteps[-t1 - 1 : -t0 - 1], + prompt_embeds=prompt_embeds, + latents=x_1_t1, + guidance_scale=guidance_scale, + callback=callback, + callback_steps=callback_steps, + extra_step_kwargs=extra_step_kwargs, + num_warmup_steps=0, + add_text_embeds=add_text_embeds, + add_time_ids=add_time_ids, + ) + + # Propagate first frame latents at time T_0 to remaining frames + x_2k_t0 = x_1_t0.repeat(video_length - 1, 1, 1, 1) + + # Add motion in latents at time T_0 + x_2k_t0 = create_motion_field_and_warp_latents( + motion_field_strength_x=motion_field_strength_x, + motion_field_strength_y=motion_field_strength_y, + latents=x_2k_t0, + frame_ids=frame_ids[1:], + ) + + # Perform forward process up to time T_1 + x_2k_t1 = self.forward_loop( + x_t0=x_2k_t0, + t0=timesteps[-t0 - 1].to(torch.long), + t1=timesteps[-t1 - 1].to(torch.long), + generator=generator, + ) + + # Perform backward process from time T_1 to 0 + latents = torch.cat([x_1_t1, x_2k_t1]) + + self.scheduler = scheduler_copy + timesteps = timesteps[-t1 - 1 :] + + b, l, d = prompt_embeds.size() + prompt_embeds = prompt_embeds[:, None].repeat(1, video_length, 1, 1).reshape(b * video_length, l, d) + + b, k = add_text_embeds.size() + add_text_embeds = add_text_embeds[:, None].repeat(1, video_length, 1).reshape(b * video_length, k) + + b, k = add_time_ids.size() + add_time_ids = add_time_ids[:, None].repeat(1, video_length, 1).reshape(b * video_length, k) + + # 7.1 Apply denoising_end + if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1: + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + x_1k_0 = self.backward_loop( + timesteps=timesteps, + prompt_embeds=prompt_embeds, + latents=latents, + guidance_scale=guidance_scale, + callback=callback, + callback_steps=callback_steps, + extra_step_kwargs=extra_step_kwargs, + num_warmup_steps=0, + add_text_embeds=add_text_embeds, + add_time_ids=add_time_ids, + ) + + latents = x_1k_0 + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + return TextToVideoSDXLPipelineOutput(images=image) + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + self.maybe_free_model_hooks() + # make sure to set the original attention processors back + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image,) + + return TextToVideoSDXLPipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/transformers_loading_utils.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/transformers_loading_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..b52d154d6ba2fb2e91a599731f3e4caf7aef5a77 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/transformers_loading_utils.py @@ -0,0 +1,121 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# +# 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 contextlib +import os +import tempfile +from typing import TYPE_CHECKING, Dict + +from huggingface_hub import DDUFEntry +from tqdm import tqdm + +from ..utils import is_safetensors_available, is_transformers_available, is_transformers_version + + +if TYPE_CHECKING: + from transformers import PreTrainedModel, PreTrainedTokenizer + +if is_transformers_available(): + from transformers import PreTrainedModel, PreTrainedTokenizer + +if is_safetensors_available(): + import safetensors.torch + + +def _load_tokenizer_from_dduf( + cls: "PreTrainedTokenizer", name: str, dduf_entries: Dict[str, DDUFEntry], **kwargs +) -> "PreTrainedTokenizer": + """ + Load a tokenizer from a DDUF archive. + + In practice, `transformers` do not provide a way to load a tokenizer from a DDUF archive. This function is a + workaround by extracting the tokenizer files from the DDUF archive and loading the tokenizer from the extracted + files. There is an extra cost of extracting the files, but of limited impact as the tokenizer files are usually + small-ish. + """ + with tempfile.TemporaryDirectory() as tmp_dir: + for entry_name, entry in dduf_entries.items(): + if entry_name.startswith(name + "/"): + tmp_entry_path = os.path.join(tmp_dir, *entry_name.split("/")) + # need to create intermediary directory if they don't exist + os.makedirs(os.path.dirname(tmp_entry_path), exist_ok=True) + with open(tmp_entry_path, "wb") as f: + with entry.as_mmap() as mm: + f.write(mm) + return cls.from_pretrained(os.path.dirname(tmp_entry_path), **kwargs) + + +def _load_transformers_model_from_dduf( + cls: "PreTrainedModel", name: str, dduf_entries: Dict[str, DDUFEntry], **kwargs +) -> "PreTrainedModel": + """ + Load a transformers model from a DDUF archive. + + In practice, `transformers` do not provide a way to load a model from a DDUF archive. This function is a workaround + by instantiating a model from the config file and loading the weights from the DDUF archive directly. + """ + config_file = dduf_entries.get(f"{name}/config.json") + if config_file is None: + raise EnvironmentError( + f"Could not find a config.json file for component {name} in DDUF file (contains {dduf_entries.keys()})." + ) + generation_config = dduf_entries.get(f"{name}/generation_config.json", None) + + weight_files = [ + entry + for entry_name, entry in dduf_entries.items() + if entry_name.startswith(f"{name}/") and entry_name.endswith(".safetensors") + ] + if not weight_files: + raise EnvironmentError( + f"Could not find any weight file for component {name} in DDUF file (contains {dduf_entries.keys()})." + ) + if not is_safetensors_available(): + raise EnvironmentError( + "Safetensors is not available, cannot load model from DDUF. Please `pip install safetensors`." + ) + if is_transformers_version("<", "4.47.0"): + raise ImportError( + "You need to install `transformers>4.47.0` in order to load a transformers model from a DDUF file. " + "You can install it with: `pip install --upgrade transformers`" + ) + + with tempfile.TemporaryDirectory() as tmp_dir: + from transformers import AutoConfig, GenerationConfig + + tmp_config_file = os.path.join(tmp_dir, "config.json") + with open(tmp_config_file, "w") as f: + f.write(config_file.read_text()) + config = AutoConfig.from_pretrained(tmp_config_file) + if generation_config is not None: + tmp_generation_config_file = os.path.join(tmp_dir, "generation_config.json") + with open(tmp_generation_config_file, "w") as f: + f.write(generation_config.read_text()) + generation_config = GenerationConfig.from_pretrained(tmp_generation_config_file) + state_dict = {} + with contextlib.ExitStack() as stack: + for entry in tqdm(weight_files, desc="Loading state_dict"): # Loop over safetensors files + # Memory-map the safetensors file + mmap = stack.enter_context(entry.as_mmap()) + # Load tensors from the memory-mapped file + tensors = safetensors.torch.load(mmap) + # Update the state dictionary with tensors + state_dict.update(tensors) + return cls.from_pretrained( + pretrained_model_name_or_path=None, + config=config, + generation_config=generation_config, + state_dict=state_dict, + **kwargs, + ) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..c89e899463beede59b8ccf02688f6168b8ee3d77 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__init__.py @@ -0,0 +1,52 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + is_torch_available, + is_transformers_available, + is_transformers_version, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import UnCLIPImageVariationPipeline, UnCLIPPipeline + + _dummy_objects.update( + {"UnCLIPImageVariationPipeline": UnCLIPImageVariationPipeline, "UnCLIPPipeline": UnCLIPPipeline} + ) +else: + _import_structure["pipeline_unclip"] = ["UnCLIPPipeline"] + _import_structure["pipeline_unclip_image_variation"] = ["UnCLIPImageVariationPipeline"] + _import_structure["text_proj"] = ["UnCLIPTextProjModel"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_unclip import UnCLIPPipeline + from .pipeline_unclip_image_variation import UnCLIPImageVariationPipeline + from .text_proj import UnCLIPTextProjModel + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6279c824a6966c8ca3c0efaa705281e5bc9fd713 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__pycache__/pipeline_unclip.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__pycache__/pipeline_unclip.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e93639a2a5e68973a9fe0797cb40ada3be9a2724 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__pycache__/pipeline_unclip.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__pycache__/pipeline_unclip_image_variation.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__pycache__/pipeline_unclip_image_variation.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0605fc333fc403b8f7066dba5111d35a73ac2bab Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__pycache__/pipeline_unclip_image_variation.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__pycache__/text_proj.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__pycache__/text_proj.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4b6e957252b1f4b98675b0a698b18f5878654e0f Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/__pycache__/text_proj.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/pipeline_unclip.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/pipeline_unclip.py new file mode 100644 index 0000000000000000000000000000000000000000..bbb9b0eb3ab27983361511da45b672933eeba472 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/pipeline_unclip.py @@ -0,0 +1,504 @@ +# Copyright 2025 Kakao Brain and The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import List, Optional, Tuple, Union + +import torch +from torch.nn import functional as F +from transformers import CLIPTextModelWithProjection, CLIPTokenizer +from transformers.models.clip.modeling_clip import CLIPTextModelOutput + +from ...models import PriorTransformer, UNet2DConditionModel, UNet2DModel +from ...schedulers import UnCLIPScheduler +from ...utils import is_torch_xla_available, logging +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput +from .text_proj import UnCLIPTextProjModel + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class UnCLIPPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + """ + Pipeline for text-to-image generation using unCLIP. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + text_encoder ([`~transformers.CLIPTextModelWithProjection`]): + Frozen text-encoder. + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + text_proj ([`UnCLIPTextProjModel`]): + Utility class to prepare and combine the embeddings before they are passed to the decoder. + decoder ([`UNet2DConditionModel`]): + The decoder to invert the image embedding into an image. + super_res_first ([`UNet2DModel`]): + Super resolution UNet. Used in all but the last step of the super resolution diffusion process. + super_res_last ([`UNet2DModel`]): + Super resolution UNet. Used in the last step of the super resolution diffusion process. + prior_scheduler ([`UnCLIPScheduler`]): + Scheduler used in the prior denoising process (a modified [`DDPMScheduler`]). + decoder_scheduler ([`UnCLIPScheduler`]): + Scheduler used in the decoder denoising process (a modified [`DDPMScheduler`]). + super_res_scheduler ([`UnCLIPScheduler`]): + Scheduler used in the super resolution denoising process (a modified [`DDPMScheduler`]). + + """ + + _last_supported_version = "0.33.1" + _exclude_from_cpu_offload = ["prior"] + + prior: PriorTransformer + decoder: UNet2DConditionModel + text_proj: UnCLIPTextProjModel + text_encoder: CLIPTextModelWithProjection + tokenizer: CLIPTokenizer + super_res_first: UNet2DModel + super_res_last: UNet2DModel + + prior_scheduler: UnCLIPScheduler + decoder_scheduler: UnCLIPScheduler + super_res_scheduler: UnCLIPScheduler + + model_cpu_offload_seq = "text_encoder->text_proj->decoder->super_res_first->super_res_last" + + def __init__( + self, + prior: PriorTransformer, + decoder: UNet2DConditionModel, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_proj: UnCLIPTextProjModel, + super_res_first: UNet2DModel, + super_res_last: UNet2DModel, + prior_scheduler: UnCLIPScheduler, + decoder_scheduler: UnCLIPScheduler, + super_res_scheduler: UnCLIPScheduler, + ): + super().__init__() + + self.register_modules( + prior=prior, + decoder=decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_proj=text_proj, + super_res_first=super_res_first, + super_res_last=super_res_last, + prior_scheduler=prior_scheduler, + decoder_scheduler=decoder_scheduler, + super_res_scheduler=super_res_scheduler, + ) + + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + text_model_output: Optional[Union[CLIPTextModelOutput, Tuple]] = None, + text_attention_mask: Optional[torch.Tensor] = None, + ): + if text_model_output is None: + batch_size = len(prompt) if isinstance(prompt, list) else 1 + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + text_mask = text_inputs.attention_mask.bool().to(device) + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] + + text_encoder_output = self.text_encoder(text_input_ids.to(device)) + + prompt_embeds = text_encoder_output.text_embeds + text_enc_hid_states = text_encoder_output.last_hidden_state + + else: + batch_size = text_model_output[0].shape[0] + prompt_embeds, text_enc_hid_states = text_model_output[0], text_model_output[1] + text_mask = text_attention_mask + + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_enc_hid_states = text_enc_hid_states.repeat_interleave(num_images_per_prompt, dim=0) + text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + uncond_tokens = [""] * batch_size + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + uncond_text_mask = uncond_input.attention_mask.bool().to(device) + negative_prompt_embeds_text_encoder_output = self.text_encoder(uncond_input.input_ids.to(device)) + + negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.text_embeds + uncond_text_enc_hid_states = negative_prompt_embeds_text_encoder_output.last_hidden_state + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len) + + seq_len = uncond_text_enc_hid_states.shape[1] + uncond_text_enc_hid_states = uncond_text_enc_hid_states.repeat(1, num_images_per_prompt, 1) + uncond_text_enc_hid_states = uncond_text_enc_hid_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + text_enc_hid_states = torch.cat([uncond_text_enc_hid_states, text_enc_hid_states]) + + text_mask = torch.cat([uncond_text_mask, text_mask]) + + return prompt_embeds, text_enc_hid_states, text_mask + + @torch.no_grad() + def __call__( + self, + prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: int = 1, + prior_num_inference_steps: int = 25, + decoder_num_inference_steps: int = 25, + super_res_num_inference_steps: int = 7, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + prior_latents: Optional[torch.Tensor] = None, + decoder_latents: Optional[torch.Tensor] = None, + super_res_latents: Optional[torch.Tensor] = None, + text_model_output: Optional[Union[CLIPTextModelOutput, Tuple]] = None, + text_attention_mask: Optional[torch.Tensor] = None, + prior_guidance_scale: float = 4.0, + decoder_guidance_scale: float = 8.0, + output_type: Optional[str] = "pil", + return_dict: bool = True, + ): + """ + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide image generation. This can only be left undefined if `text_model_output` + and `text_attention_mask` is passed. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + prior_num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps for the prior. More denoising steps usually lead to a higher quality + image at the expense of slower inference. + decoder_num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps for the decoder. More denoising steps usually lead to a higher quality + image at the expense of slower inference. + super_res_num_inference_steps (`int`, *optional*, defaults to 7): + The number of denoising steps for super resolution. More denoising steps usually lead to a higher + quality image at the expense of slower inference. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + prior_latents (`torch.Tensor` of shape (batch size, embeddings dimension), *optional*): + Pre-generated noisy latents to be used as inputs for the prior. + decoder_latents (`torch.Tensor` of shape (batch size, channels, height, width), *optional*): + Pre-generated noisy latents to be used as inputs for the decoder. + super_res_latents (`torch.Tensor` of shape (batch size, channels, super res height, super res width), *optional*): + Pre-generated noisy latents to be used as inputs for the decoder. + prior_guidance_scale (`float`, *optional*, defaults to 4.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + decoder_guidance_scale (`float`, *optional*, defaults to 4.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + text_model_output (`CLIPTextModelOutput`, *optional*): + Pre-defined [`CLIPTextModel`] outputs that can be derived from the text encoder. Pre-defined text + outputs can be passed for tasks like text embedding interpolations. Make sure to also pass + `text_attention_mask` in this case. `prompt` can the be left `None`. + text_attention_mask (`torch.Tensor`, *optional*): + Pre-defined CLIP text attention mask that can be derived from the tokenizer. Pre-defined text attention + masks are necessary when passing `text_model_output`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + if prompt is not None: + if isinstance(prompt, str): + batch_size = 1 + elif isinstance(prompt, list): + batch_size = len(prompt) + else: + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + else: + batch_size = text_model_output[0].shape[0] + + device = self._execution_device + + batch_size = batch_size * num_images_per_prompt + + do_classifier_free_guidance = prior_guidance_scale > 1.0 or decoder_guidance_scale > 1.0 + + prompt_embeds, text_enc_hid_states, text_mask = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance, text_model_output, text_attention_mask + ) + + # prior + + self.prior_scheduler.set_timesteps(prior_num_inference_steps, device=device) + prior_timesteps_tensor = self.prior_scheduler.timesteps + + embedding_dim = self.prior.config.embedding_dim + + prior_latents = self.prepare_latents( + (batch_size, embedding_dim), + prompt_embeds.dtype, + device, + generator, + prior_latents, + self.prior_scheduler, + ) + + for i, t in enumerate(self.progress_bar(prior_timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([prior_latents] * 2) if do_classifier_free_guidance else prior_latents + + predicted_image_embedding = self.prior( + latent_model_input, + timestep=t, + proj_embedding=prompt_embeds, + encoder_hidden_states=text_enc_hid_states, + attention_mask=text_mask, + ).predicted_image_embedding + + if do_classifier_free_guidance: + predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2) + predicted_image_embedding = predicted_image_embedding_uncond + prior_guidance_scale * ( + predicted_image_embedding_text - predicted_image_embedding_uncond + ) + + if i + 1 == prior_timesteps_tensor.shape[0]: + prev_timestep = None + else: + prev_timestep = prior_timesteps_tensor[i + 1] + + prior_latents = self.prior_scheduler.step( + predicted_image_embedding, + timestep=t, + sample=prior_latents, + generator=generator, + prev_timestep=prev_timestep, + ).prev_sample + + prior_latents = self.prior.post_process_latents(prior_latents) + + image_embeddings = prior_latents + + # done prior + + # decoder + + text_enc_hid_states, additive_clip_time_embeddings = self.text_proj( + image_embeddings=image_embeddings, + prompt_embeds=prompt_embeds, + text_encoder_hidden_states=text_enc_hid_states, + do_classifier_free_guidance=do_classifier_free_guidance, + ) + + if device.type == "mps": + # HACK: MPS: There is a panic when padding bool tensors, + # so cast to int tensor for the pad and back to bool afterwards + text_mask = text_mask.type(torch.int) + decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=1) + decoder_text_mask = decoder_text_mask.type(torch.bool) + else: + decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=True) + + self.decoder_scheduler.set_timesteps(decoder_num_inference_steps, device=device) + decoder_timesteps_tensor = self.decoder_scheduler.timesteps + + num_channels_latents = self.decoder.config.in_channels + height = self.decoder.config.sample_size + width = self.decoder.config.sample_size + + decoder_latents = self.prepare_latents( + (batch_size, num_channels_latents, height, width), + text_enc_hid_states.dtype, + device, + generator, + decoder_latents, + self.decoder_scheduler, + ) + + for i, t in enumerate(self.progress_bar(decoder_timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([decoder_latents] * 2) if do_classifier_free_guidance else decoder_latents + + noise_pred = self.decoder( + sample=latent_model_input, + timestep=t, + encoder_hidden_states=text_enc_hid_states, + class_labels=additive_clip_time_embeddings, + attention_mask=decoder_text_mask, + ).sample + + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred_uncond, _ = noise_pred_uncond.split(latent_model_input.shape[1], dim=1) + noise_pred_text, predicted_variance = noise_pred_text.split(latent_model_input.shape[1], dim=1) + noise_pred = noise_pred_uncond + decoder_guidance_scale * (noise_pred_text - noise_pred_uncond) + noise_pred = torch.cat([noise_pred, predicted_variance], dim=1) + + if i + 1 == decoder_timesteps_tensor.shape[0]: + prev_timestep = None + else: + prev_timestep = decoder_timesteps_tensor[i + 1] + + # compute the previous noisy sample x_t -> x_t-1 + decoder_latents = self.decoder_scheduler.step( + noise_pred, t, decoder_latents, prev_timestep=prev_timestep, generator=generator + ).prev_sample + + decoder_latents = decoder_latents.clamp(-1, 1) + + image_small = decoder_latents + + # done decoder + + # super res + + self.super_res_scheduler.set_timesteps(super_res_num_inference_steps, device=device) + super_res_timesteps_tensor = self.super_res_scheduler.timesteps + + channels = self.super_res_first.config.in_channels // 2 + height = self.super_res_first.config.sample_size + width = self.super_res_first.config.sample_size + + super_res_latents = self.prepare_latents( + (batch_size, channels, height, width), + image_small.dtype, + device, + generator, + super_res_latents, + self.super_res_scheduler, + ) + + if device.type == "mps": + # MPS does not support many interpolations + image_upscaled = F.interpolate(image_small, size=[height, width]) + else: + interpolate_antialias = {} + if "antialias" in inspect.signature(F.interpolate).parameters: + interpolate_antialias["antialias"] = True + + image_upscaled = F.interpolate( + image_small, size=[height, width], mode="bicubic", align_corners=False, **interpolate_antialias + ) + + for i, t in enumerate(self.progress_bar(super_res_timesteps_tensor)): + # no classifier free guidance + + if i == super_res_timesteps_tensor.shape[0] - 1: + unet = self.super_res_last + else: + unet = self.super_res_first + + latent_model_input = torch.cat([super_res_latents, image_upscaled], dim=1) + + noise_pred = unet( + sample=latent_model_input, + timestep=t, + ).sample + + if i + 1 == super_res_timesteps_tensor.shape[0]: + prev_timestep = None + else: + prev_timestep = super_res_timesteps_tensor[i + 1] + + # compute the previous noisy sample x_t -> x_t-1 + super_res_latents = self.super_res_scheduler.step( + noise_pred, t, super_res_latents, prev_timestep=prev_timestep, generator=generator + ).prev_sample + + if XLA_AVAILABLE: + xm.mark_step() + + image = super_res_latents + # done super res + + self.maybe_free_model_hooks() + + # post processing + image = image * 0.5 + 0.5 + image = image.clamp(0, 1) + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/pipeline_unclip_image_variation.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/pipeline_unclip_image_variation.py new file mode 100644 index 0000000000000000000000000000000000000000..31710a000e0a50242eab1623b9b9b8f8d92cd908 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/pipeline_unclip_image_variation.py @@ -0,0 +1,431 @@ +# Copyright 2025 Kakao Brain and The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import List, Optional, Union + +import PIL.Image +import torch +from torch.nn import functional as F +from transformers import ( + CLIPImageProcessor, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...models import UNet2DConditionModel, UNet2DModel +from ...schedulers import UnCLIPScheduler +from ...utils import is_torch_xla_available, logging +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput +from .text_proj import UnCLIPTextProjModel + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class UnCLIPImageVariationPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + """ + Pipeline to generate image variations from an input image using UnCLIP. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + text_encoder ([`~transformers.CLIPTextModelWithProjection`]): + Frozen text-encoder. + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `image_encoder`. + image_encoder ([`~transformers.CLIPVisionModelWithProjection`]): + Frozen CLIP image-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + text_proj ([`UnCLIPTextProjModel`]): + Utility class to prepare and combine the embeddings before they are passed to the decoder. + decoder ([`UNet2DConditionModel`]): + The decoder to invert the image embedding into an image. + super_res_first ([`UNet2DModel`]): + Super resolution UNet. Used in all but the last step of the super resolution diffusion process. + super_res_last ([`UNet2DModel`]): + Super resolution UNet. Used in the last step of the super resolution diffusion process. + decoder_scheduler ([`UnCLIPScheduler`]): + Scheduler used in the decoder denoising process (a modified [`DDPMScheduler`]). + super_res_scheduler ([`UnCLIPScheduler`]): + Scheduler used in the super resolution denoising process (a modified [`DDPMScheduler`]). + """ + + _last_supported_version = "0.33.1" + decoder: UNet2DConditionModel + text_proj: UnCLIPTextProjModel + text_encoder: CLIPTextModelWithProjection + tokenizer: CLIPTokenizer + feature_extractor: CLIPImageProcessor + image_encoder: CLIPVisionModelWithProjection + super_res_first: UNet2DModel + super_res_last: UNet2DModel + + decoder_scheduler: UnCLIPScheduler + super_res_scheduler: UnCLIPScheduler + model_cpu_offload_seq = "text_encoder->image_encoder->text_proj->decoder->super_res_first->super_res_last" + + def __init__( + self, + decoder: UNet2DConditionModel, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_proj: UnCLIPTextProjModel, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection, + super_res_first: UNet2DModel, + super_res_last: UNet2DModel, + decoder_scheduler: UnCLIPScheduler, + super_res_scheduler: UnCLIPScheduler, + ): + super().__init__() + + self.register_modules( + decoder=decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_proj=text_proj, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + super_res_first=super_res_first, + super_res_last=super_res_last, + decoder_scheduler=decoder_scheduler, + super_res_scheduler=super_res_scheduler, + ) + + # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance): + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + text_mask = text_inputs.attention_mask.bool().to(device) + text_encoder_output = self.text_encoder(text_input_ids.to(device)) + + prompt_embeds = text_encoder_output.text_embeds + text_encoder_hidden_states = text_encoder_output.last_hidden_state + + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + uncond_tokens = [""] * batch_size + + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + uncond_text_mask = uncond_input.attention_mask.bool().to(device) + negative_prompt_embeds_text_encoder_output = self.text_encoder(uncond_input.input_ids.to(device)) + + negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.text_embeds + uncond_text_encoder_hidden_states = negative_prompt_embeds_text_encoder_output.last_hidden_state + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len) + + seq_len = uncond_text_encoder_hidden_states.shape[1] + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1) + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states]) + + text_mask = torch.cat([uncond_text_mask, text_mask]) + + return prompt_embeds, text_encoder_hidden_states, text_mask + + def _encode_image(self, image, device, num_images_per_prompt, image_embeddings: Optional[torch.Tensor] = None): + dtype = next(self.image_encoder.parameters()).dtype + + if image_embeddings is None: + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(images=image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeddings = self.image_encoder(image).image_embeds + + image_embeddings = image_embeddings.repeat_interleave(num_images_per_prompt, dim=0) + + return image_embeddings + + @torch.no_grad() + def __call__( + self, + image: Optional[Union[PIL.Image.Image, List[PIL.Image.Image], torch.Tensor]] = None, + num_images_per_prompt: int = 1, + decoder_num_inference_steps: int = 25, + super_res_num_inference_steps: int = 7, + generator: Optional[torch.Generator] = None, + decoder_latents: Optional[torch.Tensor] = None, + super_res_latents: Optional[torch.Tensor] = None, + image_embeddings: Optional[torch.Tensor] = None, + decoder_guidance_scale: float = 8.0, + output_type: Optional[str] = "pil", + return_dict: bool = True, + ): + """ + The call function to the pipeline for generation. + + Args: + image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.Tensor`): + `Image` or tensor representing an image batch to be used as the starting point. If you provide a + tensor, it needs to be compatible with the [`CLIPImageProcessor`] + [configuration](https://huggingface.co/fusing/karlo-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json). + Can be left as `None` only when `image_embeddings` are passed. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + decoder_num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps for the decoder. More denoising steps usually lead to a higher quality + image at the expense of slower inference. + super_res_num_inference_steps (`int`, *optional*, defaults to 7): + The number of denoising steps for super resolution. More denoising steps usually lead to a higher + quality image at the expense of slower inference. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + decoder_latents (`torch.Tensor` of shape (batch size, channels, height, width), *optional*): + Pre-generated noisy latents to be used as inputs for the decoder. + super_res_latents (`torch.Tensor` of shape (batch size, channels, super res height, super res width), *optional*): + Pre-generated noisy latents to be used as inputs for the decoder. + decoder_guidance_scale (`float`, *optional*, defaults to 4.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + image_embeddings (`torch.Tensor`, *optional*): + Pre-defined image embeddings that can be derived from the image encoder. Pre-defined image embeddings + can be passed for tasks like image interpolations. `image` can be left as `None`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images. + """ + if image is not None: + if isinstance(image, PIL.Image.Image): + batch_size = 1 + elif isinstance(image, list): + batch_size = len(image) + else: + batch_size = image.shape[0] + else: + batch_size = image_embeddings.shape[0] + + prompt = [""] * batch_size + + device = self._execution_device + + batch_size = batch_size * num_images_per_prompt + + do_classifier_free_guidance = decoder_guidance_scale > 1.0 + + prompt_embeds, text_encoder_hidden_states, text_mask = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance + ) + + image_embeddings = self._encode_image(image, device, num_images_per_prompt, image_embeddings) + + # decoder + text_encoder_hidden_states, additive_clip_time_embeddings = self.text_proj( + image_embeddings=image_embeddings, + prompt_embeds=prompt_embeds, + text_encoder_hidden_states=text_encoder_hidden_states, + do_classifier_free_guidance=do_classifier_free_guidance, + ) + + if device.type == "mps": + # HACK: MPS: There is a panic when padding bool tensors, + # so cast to int tensor for the pad and back to bool afterwards + text_mask = text_mask.type(torch.int) + decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=1) + decoder_text_mask = decoder_text_mask.type(torch.bool) + else: + decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=True) + + self.decoder_scheduler.set_timesteps(decoder_num_inference_steps, device=device) + decoder_timesteps_tensor = self.decoder_scheduler.timesteps + + num_channels_latents = self.decoder.config.in_channels + height = self.decoder.config.sample_size + width = self.decoder.config.sample_size + + if decoder_latents is None: + decoder_latents = self.prepare_latents( + (batch_size, num_channels_latents, height, width), + text_encoder_hidden_states.dtype, + device, + generator, + decoder_latents, + self.decoder_scheduler, + ) + + for i, t in enumerate(self.progress_bar(decoder_timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([decoder_latents] * 2) if do_classifier_free_guidance else decoder_latents + + noise_pred = self.decoder( + sample=latent_model_input, + timestep=t, + encoder_hidden_states=text_encoder_hidden_states, + class_labels=additive_clip_time_embeddings, + attention_mask=decoder_text_mask, + ).sample + + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred_uncond, _ = noise_pred_uncond.split(latent_model_input.shape[1], dim=1) + noise_pred_text, predicted_variance = noise_pred_text.split(latent_model_input.shape[1], dim=1) + noise_pred = noise_pred_uncond + decoder_guidance_scale * (noise_pred_text - noise_pred_uncond) + noise_pred = torch.cat([noise_pred, predicted_variance], dim=1) + + if i + 1 == decoder_timesteps_tensor.shape[0]: + prev_timestep = None + else: + prev_timestep = decoder_timesteps_tensor[i + 1] + + # compute the previous noisy sample x_t -> x_t-1 + decoder_latents = self.decoder_scheduler.step( + noise_pred, t, decoder_latents, prev_timestep=prev_timestep, generator=generator + ).prev_sample + + decoder_latents = decoder_latents.clamp(-1, 1) + + image_small = decoder_latents + + # done decoder + + # super res + + self.super_res_scheduler.set_timesteps(super_res_num_inference_steps, device=device) + super_res_timesteps_tensor = self.super_res_scheduler.timesteps + + channels = self.super_res_first.config.in_channels // 2 + height = self.super_res_first.config.sample_size + width = self.super_res_first.config.sample_size + + if super_res_latents is None: + super_res_latents = self.prepare_latents( + (batch_size, channels, height, width), + image_small.dtype, + device, + generator, + super_res_latents, + self.super_res_scheduler, + ) + + if device.type == "mps": + # MPS does not support many interpolations + image_upscaled = F.interpolate(image_small, size=[height, width]) + else: + interpolate_antialias = {} + if "antialias" in inspect.signature(F.interpolate).parameters: + interpolate_antialias["antialias"] = True + + image_upscaled = F.interpolate( + image_small, size=[height, width], mode="bicubic", align_corners=False, **interpolate_antialias + ) + + for i, t in enumerate(self.progress_bar(super_res_timesteps_tensor)): + # no classifier free guidance + + if i == super_res_timesteps_tensor.shape[0] - 1: + unet = self.super_res_last + else: + unet = self.super_res_first + + latent_model_input = torch.cat([super_res_latents, image_upscaled], dim=1) + + noise_pred = unet( + sample=latent_model_input, + timestep=t, + ).sample + + if i + 1 == super_res_timesteps_tensor.shape[0]: + prev_timestep = None + else: + prev_timestep = super_res_timesteps_tensor[i + 1] + + # compute the previous noisy sample x_t -> x_t-1 + super_res_latents = self.super_res_scheduler.step( + noise_pred, t, super_res_latents, prev_timestep=prev_timestep, generator=generator + ).prev_sample + + if XLA_AVAILABLE: + xm.mark_step() + + image = super_res_latents + + # done super res + self.maybe_free_model_hooks() + + # post processing + + image = image * 0.5 + 0.5 + image = image.clamp(0, 1) + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/text_proj.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/text_proj.py new file mode 100644 index 0000000000000000000000000000000000000000..5e04e48ba621caa4a617e52c5b83a1573bb11975 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/unclip/text_proj.py @@ -0,0 +1,86 @@ +# Copyright 2025 Kakao Brain and The HuggingFace Team. All rights reserved. +# +# 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 torch import nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models import ModelMixin + + +class UnCLIPTextProjModel(ModelMixin, ConfigMixin): + """ + Utility class for CLIP embeddings. Used to combine the image and text embeddings into a format usable by the + decoder. + + For more details, see the original paper: https://huggingface.co/papers/2204.06125 section 2.1 + """ + + @register_to_config + def __init__( + self, + *, + clip_extra_context_tokens: int = 4, + clip_embeddings_dim: int = 768, + time_embed_dim: int, + cross_attention_dim, + ): + super().__init__() + + self.learned_classifier_free_guidance_embeddings = nn.Parameter(torch.zeros(clip_embeddings_dim)) + + # parameters for additional clip time embeddings + self.embedding_proj = nn.Linear(clip_embeddings_dim, time_embed_dim) + self.clip_image_embeddings_project_to_time_embeddings = nn.Linear(clip_embeddings_dim, time_embed_dim) + + # parameters for encoder hidden states + self.clip_extra_context_tokens = clip_extra_context_tokens + self.clip_extra_context_tokens_proj = nn.Linear( + clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim + ) + self.encoder_hidden_states_proj = nn.Linear(clip_embeddings_dim, cross_attention_dim) + self.text_encoder_hidden_states_norm = nn.LayerNorm(cross_attention_dim) + + def forward(self, *, image_embeddings, prompt_embeds, text_encoder_hidden_states, do_classifier_free_guidance): + if do_classifier_free_guidance: + # Add the classifier free guidance embeddings to the image embeddings + image_embeddings_batch_size = image_embeddings.shape[0] + classifier_free_guidance_embeddings = self.learned_classifier_free_guidance_embeddings.unsqueeze(0) + classifier_free_guidance_embeddings = classifier_free_guidance_embeddings.expand( + image_embeddings_batch_size, -1 + ) + image_embeddings = torch.cat([classifier_free_guidance_embeddings, image_embeddings], dim=0) + + # The image embeddings batch size and the text embeddings batch size are equal + assert image_embeddings.shape[0] == prompt_embeds.shape[0] + + batch_size = prompt_embeds.shape[0] + + # "Specifically, we modify the architecture described in Nichol et al. (2021) by projecting and + # adding CLIP embeddings to the existing timestep embedding, ... + time_projected_prompt_embeds = self.embedding_proj(prompt_embeds) + time_projected_image_embeddings = self.clip_image_embeddings_project_to_time_embeddings(image_embeddings) + additive_clip_time_embeddings = time_projected_image_embeddings + time_projected_prompt_embeds + + # ... and by projecting CLIP embeddings into four + # extra tokens of context that are concatenated to the sequence of outputs from the GLIDE text encoder" + clip_extra_context_tokens = self.clip_extra_context_tokens_proj(image_embeddings) + clip_extra_context_tokens = clip_extra_context_tokens.reshape(batch_size, -1, self.clip_extra_context_tokens) + clip_extra_context_tokens = clip_extra_context_tokens.permute(0, 2, 1) + + text_encoder_hidden_states = self.encoder_hidden_states_proj(text_encoder_hidden_states) + text_encoder_hidden_states = self.text_encoder_hidden_states_norm(text_encoder_hidden_states) + text_encoder_hidden_states = torch.cat([clip_extra_context_tokens, text_encoder_hidden_states], dim=1) + + return text_encoder_hidden_states, additive_clip_time_embeddings diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..1ac2b09a6e570087c80bc11bf1a8102dd4970b8f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__init__.py @@ -0,0 +1,58 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import ( + ImageTextPipelineOutput, + UniDiffuserPipeline, + ) + + _dummy_objects.update( + {"ImageTextPipelineOutput": ImageTextPipelineOutput, "UniDiffuserPipeline": UniDiffuserPipeline} + ) +else: + _import_structure["modeling_text_decoder"] = ["UniDiffuserTextDecoder"] + _import_structure["modeling_uvit"] = ["UniDiffuserModel", "UTransformer2DModel"] + _import_structure["pipeline_unidiffuser"] = ["ImageTextPipelineOutput", "UniDiffuserPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import ( + ImageTextPipelineOutput, + UniDiffuserPipeline, + ) + else: + from .modeling_text_decoder import UniDiffuserTextDecoder + from .modeling_uvit import UniDiffuserModel, UTransformer2DModel + from .pipeline_unidiffuser import ImageTextPipelineOutput, UniDiffuserPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..aacf9792405a61862d8c1fb3d751f4ec320341af Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__pycache__/modeling_text_decoder.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__pycache__/modeling_text_decoder.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0870b30ce29200ba6f94deb0a02a9ff128b65a9f Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__pycache__/modeling_text_decoder.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__pycache__/modeling_uvit.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__pycache__/modeling_uvit.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7667b7d779748c66f8bf6372e23c2de4727da7aa Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__pycache__/modeling_uvit.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__pycache__/pipeline_unidiffuser.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__pycache__/pipeline_unidiffuser.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..649c2751b33cb871591a081ea96eaa0d7c3197e0 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/__pycache__/pipeline_unidiffuser.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/modeling_text_decoder.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/modeling_text_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..0ddcbf735770aea23c08bbd3af15db9daa8cef9f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/modeling_text_decoder.py @@ -0,0 +1,296 @@ +from typing import Optional + +import numpy as np +import torch +from torch import nn +from transformers import GPT2Config, GPT2LMHeadModel +from transformers.modeling_utils import ModuleUtilsMixin + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models import ModelMixin + + +# Modified from ClipCaptionModel in https://github.com/thu-ml/unidiffuser/blob/main/libs/caption_decoder.py +class UniDiffuserTextDecoder(ModelMixin, ConfigMixin, ModuleUtilsMixin): + """ + Text decoder model for a image-text [UniDiffuser](https://huggingface.co/papers/2303.06555) model. This is used to + generate text from the UniDiffuser image-text embedding. + + Parameters: + prefix_length (`int`): + Max number of prefix tokens that will be supplied to the model. + prefix_inner_dim (`int`): + The hidden size of the incoming prefix embeddings. For UniDiffuser, this would be the hidden dim of the + CLIP text encoder. + prefix_hidden_dim (`int`, *optional*): + Hidden dim of the MLP if we encode the prefix. + vocab_size (`int`, *optional*, defaults to 50257): + Vocabulary size of the GPT-2 model. Defines the number of different tokens that can be represented by the + `inputs_ids` passed when calling [`GPT2Model`] or [`TFGPT2Model`]. + n_positions (`int`, *optional*, defaults to 1024): + The maximum sequence length that this model might ever be used with. Typically set this to something large + just in case (e.g., 512 or 1024 or 2048). + n_embd (`int`, *optional*, defaults to 768): + Dimensionality of the embeddings and hidden states. + n_layer (`int`, *optional*, defaults to 12): + Number of hidden layers in the Transformer encoder. + n_head (`int`, *optional*, defaults to 12): + Number of attention heads for each attention layer in the Transformer encoder. + n_inner (`int`, *optional*, defaults to None): + Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd + activation_function (`str`, *optional*, defaults to `"gelu"`): + Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new"]`. + resid_pdrop (`float`, *optional*, defaults to 0.1): + The dropout probability for all fully connected layers in the embeddings, encoder, and pooler. + embd_pdrop (`float`, *optional*, defaults to 0.1): + The dropout ratio for the embeddings. + attn_pdrop (`float`, *optional*, defaults to 0.1): + The dropout ratio for the attention. + layer_norm_epsilon (`float`, *optional*, defaults to 1e-5): + The epsilon to use in the layer normalization layers. + initializer_range (`float`, *optional*, defaults to 0.02): + The standard deviation of the truncated_normal_initializer for initializing all weight matrices. + scale_attn_weights (`bool`, *optional*, defaults to `True`): + Scale attention weights by dividing by sqrt(hidden_size).. + use_cache (`bool`, *optional*, defaults to `True`): + Whether or not the model should return the last key/values attentions (not used by all models). + scale_attn_by_inverse_layer_idx (`bool`, *optional*, defaults to `False`): + Whether to additionally scale attention weights by `1 / layer_idx + 1`. + reorder_and_upcast_attn (`bool`, *optional*, defaults to `False`): + Whether to scale keys (K) prior to computing attention (dot-product) and upcast attention + dot-product/softmax to float() when training with mixed precision. + """ + + _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.bias", r"h\.\d+\.attn\.masked_bias"] + + @register_to_config + def __init__( + self, + prefix_length: int, + prefix_inner_dim: int, + prefix_hidden_dim: Optional[int] = None, + vocab_size: int = 50257, # Start of GPT2 config args + n_positions: int = 1024, + n_embd: int = 768, + n_layer: int = 12, + n_head: int = 12, + n_inner: Optional[int] = None, + activation_function: str = "gelu_new", + resid_pdrop: float = 0.1, + embd_pdrop: float = 0.1, + attn_pdrop: float = 0.1, + layer_norm_epsilon: float = 1e-5, + initializer_range: float = 0.02, + scale_attn_weights: bool = True, + use_cache: bool = True, + scale_attn_by_inverse_layer_idx: bool = False, + reorder_and_upcast_attn: bool = False, + ): + super().__init__() + + self.prefix_length = prefix_length + + if prefix_inner_dim != n_embd and prefix_hidden_dim is None: + raise ValueError( + f"`prefix_hidden_dim` cannot be `None` when `prefix_inner_dim`: {prefix_hidden_dim} and" + f" `n_embd`: {n_embd} are not equal." + ) + + self.prefix_inner_dim = prefix_inner_dim + self.prefix_hidden_dim = prefix_hidden_dim + + self.encode_prefix = ( + nn.Linear(self.prefix_inner_dim, self.prefix_hidden_dim) + if self.prefix_hidden_dim is not None + else nn.Identity() + ) + self.decode_prefix = ( + nn.Linear(self.prefix_hidden_dim, n_embd) if self.prefix_hidden_dim is not None else nn.Identity() + ) + + gpt_config = GPT2Config( + vocab_size=vocab_size, + n_positions=n_positions, + n_embd=n_embd, + n_layer=n_layer, + n_head=n_head, + n_inner=n_inner, + activation_function=activation_function, + resid_pdrop=resid_pdrop, + embd_pdrop=embd_pdrop, + attn_pdrop=attn_pdrop, + layer_norm_epsilon=layer_norm_epsilon, + initializer_range=initializer_range, + scale_attn_weights=scale_attn_weights, + use_cache=use_cache, + scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx, + reorder_and_upcast_attn=reorder_and_upcast_attn, + ) + self.transformer = GPT2LMHeadModel(gpt_config) + + def forward( + self, + input_ids: torch.Tensor, + prefix_embeds: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + labels: Optional[torch.Tensor] = None, + ): + """ + Args: + input_ids (`torch.Tensor` of shape `(N, max_seq_len)`): + Text tokens to use for inference. + prefix_embeds (`torch.Tensor` of shape `(N, prefix_length, 768)`): + Prefix embedding to prepend to the embedded tokens. + attention_mask (`torch.Tensor` of shape `(N, prefix_length + max_seq_len, 768)`, *optional*): + Attention mask for the prefix embedding. + labels (`torch.Tensor`, *optional*): + Labels to use for language modeling. + """ + embedding_text = self.transformer.transformer.wte(input_ids) + hidden = self.encode_prefix(prefix_embeds) + prefix_embeds = self.decode_prefix(hidden) + embedding_cat = torch.cat((prefix_embeds, embedding_text), dim=1) + + if labels is not None: + dummy_token = self.get_dummy_token(input_ids.shape[0], input_ids.device) + labels = torch.cat((dummy_token, input_ids), dim=1) + out = self.transformer(inputs_embeds=embedding_cat, labels=labels, attention_mask=attention_mask) + if self.prefix_hidden_dim is not None: + return out, hidden + else: + return out + + def get_dummy_token(self, batch_size: int, device: torch.device) -> torch.Tensor: + return torch.zeros(batch_size, self.prefix_length, dtype=torch.int64, device=device) + + def encode(self, prefix): + return self.encode_prefix(prefix) + + @torch.no_grad() + def generate_captions(self, features, eos_token_id, device): + """ + Generate captions given text embedding features. Returns list[L]. + + Args: + features (`torch.Tensor` of shape `(B, L, D)`): + Text embedding features to generate captions from. + eos_token_id (`int`): + The token ID of the EOS token for the text decoder model. + device: + Device to perform text generation on. + + Returns: + `List[str]`: A list of strings generated from the decoder model. + """ + + features = torch.split(features, 1, dim=0) + generated_tokens = [] + generated_seq_lengths = [] + for feature in features: + feature = self.decode_prefix(feature.to(device)) # back to the clip feature + # Only support beam search for now + output_tokens, seq_lengths = self.generate_beam( + input_embeds=feature, device=device, eos_token_id=eos_token_id + ) + generated_tokens.append(output_tokens[0]) + generated_seq_lengths.append(seq_lengths[0]) + generated_tokens = torch.stack(generated_tokens) + generated_seq_lengths = torch.stack(generated_seq_lengths) + return generated_tokens, generated_seq_lengths + + @torch.no_grad() + def generate_beam( + self, + input_ids=None, + input_embeds=None, + device=None, + beam_size: int = 5, + entry_length: int = 67, + temperature: float = 1.0, + eos_token_id: Optional[int] = None, + ): + """ + Generates text using the given tokenizer and text prompt or token embedding via beam search. This + implementation is based on the beam search implementation from the [original UniDiffuser + code](https://github.com/thu-ml/unidiffuser/blob/main/libs/caption_decoder.py#L89). + + Args: + eos_token_id (`int`, *optional*): + The token ID of the EOS token for the text decoder model. + input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*): + Tokenizer indices of input sequence tokens in the vocabulary. One of `input_ids` and `input_embeds` + must be supplied. + input_embeds (`torch.Tensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*): + An embedded representation to directly pass to the transformer as a prefix for beam search. One of + `input_ids` and `input_embeds` must be supplied. + device: + The device to perform beam search on. + beam_size (`int`, *optional*, defaults to `5`): + The number of best states to store during beam search. + entry_length (`int`, *optional*, defaults to `67`): + The number of iterations to run beam search. + temperature (`float`, *optional*, defaults to 1.0): + The temperature to use when performing the softmax over logits from the decoding model. + + Returns: + `Tuple(torch.Tensor, torch.Tensor)`: A tuple of tensors where the first element is a tensor of generated + token sequences sorted by score in descending order, and the second element is the sequence lengths + corresponding to those sequences. + """ + # Generates text until stop_token is reached using beam search with the desired beam size. + stop_token_index = eos_token_id + tokens = None + scores = None + seq_lengths = torch.ones(beam_size, device=device, dtype=torch.int) + is_stopped = torch.zeros(beam_size, device=device, dtype=torch.bool) + + if input_embeds is not None: + generated = input_embeds + else: + generated = self.transformer.transformer.wte(input_ids) + + for i in range(entry_length): + outputs = self.transformer(inputs_embeds=generated) + logits = outputs.logits + logits = logits[:, -1, :] / (temperature if temperature > 0 else 1.0) + logits = logits.softmax(-1).log() + + if scores is None: + scores, next_tokens = logits.topk(beam_size, -1) + generated = generated.expand(beam_size, *generated.shape[1:]) + next_tokens, scores = next_tokens.permute(1, 0), scores.squeeze(0) + if tokens is None: + tokens = next_tokens + else: + tokens = tokens.expand(beam_size, *tokens.shape[1:]) + tokens = torch.cat((tokens, next_tokens), dim=1) + else: + logits[is_stopped] = -float(np.inf) + logits[is_stopped, 0] = 0 + scores_sum = scores[:, None] + logits + seq_lengths[~is_stopped] += 1 + scores_sum_average = scores_sum / seq_lengths[:, None] + scores_sum_average, next_tokens = scores_sum_average.view(-1).topk(beam_size, -1) + next_tokens_source = next_tokens // scores_sum.shape[1] + seq_lengths = seq_lengths[next_tokens_source] + next_tokens = next_tokens % scores_sum.shape[1] + next_tokens = next_tokens.unsqueeze(1) + tokens = tokens[next_tokens_source] + tokens = torch.cat((tokens, next_tokens), dim=1) + generated = generated[next_tokens_source] + scores = scores_sum_average * seq_lengths + is_stopped = is_stopped[next_tokens_source] + + next_token_embed = self.transformer.transformer.wte(next_tokens.squeeze()).view(generated.shape[0], 1, -1) + generated = torch.cat((generated, next_token_embed), dim=1) + is_stopped = is_stopped + next_tokens.eq(stop_token_index).squeeze() + if is_stopped.all(): + break + + scores = scores / seq_lengths + order = scores.argsort(descending=True) + # tokens tensors are already padded to max_seq_length + output_texts = [tokens[i] for i in order] + output_texts = torch.stack(output_texts, dim=0) + seq_lengths = torch.tensor([seq_lengths[i] for i in order], dtype=seq_lengths.dtype) + return output_texts, seq_lengths diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/modeling_uvit.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/modeling_uvit.py new file mode 100644 index 0000000000000000000000000000000000000000..2a04ec2e40309d41e4dea041c7f3398c351bc5dd --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/modeling_uvit.py @@ -0,0 +1,1197 @@ +import math +from typing import Optional, Union + +import torch +from torch import nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models import ModelMixin +from ...models.attention import FeedForward +from ...models.attention_processor import Attention +from ...models.embeddings import TimestepEmbedding, Timesteps, get_2d_sincos_pos_embed +from ...models.modeling_outputs import Transformer2DModelOutput +from ...models.normalization import AdaLayerNorm +from ...utils import logging + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def _no_grad_trunc_normal_(tensor, mean, std, a, b): + # Cut & paste from PyTorch official master until it's in a few official releases - RW + # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf + def norm_cdf(x): + # Computes standard normal cumulative distribution function + return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0 + + if (mean < a - 2 * std) or (mean > b + 2 * std): + logger.warning( + "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. " + "The distribution of values may be incorrect." + ) + + with torch.no_grad(): + # Values are generated by using a truncated uniform distribution and + # then using the inverse CDF for the normal distribution. + # Get upper and lower cdf values + l = norm_cdf((a - mean) / std) + u = norm_cdf((b - mean) / std) + + # Uniformly fill tensor with values from [l, u], then translate to + # [2l-1, 2u-1]. + tensor.uniform_(2 * l - 1, 2 * u - 1) + + # Use inverse cdf transform for normal distribution to get truncated + # standard normal + tensor.erfinv_() + + # Transform to proper mean, std + tensor.mul_(std * math.sqrt(2.0)) + tensor.add_(mean) + + # Clamp to ensure it's in the proper range + tensor.clamp_(min=a, max=b) + return tensor + + +def trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0): + # type: (torch.Tensor, float, float, float, float) -> torch.Tensor + r"""Fills the input Tensor with values drawn from a truncated + normal distribution. The values are effectively drawn from the normal distribution :math:`\mathcal{N}(\text{mean}, + \text{std}^2)` with values outside :math:`[a, b]` redrawn until they are within the bounds. The method used for + generating the random values works best when :math:`a \leq \text{mean} \leq b`. + + Args: + tensor: an n-dimensional `torch.Tensor` + mean: the mean of the normal distribution + std: the standard deviation of the normal distribution + a: the minimum cutoff value + b: the maximum cutoff value + Examples: + >>> w = torch.empty(3, 5) >>> nn.init.trunc_normal_(w) + """ + return _no_grad_trunc_normal_(tensor, mean, std, a, b) + + +class PatchEmbed(nn.Module): + """2D Image to Patch Embedding""" + + def __init__( + self, + height=224, + width=224, + patch_size=16, + in_channels=3, + embed_dim=768, + layer_norm=False, + flatten=True, + bias=True, + use_pos_embed=True, + ): + super().__init__() + + num_patches = (height // patch_size) * (width // patch_size) + self.flatten = flatten + self.layer_norm = layer_norm + + self.proj = nn.Conv2d( + in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias + ) + if layer_norm: + self.norm = nn.LayerNorm(embed_dim, elementwise_affine=False, eps=1e-6) + else: + self.norm = None + + self.use_pos_embed = use_pos_embed + if self.use_pos_embed: + pos_embed = get_2d_sincos_pos_embed(embed_dim, int(num_patches**0.5), output_type="pt") + self.register_buffer("pos_embed", pos_embed.float().unsqueeze(0), persistent=False) + + def forward(self, latent): + latent = self.proj(latent) + if self.flatten: + latent = latent.flatten(2).transpose(1, 2) # BCHW -> BNC + if self.layer_norm: + latent = self.norm(latent) + if self.use_pos_embed: + return latent + self.pos_embed + else: + return latent + + +class SkipBlock(nn.Module): + def __init__(self, dim: int): + super().__init__() + + self.skip_linear = nn.Linear(2 * dim, dim) + + # Use torch.nn.LayerNorm for now, following the original code + self.norm = nn.LayerNorm(dim) + + def forward(self, x, skip): + x = self.skip_linear(torch.cat([x, skip], dim=-1)) + x = self.norm(x) + + return x + + +# Modified to support both pre-LayerNorm and post-LayerNorm configurations +# Don't support AdaLayerNormZero for now +# Modified from diffusers.models.attention.BasicTransformerBlock +class UTransformerBlock(nn.Module): + r""" + A modification of BasicTransformerBlock which supports pre-LayerNorm and post-LayerNorm configurations. + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + activation_fn (`str`, *optional*, defaults to `"geglu"`): + Activation function to be used in feed-forward. + num_embeds_ada_norm (:obj: `int`, *optional*): + The number of diffusion steps used during training. See `Transformer2DModel`. + attention_bias (:obj: `bool`, *optional*, defaults to `False`): + Configure if the attentions should contain a bias parameter. + only_cross_attention (`bool`, *optional*): + Whether to use only cross-attention layers. In this case two cross attention layers are used. + double_self_attention (`bool`, *optional*): + Whether to use two self-attention layers. In this case no cross attention layers are used. + upcast_attention (`bool`, *optional*): + Whether to upcast the query and key to float32 when performing the attention calculation. + norm_elementwise_affine (`bool`, *optional*): + Whether to use learnable per-element affine parameters during layer normalization. + norm_type (`str`, defaults to `"layer_norm"`): + The layer norm implementation to use. + pre_layer_norm (`bool`, *optional*): + Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"), + as opposed to after ("post-LayerNorm"). Note that `BasicTransformerBlock` uses pre-LayerNorm, e.g. + `pre_layer_norm = True`. + final_dropout (`bool`, *optional*): + Whether to use a final Dropout layer after the feedforward network. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout=0.0, + cross_attention_dim: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + attention_bias: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_elementwise_affine: bool = True, + norm_type: str = "layer_norm", + pre_layer_norm: bool = True, + final_dropout: bool = False, + ): + super().__init__() + self.only_cross_attention = only_cross_attention + + self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm" + + self.pre_layer_norm = pre_layer_norm + + if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None: + raise ValueError( + f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to" + f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}." + ) + + # 1. Self-Attn + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim if only_cross_attention else None, + upcast_attention=upcast_attention, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None or double_self_attention: + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim if not double_self_attention else None, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + ) # is self-attn if encoder_hidden_states is none + else: + self.attn2 = None + + if self.use_ada_layer_norm: + self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) + else: + self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine) + + if cross_attention_dim is not None or double_self_attention: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + self.norm2 = ( + AdaLayerNorm(dim, num_embeds_ada_norm) + if self.use_ada_layer_norm + else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine) + ) + else: + self.norm2 = None + + # 3. Feed-forward + self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine) + self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout) + + def forward( + self, + hidden_states, + attention_mask=None, + encoder_hidden_states=None, + encoder_attention_mask=None, + timestep=None, + cross_attention_kwargs=None, + class_labels=None, + ): + # Pre-LayerNorm + if self.pre_layer_norm: + if self.use_ada_layer_norm: + norm_hidden_states = self.norm1(hidden_states, timestep) + else: + norm_hidden_states = self.norm1(hidden_states) + else: + norm_hidden_states = hidden_states + + # 1. Self-Attention + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + + # Post-LayerNorm + if not self.pre_layer_norm: + if self.use_ada_layer_norm: + attn_output = self.norm1(attn_output, timestep) + else: + attn_output = self.norm1(attn_output) + + hidden_states = attn_output + hidden_states + + if self.attn2 is not None: + # Pre-LayerNorm + if self.pre_layer_norm: + norm_hidden_states = ( + self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states) + ) + else: + norm_hidden_states = hidden_states + # TODO (Birch-San): Here we should prepare the encoder_attention mask correctly + # prepare attention mask here + + # 2. Cross-Attention + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=encoder_attention_mask, + **cross_attention_kwargs, + ) + + # Post-LayerNorm + if not self.pre_layer_norm: + attn_output = self.norm2(attn_output, timestep) if self.use_ada_layer_norm else self.norm2(attn_output) + + hidden_states = attn_output + hidden_states + + # 3. Feed-forward + # Pre-LayerNorm + if self.pre_layer_norm: + norm_hidden_states = self.norm3(hidden_states) + else: + norm_hidden_states = hidden_states + + ff_output = self.ff(norm_hidden_states) + + # Post-LayerNorm + if not self.pre_layer_norm: + ff_output = self.norm3(ff_output) + + hidden_states = ff_output + hidden_states + + return hidden_states + + +# Like UTransformerBlock except with LayerNorms on the residual backbone of the block +# Modified from diffusers.models.attention.BasicTransformerBlock +class UniDiffuserBlock(nn.Module): + r""" + A modification of BasicTransformerBlock which supports pre-LayerNorm and post-LayerNorm configurations and puts the + LayerNorms on the residual backbone of the block. This matches the transformer block in the [original UniDiffuser + implementation](https://github.com/thu-ml/unidiffuser/blob/main/libs/uvit_multi_post_ln_v1.py#L104). + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + activation_fn (`str`, *optional*, defaults to `"geglu"`): + Activation function to be used in feed-forward. + num_embeds_ada_norm (:obj: `int`, *optional*): + The number of diffusion steps used during training. See `Transformer2DModel`. + attention_bias (:obj: `bool`, *optional*, defaults to `False`): + Configure if the attentions should contain a bias parameter. + only_cross_attention (`bool`, *optional*): + Whether to use only cross-attention layers. In this case two cross attention layers are used. + double_self_attention (`bool`, *optional*): + Whether to use two self-attention layers. In this case no cross attention layers are used. + upcast_attention (`bool`, *optional*): + Whether to upcast the query and key to float() when performing the attention calculation. + norm_elementwise_affine (`bool`, *optional*): + Whether to use learnable per-element affine parameters during layer normalization. + norm_type (`str`, defaults to `"layer_norm"`): + The layer norm implementation to use. + pre_layer_norm (`bool`, *optional*): + Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"), + as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm + (`pre_layer_norm = False`). + final_dropout (`bool`, *optional*): + Whether to use a final Dropout layer after the feedforward network. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout=0.0, + cross_attention_dim: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + attention_bias: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_elementwise_affine: bool = True, + norm_type: str = "layer_norm", + pre_layer_norm: bool = False, + final_dropout: bool = True, + ): + super().__init__() + self.only_cross_attention = only_cross_attention + + self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm" + + self.pre_layer_norm = pre_layer_norm + + if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None: + raise ValueError( + f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to" + f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}." + ) + + # 1. Self-Attn + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim if only_cross_attention else None, + upcast_attention=upcast_attention, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None or double_self_attention: + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim if not double_self_attention else None, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + ) # is self-attn if encoder_hidden_states is none + else: + self.attn2 = None + + if self.use_ada_layer_norm: + self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) + else: + self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine) + + if cross_attention_dim is not None or double_self_attention: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + self.norm2 = ( + AdaLayerNorm(dim, num_embeds_ada_norm) + if self.use_ada_layer_norm + else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine) + ) + else: + self.norm2 = None + + # 3. Feed-forward + self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine) + self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout) + + def forward( + self, + hidden_states, + attention_mask=None, + encoder_hidden_states=None, + encoder_attention_mask=None, + timestep=None, + cross_attention_kwargs=None, + class_labels=None, + ): + # Following the diffusers transformer block implementation, put the LayerNorm on the + # residual backbone + # Pre-LayerNorm + if self.pre_layer_norm: + if self.use_ada_layer_norm: + hidden_states = self.norm1(hidden_states, timestep) + else: + hidden_states = self.norm1(hidden_states) + + # 1. Self-Attention + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + attn_output = self.attn1( + hidden_states, + encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + + hidden_states = attn_output + hidden_states + + # Following the diffusers transformer block implementation, put the LayerNorm on the + # residual backbone + # Post-LayerNorm + if not self.pre_layer_norm: + if self.use_ada_layer_norm: + hidden_states = self.norm1(hidden_states, timestep) + else: + hidden_states = self.norm1(hidden_states) + + if self.attn2 is not None: + # Pre-LayerNorm + if self.pre_layer_norm: + hidden_states = ( + self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states) + ) + # TODO (Birch-San): Here we should prepare the encoder_attention mask correctly + # prepare attention mask here + + # 2. Cross-Attention + attn_output = self.attn2( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=encoder_attention_mask, + **cross_attention_kwargs, + ) + + hidden_states = attn_output + hidden_states + + # Post-LayerNorm + if not self.pre_layer_norm: + hidden_states = ( + self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states) + ) + + # 3. Feed-forward + # Pre-LayerNorm + if self.pre_layer_norm: + hidden_states = self.norm3(hidden_states) + + ff_output = self.ff(hidden_states) + + hidden_states = ff_output + hidden_states + + # Post-LayerNorm + if not self.pre_layer_norm: + hidden_states = self.norm3(hidden_states) + + return hidden_states + + +# Modified from diffusers.models.transformer_2d.Transformer2DModel +# Modify the transformer block structure to be U-Net like following U-ViT +# Only supports patch-style input and torch.nn.LayerNorm currently +# https://github.com/baofff/U-ViT +class UTransformer2DModel(ModelMixin, ConfigMixin): + """ + Transformer model based on the [U-ViT](https://github.com/baofff/U-ViT) architecture for image-like data. Compared + to [`Transformer2DModel`], this model has skip connections between transformer blocks in a "U"-shaped fashion, + similar to a U-Net. Supports only continuous (actual embeddings) inputs, which are embedded via a [`PatchEmbed`] + layer and then reshaped to (b, t, d). + + Parameters: + num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. + attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. + in_channels (`int`, *optional*): + Pass if the input is continuous. The number of channels in the input. + out_channels (`int`, *optional*): + The number of output channels; if `None`, defaults to `in_channels`. + num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + norm_num_groups (`int`, *optional*, defaults to `32`): + The number of groups to use when performing Group Normalization. + cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use. + attention_bias (`bool`, *optional*): + Configure if the TransformerBlocks' attention should contain a bias parameter. + sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images. + Note that this is fixed at training time as it is used for learning a number of position embeddings. See + `ImagePositionalEmbeddings`. + num_vector_embeds (`int`, *optional*): + Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels. + Includes the class for the masked latent pixel. + patch_size (`int`, *optional*, defaults to 2): + The patch size to use in the patch embedding. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`. + The number of diffusion steps used during training. Note that this is fixed at training time as it is used + to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for + up to but not more than steps than `num_embeds_ada_norm`. + use_linear_projection (int, *optional*): TODO: Not used + only_cross_attention (`bool`, *optional*): + Whether to use only cross-attention layers. In this case two cross attention layers are used in each + transformer block. + upcast_attention (`bool`, *optional*): + Whether to upcast the query and key to float() when performing the attention calculation. + norm_type (`str`, *optional*, defaults to `"layer_norm"`): + The Layer Normalization implementation to use. Defaults to `torch.nn.LayerNorm`. + block_type (`str`, *optional*, defaults to `"unidiffuser"`): + The transformer block implementation to use. If `"unidiffuser"`, has the LayerNorms on the residual + backbone of each transformer block; otherwise has them in the attention/feedforward branches (the standard + behavior in `diffusers`.) + pre_layer_norm (`bool`, *optional*): + Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"), + as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm + (`pre_layer_norm = False`). + norm_elementwise_affine (`bool`, *optional*): + Whether to use learnable per-element affine parameters during layer normalization. + use_patch_pos_embed (`bool`, *optional*): + Whether to use position embeddings inside the patch embedding layer (`PatchEmbed`). + final_dropout (`bool`, *optional*): + Whether to use a final Dropout layer after the feedforward network. + """ + + @register_to_config + def __init__( + self, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: Optional[int] = None, + out_channels: Optional[int] = None, + num_layers: int = 1, + dropout: float = 0.0, + norm_num_groups: int = 32, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + sample_size: Optional[int] = None, + num_vector_embeds: Optional[int] = None, + patch_size: Optional[int] = 2, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + norm_type: str = "layer_norm", + block_type: str = "unidiffuser", + pre_layer_norm: bool = False, + norm_elementwise_affine: bool = True, + use_patch_pos_embed=False, + ff_final_dropout: bool = False, + ): + super().__init__() + self.use_linear_projection = use_linear_projection + self.num_attention_heads = num_attention_heads + self.attention_head_dim = attention_head_dim + inner_dim = num_attention_heads * attention_head_dim + + # 1. Input + # Only support patch input of shape (batch_size, num_channels, height, width) for now + assert in_channels is not None and patch_size is not None, "Patch input requires in_channels and patch_size." + + assert sample_size is not None, "UTransformer2DModel over patched input must provide sample_size" + + # 2. Define input layers + self.height = sample_size + self.width = sample_size + + self.patch_size = patch_size + self.pos_embed = PatchEmbed( + height=sample_size, + width=sample_size, + patch_size=patch_size, + in_channels=in_channels, + embed_dim=inner_dim, + use_pos_embed=use_patch_pos_embed, + ) + + # 3. Define transformers blocks + # Modify this to have in_blocks ("downsample" blocks, even though we don't actually downsample), a mid_block, + # and out_blocks ("upsample" blocks). Like a U-Net, there are skip connections from in_blocks to out_blocks in + # a "U"-shaped fashion (e.g. first in_block to last out_block, etc.). + # Quick hack to make the transformer block type configurable + if block_type == "unidiffuser": + block_cls = UniDiffuserBlock + else: + block_cls = UTransformerBlock + self.transformer_in_blocks = nn.ModuleList( + [ + block_cls( + inner_dim, + num_attention_heads, + attention_head_dim, + dropout=dropout, + cross_attention_dim=cross_attention_dim, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + attention_bias=attention_bias, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + norm_type=norm_type, + pre_layer_norm=pre_layer_norm, + norm_elementwise_affine=norm_elementwise_affine, + final_dropout=ff_final_dropout, + ) + for d in range(num_layers // 2) + ] + ) + + self.transformer_mid_block = block_cls( + inner_dim, + num_attention_heads, + attention_head_dim, + dropout=dropout, + cross_attention_dim=cross_attention_dim, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + attention_bias=attention_bias, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + norm_type=norm_type, + pre_layer_norm=pre_layer_norm, + norm_elementwise_affine=norm_elementwise_affine, + final_dropout=ff_final_dropout, + ) + + # For each skip connection, we use a SkipBlock (concatenation + Linear + LayerNorm) to process the inputs + # before each transformer out_block. + self.transformer_out_blocks = nn.ModuleList( + [ + nn.ModuleDict( + { + "skip": SkipBlock( + inner_dim, + ), + "block": block_cls( + inner_dim, + num_attention_heads, + attention_head_dim, + dropout=dropout, + cross_attention_dim=cross_attention_dim, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + attention_bias=attention_bias, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + norm_type=norm_type, + pre_layer_norm=pre_layer_norm, + norm_elementwise_affine=norm_elementwise_affine, + final_dropout=ff_final_dropout, + ), + } + ) + for d in range(num_layers // 2) + ] + ) + + # 4. Define output layers + self.out_channels = in_channels if out_channels is None else out_channels + + # Following the UniDiffuser U-ViT implementation, we process the transformer output with + # a LayerNorm layer with per-element affine params + self.norm_out = nn.LayerNorm(inner_dim) + + def forward( + self, + hidden_states, + encoder_hidden_states=None, + timestep=None, + class_labels=None, + cross_attention_kwargs=None, + return_dict: bool = True, + hidden_states_is_embedding: bool = False, + unpatchify: bool = True, + ): + """ + Args: + hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`. + When continuous, `torch.Tensor` of shape `(batch size, channel, height, width)`): Input hidden_states + encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*): + Conditional embeddings for cross attention layer. If not given, cross-attention defaults to + self-attention. + timestep ( `torch.long`, *optional*): + Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step. + class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*): + Optional class labels to be applied as an embedding in AdaLayerZeroNorm. Used to indicate class labels + conditioning. + cross_attention_kwargs (*optional*): + Keyword arguments to supply to the cross attention layers, if used. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + hidden_states_is_embedding (`bool`, *optional*, defaults to `False`): + Whether or not hidden_states is an embedding directly usable by the transformer. In this case we will + ignore input handling (e.g. continuous, vectorized, etc.) and directly feed hidden_states into the + transformer blocks. + unpatchify (`bool`, *optional*, defaults to `True`): + Whether to unpatchify the transformer output. + + Returns: + [`~models.transformer_2d.Transformer2DModelOutput`] or `tuple`: + [`~models.transformer_2d.Transformer2DModelOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is the sample tensor. + """ + # 0. Check inputs + + if not unpatchify and return_dict: + raise ValueError( + f"Cannot both define `unpatchify`: {unpatchify} and `return_dict`: {return_dict} since when" + f" `unpatchify` is {unpatchify} the returned output is of shape (batch_size, seq_len, hidden_dim)" + " rather than (batch_size, num_channels, height, width)." + ) + + # 1. Input + if not hidden_states_is_embedding: + hidden_states = self.pos_embed(hidden_states) + + # 2. Blocks + + # In ("downsample") blocks + skips = [] + for in_block in self.transformer_in_blocks: + hidden_states = in_block( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + timestep=timestep, + cross_attention_kwargs=cross_attention_kwargs, + class_labels=class_labels, + ) + skips.append(hidden_states) + + # Mid block + hidden_states = self.transformer_mid_block(hidden_states) + + # Out ("upsample") blocks + for out_block in self.transformer_out_blocks: + hidden_states = out_block["skip"](hidden_states, skips.pop()) + hidden_states = out_block["block"]( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + timestep=timestep, + cross_attention_kwargs=cross_attention_kwargs, + class_labels=class_labels, + ) + + # 3. Output + # Don't support AdaLayerNorm for now, so no conditioning/scale/shift logic + hidden_states = self.norm_out(hidden_states) + # hidden_states = self.proj_out(hidden_states) + + if unpatchify: + # unpatchify + height = width = int(hidden_states.shape[1] ** 0.5) + hidden_states = hidden_states.reshape( + shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels) + ) + hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states) + output = hidden_states.reshape( + shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size) + ) + else: + output = hidden_states + + if not return_dict: + return (output,) + + return Transformer2DModelOutput(sample=output) + + +class UniDiffuserModel(ModelMixin, ConfigMixin): + """ + Transformer model for a image-text [UniDiffuser](https://huggingface.co/papers/2303.06555) model. This is a + modification of [`UTransformer2DModel`] with input and output heads for the VAE-embedded latent image, the + CLIP-embedded image, and the CLIP-embedded prompt (see paper for more details). + + Parameters: + text_dim (`int`): The hidden dimension of the CLIP text model used to embed images. + clip_img_dim (`int`): The hidden dimension of the CLIP vision model used to embed prompts. + num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. + attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. + in_channels (`int`, *optional*): + Pass if the input is continuous. The number of channels in the input. + out_channels (`int`, *optional*): + The number of output channels; if `None`, defaults to `in_channels`. + num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + norm_num_groups (`int`, *optional*, defaults to `32`): + The number of groups to use when performing Group Normalization. + cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use. + attention_bias (`bool`, *optional*): + Configure if the TransformerBlocks' attention should contain a bias parameter. + sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images. + Note that this is fixed at training time as it is used for learning a number of position embeddings. See + `ImagePositionalEmbeddings`. + num_vector_embeds (`int`, *optional*): + Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels. + Includes the class for the masked latent pixel. + patch_size (`int`, *optional*, defaults to 2): + The patch size to use in the patch embedding. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`. + The number of diffusion steps used during training. Note that this is fixed at training time as it is used + to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for + up to but not more than steps than `num_embeds_ada_norm`. + use_linear_projection (int, *optional*): TODO: Not used + only_cross_attention (`bool`, *optional*): + Whether to use only cross-attention layers. In this case two cross attention layers are used in each + transformer block. + upcast_attention (`bool`, *optional*): + Whether to upcast the query and key to float32 when performing the attention calculation. + norm_type (`str`, *optional*, defaults to `"layer_norm"`): + The Layer Normalization implementation to use. Defaults to `torch.nn.LayerNorm`. + block_type (`str`, *optional*, defaults to `"unidiffuser"`): + The transformer block implementation to use. If `"unidiffuser"`, has the LayerNorms on the residual + backbone of each transformer block; otherwise has them in the attention/feedforward branches (the standard + behavior in `diffusers`.) + pre_layer_norm (`bool`, *optional*): + Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"), + as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm + (`pre_layer_norm = False`). + norm_elementwise_affine (`bool`, *optional*): + Whether to use learnable per-element affine parameters during layer normalization. + use_patch_pos_embed (`bool`, *optional*): + Whether to use position embeddings inside the patch embedding layer (`PatchEmbed`). + ff_final_dropout (`bool`, *optional*): + Whether to use a final Dropout layer after the feedforward network. + use_data_type_embedding (`bool`, *optional*): + Whether to use a data type embedding. This is only relevant for UniDiffuser-v1 style models; UniDiffuser-v1 + is continue-trained from UniDiffuser-v0 on non-publically-available data and accepts a `data_type` + argument, which can either be `1` to use the weights trained on non-publically-available data or `0` + otherwise. This argument is subsequently embedded by the data type embedding, if used. + """ + + @register_to_config + def __init__( + self, + text_dim: int = 768, + clip_img_dim: int = 512, + num_text_tokens: int = 77, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: Optional[int] = None, + out_channels: Optional[int] = None, + num_layers: int = 1, + dropout: float = 0.0, + norm_num_groups: int = 32, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + sample_size: Optional[int] = None, + num_vector_embeds: Optional[int] = None, + patch_size: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + norm_type: str = "layer_norm", + block_type: str = "unidiffuser", + pre_layer_norm: bool = False, + use_timestep_embedding=False, + norm_elementwise_affine: bool = True, + use_patch_pos_embed=False, + ff_final_dropout: bool = True, + use_data_type_embedding: bool = False, + ): + super().__init__() + + # 0. Handle dimensions + self.inner_dim = num_attention_heads * attention_head_dim + + assert sample_size is not None, "UniDiffuserModel over patched input must provide sample_size" + self.sample_size = sample_size + self.in_channels = in_channels + self.out_channels = in_channels if out_channels is None else out_channels + + self.patch_size = patch_size + # Assume image is square... + self.num_patches = (self.sample_size // patch_size) * (self.sample_size // patch_size) + + # 1. Define input layers + # 1.1 Input layers for text and image input + # For now, only support patch input for VAE latent image input + self.vae_img_in = PatchEmbed( + height=sample_size, + width=sample_size, + patch_size=patch_size, + in_channels=in_channels, + embed_dim=self.inner_dim, + use_pos_embed=use_patch_pos_embed, + ) + self.clip_img_in = nn.Linear(clip_img_dim, self.inner_dim) + self.text_in = nn.Linear(text_dim, self.inner_dim) + + # 1.2. Timestep embeddings for t_img, t_text + self.timestep_img_proj = Timesteps( + self.inner_dim, + flip_sin_to_cos=True, + downscale_freq_shift=0, + ) + self.timestep_img_embed = ( + TimestepEmbedding( + self.inner_dim, + 4 * self.inner_dim, + out_dim=self.inner_dim, + ) + if use_timestep_embedding + else nn.Identity() + ) + + self.timestep_text_proj = Timesteps( + self.inner_dim, + flip_sin_to_cos=True, + downscale_freq_shift=0, + ) + self.timestep_text_embed = ( + TimestepEmbedding( + self.inner_dim, + 4 * self.inner_dim, + out_dim=self.inner_dim, + ) + if use_timestep_embedding + else nn.Identity() + ) + + # 1.3. Positional embedding + self.num_text_tokens = num_text_tokens + self.num_tokens = 1 + 1 + num_text_tokens + 1 + self.num_patches + self.pos_embed = nn.Parameter(torch.zeros(1, self.num_tokens, self.inner_dim)) + self.pos_embed_drop = nn.Dropout(p=dropout) + trunc_normal_(self.pos_embed, std=0.02) + + # 1.4. Handle data type token embeddings for UniDiffuser-V1, if necessary + self.use_data_type_embedding = use_data_type_embedding + if self.use_data_type_embedding: + self.data_type_token_embedding = nn.Embedding(2, self.inner_dim) + self.data_type_pos_embed_token = nn.Parameter(torch.zeros(1, 1, self.inner_dim)) + + # 2. Define transformer blocks + self.transformer = UTransformer2DModel( + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + in_channels=in_channels, + out_channels=out_channels, + num_layers=num_layers, + dropout=dropout, + norm_num_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim, + attention_bias=attention_bias, + sample_size=sample_size, + num_vector_embeds=num_vector_embeds, + patch_size=patch_size, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + norm_type=norm_type, + block_type=block_type, + pre_layer_norm=pre_layer_norm, + norm_elementwise_affine=norm_elementwise_affine, + use_patch_pos_embed=use_patch_pos_embed, + ff_final_dropout=ff_final_dropout, + ) + + # 3. Define output layers + patch_dim = (patch_size**2) * out_channels + self.vae_img_out = nn.Linear(self.inner_dim, patch_dim) + self.clip_img_out = nn.Linear(self.inner_dim, clip_img_dim) + self.text_out = nn.Linear(self.inner_dim, text_dim) + + @torch.jit.ignore + def no_weight_decay(self): + return {"pos_embed"} + + def forward( + self, + latent_image_embeds: torch.Tensor, + image_embeds: torch.Tensor, + prompt_embeds: torch.Tensor, + timestep_img: Union[torch.Tensor, float, int], + timestep_text: Union[torch.Tensor, float, int], + data_type: Optional[Union[torch.Tensor, float, int]] = 1, + encoder_hidden_states=None, + cross_attention_kwargs=None, + ): + """ + Args: + latent_image_embeds (`torch.Tensor` of shape `(batch size, latent channels, height, width)`): + Latent image representation from the VAE encoder. + image_embeds (`torch.Tensor` of shape `(batch size, 1, clip_img_dim)`): + CLIP-embedded image representation (unsqueezed in the first dimension). + prompt_embeds (`torch.Tensor` of shape `(batch size, seq_len, text_dim)`): + CLIP-embedded text representation. + timestep_img (`torch.long` or `float` or `int`): + Current denoising step for the image. + timestep_text (`torch.long` or `float` or `int`): + Current denoising step for the text. + data_type: (`torch.int` or `float` or `int`, *optional*, defaults to `1`): + Only used in UniDiffuser-v1-style models. Can be either `1`, to use weights trained on nonpublic data, + or `0` otherwise. + encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*): + Conditional embeddings for cross attention layer. If not given, cross-attention defaults to + self-attention. + cross_attention_kwargs (*optional*): + Keyword arguments to supply to the cross attention layers, if used. + + + Returns: + `tuple`: Returns relevant parts of the model's noise prediction: the first element of the tuple is tbe VAE + image embedding, the second element is the CLIP image embedding, and the third element is the CLIP text + embedding. + """ + batch_size = latent_image_embeds.shape[0] + + # 1. Input + # 1.1. Map inputs to shape (B, N, inner_dim) + vae_hidden_states = self.vae_img_in(latent_image_embeds) + clip_hidden_states = self.clip_img_in(image_embeds) + text_hidden_states = self.text_in(prompt_embeds) + + num_text_tokens, num_img_tokens = text_hidden_states.size(1), vae_hidden_states.size(1) + + # 1.2. Encode image timesteps to single token (B, 1, inner_dim) + if not torch.is_tensor(timestep_img): + timestep_img = torch.tensor([timestep_img], dtype=torch.long, device=vae_hidden_states.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep_img = timestep_img * torch.ones(batch_size, dtype=timestep_img.dtype, device=timestep_img.device) + + timestep_img_token = self.timestep_img_proj(timestep_img) + # t_img_token does not contain any weights and will always return f32 tensors + # but time_embedding might be fp16, so we need to cast here. + timestep_img_token = timestep_img_token.to(dtype=self.dtype) + timestep_img_token = self.timestep_img_embed(timestep_img_token) + timestep_img_token = timestep_img_token.unsqueeze(dim=1) + + # 1.3. Encode text timesteps to single token (B, 1, inner_dim) + if not torch.is_tensor(timestep_text): + timestep_text = torch.tensor([timestep_text], dtype=torch.long, device=vae_hidden_states.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep_text = timestep_text * torch.ones(batch_size, dtype=timestep_text.dtype, device=timestep_text.device) + + timestep_text_token = self.timestep_text_proj(timestep_text) + # t_text_token does not contain any weights and will always return f32 tensors + # but time_embedding might be fp16, so we need to cast here. + timestep_text_token = timestep_text_token.to(dtype=self.dtype) + timestep_text_token = self.timestep_text_embed(timestep_text_token) + timestep_text_token = timestep_text_token.unsqueeze(dim=1) + + # 1.4. Concatenate all of the embeddings together. + if self.use_data_type_embedding: + assert data_type is not None, "data_type must be supplied if the model uses a data type embedding" + if not torch.is_tensor(data_type): + data_type = torch.tensor([data_type], dtype=torch.int, device=vae_hidden_states.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + data_type = data_type * torch.ones(batch_size, dtype=data_type.dtype, device=data_type.device) + + data_type_token = self.data_type_token_embedding(data_type).unsqueeze(dim=1) + hidden_states = torch.cat( + [ + timestep_img_token, + timestep_text_token, + data_type_token, + text_hidden_states, + clip_hidden_states, + vae_hidden_states, + ], + dim=1, + ) + else: + hidden_states = torch.cat( + [timestep_img_token, timestep_text_token, text_hidden_states, clip_hidden_states, vae_hidden_states], + dim=1, + ) + + # 1.5. Prepare the positional embeddings and add to hidden states + # Note: I think img_vae should always have the proper shape, so there's no need to interpolate + # the position embeddings. + if self.use_data_type_embedding: + pos_embed = torch.cat( + [self.pos_embed[:, : 1 + 1, :], self.data_type_pos_embed_token, self.pos_embed[:, 1 + 1 :, :]], dim=1 + ) + else: + pos_embed = self.pos_embed + hidden_states = hidden_states + pos_embed + hidden_states = self.pos_embed_drop(hidden_states) + + # 2. Blocks + hidden_states = self.transformer( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + timestep=None, + class_labels=None, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + hidden_states_is_embedding=True, + unpatchify=False, + )[0] + + # 3. Output + # Split out the predicted noise representation. + if self.use_data_type_embedding: + ( + t_img_token_out, + t_text_token_out, + data_type_token_out, + text_out, + img_clip_out, + img_vae_out, + ) = hidden_states.split((1, 1, 1, num_text_tokens, 1, num_img_tokens), dim=1) + else: + t_img_token_out, t_text_token_out, text_out, img_clip_out, img_vae_out = hidden_states.split( + (1, 1, num_text_tokens, 1, num_img_tokens), dim=1 + ) + + img_vae_out = self.vae_img_out(img_vae_out) + + # unpatchify + height = width = int(img_vae_out.shape[1] ** 0.5) + img_vae_out = img_vae_out.reshape( + shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels) + ) + img_vae_out = torch.einsum("nhwpqc->nchpwq", img_vae_out) + img_vae_out = img_vae_out.reshape( + shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size) + ) + + img_clip_out = self.clip_img_out(img_clip_out) + + text_out = self.text_out(text_out) + + return img_vae_out, img_clip_out, text_out diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/pipeline_unidiffuser.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/pipeline_unidiffuser.py new file mode 100644 index 0000000000000000000000000000000000000000..40fd3b337301b3a08a955e34990f0cbc1fbf8ce6 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/unidiffuser/pipeline_unidiffuser.py @@ -0,0 +1,1438 @@ +import inspect +from dataclasses import dataclass +from typing import Callable, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTokenizer, + CLIPVisionModelWithProjection, + GPT2Tokenizer, +) + +from ...image_processor import VaeImageProcessor +from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.outputs import BaseOutput +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline +from .modeling_text_decoder import UniDiffuserTextDecoder +from .modeling_uvit import UniDiffuserModel + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# New BaseOutput child class for joint image-text output +@dataclass +class ImageTextPipelineOutput(BaseOutput): + """ + Output class for joint image-text pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + text (`List[str]` or `List[List[str]]`) + List of generated text strings of length `batch_size` or a list of list of strings whose outer list has + length `batch_size`. + """ + + images: Optional[Union[List[PIL.Image.Image], np.ndarray]] + text: Optional[Union[List[str], List[List[str]]]] + + +class UniDiffuserPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + r""" + Pipeline for a bimodal image-text model which supports unconditional text and image generation, text-conditioned + image generation, image-conditioned text generation, and joint image-text generation. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. This + is part of the UniDiffuser image representation along with the CLIP vision encoding. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + image_encoder ([`CLIPVisionModel`]): + A [`~transformers.CLIPVisionModel`] to encode images as part of its image representation along with the VAE + latent representation. + image_processor ([`CLIPImageProcessor`]): + [`~transformers.CLIPImageProcessor`] to preprocess an image before CLIP encoding it with `image_encoder`. + clip_tokenizer ([`CLIPTokenizer`]): + A [`~transformers.CLIPTokenizer`] to tokenize the prompt before encoding it with `text_encoder`. + text_decoder ([`UniDiffuserTextDecoder`]): + Frozen text decoder. This is a GPT-style model which is used to generate text from the UniDiffuser + embedding. + text_tokenizer ([`GPT2Tokenizer`]): + A [`~transformers.GPT2Tokenizer`] to decode text for text generation; used along with the `text_decoder`. + unet ([`UniDiffuserModel`]): + A [U-ViT](https://github.com/baofff/U-ViT) model with UNNet-style skip connections between transformer + layers to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image and/or text latents. The + original UniDiffuser paper uses the [`DPMSolverMultistepScheduler`] scheduler. + """ + + _last_supported_version = "0.33.1" + # TODO: support for moving submodules for components with enable_model_cpu_offload + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae->text_decoder" + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + image_encoder: CLIPVisionModelWithProjection, + clip_image_processor: CLIPImageProcessor, + clip_tokenizer: CLIPTokenizer, + text_decoder: UniDiffuserTextDecoder, + text_tokenizer: GPT2Tokenizer, + unet: UniDiffuserModel, + scheduler: KarrasDiffusionSchedulers, + ): + super().__init__() + + if text_encoder.config.hidden_size != text_decoder.prefix_inner_dim: + raise ValueError( + f"The text encoder hidden size and text decoder prefix inner dim must be the same, but" + f" `text_encoder.config.hidden_size`: {text_encoder.config.hidden_size} and `text_decoder.prefix_inner_dim`: {text_decoder.prefix_inner_dim}" + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + image_encoder=image_encoder, + clip_image_processor=clip_image_processor, + clip_tokenizer=clip_tokenizer, + text_decoder=text_decoder, + text_tokenizer=text_tokenizer, + unet=unet, + scheduler=scheduler, + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.num_channels_latents = vae.config.latent_channels + self.text_encoder_seq_len = text_encoder.config.max_position_embeddings + self.text_encoder_hidden_size = text_encoder.config.hidden_size + self.image_encoder_projection_dim = image_encoder.config.projection_dim + self.unet_resolution = unet.config.sample_size + + self.text_intermediate_dim = self.text_encoder_hidden_size + if self.text_decoder.prefix_hidden_dim is not None: + self.text_intermediate_dim = self.text_decoder.prefix_hidden_dim + + self.mode = None + + # TODO: handle safety checking? + self.safety_checker = None + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def _infer_mode(self, prompt, prompt_embeds, image, latents, prompt_latents, vae_latents, clip_latents): + r""" + Infer the generation task ('mode') from the inputs to `__call__`. If the mode has been manually set, the set + mode will be used. + """ + prompt_available = (prompt is not None) or (prompt_embeds is not None) + image_available = image is not None + input_available = prompt_available or image_available + + prompt_latents_available = prompt_latents is not None + vae_latents_available = vae_latents is not None + clip_latents_available = clip_latents is not None + full_latents_available = latents is not None + image_latents_available = vae_latents_available and clip_latents_available + all_indv_latents_available = prompt_latents_available and image_latents_available + + if self.mode is not None: + # Preferentially use the mode set by the user + mode = self.mode + elif prompt_available: + mode = "text2img" + elif image_available: + mode = "img2text" + else: + # Neither prompt nor image supplied, infer based on availability of latents + if full_latents_available or all_indv_latents_available: + mode = "joint" + elif prompt_latents_available: + mode = "text" + elif image_latents_available: + mode = "img" + else: + # No inputs or latents available + mode = "joint" + + # Give warnings for ambiguous cases + if self.mode is None and prompt_available and image_available: + logger.warning( + f"You have supplied both a text prompt and image to the pipeline and mode has not been set manually," + f" defaulting to mode '{mode}'." + ) + + if self.mode is None and not input_available: + if vae_latents_available != clip_latents_available: + # Exactly one of vae_latents and clip_latents is supplied + logger.warning( + f"You have supplied exactly one of `vae_latents` and `clip_latents`, whereas either both or none" + f" are expected to be supplied. Defaulting to mode '{mode}'." + ) + elif not prompt_latents_available and not vae_latents_available and not clip_latents_available: + # No inputs or latents supplied + logger.warning( + f"No inputs or latents have been supplied, and mode has not been manually set," + f" defaulting to mode '{mode}'." + ) + + return mode + + # Copied from diffusers.pipelines.pipeline_utils.StableDiffusionMixin.enable_vae_slicing + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + # Copied from diffusers.pipelines.pipeline_utils.StableDiffusionMixin.disable_vae_slicing + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + # Copied from diffusers.pipelines.pipeline_utils.StableDiffusionMixin.enable_vae_tiling + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + # Copied from diffusers.pipelines.pipeline_utils.StableDiffusionMixin.disable_vae_tiling + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + # Functions to manually set the mode + def set_text_mode(self): + r"""Manually set the generation mode to unconditional ("marginal") text generation.""" + self.mode = "text" + + def set_image_mode(self): + r"""Manually set the generation mode to unconditional ("marginal") image generation.""" + self.mode = "img" + + def set_text_to_image_mode(self): + r"""Manually set the generation mode to text-conditioned image generation.""" + self.mode = "text2img" + + def set_image_to_text_mode(self): + r"""Manually set the generation mode to image-conditioned text generation.""" + self.mode = "img2text" + + def set_joint_mode(self): + r"""Manually set the generation mode to unconditional joint image-text generation.""" + self.mode = "joint" + + def reset_mode(self): + r"""Removes a manually set mode; after calling this, the pipeline will infer the mode from inputs.""" + self.mode = None + + def _infer_batch_size( + self, + mode, + prompt, + prompt_embeds, + image, + num_images_per_prompt, + num_prompts_per_image, + latents, + prompt_latents, + vae_latents, + clip_latents, + ): + r"""Infers the batch size and multiplier depending on mode and supplied arguments to `__call__`.""" + if num_images_per_prompt is None: + num_images_per_prompt = 1 + if num_prompts_per_image is None: + num_prompts_per_image = 1 + + assert num_images_per_prompt > 0, "num_images_per_prompt must be a positive integer" + assert num_prompts_per_image > 0, "num_prompts_per_image must be a positive integer" + + if mode in ["text2img"]: + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + # Either prompt or prompt_embeds must be present for text2img. + batch_size = prompt_embeds.shape[0] + multiplier = num_images_per_prompt + elif mode in ["img2text"]: + if isinstance(image, PIL.Image.Image): + batch_size = 1 + else: + # Image must be available and type either PIL.Image.Image or torch.Tensor. + # Not currently supporting something like image_embeds. + batch_size = image.shape[0] + multiplier = num_prompts_per_image + elif mode in ["img"]: + if vae_latents is not None: + batch_size = vae_latents.shape[0] + elif clip_latents is not None: + batch_size = clip_latents.shape[0] + else: + batch_size = 1 + multiplier = num_images_per_prompt + elif mode in ["text"]: + if prompt_latents is not None: + batch_size = prompt_latents.shape[0] + else: + batch_size = 1 + multiplier = num_prompts_per_image + elif mode in ["joint"]: + if latents is not None: + batch_size = latents.shape[0] + elif prompt_latents is not None: + batch_size = prompt_latents.shape[0] + elif vae_latents is not None: + batch_size = vae_latents.shape[0] + elif clip_latents is not None: + batch_size = clip_latents.shape[0] + else: + batch_size = 1 + + if num_images_per_prompt == num_prompts_per_image: + multiplier = num_images_per_prompt + else: + multiplier = min(num_images_per_prompt, num_prompts_per_image) + logger.warning( + f"You are using mode `{mode}` and `num_images_per_prompt`: {num_images_per_prompt} and" + f" num_prompts_per_image: {num_prompts_per_image} are not equal. Using batch size equal to" + f" `min(num_images_per_prompt, num_prompts_per_image) = {batch_size}." + ) + return batch_size, multiplier + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt with self.tokenizer->self.clip_tokenizer + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.clip_tokenizer) + + text_inputs = self.clip_tokenizer( + prompt, + padding="max_length", + max_length=self.clip_tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.clip_tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.clip_tokenizer.batch_decode( + untruncated_ids[:, self.clip_tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.clip_tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.clip_tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.clip_tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_instruct_pix2pix.StableDiffusionInstructPix2PixPipeline.prepare_image_latents + # Add num_prompts_per_image argument, sample from autoencoder moment distribution + def encode_image_vae_latents( + self, + image, + batch_size, + num_prompts_per_image, + dtype, + device, + do_classifier_free_guidance, + generator=None, + ): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_prompts_per_image + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if isinstance(generator, list): + image_latents = [ + self.vae.encode(image[i : i + 1]).latent_dist.sample(generator=generator[i]) + * self.vae.config.scaling_factor + for i in range(batch_size) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = self.vae.encode(image).latent_dist.sample(generator=generator) + # Scale image_latents by the VAE's scaling factor + image_latents = image_latents * self.vae.config.scaling_factor + + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand image_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {image_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + if do_classifier_free_guidance: + uncond_image_latents = torch.zeros_like(image_latents) + image_latents = torch.cat([image_latents, image_latents, uncond_image_latents], dim=0) + + return image_latents + + def encode_image_clip_latents( + self, + image, + batch_size, + num_prompts_per_image, + dtype, + device, + generator=None, + ): + # Map image to CLIP embedding. + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + preprocessed_image = self.clip_image_processor.preprocess( + image, + return_tensors="pt", + ) + preprocessed_image = preprocessed_image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_prompts_per_image + if isinstance(generator, list): + image_latents = [ + self.image_encoder(**preprocessed_image[i : i + 1]).image_embeds for i in range(batch_size) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = self.image_encoder(**preprocessed_image).image_embeds + + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand image_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {image_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + return image_latents + + def prepare_text_latents( + self, batch_size, num_images_per_prompt, seq_len, hidden_size, dtype, device, generator, latents=None + ): + # Prepare latents for the CLIP embedded prompt. + shape = (batch_size * num_images_per_prompt, seq_len, hidden_size) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + # latents is assumed to have shace (B, L, D) + latents = latents.repeat(num_images_per_prompt, 1, 1) + latents = latents.to(device=device, dtype=dtype) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + # Rename prepare_latents -> prepare_image_vae_latents and add num_prompts_per_image argument. + def prepare_image_vae_latents( + self, + batch_size, + num_prompts_per_image, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + shape = ( + batch_size * num_prompts_per_image, + num_channels_latents, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + # latents is assumed to have shape (B, C, H, W) + latents = latents.repeat(num_prompts_per_image, 1, 1, 1) + latents = latents.to(device=device, dtype=dtype) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def prepare_image_clip_latents( + self, batch_size, num_prompts_per_image, clip_img_dim, dtype, device, generator, latents=None + ): + # Prepare latents for the CLIP embedded image. + shape = (batch_size * num_prompts_per_image, 1, clip_img_dim) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + # latents is assumed to have shape (B, L, D) + latents = latents.repeat(num_prompts_per_image, 1, 1) + latents = latents.to(device=device, dtype=dtype) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def decode_text_latents(self, text_latents, device): + output_token_list, seq_lengths = self.text_decoder.generate_captions( + text_latents, self.text_tokenizer.eos_token_id, device=device + ) + output_list = output_token_list.cpu().numpy() + generated_text = [ + self.text_tokenizer.decode(output[: int(length)], skip_special_tokens=True) + for output, length in zip(output_list, seq_lengths) + ] + return generated_text + + def _split(self, x, height, width): + r""" + Splits a flattened embedding x of shape (B, C * H * W + clip_img_dim) into two tensors of shape (B, C, H, W) + and (B, 1, clip_img_dim) + """ + batch_size = x.shape[0] + latent_height = height // self.vae_scale_factor + latent_width = width // self.vae_scale_factor + img_vae_dim = self.num_channels_latents * latent_height * latent_width + + img_vae, img_clip = x.split([img_vae_dim, self.image_encoder_projection_dim], dim=1) + + img_vae = torch.reshape(img_vae, (batch_size, self.num_channels_latents, latent_height, latent_width)) + img_clip = torch.reshape(img_clip, (batch_size, 1, self.image_encoder_projection_dim)) + return img_vae, img_clip + + def _combine(self, img_vae, img_clip): + r""" + Combines a latent image img_vae of shape (B, C, H, W) and a CLIP-embedded image img_clip of shape (B, 1, + clip_img_dim) into a single tensor of shape (B, C * H * W + clip_img_dim). + """ + img_vae = torch.reshape(img_vae, (img_vae.shape[0], -1)) + img_clip = torch.reshape(img_clip, (img_clip.shape[0], -1)) + return torch.concat([img_vae, img_clip], dim=-1) + + def _split_joint(self, x, height, width): + r""" + Splits a flattened embedding x of shape (B, C * H * W + clip_img_dim + text_seq_len * text_dim] into (img_vae, + img_clip, text) where img_vae is of shape (B, C, H, W), img_clip is of shape (B, 1, clip_img_dim), and text is + of shape (B, text_seq_len, text_dim). + """ + batch_size = x.shape[0] + latent_height = height // self.vae_scale_factor + latent_width = width // self.vae_scale_factor + img_vae_dim = self.num_channels_latents * latent_height * latent_width + text_dim = self.text_encoder_seq_len * self.text_intermediate_dim + + img_vae, img_clip, text = x.split([img_vae_dim, self.image_encoder_projection_dim, text_dim], dim=1) + + img_vae = torch.reshape(img_vae, (batch_size, self.num_channels_latents, latent_height, latent_width)) + img_clip = torch.reshape(img_clip, (batch_size, 1, self.image_encoder_projection_dim)) + text = torch.reshape(text, (batch_size, self.text_encoder_seq_len, self.text_intermediate_dim)) + return img_vae, img_clip, text + + def _combine_joint(self, img_vae, img_clip, text): + r""" + Combines a latent image img_vae of shape (B, C, H, W), a CLIP-embedded image img_clip of shape (B, L_img, + clip_img_dim), and a text embedding text of shape (B, L_text, text_dim) into a single embedding x of shape (B, + C * H * W + L_img * clip_img_dim + L_text * text_dim). + """ + img_vae = torch.reshape(img_vae, (img_vae.shape[0], -1)) + img_clip = torch.reshape(img_clip, (img_clip.shape[0], -1)) + text = torch.reshape(text, (text.shape[0], -1)) + return torch.concat([img_vae, img_clip, text], dim=-1) + + def _get_noise_pred( + self, + mode, + latents, + t, + prompt_embeds, + img_vae, + img_clip, + max_timestep, + data_type, + guidance_scale, + generator, + device, + height, + width, + ): + r""" + Gets the noise prediction using the `unet` and performs classifier-free guidance, if necessary. + """ + if mode == "joint": + # Joint text-image generation + img_vae_latents, img_clip_latents, text_latents = self._split_joint(latents, height, width) + + img_vae_out, img_clip_out, text_out = self.unet( + img_vae_latents, img_clip_latents, text_latents, timestep_img=t, timestep_text=t, data_type=data_type + ) + + x_out = self._combine_joint(img_vae_out, img_clip_out, text_out) + + if guidance_scale <= 1.0: + return x_out + + # Classifier-free guidance + img_vae_T = randn_tensor(img_vae.shape, generator=generator, device=device, dtype=img_vae.dtype) + img_clip_T = randn_tensor(img_clip.shape, generator=generator, device=device, dtype=img_clip.dtype) + text_T = randn_tensor(prompt_embeds.shape, generator=generator, device=device, dtype=prompt_embeds.dtype) + + _, _, text_out_uncond = self.unet( + img_vae_T, img_clip_T, text_latents, timestep_img=max_timestep, timestep_text=t, data_type=data_type + ) + + img_vae_out_uncond, img_clip_out_uncond, _ = self.unet( + img_vae_latents, + img_clip_latents, + text_T, + timestep_img=t, + timestep_text=max_timestep, + data_type=data_type, + ) + + x_out_uncond = self._combine_joint(img_vae_out_uncond, img_clip_out_uncond, text_out_uncond) + + return guidance_scale * x_out + (1.0 - guidance_scale) * x_out_uncond + elif mode == "text2img": + # Text-conditioned image generation + img_vae_latents, img_clip_latents = self._split(latents, height, width) + + img_vae_out, img_clip_out, text_out = self.unet( + img_vae_latents, img_clip_latents, prompt_embeds, timestep_img=t, timestep_text=0, data_type=data_type + ) + + img_out = self._combine(img_vae_out, img_clip_out) + + if guidance_scale <= 1.0: + return img_out + + # Classifier-free guidance + text_T = randn_tensor(prompt_embeds.shape, generator=generator, device=device, dtype=prompt_embeds.dtype) + + img_vae_out_uncond, img_clip_out_uncond, text_out_uncond = self.unet( + img_vae_latents, + img_clip_latents, + text_T, + timestep_img=t, + timestep_text=max_timestep, + data_type=data_type, + ) + + img_out_uncond = self._combine(img_vae_out_uncond, img_clip_out_uncond) + + return guidance_scale * img_out + (1.0 - guidance_scale) * img_out_uncond + elif mode == "img2text": + # Image-conditioned text generation + img_vae_out, img_clip_out, text_out = self.unet( + img_vae, img_clip, latents, timestep_img=0, timestep_text=t, data_type=data_type + ) + + if guidance_scale <= 1.0: + return text_out + + # Classifier-free guidance + img_vae_T = randn_tensor(img_vae.shape, generator=generator, device=device, dtype=img_vae.dtype) + img_clip_T = randn_tensor(img_clip.shape, generator=generator, device=device, dtype=img_clip.dtype) + + img_vae_out_uncond, img_clip_out_uncond, text_out_uncond = self.unet( + img_vae_T, img_clip_T, latents, timestep_img=max_timestep, timestep_text=t, data_type=data_type + ) + + return guidance_scale * text_out + (1.0 - guidance_scale) * text_out_uncond + elif mode == "text": + # Unconditional ("marginal") text generation (no CFG) + img_vae_out, img_clip_out, text_out = self.unet( + img_vae, img_clip, latents, timestep_img=max_timestep, timestep_text=t, data_type=data_type + ) + + return text_out + elif mode == "img": + # Unconditional ("marginal") image generation (no CFG) + img_vae_latents, img_clip_latents = self._split(latents, height, width) + + img_vae_out, img_clip_out, text_out = self.unet( + img_vae_latents, + img_clip_latents, + prompt_embeds, + timestep_img=t, + timestep_text=max_timestep, + data_type=data_type, + ) + + img_out = self._combine(img_vae_out, img_clip_out) + return img_out + + def check_latents_shape(self, latents_name, latents, expected_shape): + latents_shape = latents.shape + expected_num_dims = len(expected_shape) + 1 # expected dimensions plus the batch dimension + expected_shape_str = ", ".join(str(dim) for dim in expected_shape) + if len(latents_shape) != expected_num_dims: + raise ValueError( + f"`{latents_name}` should have shape (batch_size, {expected_shape_str}), but the current shape" + f" {latents_shape} has {len(latents_shape)} dimensions." + ) + for i in range(1, expected_num_dims): + if latents_shape[i] != expected_shape[i - 1]: + raise ValueError( + f"`{latents_name}` should have shape (batch_size, {expected_shape_str}), but the current shape" + f" {latents_shape} has {latents_shape[i]} != {expected_shape[i - 1]} at dimension {i}." + ) + + def check_inputs( + self, + mode, + prompt, + image, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + latents=None, + prompt_latents=None, + vae_latents=None, + clip_latents=None, + ): + # Check inputs before running the generative process. + if height % self.vae_scale_factor != 0 or width % self.vae_scale_factor != 0: + raise ValueError( + f"`height` and `width` have to be divisible by {self.vae_scale_factor} but are {height} and {width}." + ) + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if mode == "text2img": + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if mode == "img2text": + if image is None: + raise ValueError("`img2text` mode requires an image to be provided.") + + # Check provided latents + latent_height = height // self.vae_scale_factor + latent_width = width // self.vae_scale_factor + full_latents_available = latents is not None + prompt_latents_available = prompt_latents is not None + vae_latents_available = vae_latents is not None + clip_latents_available = clip_latents is not None + + if full_latents_available: + individual_latents_available = ( + prompt_latents is not None or vae_latents is not None or clip_latents is not None + ) + if individual_latents_available: + logger.warning( + "You have supplied both `latents` and at least one of `prompt_latents`, `vae_latents`, and" + " `clip_latents`. The value of `latents` will override the value of any individually supplied latents." + ) + # Check shape of full latents + img_vae_dim = self.num_channels_latents * latent_height * latent_width + text_dim = self.text_encoder_seq_len * self.text_encoder_hidden_size + latents_dim = img_vae_dim + self.image_encoder_projection_dim + text_dim + latents_expected_shape = (latents_dim,) + self.check_latents_shape("latents", latents, latents_expected_shape) + + # Check individual latent shapes, if present + if prompt_latents_available: + prompt_latents_expected_shape = (self.text_encoder_seq_len, self.text_encoder_hidden_size) + self.check_latents_shape("prompt_latents", prompt_latents, prompt_latents_expected_shape) + + if vae_latents_available: + vae_latents_expected_shape = (self.num_channels_latents, latent_height, latent_width) + self.check_latents_shape("vae_latents", vae_latents, vae_latents_expected_shape) + + if clip_latents_available: + clip_latents_expected_shape = (1, self.image_encoder_projection_dim) + self.check_latents_shape("clip_latents", clip_latents, clip_latents_expected_shape) + + if mode in ["text2img", "img"] and vae_latents_available and clip_latents_available: + if vae_latents.shape[0] != clip_latents.shape[0]: + raise ValueError( + f"Both `vae_latents` and `clip_latents` are supplied, but their batch dimensions are not equal:" + f" {vae_latents.shape[0]} != {clip_latents.shape[0]}." + ) + + if mode == "joint" and prompt_latents_available and vae_latents_available and clip_latents_available: + if prompt_latents.shape[0] != vae_latents.shape[0] or prompt_latents.shape[0] != clip_latents.shape[0]: + raise ValueError( + f"All of `prompt_latents`, `vae_latents`, and `clip_latents` are supplied, but their batch" + f" dimensions are not equal: {prompt_latents.shape[0]} != {vae_latents.shape[0]}" + f" != {clip_latents.shape[0]}." + ) + + @torch.no_grad() + def __call__( + self, + prompt: Optional[Union[str, List[str]]] = None, + image: Optional[Union[torch.Tensor, PIL.Image.Image]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + data_type: Optional[int] = 1, + num_inference_steps: int = 50, + guidance_scale: float = 8.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + num_prompts_per_image: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_latents: Optional[torch.Tensor] = None, + vae_latents: Optional[torch.Tensor] = None, + clip_latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.Tensor], None]] = None, + callback_steps: int = 1, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + Required for text-conditioned image generation (`text2img`) mode. + image (`torch.Tensor` or `PIL.Image.Image`, *optional*): + `Image` or tensor representing an image batch. Required for image-conditioned text generation + (`img2text`) mode. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + data_type (`int`, *optional*, defaults to 1): + The data type (either 0 or 1). Only used if you are loading a checkpoint which supports a data type + embedding; this is added for compatibility with the + [UniDiffuser-v1](https://huggingface.co/thu-ml/unidiffuser-v1) checkpoint. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 8.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). Used in + text-conditioned image generation (`text2img`) mode. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. Used in `text2img` (text-conditioned image generation) and + `img` mode. If the mode is joint and both `num_images_per_prompt` and `num_prompts_per_image` are + supplied, `min(num_images_per_prompt, num_prompts_per_image)` samples are generated. + num_prompts_per_image (`int`, *optional*, defaults to 1): + The number of prompts to generate per image. Used in `img2text` (image-conditioned text generation) and + `text` mode. If the mode is joint and both `num_images_per_prompt` and `num_prompts_per_image` are + supplied, `min(num_images_per_prompt, num_prompts_per_image)` samples are generated. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for joint + image-text generation. Can be used to tweak the same generation with different prompts. If not + provided, a latents tensor is generated by sampling using the supplied random `generator`. This assumes + a full set of VAE, CLIP, and text latents, if supplied, overrides the value of `prompt_latents`, + `vae_latents`, and `clip_latents`. + prompt_latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for text + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + vae_latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + clip_latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. Used in text-conditioned + image generation (`text2img`) mode. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are be generated from the `negative_prompt` input argument. Used + in text-conditioned image generation (`text2img`) mode. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImageTextPipelineOutput`] instead of a plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + + Returns: + [`~pipelines.unidiffuser.ImageTextPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.unidiffuser.ImageTextPipelineOutput`] is returned, otherwise a + `tuple` is returned where the first element is a list with the generated images and the second element + is a list of generated texts. + """ + + # 0. Default height and width to unet + height = height or self.unet_resolution * self.vae_scale_factor + width = width or self.unet_resolution * self.vae_scale_factor + + # 1. Check inputs + # Recalculate mode for each call to the pipeline. + mode = self._infer_mode(prompt, prompt_embeds, image, latents, prompt_latents, vae_latents, clip_latents) + self.check_inputs( + mode, + prompt, + image, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + latents, + prompt_latents, + vae_latents, + clip_latents, + ) + + # 2. Define call parameters + batch_size, multiplier = self._infer_batch_size( + mode, + prompt, + prompt_embeds, + image, + num_images_per_prompt, + num_prompts_per_image, + latents, + prompt_latents, + vae_latents, + clip_latents, + ) + device = self._execution_device + reduce_text_emb_dim = self.text_intermediate_dim < self.text_encoder_hidden_size or self.mode != "text2img" + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + # Note that this differs from the formulation in the unidiffusers paper! + do_classifier_free_guidance = guidance_scale > 1.0 + + # check if scheduler is in sigmas space + # scheduler_is_in_sigma_space = hasattr(self.scheduler, "sigmas") + + # 3. Encode input prompt, if available; otherwise prepare text latents + if latents is not None: + # Overwrite individual latents + vae_latents, clip_latents, prompt_latents = self._split_joint(latents, height, width) + + if mode in ["text2img"]: + # 3.1. Encode input prompt, if available + assert prompt is not None or prompt_embeds is not None + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=multiplier, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + # if do_classifier_free_guidance: + # prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + else: + # 3.2. Prepare text latent variables, if input not available + prompt_embeds = self.prepare_text_latents( + batch_size=batch_size, + num_images_per_prompt=multiplier, + seq_len=self.text_encoder_seq_len, + hidden_size=self.text_encoder_hidden_size, + dtype=self.text_encoder.dtype, # Should work with both full precision and mixed precision + device=device, + generator=generator, + latents=prompt_latents, + ) + + if reduce_text_emb_dim: + prompt_embeds = self.text_decoder.encode(prompt_embeds) + + # 4. Encode image, if available; otherwise prepare image latents + if mode in ["img2text"]: + # 4.1. Encode images, if available + assert image is not None, "`img2text` requires a conditioning image" + # Encode image using VAE + image_vae = self.image_processor.preprocess(image) + height, width = image_vae.shape[-2:] + image_vae_latents = self.encode_image_vae_latents( + image=image_vae, + batch_size=batch_size, + num_prompts_per_image=multiplier, + dtype=prompt_embeds.dtype, + device=device, + do_classifier_free_guidance=False, # Copied from InstructPix2Pix, don't use their version of CFG + generator=generator, + ) + + # Encode image using CLIP + image_clip_latents = self.encode_image_clip_latents( + image=image, + batch_size=batch_size, + num_prompts_per_image=multiplier, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + ) + # (batch_size, clip_hidden_size) => (batch_size, 1, clip_hidden_size) + image_clip_latents = image_clip_latents.unsqueeze(1) + else: + # 4.2. Prepare image latent variables, if input not available + # Prepare image VAE latents in latent space + image_vae_latents = self.prepare_image_vae_latents( + batch_size=batch_size, + num_prompts_per_image=multiplier, + num_channels_latents=self.num_channels_latents, + height=height, + width=width, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=vae_latents, + ) + + # Prepare image CLIP latents + image_clip_latents = self.prepare_image_clip_latents( + batch_size=batch_size, + num_prompts_per_image=multiplier, + clip_img_dim=self.image_encoder_projection_dim, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=clip_latents, + ) + + # 5. Set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + # max_timestep = timesteps[0] + max_timestep = self.scheduler.config.num_train_timesteps + + # 6. Prepare latent variables + if mode == "joint": + latents = self._combine_joint(image_vae_latents, image_clip_latents, prompt_embeds) + elif mode in ["text2img", "img"]: + latents = self._combine(image_vae_latents, image_clip_latents) + elif mode in ["img2text", "text"]: + latents = prompt_embeds + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + logger.debug(f"Scheduler extra step kwargs: {extra_step_kwargs}") + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # predict the noise residual + # Also applies classifier-free guidance as described in the UniDiffuser paper + noise_pred = self._get_noise_pred( + mode, + latents, + t, + prompt_embeds, + image_vae_latents, + image_clip_latents, + max_timestep, + data_type, + guidance_scale, + generator, + device, + height, + width, + ) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 9. Post-processing + image = None + text = None + if mode == "joint": + image_vae_latents, image_clip_latents, text_latents = self._split_joint(latents, height, width) + + if not output_type == "latent": + # Map latent VAE image back to pixel space + image = self.vae.decode(image_vae_latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = image_vae_latents + + text = self.decode_text_latents(text_latents, device) + elif mode in ["text2img", "img"]: + image_vae_latents, image_clip_latents = self._split(latents, height, width) + + if not output_type == "latent": + # Map latent VAE image back to pixel space + image = self.vae.decode(image_vae_latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = image_vae_latents + elif mode in ["img2text", "text"]: + text_latents = latents + text = self.decode_text_latents(text_latents, device) + + self.maybe_free_model_hooks() + + # 10. Postprocess the image, if necessary + if image is not None: + do_denormalize = [True] * image.shape[0] + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload last model to CPU + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.final_offload_hook.offload() + + if not return_dict: + return (image, text) + + return ImageTextPipelineOutput(images=image, text=text) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..ab765a1bbad9ae0b8c7f2e7dec5019e3a5e4ea01 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__init__.py @@ -0,0 +1,52 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_visualcloze_combined"] = ["VisualClozePipeline"] + _import_structure["pipeline_visualcloze_generation"] = ["VisualClozeGenerationPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_visualcloze_combined import VisualClozePipeline + from .pipeline_visualcloze_generation import VisualClozeGenerationPipeline + + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0dd0bcec854c6e2bcaf8d57a2950281f20de73c9 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__pycache__/pipeline_visualcloze_combined.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__pycache__/pipeline_visualcloze_combined.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bca8fd0038dd3f1384f893f727787270d3f13eea Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__pycache__/pipeline_visualcloze_combined.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__pycache__/pipeline_visualcloze_generation.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__pycache__/pipeline_visualcloze_generation.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d5f26e480921fe305dc18e916711cd02fd04d0c0 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__pycache__/pipeline_visualcloze_generation.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__pycache__/visualcloze_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__pycache__/visualcloze_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3ed6d8b20db6ef16481655b614f531ba5c34f52f Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/__pycache__/visualcloze_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/pipeline_visualcloze_combined.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/pipeline_visualcloze_combined.py new file mode 100644 index 0000000000000000000000000000000000000000..68130baad7093b892ad8895bb6b65311c59a3cde --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/pipeline_visualcloze_combined.py @@ -0,0 +1,444 @@ +# Copyright 2025 VisualCloze team and The HuggingFace Team. All rights reserved. +# +# 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 Any, Callable, Dict, List, Optional, Union + +import torch +from PIL import Image +from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast + +from ...loaders import FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ..flux.pipeline_flux_fill import FluxFillPipeline as VisualClozeUpsamplingPipeline +from ..flux.pipeline_output import FluxPipelineOutput +from ..pipeline_utils import DiffusionPipeline +from .pipeline_visualcloze_generation import VisualClozeGenerationPipeline + + +if is_torch_xla_available(): + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers import VisualClozePipeline + >>> from diffusers.utils import load_image + + >>> image_paths = [ + ... # in-context examples + ... [ + ... load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/visualcloze/visualcloze_mask2image_incontext-example-1_mask.jpg" + ... ), + ... load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/visualcloze/visualcloze_mask2image_incontext-example-1_image.jpg" + ... ), + ... ], + ... # query with the target image + ... [ + ... load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/visualcloze/visualcloze_mask2image_query_mask.jpg" + ... ), + ... None, # No image needed for the target image + ... ], + ... ] + >>> task_prompt = "In each row, a logical task is demonstrated to achieve [IMAGE2] an aesthetically pleasing photograph based on [IMAGE1] sam 2-generated masks with rich color coding." + >>> content_prompt = "Majestic photo of a golden eagle perched on a rocky outcrop in a mountainous landscape. The eagle is positioned in the right foreground, facing left, with its sharp beak and keen eyes prominently visible. Its plumage is a mix of dark brown and golden hues, with intricate feather details. The background features a soft-focus view of snow-capped mountains under a cloudy sky, creating a serene and grandiose atmosphere. The foreground includes rugged rocks and patches of green moss. Photorealistic, medium depth of field, soft natural lighting, cool color palette, high contrast, sharp focus on the eagle, blurred background, tranquil, majestic, wildlife photography." + >>> pipe = VisualClozePipeline.from_pretrained( + ... "VisualCloze/VisualClozePipeline-384", resolution=384, torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> image = pipe( + ... task_prompt=task_prompt, + ... content_prompt=content_prompt, + ... image=image_paths, + ... upsampling_width=1344, + ... upsampling_height=768, + ... upsampling_strength=0.4, + ... guidance_scale=30, + ... num_inference_steps=30, + ... max_sequence_length=512, + ... generator=torch.Generator("cpu").manual_seed(0), + ... ).images[0][0] + >>> image.save("visualcloze.png") + ``` +""" + + +class VisualClozePipeline( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, +): + r""" + The VisualCloze pipeline for image generation with visual context. Reference: + https://github.com/lzyhha/VisualCloze/tree/main. This pipeline is designed to generate images based on visual + in-context examples. + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + resolution (`int`, *optional*, defaults to 384): + The resolution of each image when concatenating images from the query and in-context examples. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + resolution: int = 384, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + ) + + self.generation_pipe = VisualClozeGenerationPipeline( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + resolution=resolution, + ) + self.upsampling_pipe = VisualClozeUpsamplingPipeline( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + ) + + def check_inputs( + self, + image, + task_prompt, + content_prompt, + upsampling_height, + upsampling_width, + strength, + prompt_embeds=None, + pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if upsampling_height is not None and upsampling_height % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`upsampling_height`has to be divisible by {self.vae_scale_factor * 2} but are {upsampling_height}. Dimensions will be resized accordingly" + ) + if upsampling_width is not None and upsampling_width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`upsampling_width` have to be divisible by {self.vae_scale_factor * 2} but are {upsampling_width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + # Validate prompt inputs + if (task_prompt is not None or content_prompt is not None) and prompt_embeds is not None: + raise ValueError("Cannot provide both text `task_prompt` + `content_prompt` and `prompt_embeds`. ") + + if task_prompt is None and content_prompt is None and prompt_embeds is None: + raise ValueError("Must provide either `task_prompt` + `content_prompt` or pre-computed `prompt_embeds`. ") + + # Validate prompt types and consistency + if task_prompt is None: + raise ValueError("`task_prompt` is missing.") + + if task_prompt is not None and not isinstance(task_prompt, (str, list)): + raise ValueError(f"`task_prompt` must be str or list, got {type(task_prompt)}") + + if content_prompt is not None and not isinstance(content_prompt, (str, list)): + raise ValueError(f"`content_prompt` must be str or list, got {type(content_prompt)}") + + if isinstance(task_prompt, list) or isinstance(content_prompt, list): + if not isinstance(task_prompt, list) or not isinstance(content_prompt, list): + raise ValueError( + f"`task_prompt` and `content_prompt` must both be lists, or both be of type str or None, " + f"got {type(task_prompt)} and {type(content_prompt)}" + ) + if len(content_prompt) != len(task_prompt): + raise ValueError("`task_prompt` and `content_prompt` must have the same length whe they are lists.") + + for sample in image: + if not isinstance(sample, list) or not isinstance(sample[0], list): + raise ValueError("Each sample in the batch must have a 2D list of images.") + if len({len(row) for row in sample}) != 1: + raise ValueError("Each in-context example and query should contain the same number of images.") + if not any(img is None for img in sample[-1]): + raise ValueError("There are no targets in the query, which should be represented as None.") + for row in sample[:-1]: + if any(img is None for img in row): + raise ValueError("Images are missing in in-context examples.") + + # Validate embeddings + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + # Validate sequence length + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"max_sequence_length cannot exceed 512, got {max_sequence_length}") + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + task_prompt: Union[str, List[str]] = None, + content_prompt: Union[str, List[str]] = None, + image: Optional[torch.FloatTensor] = None, + upsampling_height: Optional[int] = None, + upsampling_width: Optional[int] = None, + num_inference_steps: int = 50, + sigmas: Optional[List[float]] = None, + guidance_scale: float = 30.0, + num_images_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 512, + upsampling_strength: float = 1.0, + ): + r""" + Function invoked when calling the VisualCloze pipeline for generation. + + Args: + task_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to define the task intention. + content_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to define the content or caption of the target image to be generated. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)`. + upsampling_height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image (i.e., output image) after upsampling via SDEdit. By + default, the image is upsampled by a factor of three, and the base resolution is determined by the + resolution parameter of the pipeline. When only one of `upsampling_height` or `upsampling_width` is + specified, the other will be automatically set based on the aspect ratio. + upsampling_width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image (i.e., output image) after upsampling via SDEdit. By + default, the image is upsampled by a factor of three, and the base resolution is determined by the + resolution parameter of the pipeline. When only one of `upsampling_height` or `upsampling_width` is + specified, the other will be automatically set based on the aspect ratio. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 30.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + upsampling_strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image` when upsampling the results. Must be between 0 and + 1. The generated image is used as a starting point and more noise is added the higher the + `upsampling_strength`. The number of denoising steps depends on the amount of noise initially added. + When `upsampling_strength` is 1, added noise is maximum and the denoising process runs for the full + number of iterations specified in `num_inference_steps`. A value of 0 skips the upsampling step and + output the results at the resolution of `self.resolution`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + generation_output = self.generation_pipe( + task_prompt=task_prompt, + content_prompt=content_prompt, + image=image, + num_inference_steps=num_inference_steps, + sigmas=sigmas, + guidance_scale=guidance_scale, + num_images_per_prompt=num_images_per_prompt, + generator=generator, + latents=latents, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + joint_attention_kwargs=joint_attention_kwargs, + callback_on_step_end=callback_on_step_end, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + output_type=output_type if upsampling_strength == 0 else "pil", + ) + if upsampling_strength == 0: + if not return_dict: + return (generation_output,) + + return FluxPipelineOutput(images=generation_output) + + # Upsampling the generated images + # 1. Prepare the input images and prompts + if not isinstance(content_prompt, (list)): + content_prompt = [content_prompt] + n_target_per_sample = [] + upsampling_image = [] + upsampling_mask = [] + upsampling_prompt = [] + upsampling_generator = generator if isinstance(generator, (torch.Generator,)) else [] + for i in range(len(generation_output.images)): + n_target_per_sample.append(len(generation_output.images[i])) + for image in generation_output.images[i]: + upsampling_image.append(image) + upsampling_mask.append(Image.new("RGB", image.size, (255, 255, 255))) + upsampling_prompt.append( + content_prompt[i % len(content_prompt)] if content_prompt[i % len(content_prompt)] else "" + ) + if not isinstance(generator, (torch.Generator,)): + upsampling_generator.append(generator[i % len(content_prompt)]) + + # 2. Apply the denosing loop + upsampling_output = self.upsampling_pipe( + prompt=upsampling_prompt, + image=upsampling_image, + mask_image=upsampling_mask, + height=upsampling_height, + width=upsampling_width, + strength=upsampling_strength, + num_inference_steps=num_inference_steps, + sigmas=sigmas, + guidance_scale=guidance_scale, + generator=upsampling_generator, + output_type=output_type, + joint_attention_kwargs=joint_attention_kwargs, + callback_on_step_end=callback_on_step_end, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + image = upsampling_output.images + + output = [] + if output_type == "pil": + # Each sample in the batch may have multiple output images. When returning as PIL images, + # these images cannot be concatenated. Therefore, for each sample, + # a list is used to represent all the output images. + output = [] + start = 0 + for n in n_target_per_sample: + output.append(image[start : start + n]) + start += n + else: + output = image + + if not return_dict: + return (output,) + + return FluxPipelineOutput(images=output) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/pipeline_visualcloze_generation.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/pipeline_visualcloze_generation.py new file mode 100644 index 0000000000000000000000000000000000000000..e7a1d4a4b24873c9f6a69c437c3b9ef272fc6b19 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/pipeline_visualcloze_generation.py @@ -0,0 +1,952 @@ +# Copyright 2025 VisualCloze team and The HuggingFace Team. All rights reserved. +# +# 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 Any, Callable, Dict, List, Optional, Union + +import numpy as np +import torch +from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast + +from ...loaders import FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..flux.pipeline_flux_fill import calculate_shift, retrieve_latents, retrieve_timesteps +from ..flux.pipeline_output import FluxPipelineOutput +from ..pipeline_utils import DiffusionPipeline +from .visualcloze_utils import VisualClozeProcessor + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers import VisualClozeGenerationPipeline, FluxFillPipeline as VisualClozeUpsamplingPipeline + >>> from diffusers.utils import load_image + >>> from PIL import Image + + >>> image_paths = [ + ... # in-context examples + ... [ + ... load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/visualcloze/visualcloze_mask2image_incontext-example-1_mask.jpg" + ... ), + ... load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/visualcloze/visualcloze_mask2image_incontext-example-1_image.jpg" + ... ), + ... ], + ... # query with the target image + ... [ + ... load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/visualcloze/visualcloze_mask2image_query_mask.jpg" + ... ), + ... None, # No image needed for the target image + ... ], + ... ] + >>> task_prompt = "In each row, a logical task is demonstrated to achieve [IMAGE2] an aesthetically pleasing photograph based on [IMAGE1] sam 2-generated masks with rich color coding." + >>> content_prompt = "Majestic photo of a golden eagle perched on a rocky outcrop in a mountainous landscape. The eagle is positioned in the right foreground, facing left, with its sharp beak and keen eyes prominently visible. Its plumage is a mix of dark brown and golden hues, with intricate feather details. The background features a soft-focus view of snow-capped mountains under a cloudy sky, creating a serene and grandiose atmosphere. The foreground includes rugged rocks and patches of green moss. Photorealistic, medium depth of field, soft natural lighting, cool color palette, high contrast, sharp focus on the eagle, blurred background, tranquil, majestic, wildlife photography." + >>> pipe = VisualClozeGenerationPipeline.from_pretrained( + ... "VisualCloze/VisualClozePipeline-384", resolution=384, torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> image = pipe( + ... task_prompt=task_prompt, + ... content_prompt=content_prompt, + ... image=image_paths, + ... guidance_scale=30, + ... num_inference_steps=30, + ... max_sequence_length=512, + ... generator=torch.Generator("cpu").manual_seed(0), + ... ).images[0][0] + + >>> # optional, upsampling the generated image + >>> pipe_upsample = VisualClozeUpsamplingPipeline.from_pipe(pipe) + >>> pipe_upsample.to("cuda") + + >>> mask_image = Image.new("RGB", image.size, (255, 255, 255)) + + >>> image = pipe_upsample( + ... image=image, + ... mask_image=mask_image, + ... prompt=content_prompt, + ... width=1344, + ... height=768, + ... strength=0.4, + ... guidance_scale=30, + ... num_inference_steps=30, + ... max_sequence_length=512, + ... generator=torch.Generator("cpu").manual_seed(0), + ... ).images[0] + + >>> image.save("visualcloze.png") + ``` +""" + + +class VisualClozeGenerationPipeline( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, +): + r""" + The VisualCloze pipeline for image generation with visual context. Reference: + https://github.com/lzyhha/VisualCloze/tree/main This pipeline is designed to generate images based on visual + in-context examples. + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + resolution (`int`, *optional*, defaults to 384): + The resolution of each image when concatenating images from the query and in-context examples. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + resolution: int = 384, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + ) + self.resolution = resolution + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VisualClozeProcessor( + vae_scale_factor=self.vae_scale_factor * 2, vae_latent_channels=self.latent_channels, resolution=resolution + ) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: Union[str, List[str]], + num_images_per_prompt: int = 1, + device: Optional[torch.device] = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + # Modified from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt + def encode_prompt( + self, + layout_prompt: Union[str, List[str]], + task_prompt: Union[str, List[str]], + content_prompt: Union[str, List[str]], + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + max_sequence_length: int = 512, + lora_scale: Optional[float] = None, + ): + r""" + + Args: + layout_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to define the number of in-context examples and the number of images involved in + the task. + task_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to define the task intention. + content_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to define the content or caption of the target image to be generated. + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + if isinstance(layout_prompt, str): + layout_prompt = [layout_prompt] + task_prompt = [task_prompt] + content_prompt = [content_prompt] + + def _preprocess(prompt, content=False): + if prompt is not None: + return f"The last image of the last row depicts: {prompt}" if content else prompt + else: + return "" + + prompt = [ + f"{_preprocess(layout_prompt[i])} {_preprocess(task_prompt[i])} {_preprocess(content_prompt[i], content=True)}".strip() + for i in range(len(layout_prompt)) + ] + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_inpaint.StableDiffusion3InpaintPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def check_inputs( + self, + image, + task_prompt, + content_prompt, + prompt_embeds=None, + pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + # Validate prompt inputs + if (task_prompt is not None or content_prompt is not None) and prompt_embeds is not None: + raise ValueError("Cannot provide both text `task_prompt` + `content_prompt` and `prompt_embeds`. ") + + if task_prompt is None and content_prompt is None and prompt_embeds is None: + raise ValueError("Must provide either `task_prompt` + `content_prompt` or pre-computed `prompt_embeds`. ") + + # Validate prompt types and consistency + if task_prompt is None: + raise ValueError("`task_prompt` is missing.") + + if task_prompt is not None and not isinstance(task_prompt, (str, list)): + raise ValueError(f"`task_prompt` must be str or list, got {type(task_prompt)}") + + if content_prompt is not None and not isinstance(content_prompt, (str, list)): + raise ValueError(f"`content_prompt` must be str or list, got {type(content_prompt)}") + + if isinstance(task_prompt, list) or isinstance(content_prompt, list): + if not isinstance(task_prompt, list) or not isinstance(content_prompt, list): + raise ValueError( + f"`task_prompt` and `content_prompt` must both be lists, or both be of type str or None, " + f"got {type(task_prompt)} and {type(content_prompt)}" + ) + if len(content_prompt) != len(task_prompt): + raise ValueError("`task_prompt` and `content_prompt` must have the same length whe they are lists.") + + for sample in image: + if not isinstance(sample, list) or not isinstance(sample[0], list): + raise ValueError("Each sample in the batch must have a 2D list of images.") + if len({len(row) for row in sample}) != 1: + raise ValueError("Each in-context example and query should contain the same number of images.") + if not any(img is None for img in sample[-1]): + raise ValueError("There are no targets in the query, which should be represented as None.") + for row in sample[:-1]: + if any(img is None for img in row): + raise ValueError("Images are missing in in-context examples.") + + # Validate embeddings + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + # Validate sequence length + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"max_sequence_length cannot exceed 512, got {max_sequence_length}") + + @staticmethod + def _prepare_latent_image_ids(image, vae_scale_factor, device, dtype): + latent_image_ids = [] + + for idx, img in enumerate(image, start=1): + img = img.squeeze(0) + channels, height, width = img.shape + + num_patches_h = height // vae_scale_factor // 2 + num_patches_w = width // vae_scale_factor // 2 + + patch_ids = torch.zeros(num_patches_h, num_patches_w, 3, device=device, dtype=dtype) + patch_ids[..., 0] = idx + patch_ids[..., 1] = torch.arange(num_patches_h, device=device, dtype=dtype)[:, None] + patch_ids[..., 2] = torch.arange(num_patches_w, device=device, dtype=dtype)[None, :] + + patch_ids = patch_ids.reshape(-1, 3) + latent_image_ids.append(patch_ids) + + return torch.cat(latent_image_ids, dim=0) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + def _unpack_latents(latents, sizes, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + start = 0 + unpacked_latents = [] + for i in range(len(sizes)): + cur_size = sizes[i] + height = cur_size[0][0] // vae_scale_factor + width = sum([size[1] for size in cur_size]) // vae_scale_factor + + end = start + (height * width) // 4 + + cur_latents = latents[:, start:end] + cur_latents = cur_latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + cur_latents = cur_latents.permute(0, 3, 1, 4, 2, 5) + cur_latents = cur_latents.reshape(batch_size, channels // (2 * 2), height, width) + + unpacked_latents.append(cur_latents) + + start = end + + return unpacked_latents + + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + def _prepare_latents(self, image, mask, gen, vae_scale_factor, device, dtype): + """Helper function to prepare latents for a single batch.""" + # Concatenate images and masks along width dimension + image = [torch.cat(img, dim=3).to(device=device, dtype=dtype) for img in image] + mask = [torch.cat(m, dim=3).to(device=device, dtype=dtype) for m in mask] + + # Generate latent image IDs + latent_image_ids = self._prepare_latent_image_ids(image, vae_scale_factor, device, dtype) + + # For initial encoding, use actual images + image_latent = [self._encode_vae_image(img, gen) for img in image] + masked_image_latent = [img.clone() for img in image_latent] + + for i in range(len(image_latent)): + # Rearrange latents and masks for patch processing + num_channels_latents, height, width = image_latent[i].shape[1:] + image_latent[i] = self._pack_latents(image_latent[i], 1, num_channels_latents, height, width) + masked_image_latent[i] = self._pack_latents(masked_image_latent[i], 1, num_channels_latents, height, width) + + # Rearrange masks for patch processing + num_channels_latents, height, width = mask[i].shape[1:] + mask[i] = mask[i].view( + 1, + num_channels_latents, + height // vae_scale_factor, + vae_scale_factor, + width // vae_scale_factor, + vae_scale_factor, + ) + mask[i] = mask[i].permute(0, 1, 3, 5, 2, 4) + mask[i] = mask[i].reshape( + 1, + num_channels_latents * (vae_scale_factor**2), + height // vae_scale_factor, + width // vae_scale_factor, + ) + mask[i] = self._pack_latents( + mask[i], + 1, + num_channels_latents * (vae_scale_factor**2), + height // vae_scale_factor, + width // vae_scale_factor, + ) + + # Concatenate along batch dimension + image_latent = torch.cat(image_latent, dim=1) + masked_image_latent = torch.cat(masked_image_latent, dim=1) + mask = torch.cat(mask, dim=1) + + return image_latent, masked_image_latent, mask, latent_image_ids + + def prepare_latents( + self, + input_image, + input_mask, + timestep, + batch_size, + dtype, + device, + generator, + vae_scale_factor, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # Process each batch + masked_image_latents = [] + image_latents = [] + masks = [] + latent_image_ids = [] + + for i in range(len(input_image)): + _image_latent, _masked_image_latent, _mask, _latent_image_ids = self._prepare_latents( + input_image[i], + input_mask[i], + generator if isinstance(generator, torch.Generator) else generator[i], + vae_scale_factor, + device, + dtype, + ) + masked_image_latents.append(_masked_image_latent) + image_latents.append(_image_latent) + masks.append(_mask) + latent_image_ids.append(_latent_image_ids) + + # Concatenate all batches + masked_image_latents = torch.cat(masked_image_latents, dim=0) + image_latents = torch.cat(image_latents, dim=0) + masks = torch.cat(masks, dim=0) + + # Handle batch size expansion + if batch_size > masked_image_latents.shape[0]: + if batch_size % masked_image_latents.shape[0] == 0: + # Expand batches by repeating + additional_image_per_prompt = batch_size // masked_image_latents.shape[0] + masked_image_latents = torch.cat([masked_image_latents] * additional_image_per_prompt, dim=0) + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + masks = torch.cat([masks] * additional_image_per_prompt, dim=0) + else: + raise ValueError( + f"Cannot expand batch size from {masked_image_latents.shape[0]} to {batch_size}. " + "Batch sizes must be multiples of each other." + ) + + # Add noise to latents + noises = randn_tensor(image_latents.shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noises).to(dtype=dtype) + + # Combine masked latents with masks + masked_image_latents = torch.cat((masked_image_latents, masks), dim=-1).to(dtype=dtype) + + return latents, masked_image_latents, latent_image_ids[0] + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + task_prompt: Union[str, List[str]] = None, + content_prompt: Union[str, List[str]] = None, + image: Optional[torch.FloatTensor] = None, + num_inference_steps: int = 50, + sigmas: Optional[List[float]] = None, + guidance_scale: float = 30.0, + num_images_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the VisualCloze pipeline for generation. + + Args: + task_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to define the task intention. + content_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to define the content or caption of the target image to be generated. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 30.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + image, + task_prompt, + content_prompt, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + processor_output = self.image_processor.preprocess( + task_prompt, content_prompt, image, vae_scale_factor=self.vae_scale_factor + ) + + # 2. Define call parameters + if processor_output["task_prompt"] is not None and isinstance(processor_output["task_prompt"], str): + batch_size = 1 + elif processor_output["task_prompt"] is not None and isinstance(processor_output["task_prompt"], list): + batch_size = len(processor_output["task_prompt"]) + + device = self._execution_device + + # 3. Prepare prompt embeddings + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + prompt_embeds, pooled_prompt_embeds, text_ids = self.encode_prompt( + layout_prompt=processor_output["layout_prompt"], + task_prompt=processor_output["task_prompt"], + content_prompt=processor_output["content_prompt"], + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 4. Prepare timesteps + # Calculate sequence length and shift factor + image_seq_len = sum( + (size[0] // self.vae_scale_factor // 2) * (size[1] // self.vae_scale_factor // 2) + for sample in processor_output["image_size"][0] + for size in sample + ) + + # Calculate noise schedule parameters + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + + # Get timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, 1.0, device) + + # 5. Prepare latent variables + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + latents, masked_image_latents, latent_image_ids = self.prepare_latents( + processor_output["init_image"], + processor_output["mask"], + latent_timestep, + batch_size * num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + vae_scale_factor=self.vae_scale_factor, + ) + + # Calculate warmup steps + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # Prepare guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # Broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + latent_model_input = torch.cat((latents, masked_image_latents), dim=2) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # Compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # Some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # Call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + # XLA optimization + if XLA_AVAILABLE: + xm.mark_step() + + # 7. Post-process the image + # Crop the target image + # Since the generated image is a concatenation of the conditional and target regions, + # we need to extract only the target regions based on their positions + image = [] + if output_type == "latent": + image = latents + else: + for b in range(len(latents)): + cur_image_size = processor_output["image_size"][b % batch_size] + cur_target_position = processor_output["target_position"][b % batch_size] + cur_latent = self._unpack_latents(latents[b].unsqueeze(0), cur_image_size, self.vae_scale_factor)[-1] + cur_latent = (cur_latent / self.vae.config.scaling_factor) + self.vae.config.shift_factor + cur_image = self.vae.decode(cur_latent, return_dict=False)[0] + cur_image = self.image_processor.postprocess(cur_image, output_type=output_type)[0] + + start = 0 + cropped = [] + for i, size in enumerate(cur_image_size[-1]): + if cur_target_position[i]: + if output_type == "pil": + cropped.append(cur_image.crop((start, 0, start + size[1], size[0]))) + else: + cropped.append(cur_image[0 : size[0], start : start + size[1]]) + start += size[1] + image.append(cropped) + if output_type != "pil": + image = np.concatenate([arr[None] for sub_image in image for arr in sub_image], axis=0) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/visualcloze_utils.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/visualcloze_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..5d221bc1e8b122642285760f137464b59c3f747d --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/visualcloze/visualcloze_utils.py @@ -0,0 +1,251 @@ +# Copyright 2025 VisualCloze team and The HuggingFace Team. All rights reserved. +# +# 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 Dict, List, Optional, Tuple, Union + +import torch +from PIL import Image + +from ...image_processor import VaeImageProcessor + + +class VisualClozeProcessor(VaeImageProcessor): + """ + Image processor for the VisualCloze pipeline. + + This processor handles the preprocessing of images for visual cloze tasks, including resizing, normalization, and + mask generation. + + Args: + resolution (int, optional): + Target resolution for processing images. Each image will be resized to this resolution before being + concatenated to avoid the out-of-memory error. Defaults to 384. + *args: Additional arguments passed to [~image_processor.VaeImageProcessor] + **kwargs: Additional keyword arguments passed to [~image_processor.VaeImageProcessor] + """ + + def __init__(self, *args, resolution: int = 384, **kwargs): + super().__init__(*args, **kwargs) + self.resolution = resolution + + def preprocess_image( + self, input_images: List[List[Optional[Image.Image]]], vae_scale_factor: int + ) -> Tuple[List[List[torch.Tensor]], List[List[List[int]]], List[int]]: + """ + Preprocesses input images for the VisualCloze pipeline. + + This function handles the preprocessing of input images by: + 1. Resizing and cropping images to maintain consistent dimensions + 2. Converting images to the Tensor format for the VAE + 3. Normalizing pixel values + 4. Tracking image sizes and positions of target images + + Args: + input_images (List[List[Optional[Image.Image]]]): + A nested list of PIL Images where: + - Outer list represents different samples, including in-context examples and the query + - Inner list contains images for the task + - In the last row, condition images are provided and the target images are placed as None + vae_scale_factor (int): + The scale factor used by the VAE for resizing images + + Returns: + Tuple containing: + - List[List[torch.Tensor]]: Preprocessed images in tensor format + - List[List[List[int]]]: Dimensions of each processed image [height, width] + - List[int]: Target positions indicating which images are to be generated + """ + n_samples, n_task_images = len(input_images), len(input_images[0]) + divisible = 2 * vae_scale_factor + + processed_images: List[List[Image.Image]] = [[] for _ in range(n_samples)] + resize_size: List[Optional[Tuple[int, int]]] = [None for _ in range(n_samples)] + target_position: List[int] = [] + + # Process each sample + for i in range(n_samples): + # Determine size from first non-None image + for j in range(n_task_images): + if input_images[i][j] is not None: + aspect_ratio = input_images[i][j].width / input_images[i][j].height + target_area = self.resolution * self.resolution + new_h = int((target_area / aspect_ratio) ** 0.5) + new_w = int(new_h * aspect_ratio) + + new_w = max(new_w // divisible, 1) * divisible + new_h = max(new_h // divisible, 1) * divisible + resize_size[i] = (new_w, new_h) + break + + # Process all images in the sample + for j in range(n_task_images): + if input_images[i][j] is not None: + target = self._resize_and_crop(input_images[i][j], resize_size[i][0], resize_size[i][1]) + processed_images[i].append(target) + if i == n_samples - 1: + target_position.append(0) + else: + blank = Image.new("RGB", resize_size[i] or (self.resolution, self.resolution), (0, 0, 0)) + processed_images[i].append(blank) + if i == n_samples - 1: + target_position.append(1) + + # Ensure consistent width for multiple target images when there are multiple target images + if len(target_position) > 1 and sum(target_position) > 1: + new_w = resize_size[n_samples - 1][0] or 384 + for i in range(len(processed_images)): + for j in range(len(processed_images[i])): + if processed_images[i][j] is not None: + new_h = int(processed_images[i][j].height * (new_w / processed_images[i][j].width)) + new_w = int(new_w / 16) * 16 + new_h = int(new_h / 16) * 16 + processed_images[i][j] = self.height(processed_images[i][j], new_h, new_w) + + # Convert to tensors and normalize + image_sizes = [] + for i in range(len(processed_images)): + image_sizes.append([[img.height, img.width] for img in processed_images[i]]) + for j, image in enumerate(processed_images[i]): + image = self.pil_to_numpy(image) + image = self.numpy_to_pt(image) + image = self.normalize(image) + processed_images[i][j] = image + + return processed_images, image_sizes, target_position + + def preprocess_mask( + self, input_images: List[List[Image.Image]], target_position: List[int] + ) -> List[List[torch.Tensor]]: + """ + Generate masks for the VisualCloze pipeline. + + Args: + input_images (List[List[Image.Image]]): + Processed images from preprocess_image + target_position (List[int]): + Binary list marking the positions of target images (1 for target, 0 for condition) + + Returns: + List[List[torch.Tensor]]: + A nested list of mask tensors (1 for target positions, 0 for condition images) + """ + mask = [] + for i, row in enumerate(input_images): + if i == len(input_images) - 1: # Query row + row_masks = [ + torch.full((1, 1, row[0].shape[2], row[0].shape[3]), fill_value=m) for m in target_position + ] + else: # In-context examples + row_masks = [ + torch.full((1, 1, row[0].shape[2], row[0].shape[3]), fill_value=0) for _ in target_position + ] + mask.append(row_masks) + return mask + + def preprocess_image_upsampling( + self, + input_images: List[List[Image.Image]], + height: int, + width: int, + ) -> Tuple[List[List[Image.Image]], List[List[List[int]]]]: + """Process images for the upsampling stage in the VisualCloze pipeline. + + Args: + input_images: Input image to process + height: Target height + width: Target width + + Returns: + Tuple of processed image and its size + """ + image = self.resize(input_images[0][0], height, width) + image = self.pil_to_numpy(image) # to np + image = self.numpy_to_pt(image) # to pt + image = self.normalize(image) + + input_images[0][0] = image + image_sizes = [[[height, width]]] + return input_images, image_sizes + + def preprocess_mask_upsampling(self, input_images: List[List[Image.Image]]) -> List[List[torch.Tensor]]: + return [[torch.ones((1, 1, input_images[0][0].shape[2], input_images[0][0].shape[3]))]] + + def get_layout_prompt(self, size: Tuple[int, int]) -> str: + layout_instruction = ( + f"A grid layout with {size[0]} rows and {size[1]} columns, displaying {size[0] * size[1]} images arranged side by side.", + ) + return layout_instruction + + def preprocess( + self, + task_prompt: Union[str, List[str]], + content_prompt: Union[str, List[str]], + input_images: Optional[List[List[List[Optional[str]]]]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + upsampling: bool = False, + vae_scale_factor: int = 16, + ) -> Dict: + """Process visual cloze inputs. + + Args: + task_prompt: Task description(s) + content_prompt: Content description(s) + input_images: List of images or None for the target images + height: Optional target height for upsampling stage + width: Optional target width for upsampling stage + upsampling: Whether this is in the upsampling processing stage + + Returns: + Dictionary containing processed images, masks, prompts and metadata + """ + if isinstance(task_prompt, str): + task_prompt = [task_prompt] + content_prompt = [content_prompt] + input_images = [input_images] + + output = { + "init_image": [], + "mask": [], + "task_prompt": task_prompt if not upsampling else [None for _ in range(len(task_prompt))], + "content_prompt": content_prompt, + "layout_prompt": [], + "target_position": [], + "image_size": [], + } + for i in range(len(task_prompt)): + if upsampling: + layout_prompt = None + else: + layout_prompt = self.get_layout_prompt((len(input_images[i]), len(input_images[i][0]))) + + if upsampling: + cur_processed_images, cur_image_size = self.preprocess_image_upsampling( + input_images[i], height=height, width=width + ) + cur_mask = self.preprocess_mask_upsampling(cur_processed_images) + else: + cur_processed_images, cur_image_size, cur_target_position = self.preprocess_image( + input_images[i], vae_scale_factor=vae_scale_factor + ) + cur_mask = self.preprocess_mask(cur_processed_images, cur_target_position) + + output["target_position"].append(cur_target_position) + + output["image_size"].append(cur_image_size) + output["init_image"].append(cur_processed_images) + output["mask"].append(cur_mask) + output["layout_prompt"].append(layout_prompt) + + return output diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..bb96372b1db2b2c29f86cd7d75428161d4bbc824 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__init__.py @@ -0,0 +1,53 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_wan"] = ["WanPipeline"] + _import_structure["pipeline_wan_i2v"] = ["WanImageToVideoPipeline"] + _import_structure["pipeline_wan_vace"] = ["WanVACEPipeline"] + _import_structure["pipeline_wan_video2video"] = ["WanVideoToVideoPipeline"] +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_wan import WanPipeline + from .pipeline_wan_i2v import WanImageToVideoPipeline + from .pipeline_wan_vace import WanVACEPipeline + from .pipeline_wan_video2video import WanVideoToVideoPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..148b697e940ca0511560c7a2b3307a870ce07be2 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_output.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_output.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6fa4bc0e8d84b3ea76d7b25e2a3de4e40465cca9 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_output.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_wan.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_wan.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..303475ab7877949883bec6c0b3ef6677737b6880 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_wan.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_wan_i2v.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_wan_i2v.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c02de1ac80b7b938ae89e937780b43194037f2f9 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_wan_i2v.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_wan_vace.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_wan_vace.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..edb9ccd0d419d1512c62fe2796a4ea867065c69e Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_wan_vace.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_wan_video2video.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_wan_video2video.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2bcdf54f4d52e25e23f2c590f351ecd3f95c6140 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/__pycache__/pipeline_wan_video2video.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_output.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..88907ad0f0a1dfc828879f30bdd31b9043396144 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import torch + +from diffusers.utils import BaseOutput + + +@dataclass +class WanPipelineOutput(BaseOutput): + r""" + Output class for Wan pipelines. + + Args: + frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]): + List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)`. + """ + + frames: torch.Tensor diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_wan.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_wan.py new file mode 100644 index 0000000000000000000000000000000000000000..6df66118b068d6f802382b54739c93f436ff3fa4 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_wan.py @@ -0,0 +1,598 @@ +# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved. +# +# 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 html +from typing import Any, Callable, Dict, List, Optional, Union + +import regex as re +import torch +from transformers import AutoTokenizer, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import WanLoraLoaderMixin +from ...models import AutoencoderKLWan, WanTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import WanPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers.utils import export_to_video + >>> from diffusers import AutoencoderKLWan, WanPipeline + >>> from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler + + >>> # Available models: Wan-AI/Wan2.1-T2V-14B-Diffusers, Wan-AI/Wan2.1-T2V-1.3B-Diffusers + >>> model_id = "Wan-AI/Wan2.1-T2V-14B-Diffusers" + >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32) + >>> pipe = WanPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16) + >>> flow_shift = 5.0 # 5.0 for 720P, 3.0 for 480P + >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift) + >>> pipe.to("cuda") + + >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window." + >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards" + + >>> output = pipe( + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... height=720, + ... width=1280, + ... num_frames=81, + ... guidance_scale=5.0, + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=16) + ``` +""" + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +class WanPipeline(DiffusionPipeline, WanLoraLoaderMixin): + r""" + Pipeline for text-to-video generation using Wan. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + tokenizer ([`T5Tokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + transformer ([`WanTransformer3DModel`]): + Conditional Transformer to denoise the input latents. + scheduler ([`UniPCMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: UMT5EncoderModel, + transformer: WanTransformer3DModel, + vae: AutoencoderKLWan, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + + self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 226, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds + + def encode_prompt( + self, + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + max_sequence_length: int = 226, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + height, + width, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 81, + dtype: Optional[torch.dtype] = None, + device: Optional[torch.device] = None, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + shape = ( + batch_size, + num_channels_latents, + num_latent_frames, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + negative_prompt: Union[str, List[str]] = None, + height: int = 480, + width: int = 832, + num_frames: int = 81, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + num_videos_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "np", + return_dict: bool = True, + attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, pass `prompt_embeds` instead. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to avoid during image generation. If not defined, pass `negative_prompt_embeds` + instead. Ignored when not using guidance (`guidance_scale` < `1`). + height (`int`, defaults to `480`): + The height in pixels of the generated image. + width (`int`, defaults to `832`): + The width in pixels of the generated image. + num_frames (`int`, defaults to `81`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, defaults to `512`): + The maximum sequence length of the text encoder. If the prompt is longer than this, it will be + truncated. If the prompt is shorter, it will be padded to this length. + + Examples: + + Returns: + [`~WanPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where + the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + height, + width, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + ) + + if num_frames % self.vae_scale_factor_temporal != 1: + logger.warning( + f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number." + ) + num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + num_frames = max(num_frames, 1) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + transformer_dtype = self.transformer.dtype + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + torch.float32, + device, + generator, + latents, + ) + + # 6. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + latent_model_input = latents.to(transformer_dtype) + timestep = t.expand(latents.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + noise_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if not output_type == "latent": + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return WanPipelineOutput(frames=video) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_wan_i2v.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_wan_i2v.py new file mode 100644 index 0000000000000000000000000000000000000000..c71138a97dd9ade607fc10a5f8a6ca05447f1653 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_wan_i2v.py @@ -0,0 +1,744 @@ +# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved. +# +# 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 html +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import PIL +import regex as re +import torch +from transformers import AutoTokenizer, CLIPImageProcessor, CLIPVisionModel, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import WanLoraLoaderMixin +from ...models import AutoencoderKLWan, WanTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import WanPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> import numpy as np + >>> from diffusers import AutoencoderKLWan, WanImageToVideoPipeline + >>> from diffusers.utils import export_to_video, load_image + >>> from transformers import CLIPVisionModel + + >>> # Available models: Wan-AI/Wan2.1-I2V-14B-480P-Diffusers, Wan-AI/Wan2.1-I2V-14B-720P-Diffusers + >>> model_id = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers" + >>> image_encoder = CLIPVisionModel.from_pretrained( + ... model_id, subfolder="image_encoder", torch_dtype=torch.float32 + ... ) + >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32) + >>> pipe = WanImageToVideoPipeline.from_pretrained( + ... model_id, vae=vae, image_encoder=image_encoder, torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg" + ... ) + >>> max_area = 480 * 832 + >>> aspect_ratio = image.height / image.width + >>> mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1] + >>> height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value + >>> width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value + >>> image = image.resize((width, height)) + >>> prompt = ( + ... "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in " + ... "the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot." + ... ) + >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards" + + >>> output = pipe( + ... image=image, + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... height=height, + ... width=width, + ... num_frames=81, + ... guidance_scale=5.0, + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=16) + ``` +""" + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class WanImageToVideoPipeline(DiffusionPipeline, WanLoraLoaderMixin): + r""" + Pipeline for image-to-video generation using Wan. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + tokenizer ([`T5Tokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + image_encoder ([`CLIPVisionModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPVisionModel), specifically + the + [clip-vit-huge-patch14](https://github.com/mlfoundations/open_clip/blob/main/docs/PRETRAINED.md#vit-h14-xlm-roberta-large) + variant. + transformer ([`WanTransformer3DModel`]): + Conditional Transformer to denoise the input latents. + scheduler ([`UniPCMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: UMT5EncoderModel, + image_encoder: CLIPVisionModel, + image_processor: CLIPImageProcessor, + transformer: WanTransformer3DModel, + vae: AutoencoderKLWan, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + image_encoder=image_encoder, + transformer=transformer, + scheduler=scheduler, + image_processor=image_processor, + ) + + self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + self.image_processor = image_processor + + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 512, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds + + def encode_image( + self, + image: PipelineImageInput, + device: Optional[torch.device] = None, + ): + device = device or self._execution_device + image = self.image_processor(images=image, return_tensors="pt").to(device) + image_embeds = self.image_encoder(**image, output_hidden_states=True) + return image_embeds.hidden_states[-2] + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.encode_prompt + def encode_prompt( + self, + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + max_sequence_length: int = 226, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + image, + height, + width, + prompt_embeds=None, + negative_prompt_embeds=None, + image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if image is not None and image_embeds is not None: + raise ValueError( + f"Cannot forward both `image`: {image} and `image_embeds`: {image_embeds}. Please make sure to" + " only forward one of the two." + ) + if image is None and image_embeds is None: + raise ValueError( + "Provide either `image` or `prompt_embeds`. Cannot leave both `image` and `image_embeds` undefined." + ) + if image is not None and not isinstance(image, torch.Tensor) and not isinstance(image, PIL.Image.Image): + raise ValueError(f"`image` has to be of type `torch.Tensor` or `PIL.Image.Image` but is {type(image)}") + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + def prepare_latents( + self, + image: PipelineImageInput, + batch_size: int, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 81, + dtype: Optional[torch.dtype] = None, + device: Optional[torch.device] = None, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + last_image: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + latent_height = height // self.vae_scale_factor_spatial + latent_width = width // self.vae_scale_factor_spatial + + shape = (batch_size, num_channels_latents, num_latent_frames, latent_height, latent_width) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device=device, dtype=dtype) + + image = image.unsqueeze(2) + if last_image is None: + video_condition = torch.cat( + [image, image.new_zeros(image.shape[0], image.shape[1], num_frames - 1, height, width)], dim=2 + ) + else: + last_image = last_image.unsqueeze(2) + video_condition = torch.cat( + [image, image.new_zeros(image.shape[0], image.shape[1], num_frames - 2, height, width), last_image], + dim=2, + ) + video_condition = video_condition.to(device=device, dtype=self.vae.dtype) + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + + if isinstance(generator, list): + latent_condition = [ + retrieve_latents(self.vae.encode(video_condition), sample_mode="argmax") for _ in generator + ] + latent_condition = torch.cat(latent_condition) + else: + latent_condition = retrieve_latents(self.vae.encode(video_condition), sample_mode="argmax") + latent_condition = latent_condition.repeat(batch_size, 1, 1, 1, 1) + + latent_condition = latent_condition.to(dtype) + latent_condition = (latent_condition - latents_mean) * latents_std + + mask_lat_size = torch.ones(batch_size, 1, num_frames, latent_height, latent_width) + + if last_image is None: + mask_lat_size[:, :, list(range(1, num_frames))] = 0 + else: + mask_lat_size[:, :, list(range(1, num_frames - 1))] = 0 + first_frame_mask = mask_lat_size[:, :, 0:1] + first_frame_mask = torch.repeat_interleave(first_frame_mask, dim=2, repeats=self.vae_scale_factor_temporal) + mask_lat_size = torch.concat([first_frame_mask, mask_lat_size[:, :, 1:, :]], dim=2) + mask_lat_size = mask_lat_size.view(batch_size, -1, self.vae_scale_factor_temporal, latent_height, latent_width) + mask_lat_size = mask_lat_size.transpose(1, 2) + mask_lat_size = mask_lat_size.to(latent_condition.device) + + return latents, torch.concat([mask_lat_size, latent_condition], dim=1) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput, + prompt: Union[str, List[str]] = None, + negative_prompt: Union[str, List[str]] = None, + height: int = 480, + width: int = 832, + num_frames: int = 81, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + num_videos_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + image_embeds: Optional[torch.Tensor] = None, + last_image: Optional[torch.Tensor] = None, + output_type: Optional[str] = "np", + return_dict: bool = True, + attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PipelineImageInput`): + The input image to condition the generation on. Must be an image, a list of images or a `torch.Tensor`. + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + height (`int`, defaults to `480`): + The height of the generated video. + width (`int`, defaults to `832`): + The width of the generated video. + num_frames (`int`, defaults to `81`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `negative_prompt` input argument. + image_embeds (`torch.Tensor`, *optional*): + Pre-generated image embeddings. Can be used to easily tweak image inputs (weighting). If not provided, + image embeddings are generated from the `image` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, defaults to `512`): + The maximum sequence length of the text encoder. If the prompt is longer than this, it will be + truncated. If the prompt is shorter, it will be padded to this length. + + Examples: + + Returns: + [`~WanPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where + the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + image, + height, + width, + prompt_embeds, + negative_prompt_embeds, + image_embeds, + callback_on_step_end_tensor_inputs, + ) + + if num_frames % self.vae_scale_factor_temporal != 1: + logger.warning( + f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number." + ) + num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + num_frames = max(num_frames, 1) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + # Encode image embedding + transformer_dtype = self.transformer.dtype + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + if image_embeds is None: + if last_image is None: + image_embeds = self.encode_image(image, device) + else: + image_embeds = self.encode_image([image, last_image], device) + image_embeds = image_embeds.repeat(batch_size, 1, 1) + image_embeds = image_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.vae.config.z_dim + image = self.video_processor.preprocess(image, height=height, width=width).to(device, dtype=torch.float32) + if last_image is not None: + last_image = self.video_processor.preprocess(last_image, height=height, width=width).to( + device, dtype=torch.float32 + ) + latents, condition = self.prepare_latents( + image, + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + torch.float32, + device, + generator, + latents, + last_image, + ) + + # 6. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + latent_model_input = torch.cat([latents, condition], dim=1).to(transformer_dtype) + timestep = t.expand(latents.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + encoder_hidden_states_image=image_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + noise_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + encoder_hidden_states_image=image_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if not output_type == "latent": + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return WanPipelineOutput(frames=video) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_wan_vace.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_wan_vace.py new file mode 100644 index 0000000000000000000000000000000000000000..e5f83dd401ad0406e709f75b6645dc20316b857e --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_wan_vace.py @@ -0,0 +1,976 @@ +# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved. +# +# 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 html +from typing import Any, Callable, Dict, List, Optional, Union + +import PIL.Image +import regex as re +import torch +from transformers import AutoTokenizer, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import WanLoraLoaderMixin +from ...models import AutoencoderKLWan, WanVACETransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import WanPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> import PIL.Image + >>> from diffusers import AutoencoderKLWan, WanVACEPipeline + >>> from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler + >>> from diffusers.utils import export_to_video, load_image + def prepare_video_and_mask(first_img: PIL.Image.Image, last_img: PIL.Image.Image, height: int, width: int, num_frames: int): + first_img = first_img.resize((width, height)) + last_img = last_img.resize((width, height)) + frames = [] + frames.append(first_img) + # Ideally, this should be 127.5 to match original code, but they perform computation on numpy arrays + # whereas we are passing PIL images. If you choose to pass numpy arrays, you can set it to 127.5 to + # match the original code. + frames.extend([PIL.Image.new("RGB", (width, height), (128, 128, 128))] * (num_frames - 2)) + frames.append(last_img) + mask_black = PIL.Image.new("L", (width, height), 0) + mask_white = PIL.Image.new("L", (width, height), 255) + mask = [mask_black, *[mask_white] * (num_frames - 2), mask_black] + return frames, mask + + >>> # Available checkpoints: Wan-AI/Wan2.1-VACE-1.3B-diffusers, Wan-AI/Wan2.1-VACE-14B-diffusers + >>> model_id = "Wan-AI/Wan2.1-VACE-1.3B-diffusers" + >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32) + >>> pipe = WanVACEPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16) + >>> flow_shift = 3.0 # 5.0 for 720P, 3.0 for 480P + >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift) + >>> pipe.to("cuda") + + >>> prompt = "CG animation style, a small blue bird takes off from the ground, flapping its wings. The bird's feathers are delicate, with a unique pattern on its chest. The background shows a blue sky with white clouds under bright sunshine. The camera follows the bird upward, capturing its flight and the vastness of the sky from a close-up, low-angle perspective." + >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards" + >>> first_frame = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_first_frame.png" + ... ) + >>> last_frame = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_last_frame.png>>> " + ... ) + + >>> height = 512 + >>> width = 512 + >>> num_frames = 81 + >>> video, mask = prepare_video_and_mask(first_frame, last_frame, height, width, num_frames) + + >>> output = pipe( + ... video=video, + ... mask=mask, + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... height=height, + ... width=width, + ... num_frames=num_frames, + ... num_inference_steps=30, + ... guidance_scale=5.0, + ... generator=torch.Generator().manual_seed(42), + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=16) + ``` +""" + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class WanVACEPipeline(DiffusionPipeline, WanLoraLoaderMixin): + r""" + Pipeline for controllable generation using Wan. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + tokenizer ([`T5Tokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + transformer ([`WanTransformer3DModel`]): + Conditional Transformer to denoise the input latents. + scheduler ([`UniPCMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: UMT5EncoderModel, + transformer: WanVACETransformer3DModel, + vae: AutoencoderKLWan, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + + self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 226, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.encode_prompt + def encode_prompt( + self, + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + max_sequence_length: int = 226, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + height, + width, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + video=None, + mask=None, + reference_images=None, + ): + base = self.vae_scale_factor_spatial * self.transformer.config.patch_size[1] + if height % base != 0 or width % base != 0: + raise ValueError(f"`height` and `width` have to be divisible by {base} but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + if video is not None: + if mask is not None: + if len(video) != len(mask): + raise ValueError( + f"Length of `video` {len(video)} and `mask` {len(mask)} do not match. Please make sure that" + " they have the same length." + ) + if reference_images is not None: + is_pil_image = isinstance(reference_images, PIL.Image.Image) + is_list_of_pil_images = isinstance(reference_images, list) and all( + isinstance(ref_img, PIL.Image.Image) for ref_img in reference_images + ) + is_list_of_list_of_pil_images = isinstance(reference_images, list) and all( + isinstance(ref_img, list) and all(isinstance(ref_img_, PIL.Image.Image) for ref_img_ in ref_img) + for ref_img in reference_images + ) + if not (is_pil_image or is_list_of_pil_images or is_list_of_list_of_pil_images): + raise ValueError( + "`reference_images` has to be of type `PIL.Image.Image` or `list` of `PIL.Image.Image`, or " + "`list` of `list` of `PIL.Image.Image`, but is {type(reference_images)}" + ) + if is_list_of_list_of_pil_images and len(reference_images) != 1: + raise ValueError( + "The pipeline only supports generating one video at a time at the moment. When passing a list " + "of list of reference images, where the outer list corresponds to the batch size and the inner " + "list corresponds to list of conditioning images per video, please make sure to only pass " + "one inner list of reference images (i.e., `[[, , ...]]`" + ) + elif mask is not None: + raise ValueError("`mask` can only be passed if `video` is passed as well.") + + def preprocess_conditions( + self, + video: Optional[List[PipelineImageInput]] = None, + mask: Optional[List[PipelineImageInput]] = None, + reference_images: Optional[Union[PIL.Image.Image, List[PIL.Image.Image], List[List[PIL.Image.Image]]]] = None, + batch_size: int = 1, + height: int = 480, + width: int = 832, + num_frames: int = 81, + dtype: Optional[torch.dtype] = None, + device: Optional[torch.device] = None, + ): + if video is not None: + base = self.vae_scale_factor_spatial * self.transformer.config.patch_size[1] + video_height, video_width = self.video_processor.get_default_height_width(video[0]) + + if video_height * video_width > height * width: + scale = min(width / video_width, height / video_height) + video_height, video_width = int(video_height * scale), int(video_width * scale) + + if video_height % base != 0 or video_width % base != 0: + logger.warning( + f"Video height and width should be divisible by {base}, but got {video_height} and {video_width}. " + ) + video_height = (video_height // base) * base + video_width = (video_width // base) * base + + assert video_height * video_width <= height * width + + video = self.video_processor.preprocess_video(video, video_height, video_width) + image_size = (video_height, video_width) # Use the height/width of video (with possible rescaling) + else: + video = torch.zeros(batch_size, 3, num_frames, height, width, dtype=dtype, device=device) + image_size = (height, width) # Use the height/width provider by user + + if mask is not None: + mask = self.video_processor.preprocess_video(mask, image_size[0], image_size[1]) + mask = torch.clamp((mask + 1) / 2, min=0, max=1) + else: + mask = torch.ones_like(video) + + video = video.to(dtype=dtype, device=device) + mask = mask.to(dtype=dtype, device=device) + + # Make a list of list of images where the outer list corresponds to video batch size and the inner list + # corresponds to list of conditioning images per video + if reference_images is None or isinstance(reference_images, PIL.Image.Image): + reference_images = [[reference_images] for _ in range(video.shape[0])] + elif isinstance(reference_images, (list, tuple)) and isinstance(next(iter(reference_images)), PIL.Image.Image): + reference_images = [reference_images] + elif ( + isinstance(reference_images, (list, tuple)) + and isinstance(next(iter(reference_images)), list) + and isinstance(next(iter(reference_images[0])), PIL.Image.Image) + ): + reference_images = reference_images + else: + raise ValueError( + "`reference_images` has to be of type `PIL.Image.Image` or `list` of `PIL.Image.Image`, or " + "`list` of `list` of `PIL.Image.Image`, but is {type(reference_images)}" + ) + + if video.shape[0] != len(reference_images): + raise ValueError( + f"Batch size of `video` {video.shape[0]} and length of `reference_images` {len(reference_images)} does not match." + ) + + ref_images_lengths = [len(reference_images_batch) for reference_images_batch in reference_images] + if any(l != ref_images_lengths[0] for l in ref_images_lengths): + raise ValueError( + f"All batches of `reference_images` should have the same length, but got {ref_images_lengths}. Support for this " + "may be added in the future." + ) + + reference_images_preprocessed = [] + for i, reference_images_batch in enumerate(reference_images): + preprocessed_images = [] + for j, image in enumerate(reference_images_batch): + if image is None: + continue + image = self.video_processor.preprocess(image, None, None) + img_height, img_width = image.shape[-2:] + scale = min(image_size[0] / img_height, image_size[1] / img_width) + new_height, new_width = int(img_height * scale), int(img_width * scale) + resized_image = torch.nn.functional.interpolate( + image, size=(new_height, new_width), mode="bilinear", align_corners=False + ).squeeze(0) # [C, H, W] + top = (image_size[0] - new_height) // 2 + left = (image_size[1] - new_width) // 2 + canvas = torch.ones(3, *image_size, device=device, dtype=dtype) + canvas[:, top : top + new_height, left : left + new_width] = resized_image + preprocessed_images.append(canvas) + reference_images_preprocessed.append(preprocessed_images) + + return video, mask, reference_images_preprocessed + + def prepare_video_latents( + self, + video: torch.Tensor, + mask: torch.Tensor, + reference_images: Optional[List[List[torch.Tensor]]] = None, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + device: Optional[torch.device] = None, + ) -> torch.Tensor: + device = device or self._execution_device + + if isinstance(generator, list): + # TODO: support this + raise ValueError("Passing a list of generators is not yet supported. This may be supported in the future.") + + if reference_images is None: + # For each batch of video, we set no re + # ference image (as one or more can be passed by user) + reference_images = [[None] for _ in range(video.shape[0])] + else: + if video.shape[0] != len(reference_images): + raise ValueError( + f"Batch size of `video` {video.shape[0]} and length of `reference_images` {len(reference_images)} does not match." + ) + + if video.shape[0] != 1: + # TODO: support this + raise ValueError( + "Generating with more than one video is not yet supported. This may be supported in the future." + ) + + vae_dtype = self.vae.dtype + video = video.to(dtype=vae_dtype) + + latents_mean = torch.tensor(self.vae.config.latents_mean, device=device, dtype=torch.float32).view( + 1, self.vae.config.z_dim, 1, 1, 1 + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std, device=device, dtype=torch.float32).view( + 1, self.vae.config.z_dim, 1, 1, 1 + ) + + if mask is None: + latents = retrieve_latents(self.vae.encode(video), generator, sample_mode="argmax").unbind(0) + latents = ((latents.float() - latents_mean) * latents_std).to(vae_dtype) + else: + mask = mask.to(dtype=vae_dtype) + mask = torch.where(mask > 0.5, 1.0, 0.0) + inactive = video * (1 - mask) + reactive = video * mask + inactive = retrieve_latents(self.vae.encode(inactive), generator, sample_mode="argmax") + reactive = retrieve_latents(self.vae.encode(reactive), generator, sample_mode="argmax") + inactive = ((inactive.float() - latents_mean) * latents_std).to(vae_dtype) + reactive = ((reactive.float() - latents_mean) * latents_std).to(vae_dtype) + latents = torch.cat([inactive, reactive], dim=1) + + latent_list = [] + for latent, reference_images_batch in zip(latents, reference_images): + for reference_image in reference_images_batch: + assert reference_image.ndim == 3 + reference_image = reference_image.to(dtype=vae_dtype) + reference_image = reference_image[None, :, None, :, :] # [1, C, 1, H, W] + reference_latent = retrieve_latents(self.vae.encode(reference_image), generator, sample_mode="argmax") + reference_latent = ((reference_latent.float() - latents_mean) * latents_std).to(vae_dtype) + reference_latent = reference_latent.squeeze(0) # [C, 1, H, W] + reference_latent = torch.cat([reference_latent, torch.zeros_like(reference_latent)], dim=0) + latent = torch.cat([reference_latent.squeeze(0), latent], dim=1) + latent_list.append(latent) + return torch.stack(latent_list) + + def prepare_masks( + self, + mask: torch.Tensor, + reference_images: Optional[List[torch.Tensor]] = None, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + ) -> torch.Tensor: + if isinstance(generator, list): + # TODO: support this + raise ValueError("Passing a list of generators is not yet supported. This may be supported in the future.") + + if reference_images is None: + # For each batch of video, we set no reference image (as one or more can be passed by user) + reference_images = [[None] for _ in range(mask.shape[0])] + else: + if mask.shape[0] != len(reference_images): + raise ValueError( + f"Batch size of `mask` {mask.shape[0]} and length of `reference_images` {len(reference_images)} does not match." + ) + + if mask.shape[0] != 1: + # TODO: support this + raise ValueError( + "Generating with more than one video is not yet supported. This may be supported in the future." + ) + + transformer_patch_size = self.transformer.config.patch_size[1] + + mask_list = [] + for mask_, reference_images_batch in zip(mask, reference_images): + num_channels, num_frames, height, width = mask_.shape + new_num_frames = (num_frames + self.vae_scale_factor_temporal - 1) // self.vae_scale_factor_temporal + new_height = height // (self.vae_scale_factor_spatial * transformer_patch_size) * transformer_patch_size + new_width = width // (self.vae_scale_factor_spatial * transformer_patch_size) * transformer_patch_size + mask_ = mask_[0, :, :, :] + mask_ = mask_.view( + num_frames, new_height, self.vae_scale_factor_spatial, new_width, self.vae_scale_factor_spatial + ) + mask_ = mask_.permute(2, 4, 0, 1, 3).flatten(0, 1) # [8x8, num_frames, new_height, new_width] + mask_ = torch.nn.functional.interpolate( + mask_.unsqueeze(0), size=(new_num_frames, new_height, new_width), mode="nearest-exact" + ).squeeze(0) + num_ref_images = len(reference_images_batch) + if num_ref_images > 0: + mask_padding = torch.zeros_like(mask_[:, :num_ref_images, :, :]) + mask_ = torch.cat([mask_padding, mask_], dim=1) + mask_list.append(mask_) + return torch.stack(mask_list) + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 81, + dtype: Optional[torch.dtype] = None, + device: Optional[torch.device] = None, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + shape = ( + batch_size, + num_channels_latents, + num_latent_frames, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + negative_prompt: Union[str, List[str]] = None, + video: Optional[List[PipelineImageInput]] = None, + mask: Optional[List[PipelineImageInput]] = None, + reference_images: Optional[List[PipelineImageInput]] = None, + conditioning_scale: Union[float, List[float], torch.Tensor] = 1.0, + height: int = 480, + width: int = 832, + num_frames: int = 81, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + num_videos_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "np", + return_dict: bool = True, + attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds` + instead. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + video (`List[PIL.Image.Image]`, *optional*): + The input video or videos to be used as a starting point for the generation. The video should be a list + of PIL images, a numpy array, or a torch tensor. Currently, the pipeline only supports generating one + video at a time. + mask (`List[PIL.Image.Image]`, *optional*): + The input mask defines which video regions to condition on and which to generate. Black areas in the + mask indicate conditioning regions, while white areas indicate regions for generation. The mask should + be a list of PIL images, a numpy array, or a torch tensor. Currently supports generating a single video + at a time. + reference_images (`List[PIL.Image.Image]`, *optional*): + A list of one or more reference images as extra conditioning for the generation. For example, if you + are trying to inpaint a video to change the character, you can pass reference images of the new + character here. Refer to the Diffusers [examples](https://github.com/huggingface/diffusers/pull/11582) + and original [user + guide](https://github.com/ali-vilab/VACE/blob/0897c6d055d7d9ea9e191dce763006664d9780f8/UserGuide.md) + for a full list of supported tasks and use cases. + conditioning_scale (`float`, `List[float]`, `torch.Tensor`, defaults to `1.0`): + The conditioning scale to be applied when adding the control conditioning latent stream to the + denoising latent stream in each control layer of the model. If a float is provided, it will be applied + uniformly to all layers. If a list or tensor is provided, it should have the same length as the number + of control layers in the model (`len(transformer.config.vace_layers)`). + height (`int`, defaults to `480`): + The height in pixels of the generated image. + width (`int`, defaults to `832`): + The width in pixels of the generated image. + num_frames (`int`, defaults to `81`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, defaults to `512`): + The maximum sequence length of the text encoder. If the prompt is longer than this, it will be + truncated. If the prompt is shorter, it will be padded to this length. + + Examples: + + Returns: + [`~WanPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where + the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # Simplification of implementation for now + if not isinstance(prompt, str): + raise ValueError("Passing a list of prompts is not yet supported. This may be supported in the future.") + if num_videos_per_prompt != 1: + raise ValueError( + "Generating multiple videos per prompt is not yet supported. This may be supported in the future." + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + height, + width, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + video, + mask, + reference_images, + ) + + if num_frames % self.vae_scale_factor_temporal != 1: + logger.warning( + f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number." + ) + num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + num_frames = max(num_frames, 1) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + vae_dtype = self.vae.dtype + transformer_dtype = self.transformer.dtype + + if isinstance(conditioning_scale, (int, float)): + conditioning_scale = [conditioning_scale] * len(self.transformer.config.vace_layers) + if isinstance(conditioning_scale, list): + if len(conditioning_scale) != len(self.transformer.config.vace_layers): + raise ValueError( + f"Length of `conditioning_scale` {len(conditioning_scale)} does not match number of layers {len(self.transformer.config.vace_layers)}." + ) + conditioning_scale = torch.tensor(conditioning_scale) + if isinstance(conditioning_scale, torch.Tensor): + if conditioning_scale.size(0) != len(self.transformer.config.vace_layers): + raise ValueError( + f"Length of `conditioning_scale` {conditioning_scale.size(0)} does not match number of layers {len(self.transformer.config.vace_layers)}." + ) + conditioning_scale = conditioning_scale.to(device=device, dtype=transformer_dtype) + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + video, mask, reference_images = self.preprocess_conditions( + video, + mask, + reference_images, + batch_size, + height, + width, + num_frames, + torch.float32, + device, + ) + num_reference_images = len(reference_images[0]) + + conditioning_latents = self.prepare_video_latents(video, mask, reference_images, generator, device) + mask = self.prepare_masks(mask, reference_images, generator) + conditioning_latents = torch.cat([conditioning_latents, mask], dim=1) + conditioning_latents = conditioning_latents.to(transformer_dtype) + + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames + num_reference_images * self.vae_scale_factor_temporal, + torch.float32, + device, + generator, + latents, + ) + + if conditioning_latents.shape[2] != latents.shape[2]: + logger.warning( + "The number of frames in the conditioning latents does not match the number of frames to be generated. Generation quality may be affected." + ) + + # 6. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + latent_model_input = latents.to(transformer_dtype) + timestep = t.expand(latents.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + control_hidden_states=conditioning_latents, + control_hidden_states_scale=conditioning_scale, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + noise_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + control_hidden_states=conditioning_latents, + control_hidden_states_scale=conditioning_scale, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if not output_type == "latent": + latents = latents[:, :, num_reference_images:] + latents = latents.to(vae_dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return WanPipelineOutput(frames=video) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_wan_video2video.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_wan_video2video.py new file mode 100644 index 0000000000000000000000000000000000000000..1a2d2e9c2232396ea55b3255fc5e72c0b5ea5c1c --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/wan/pipeline_wan_video2video.py @@ -0,0 +1,725 @@ +# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved. +# +# 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 html +import inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import regex as re +import torch +from PIL import Image +from transformers import AutoTokenizer, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import WanLoraLoaderMixin +from ...models import AutoencoderKLWan, WanTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import WanPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers.utils import export_to_video + >>> from diffusers import AutoencoderKLWan, WanVideoToVideoPipeline + >>> from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler + + >>> # Available models: Wan-AI/Wan2.1-T2V-14B-Diffusers, Wan-AI/Wan2.1-T2V-1.3B-Diffusers + >>> model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers" + >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32) + >>> pipe = WanVideoToVideoPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16) + >>> flow_shift = 3.0 # 5.0 for 720P, 3.0 for 480P + >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift) + >>> pipe.to("cuda") + + >>> prompt = "A robot standing on a mountain top. The sun is setting in the background" + >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards" + >>> video = load_video( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/hiker.mp4" + ... ) + >>> output = pipe( + ... video=video, + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... height=480, + ... width=720, + ... guidance_scale=5.0, + ... strength=0.7, + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=16) + ``` +""" + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class WanVideoToVideoPipeline(DiffusionPipeline, WanLoraLoaderMixin): + r""" + Pipeline for video-to-video generation using Wan. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + tokenizer ([`T5Tokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + transformer ([`WanTransformer3DModel`]): + Conditional Transformer to denoise the input latents. + scheduler ([`UniPCMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: UMT5EncoderModel, + transformer: WanTransformer3DModel, + vae: AutoencoderKLWan, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + + self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 226, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.encode_prompt + def encode_prompt( + self, + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + max_sequence_length: int = 226, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + height, + width, + video=None, + latents=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + if video is not None and latents is not None: + raise ValueError("Only one of `video` or `latents` should be provided") + + def prepare_latents( + self, + video: Optional[torch.Tensor] = None, + batch_size: int = 1, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + dtype: Optional[torch.dtype] = None, + device: Optional[torch.device] = None, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.Tensor] = None, + timestep: Optional[torch.Tensor] = None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + num_latent_frames = ( + (video.size(2) - 1) // self.vae_scale_factor_temporal + 1 if latents is None else latents.size(1) + ) + shape = ( + batch_size, + num_channels_latents, + num_latent_frames, + height // self.vae_scale_factor_spatial, + width // self.vae_scale_factor_spatial, + ) + + if latents is None: + init_latents = [retrieve_latents(self.vae.encode(vid.unsqueeze(0)), sample_mode="argmax") for vid in video] + + init_latents = torch.cat(init_latents, dim=0).to(dtype) + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, self.vae.config.z_dim, 1, 1, 1).to(device, dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + device, dtype + ) + + init_latents = (init_latents - latents_mean) * latents_std + + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + if hasattr(self.scheduler, "add_noise"): + latents = self.scheduler.add_noise(init_latents, noise, timestep) + else: + latents = self.scheduler.scale_noise(init_latents, timestep, noise) + else: + latents = latents.to(device) + + return latents + + # Copied from diffusers.pipelines.animatediff.pipeline_animatediff_video2video.AnimateDiffVideoToVideoPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, timesteps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = timesteps[t_start * self.scheduler.order :] + + return timesteps, num_inference_steps - t_start + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + video: List[Image.Image] = None, + prompt: Union[str, List[str]] = None, + negative_prompt: Union[str, List[str]] = None, + height: int = 480, + width: int = 832, + num_inference_steps: int = 50, + timesteps: Optional[List[int]] = None, + guidance_scale: float = 5.0, + strength: float = 0.8, + num_videos_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + output_type: Optional[str] = "np", + return_dict: bool = True, + attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end: Optional[ + Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] + ] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds` + instead. + height (`int`, defaults to `480`): + The height in pixels of the generated image. + width (`int`, defaults to `832`): + The width in pixels of the generated image. + num_frames (`int`, defaults to `81`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + strength (`float`, defaults to `0.8`): + Higher strength leads to more differences between original image and generated video. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, defaults to `512`): + The maximum sequence length of the text encoder. If the prompt is longer than this, it will be + truncated. If the prompt is shorter, it will be padded to this length. + + Examples: + + Returns: + [`~WanPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where + the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + height = height or self.transformer.config.sample_height * self.vae_scale_factor_spatial + width = width or self.transformer.config.sample_width * self.vae_scale_factor_spatial + num_videos_per_prompt = 1 + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + height, + width, + video, + latents, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + transformer_dtype = self.transformer.dtype + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, timesteps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_videos_per_prompt) + self._num_timesteps = len(timesteps) + + if latents is None: + video = self.video_processor.preprocess_video(video, height=height, width=width).to( + device, dtype=torch.float32 + ) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + video, + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + torch.float32, + device, + generator, + latents, + latent_timestep, + ) + + # 6. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + latent_model_input = latents.to(transformer_dtype) + timestep = t.expand(latents.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + noise_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if not output_type == "latent": + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return WanPipelineOutput(frames=video) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__init__.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..ddb852d1931558fe0948e81e16cf9a92fc2a114b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__init__.py @@ -0,0 +1,56 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["modeling_paella_vq_model"] = ["PaellaVQModel"] + _import_structure["modeling_wuerstchen_diffnext"] = ["WuerstchenDiffNeXt"] + _import_structure["modeling_wuerstchen_prior"] = ["WuerstchenPrior"] + _import_structure["pipeline_wuerstchen"] = ["WuerstchenDecoderPipeline"] + _import_structure["pipeline_wuerstchen_combined"] = ["WuerstchenCombinedPipeline"] + _import_structure["pipeline_wuerstchen_prior"] = ["DEFAULT_STAGE_C_TIMESTEPS", "WuerstchenPriorPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .modeling_paella_vq_model import PaellaVQModel + from .modeling_wuerstchen_diffnext import WuerstchenDiffNeXt + from .modeling_wuerstchen_prior import WuerstchenPrior + from .pipeline_wuerstchen import WuerstchenDecoderPipeline + from .pipeline_wuerstchen_combined import WuerstchenCombinedPipeline + from .pipeline_wuerstchen_prior import DEFAULT_STAGE_C_TIMESTEPS, WuerstchenPriorPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..064ef6e0f6ce571f0af2f0ab0be5ade106d84982 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/modeling_paella_vq_model.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/modeling_paella_vq_model.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..27917068256ae1ce7b039b106d91a1fd5aefad5e Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/modeling_paella_vq_model.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_common.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_common.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..48b4a8330b017915e6ec0dc0472c9b6f6a5e3b9d Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_common.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_diffnext.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_diffnext.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e8c8fe30898853fc9d1a2f39b7c65c064393c0cb Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_diffnext.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_prior.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_prior.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..13b530928eff125fc2a8f2c76364cd11e838fcc0 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_prior.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fd7f9997f6b710388aa502e693b0080f66746ab4 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen_combined.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen_combined.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..077a75e79f56b548fa71a073e03fa88d7b32b2e9 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen_combined.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen_prior.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen_prior.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..322f19e9dba9be59a318376fd38240cff3930246 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen_prior.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/modeling_paella_vq_model.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/modeling_paella_vq_model.py new file mode 100644 index 0000000000000000000000000000000000000000..5ab206b15176e5ab9cb636762d38d3c38618124a --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/modeling_paella_vq_model.py @@ -0,0 +1,172 @@ +# Copyright (c) 2022 Dominic Rampas MIT License +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 +import torch.nn as nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models.autoencoders.vae import DecoderOutput, VectorQuantizer +from ...models.modeling_utils import ModelMixin +from ...models.vq_model import VQEncoderOutput +from ...utils.accelerate_utils import apply_forward_hook + + +class MixingResidualBlock(nn.Module): + """ + Residual block with mixing used by Paella's VQ-VAE. + """ + + def __init__(self, inp_channels, embed_dim): + super().__init__() + # depthwise + self.norm1 = nn.LayerNorm(inp_channels, elementwise_affine=False, eps=1e-6) + self.depthwise = nn.Sequential( + nn.ReplicationPad2d(1), nn.Conv2d(inp_channels, inp_channels, kernel_size=3, groups=inp_channels) + ) + + # channelwise + self.norm2 = nn.LayerNorm(inp_channels, elementwise_affine=False, eps=1e-6) + self.channelwise = nn.Sequential( + nn.Linear(inp_channels, embed_dim), nn.GELU(), nn.Linear(embed_dim, inp_channels) + ) + + self.gammas = nn.Parameter(torch.zeros(6), requires_grad=True) + + def forward(self, x): + mods = self.gammas + x_temp = self.norm1(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) * (1 + mods[0]) + mods[1] + x = x + self.depthwise(x_temp) * mods[2] + x_temp = self.norm2(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) * (1 + mods[3]) + mods[4] + x = x + self.channelwise(x_temp.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) * mods[5] + return x + + +class PaellaVQModel(ModelMixin, ConfigMixin): + r"""VQ-VAE model from Paella model. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library + implements for all the model (such as downloading or saving, etc.) + + Parameters: + in_channels (int, *optional*, defaults to 3): Number of channels in the input image. + out_channels (int, *optional*, defaults to 3): Number of channels in the output. + up_down_scale_factor (int, *optional*, defaults to 2): Up and Downscale factor of the input image. + levels (int, *optional*, defaults to 2): Number of levels in the model. + bottleneck_blocks (int, *optional*, defaults to 12): Number of bottleneck blocks in the model. + embed_dim (int, *optional*, defaults to 384): Number of hidden channels in the model. + latent_channels (int, *optional*, defaults to 4): Number of latent channels in the VQ-VAE model. + num_vq_embeddings (int, *optional*, defaults to 8192): Number of codebook vectors in the VQ-VAE. + scale_factor (float, *optional*, defaults to 0.3764): Scaling factor of the latent space. + """ + + @register_to_config + def __init__( + self, + in_channels: int = 3, + out_channels: int = 3, + up_down_scale_factor: int = 2, + levels: int = 2, + bottleneck_blocks: int = 12, + embed_dim: int = 384, + latent_channels: int = 4, + num_vq_embeddings: int = 8192, + scale_factor: float = 0.3764, + ): + super().__init__() + + c_levels = [embed_dim // (2**i) for i in reversed(range(levels))] + # Encoder blocks + self.in_block = nn.Sequential( + nn.PixelUnshuffle(up_down_scale_factor), + nn.Conv2d(in_channels * up_down_scale_factor**2, c_levels[0], kernel_size=1), + ) + down_blocks = [] + for i in range(levels): + if i > 0: + down_blocks.append(nn.Conv2d(c_levels[i - 1], c_levels[i], kernel_size=4, stride=2, padding=1)) + block = MixingResidualBlock(c_levels[i], c_levels[i] * 4) + down_blocks.append(block) + down_blocks.append( + nn.Sequential( + nn.Conv2d(c_levels[-1], latent_channels, kernel_size=1, bias=False), + nn.BatchNorm2d(latent_channels), # then normalize them to have mean 0 and std 1 + ) + ) + self.down_blocks = nn.Sequential(*down_blocks) + + # Vector Quantizer + self.vquantizer = VectorQuantizer(num_vq_embeddings, vq_embed_dim=latent_channels, legacy=False, beta=0.25) + + # Decoder blocks + up_blocks = [nn.Sequential(nn.Conv2d(latent_channels, c_levels[-1], kernel_size=1))] + for i in range(levels): + for j in range(bottleneck_blocks if i == 0 else 1): + block = MixingResidualBlock(c_levels[levels - 1 - i], c_levels[levels - 1 - i] * 4) + up_blocks.append(block) + if i < levels - 1: + up_blocks.append( + nn.ConvTranspose2d( + c_levels[levels - 1 - i], c_levels[levels - 2 - i], kernel_size=4, stride=2, padding=1 + ) + ) + self.up_blocks = nn.Sequential(*up_blocks) + self.out_block = nn.Sequential( + nn.Conv2d(c_levels[0], out_channels * up_down_scale_factor**2, kernel_size=1), + nn.PixelShuffle(up_down_scale_factor), + ) + + @apply_forward_hook + def encode(self, x: torch.Tensor, return_dict: bool = True) -> VQEncoderOutput: + h = self.in_block(x) + h = self.down_blocks(h) + + if not return_dict: + return (h,) + + return VQEncoderOutput(latents=h) + + @apply_forward_hook + def decode( + self, h: torch.Tensor, force_not_quantize: bool = True, return_dict: bool = True + ) -> Union[DecoderOutput, torch.Tensor]: + if not force_not_quantize: + quant, _, _ = self.vquantizer(h) + else: + quant = h + + x = self.up_blocks(quant) + dec = self.out_block(x) + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + def forward(self, sample: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]: + r""" + Args: + sample (`torch.Tensor`): Input sample. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`DecoderOutput`] instead of a plain tuple. + """ + x = sample + h = self.encode(x).latents + dec = self.decode(h).sample + + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/modeling_wuerstchen_common.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/modeling_wuerstchen_common.py new file mode 100644 index 0000000000000000000000000000000000000000..73e71b3076fbca259ae76138bc4ab3d3797e2755 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/modeling_wuerstchen_common.py @@ -0,0 +1,76 @@ +import torch +import torch.nn as nn + +from ...models.attention_processor import Attention + + +class WuerstchenLayerNorm(nn.LayerNorm): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, x): + x = x.permute(0, 2, 3, 1) + x = super().forward(x) + return x.permute(0, 3, 1, 2) + + +class TimestepBlock(nn.Module): + def __init__(self, c, c_timestep): + super().__init__() + + self.mapper = nn.Linear(c_timestep, c * 2) + + def forward(self, x, t): + a, b = self.mapper(t)[:, :, None, None].chunk(2, dim=1) + return x * (1 + a) + b + + +class ResBlock(nn.Module): + def __init__(self, c, c_skip=0, kernel_size=3, dropout=0.0): + super().__init__() + + self.depthwise = nn.Conv2d(c + c_skip, c, kernel_size=kernel_size, padding=kernel_size // 2, groups=c) + self.norm = WuerstchenLayerNorm(c, elementwise_affine=False, eps=1e-6) + self.channelwise = nn.Sequential( + nn.Linear(c, c * 4), nn.GELU(), GlobalResponseNorm(c * 4), nn.Dropout(dropout), nn.Linear(c * 4, c) + ) + + def forward(self, x, x_skip=None): + x_res = x + if x_skip is not None: + x = torch.cat([x, x_skip], dim=1) + x = self.norm(self.depthwise(x)).permute(0, 2, 3, 1) + x = self.channelwise(x).permute(0, 3, 1, 2) + return x + x_res + + +# from https://github.com/facebookresearch/ConvNeXt-V2/blob/3608f67cc1dae164790c5d0aead7bf2d73d9719b/models/utils.py#L105 +class GlobalResponseNorm(nn.Module): + def __init__(self, dim): + super().__init__() + self.gamma = nn.Parameter(torch.zeros(1, 1, 1, dim)) + self.beta = nn.Parameter(torch.zeros(1, 1, 1, dim)) + + def forward(self, x): + agg_norm = torch.norm(x, p=2, dim=(1, 2), keepdim=True) + stand_div_norm = agg_norm / (agg_norm.mean(dim=-1, keepdim=True) + 1e-6) + return self.gamma * (x * stand_div_norm) + self.beta + x + + +class AttnBlock(nn.Module): + def __init__(self, c, c_cond, nhead, self_attn=True, dropout=0.0): + super().__init__() + + self.self_attn = self_attn + self.norm = WuerstchenLayerNorm(c, elementwise_affine=False, eps=1e-6) + self.attention = Attention(query_dim=c, heads=nhead, dim_head=c // nhead, dropout=dropout, bias=True) + self.kv_mapper = nn.Sequential(nn.SiLU(), nn.Linear(c_cond, c)) + + def forward(self, x, kv): + kv = self.kv_mapper(kv) + norm_x = self.norm(x) + if self.self_attn: + batch_size, channel, _, _ = x.shape + kv = torch.cat([norm_x.view(batch_size, channel, -1).transpose(1, 2), kv], dim=1) + x = x + self.attention(norm_x, encoder_hidden_states=kv) + return x diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/modeling_wuerstchen_diffnext.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/modeling_wuerstchen_diffnext.py new file mode 100644 index 0000000000000000000000000000000000000000..77ae597655d164e25add03a641a6b8eca5396200 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/modeling_wuerstchen_diffnext.py @@ -0,0 +1,254 @@ +# Copyright (c) 2023 Dominic Rampas MIT License +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 numpy as np +import torch +import torch.nn as nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models.modeling_utils import ModelMixin +from .modeling_wuerstchen_common import AttnBlock, GlobalResponseNorm, TimestepBlock, WuerstchenLayerNorm + + +class WuerstchenDiffNeXt(ModelMixin, ConfigMixin): + @register_to_config + def __init__( + self, + c_in=4, + c_out=4, + c_r=64, + patch_size=2, + c_cond=1024, + c_hidden=[320, 640, 1280, 1280], + nhead=[-1, 10, 20, 20], + blocks=[4, 4, 14, 4], + level_config=["CT", "CTA", "CTA", "CTA"], + inject_effnet=[False, True, True, True], + effnet_embd=16, + clip_embd=1024, + kernel_size=3, + dropout=0.1, + ): + super().__init__() + self.c_r = c_r + self.c_cond = c_cond + if not isinstance(dropout, list): + dropout = [dropout] * len(c_hidden) + + # CONDITIONING + self.clip_mapper = nn.Linear(clip_embd, c_cond) + self.effnet_mappers = nn.ModuleList( + [ + nn.Conv2d(effnet_embd, c_cond, kernel_size=1) if inject else None + for inject in inject_effnet + list(reversed(inject_effnet)) + ] + ) + self.seq_norm = nn.LayerNorm(c_cond, elementwise_affine=False, eps=1e-6) + + self.embedding = nn.Sequential( + nn.PixelUnshuffle(patch_size), + nn.Conv2d(c_in * (patch_size**2), c_hidden[0], kernel_size=1), + WuerstchenLayerNorm(c_hidden[0], elementwise_affine=False, eps=1e-6), + ) + + def get_block(block_type, c_hidden, nhead, c_skip=0, dropout=0): + if block_type == "C": + return ResBlockStageB(c_hidden, c_skip, kernel_size=kernel_size, dropout=dropout) + elif block_type == "A": + return AttnBlock(c_hidden, c_cond, nhead, self_attn=True, dropout=dropout) + elif block_type == "T": + return TimestepBlock(c_hidden, c_r) + else: + raise ValueError(f"Block type {block_type} not supported") + + # BLOCKS + # -- down blocks + self.down_blocks = nn.ModuleList() + for i in range(len(c_hidden)): + down_block = nn.ModuleList() + if i > 0: + down_block.append( + nn.Sequential( + WuerstchenLayerNorm(c_hidden[i - 1], elementwise_affine=False, eps=1e-6), + nn.Conv2d(c_hidden[i - 1], c_hidden[i], kernel_size=2, stride=2), + ) + ) + for _ in range(blocks[i]): + for block_type in level_config[i]: + c_skip = c_cond if inject_effnet[i] else 0 + down_block.append(get_block(block_type, c_hidden[i], nhead[i], c_skip=c_skip, dropout=dropout[i])) + self.down_blocks.append(down_block) + + # -- up blocks + self.up_blocks = nn.ModuleList() + for i in reversed(range(len(c_hidden))): + up_block = nn.ModuleList() + for j in range(blocks[i]): + for k, block_type in enumerate(level_config[i]): + c_skip = c_hidden[i] if i < len(c_hidden) - 1 and j == k == 0 else 0 + c_skip += c_cond if inject_effnet[i] else 0 + up_block.append(get_block(block_type, c_hidden[i], nhead[i], c_skip=c_skip, dropout=dropout[i])) + if i > 0: + up_block.append( + nn.Sequential( + WuerstchenLayerNorm(c_hidden[i], elementwise_affine=False, eps=1e-6), + nn.ConvTranspose2d(c_hidden[i], c_hidden[i - 1], kernel_size=2, stride=2), + ) + ) + self.up_blocks.append(up_block) + + # OUTPUT + self.clf = nn.Sequential( + WuerstchenLayerNorm(c_hidden[0], elementwise_affine=False, eps=1e-6), + nn.Conv2d(c_hidden[0], 2 * c_out * (patch_size**2), kernel_size=1), + nn.PixelShuffle(patch_size), + ) + + # --- WEIGHT INIT --- + self.apply(self._init_weights) + + def _init_weights(self, m): + # General init + if isinstance(m, (nn.Conv2d, nn.Linear)): + nn.init.xavier_uniform_(m.weight) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + + for mapper in self.effnet_mappers: + if mapper is not None: + nn.init.normal_(mapper.weight, std=0.02) # conditionings + nn.init.normal_(self.clip_mapper.weight, std=0.02) # conditionings + nn.init.xavier_uniform_(self.embedding[1].weight, 0.02) # inputs + nn.init.constant_(self.clf[1].weight, 0) # outputs + + # blocks + for level_block in self.down_blocks + self.up_blocks: + for block in level_block: + if isinstance(block, ResBlockStageB): + block.channelwise[-1].weight.data *= np.sqrt(1 / sum(self.config.blocks)) + elif isinstance(block, TimestepBlock): + nn.init.constant_(block.mapper.weight, 0) + + def gen_r_embedding(self, r, max_positions=10000): + r = r * max_positions + half_dim = self.c_r // 2 + emb = math.log(max_positions) / (half_dim - 1) + emb = torch.arange(half_dim, device=r.device).float().mul(-emb).exp() + emb = r[:, None] * emb[None, :] + emb = torch.cat([emb.sin(), emb.cos()], dim=1) + if self.c_r % 2 == 1: # zero pad + emb = nn.functional.pad(emb, (0, 1), mode="constant") + return emb.to(dtype=r.dtype) + + def gen_c_embeddings(self, clip): + clip = self.clip_mapper(clip) + clip = self.seq_norm(clip) + return clip + + def _down_encode(self, x, r_embed, effnet, clip=None): + level_outputs = [] + for i, down_block in enumerate(self.down_blocks): + effnet_c = None + for block in down_block: + if isinstance(block, ResBlockStageB): + if effnet_c is None and self.effnet_mappers[i] is not None: + dtype = effnet.dtype + effnet_c = self.effnet_mappers[i]( + nn.functional.interpolate( + effnet.float(), size=x.shape[-2:], mode="bicubic", antialias=True, align_corners=True + ).to(dtype) + ) + skip = effnet_c if self.effnet_mappers[i] is not None else None + x = block(x, skip) + elif isinstance(block, AttnBlock): + x = block(x, clip) + elif isinstance(block, TimestepBlock): + x = block(x, r_embed) + else: + x = block(x) + level_outputs.insert(0, x) + return level_outputs + + def _up_decode(self, level_outputs, r_embed, effnet, clip=None): + x = level_outputs[0] + for i, up_block in enumerate(self.up_blocks): + effnet_c = None + for j, block in enumerate(up_block): + if isinstance(block, ResBlockStageB): + if effnet_c is None and self.effnet_mappers[len(self.down_blocks) + i] is not None: + dtype = effnet.dtype + effnet_c = self.effnet_mappers[len(self.down_blocks) + i]( + nn.functional.interpolate( + effnet.float(), size=x.shape[-2:], mode="bicubic", antialias=True, align_corners=True + ).to(dtype) + ) + skip = level_outputs[i] if j == 0 and i > 0 else None + if effnet_c is not None: + if skip is not None: + skip = torch.cat([skip, effnet_c], dim=1) + else: + skip = effnet_c + x = block(x, skip) + elif isinstance(block, AttnBlock): + x = block(x, clip) + elif isinstance(block, TimestepBlock): + x = block(x, r_embed) + else: + x = block(x) + return x + + def forward(self, x, r, effnet, clip=None, x_cat=None, eps=1e-3, return_noise=True): + if x_cat is not None: + x = torch.cat([x, x_cat], dim=1) + # Process the conditioning embeddings + r_embed = self.gen_r_embedding(r) + if clip is not None: + clip = self.gen_c_embeddings(clip) + + # Model Blocks + x_in = x + x = self.embedding(x) + level_outputs = self._down_encode(x, r_embed, effnet, clip) + x = self._up_decode(level_outputs, r_embed, effnet, clip) + a, b = self.clf(x).chunk(2, dim=1) + b = b.sigmoid() * (1 - eps * 2) + eps + if return_noise: + return (x_in - a) / b + else: + return a, b + + +class ResBlockStageB(nn.Module): + def __init__(self, c, c_skip=0, kernel_size=3, dropout=0.0): + super().__init__() + self.depthwise = nn.Conv2d(c, c, kernel_size=kernel_size, padding=kernel_size // 2, groups=c) + self.norm = WuerstchenLayerNorm(c, elementwise_affine=False, eps=1e-6) + self.channelwise = nn.Sequential( + nn.Linear(c + c_skip, c * 4), + nn.GELU(), + GlobalResponseNorm(c * 4), + nn.Dropout(dropout), + nn.Linear(c * 4, c), + ) + + def forward(self, x, x_skip=None): + x_res = x + x = self.norm(self.depthwise(x)) + if x_skip is not None: + x = torch.cat([x, x_skip], dim=1) + x = self.channelwise(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) + return x + x_res diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/modeling_wuerstchen_prior.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/modeling_wuerstchen_prior.py new file mode 100644 index 0000000000000000000000000000000000000000..ae821a55788aeddf49d8a7299c6c82da467f2832 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/modeling_wuerstchen_prior.py @@ -0,0 +1,174 @@ +# Copyright (c) 2023 Dominic Rampas MIT License +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 typing import Dict, Union + +import torch +import torch.nn as nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin +from ...models.attention_processor import ( + ADDED_KV_ATTENTION_PROCESSORS, + CROSS_ATTENTION_PROCESSORS, + AttentionProcessor, + AttnAddedKVProcessor, + AttnProcessor, +) +from ...models.modeling_utils import ModelMixin +from .modeling_wuerstchen_common import AttnBlock, ResBlock, TimestepBlock, WuerstchenLayerNorm + + +class WuerstchenPrior(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin): + unet_name = "prior" + _supports_gradient_checkpointing = True + + @register_to_config + def __init__(self, c_in=16, c=1280, c_cond=1024, c_r=64, depth=16, nhead=16, dropout=0.1): + super().__init__() + + self.c_r = c_r + self.projection = nn.Conv2d(c_in, c, kernel_size=1) + self.cond_mapper = nn.Sequential( + nn.Linear(c_cond, c), + nn.LeakyReLU(0.2), + nn.Linear(c, c), + ) + + self.blocks = nn.ModuleList() + for _ in range(depth): + self.blocks.append(ResBlock(c, dropout=dropout)) + self.blocks.append(TimestepBlock(c, c_r)) + self.blocks.append(AttnBlock(c, c, nhead, self_attn=True, dropout=dropout)) + self.out = nn.Sequential( + WuerstchenLayerNorm(c, elementwise_affine=False, eps=1e-6), + nn.Conv2d(c, c_in * 2, kernel_size=1), + ) + + self.gradient_checkpointing = False + self.set_default_attn_processor() + + @property + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor() + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor + def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor + def set_default_attn_processor(self): + """ + Disables custom attention processors and sets the default attention implementation. + """ + if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnAddedKVProcessor() + elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnProcessor() + else: + raise ValueError( + f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}" + ) + + self.set_attn_processor(processor) + + def gen_r_embedding(self, r, max_positions=10000): + r = r * max_positions + half_dim = self.c_r // 2 + emb = math.log(max_positions) / (half_dim - 1) + emb = torch.arange(half_dim, device=r.device).float().mul(-emb).exp() + emb = r[:, None] * emb[None, :] + emb = torch.cat([emb.sin(), emb.cos()], dim=1) + if self.c_r % 2 == 1: # zero pad + emb = nn.functional.pad(emb, (0, 1), mode="constant") + return emb.to(dtype=r.dtype) + + def forward(self, x, r, c): + x_in = x + x = self.projection(x) + c_embed = self.cond_mapper(c) + r_embed = self.gen_r_embedding(r) + + if torch.is_grad_enabled() and self.gradient_checkpointing: + for block in self.blocks: + if isinstance(block, AttnBlock): + x = self._gradient_checkpointing_func(block, x, c_embed) + elif isinstance(block, TimestepBlock): + x = self._gradient_checkpointing_func(block, x, r_embed) + else: + x = self._gradient_checkpointing_func(block, x) + else: + for block in self.blocks: + if isinstance(block, AttnBlock): + x = block(x, c_embed) + elif isinstance(block, TimestepBlock): + x = block(x, r_embed) + else: + x = block(x) + a, b = self.out(x).chunk(2, dim=1) + return (x_in - a) / ((1 - b).abs() + 1e-5) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/pipeline_wuerstchen.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/pipeline_wuerstchen.py new file mode 100644 index 0000000000000000000000000000000000000000..b9b02a6dd38a0e1fc28c1136f0419acb4aad6dd5 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/pipeline_wuerstchen.py @@ -0,0 +1,449 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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, Dict, List, Optional, Union + +import numpy as np +import torch +from transformers import CLIPTextModel, CLIPTokenizer + +from ...schedulers import DDPMWuerstchenScheduler +from ...utils import deprecate, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput +from .modeling_paella_vq_model import PaellaVQModel +from .modeling_wuerstchen_diffnext import WuerstchenDiffNeXt + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import WuerstchenPriorPipeline, WuerstchenDecoderPipeline + + >>> prior_pipe = WuerstchenPriorPipeline.from_pretrained( + ... "warp-ai/wuerstchen-prior", torch_dtype=torch.float16 + ... ).to("cuda") + >>> gen_pipe = WuerstchenDecoderPipeline.from_pretrain("warp-ai/wuerstchen", torch_dtype=torch.float16).to( + ... "cuda" + ... ) + + >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet" + >>> prior_output = pipe(prompt) + >>> images = gen_pipe(prior_output.image_embeddings, prompt=prompt) + ``` +""" + + +class WuerstchenDecoderPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + """ + Pipeline for generating images from the Wuerstchen model. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + tokenizer (`CLIPTokenizer`): + The CLIP tokenizer. + text_encoder (`CLIPTextModel`): + The CLIP text encoder. + decoder ([`WuerstchenDiffNeXt`]): + The WuerstchenDiffNeXt unet decoder. + vqgan ([`PaellaVQModel`]): + The VQGAN model. + scheduler ([`DDPMWuerstchenScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + latent_dim_scale (float, `optional`, defaults to 10.67): + Multiplier to determine the VQ latent space size from the image embeddings. If the image embeddings are + height=24 and width=24, the VQ latent shape needs to be height=int(24*10.67)=256 and + width=int(24*10.67)=256 in order to match the training conditions. + """ + + model_cpu_offload_seq = "text_encoder->decoder->vqgan" + _callback_tensor_inputs = [ + "latents", + "text_encoder_hidden_states", + "negative_prompt_embeds", + "image_embeddings", + ] + + def __init__( + self, + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModel, + decoder: WuerstchenDiffNeXt, + scheduler: DDPMWuerstchenScheduler, + vqgan: PaellaVQModel, + latent_dim_scale: float = 10.67, + ) -> None: + super().__init__() + self.register_modules( + tokenizer=tokenizer, + text_encoder=text_encoder, + decoder=decoder, + scheduler=scheduler, + vqgan=vqgan, + ) + self.register_to_config(latent_dim_scale=latent_dim_scale) + + # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + ): + batch_size = len(prompt) if isinstance(prompt, list) else 1 + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + attention_mask = text_inputs.attention_mask + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] + attention_mask = attention_mask[:, : self.tokenizer.model_max_length] + + text_encoder_output = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask.to(device)) + text_encoder_hidden_states = text_encoder_output.last_hidden_state + text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + + uncond_text_encoder_hidden_states = None + if do_classifier_free_guidance: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + negative_prompt_embeds_text_encoder_output = self.text_encoder( + uncond_input.input_ids.to(device), attention_mask=uncond_input.attention_mask.to(device) + ) + + uncond_text_encoder_hidden_states = negative_prompt_embeds_text_encoder_output.last_hidden_state + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = uncond_text_encoder_hidden_states.shape[1] + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1) + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + return text_encoder_hidden_states, uncond_text_encoder_hidden_states + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image_embeddings: Union[torch.Tensor, List[torch.Tensor]], + prompt: Union[str, List[str]] = None, + num_inference_steps: int = 12, + timesteps: Optional[List[float]] = None, + guidance_scale: float = 0.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: int = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + image_embedding (`torch.Tensor` or `List[torch.Tensor]`): + Image Embeddings either extracted from an image or generated by a Prior Model. + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + num_inference_steps (`int`, *optional*, defaults to 12): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps` + timesteps are used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to 0.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `decoder_guidance_scale` is defined as `w` of + equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by + setting `decoder_guidance_scale > 1`. Higher guidance scale encourages to generate images that are + closely linked to the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `decoder_guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` [`~pipelines.ImagePipelineOutput`] if `return_dict` is True, + otherwise a `tuple`. When returning a tuple, the first element is a list with the generated image + embeddings. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + # 0. Define commonly used variables + device = self._execution_device + dtype = self.decoder.dtype + self._guidance_scale = guidance_scale + + # 1. Check inputs. Raise error if not correct + if not isinstance(prompt, list): + if isinstance(prompt, str): + prompt = [prompt] + else: + raise TypeError(f"'prompt' must be of type 'list' or 'str', but got {type(prompt)}.") + + if self.do_classifier_free_guidance: + if negative_prompt is not None and not isinstance(negative_prompt, list): + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] + else: + raise TypeError( + f"'negative_prompt' must be of type 'list' or 'str', but got {type(negative_prompt)}." + ) + + if isinstance(image_embeddings, list): + image_embeddings = torch.cat(image_embeddings, dim=0) + if isinstance(image_embeddings, np.ndarray): + image_embeddings = torch.Tensor(image_embeddings, device=device).to(dtype=dtype) + if not isinstance(image_embeddings, torch.Tensor): + raise TypeError( + f"'image_embeddings' must be of type 'torch.Tensor' or 'np.array', but got {type(image_embeddings)}." + ) + + if not isinstance(num_inference_steps, int): + raise TypeError( + f"'num_inference_steps' must be of type 'int', but got {type(num_inference_steps)}\ + In Case you want to provide explicit timesteps, please use the 'timesteps' argument." + ) + + # 2. Encode caption + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + image_embeddings.size(0) * num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + ) + text_encoder_hidden_states = ( + torch.cat([prompt_embeds, negative_prompt_embeds]) if negative_prompt_embeds is not None else prompt_embeds + ) + effnet = ( + torch.cat([image_embeddings, torch.zeros_like(image_embeddings)]) + if self.do_classifier_free_guidance + else image_embeddings + ) + + # 3. Determine latent shape of latents + latent_height = int(image_embeddings.size(2) * self.config.latent_dim_scale) + latent_width = int(image_embeddings.size(3) * self.config.latent_dim_scale) + latent_features_shape = (image_embeddings.size(0) * num_images_per_prompt, 4, latent_height, latent_width) + + # 4. Prepare and set timesteps + if timesteps is not None: + self.scheduler.set_timesteps(timesteps=timesteps, device=device) + timesteps = self.scheduler.timesteps + num_inference_steps = len(timesteps) + else: + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latents + latents = self.prepare_latents(latent_features_shape, dtype, device, generator, latents, self.scheduler) + + # 6. Run denoising loop + self._num_timesteps = len(timesteps[:-1]) + for i, t in enumerate(self.progress_bar(timesteps[:-1])): + ratio = t.expand(latents.size(0)).to(dtype) + # 7. Denoise latents + predicted_latents = self.decoder( + torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents, + r=torch.cat([ratio] * 2) if self.do_classifier_free_guidance else ratio, + effnet=effnet, + clip=text_encoder_hidden_states, + ) + + # 8. Check for classifier free guidance and apply it + if self.do_classifier_free_guidance: + predicted_latents_text, predicted_latents_uncond = predicted_latents.chunk(2) + predicted_latents = torch.lerp(predicted_latents_uncond, predicted_latents_text, self.guidance_scale) + + # 9. Renoise latents to next timestep + latents = self.scheduler.step( + model_output=predicted_latents, + timestep=ratio, + sample=latents, + generator=generator, + ).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + image_embeddings = callback_outputs.pop("image_embeddings", image_embeddings) + text_encoder_hidden_states = callback_outputs.pop( + "text_encoder_hidden_states", text_encoder_hidden_states + ) + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type not in ["pt", "np", "pil", "latent"]: + raise ValueError( + f"Only the output types `pt`, `np`, `pil` and `latent` are supported not output_type={output_type}" + ) + + if not output_type == "latent": + # 10. Scale and decode the image latents with vq-vae + latents = self.vqgan.config.scale_factor * latents + images = self.vqgan.decode(latents).sample.clamp(0, 1) + if output_type == "np": + images = images.permute(0, 2, 3, 1).cpu().float().numpy() + elif output_type == "pil": + images = images.permute(0, 2, 3, 1).cpu().float().numpy() + images = self.numpy_to_pil(images) + else: + images = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return images + return ImagePipelineOutput(images) diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/pipeline_wuerstchen_combined.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/pipeline_wuerstchen_combined.py new file mode 100644 index 0000000000000000000000000000000000000000..00a88ce34ed2a5a64591624c54b0df425e3179cd --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/pipeline_wuerstchen_combined.py @@ -0,0 +1,307 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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, Dict, List, Optional, Union + +import torch +from transformers import CLIPTextModel, CLIPTokenizer + +from ...schedulers import DDPMWuerstchenScheduler +from ...utils import deprecate, replace_example_docstring +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline +from .modeling_paella_vq_model import PaellaVQModel +from .modeling_wuerstchen_diffnext import WuerstchenDiffNeXt +from .modeling_wuerstchen_prior import WuerstchenPrior +from .pipeline_wuerstchen import WuerstchenDecoderPipeline +from .pipeline_wuerstchen_prior import WuerstchenPriorPipeline + + +TEXT2IMAGE_EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusions import WuerstchenCombinedPipeline + + >>> pipe = WuerstchenCombinedPipeline.from_pretrained("warp-ai/Wuerstchen", torch_dtype=torch.float16).to( + ... "cuda" + ... ) + >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet" + >>> images = pipe(prompt=prompt) + ``` +""" + + +class WuerstchenCombinedPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + """ + Combined Pipeline for text-to-image generation using Wuerstchen + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + tokenizer (`CLIPTokenizer`): + The decoder tokenizer to be used for text inputs. + text_encoder (`CLIPTextModel`): + The decoder text encoder to be used for text inputs. + decoder (`WuerstchenDiffNeXt`): + The decoder model to be used for decoder image generation pipeline. + scheduler (`DDPMWuerstchenScheduler`): + The scheduler to be used for decoder image generation pipeline. + vqgan (`PaellaVQModel`): + The VQGAN model to be used for decoder image generation pipeline. + prior_tokenizer (`CLIPTokenizer`): + The prior tokenizer to be used for text inputs. + prior_text_encoder (`CLIPTextModel`): + The prior text encoder to be used for text inputs. + prior_prior (`WuerstchenPrior`): + The prior model to be used for prior pipeline. + prior_scheduler (`DDPMWuerstchenScheduler`): + The scheduler to be used for prior pipeline. + """ + + _last_supported_version = "0.33.1" + _load_connected_pipes = True + + def __init__( + self, + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModel, + decoder: WuerstchenDiffNeXt, + scheduler: DDPMWuerstchenScheduler, + vqgan: PaellaVQModel, + prior_tokenizer: CLIPTokenizer, + prior_text_encoder: CLIPTextModel, + prior_prior: WuerstchenPrior, + prior_scheduler: DDPMWuerstchenScheduler, + ): + super().__init__() + + self.register_modules( + text_encoder=text_encoder, + tokenizer=tokenizer, + decoder=decoder, + scheduler=scheduler, + vqgan=vqgan, + prior_prior=prior_prior, + prior_text_encoder=prior_text_encoder, + prior_tokenizer=prior_tokenizer, + prior_scheduler=prior_scheduler, + ) + self.prior_pipe = WuerstchenPriorPipeline( + prior=prior_prior, + text_encoder=prior_text_encoder, + tokenizer=prior_tokenizer, + scheduler=prior_scheduler, + ) + self.decoder_pipe = WuerstchenDecoderPipeline( + text_encoder=text_encoder, + tokenizer=tokenizer, + decoder=decoder, + scheduler=scheduler, + vqgan=vqgan, + ) + + def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None): + self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op) + + def enable_model_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None): + r""" + Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared + to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward` + method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with + `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`. + """ + self.prior_pipe.enable_model_cpu_offload(gpu_id=gpu_id, device=device) + self.decoder_pipe.enable_model_cpu_offload(gpu_id=gpu_id, device=device) + + def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None): + r""" + Offloads all models (`unet`, `text_encoder`, `vae`, and `safety checker` state dicts) to CPU using 🤗 + Accelerate, significantly reducing memory usage. Models are moved to a `torch.device('meta')` and loaded on a + GPU only when their specific submodule's `forward` method is called. Offloading happens on a submodule basis. + Memory savings are higher than using `enable_model_cpu_offload`, but performance is lower. + """ + self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + + def progress_bar(self, iterable=None, total=None): + self.prior_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.progress_bar(iterable=iterable, total=total) + + def set_progress_bar_config(self, **kwargs): + self.prior_pipe.set_progress_bar_config(**kwargs) + self.decoder_pipe.set_progress_bar_config(**kwargs) + + @torch.no_grad() + @replace_example_docstring(TEXT2IMAGE_EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Optional[Union[str, List[str]]] = None, + height: int = 512, + width: int = 512, + prior_num_inference_steps: int = 60, + prior_timesteps: Optional[List[float]] = None, + prior_guidance_scale: float = 4.0, + num_inference_steps: int = 12, + decoder_timesteps: Optional[List[float]] = None, + decoder_guidance_scale: float = 0.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + num_images_per_prompt: int = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + prior_callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + prior_callback_on_step_end_tensor_inputs: List[str] = ["latents"], + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation for the prior and decoder. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings for the prior. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings for the prior. Can be used to easily tweak text inputs, *e.g.* + prompt weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` + input argument. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + prior_guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `prior_guidance_scale` is defined as `w` of + equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by + setting `prior_guidance_scale > 1`. Higher guidance scale encourages to generate images that are + closely linked to the text `prompt`, usually at the expense of lower image quality. + prior_num_inference_steps (`Union[int, Dict[float, int]]`, *optional*, defaults to 60): + The number of prior denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. For more specific timestep spacing, you can pass customized + `prior_timesteps` + num_inference_steps (`int`, *optional*, defaults to 12): + The number of decoder denoising steps. More denoising steps usually lead to a higher quality image at + the expense of slower inference. For more specific timestep spacing, you can pass customized + `timesteps` + prior_timesteps (`List[float]`, *optional*): + Custom timesteps to use for the denoising process for the prior. If not defined, equal spaced + `prior_num_inference_steps` timesteps are used. Must be in descending order. + decoder_timesteps (`List[float]`, *optional*): + Custom timesteps to use for the denoising process for the decoder. If not defined, equal spaced + `num_inference_steps` timesteps are used. Must be in descending order. + decoder_guidance_scale (`float`, *optional*, defaults to 0.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + prior_callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `prior_callback_on_step_end(self: DiffusionPipeline, step: int, timestep: + int, callback_kwargs: Dict)`. + prior_callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `prior_callback_on_step_end` function. The tensors specified in the + list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in + the `._callback_tensor_inputs` attribute of your pipeline class. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` [`~pipelines.ImagePipelineOutput`] if `return_dict` is True, + otherwise a `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + prior_kwargs = {} + if kwargs.get("prior_callback", None) is not None: + prior_kwargs["callback"] = kwargs.pop("prior_callback") + deprecate( + "prior_callback", + "1.0.0", + "Passing `prior_callback` as an input argument to `__call__` is deprecated, consider use `prior_callback_on_step_end`", + ) + if kwargs.get("prior_callback_steps", None) is not None: + deprecate( + "prior_callback_steps", + "1.0.0", + "Passing `prior_callback_steps` as an input argument to `__call__` is deprecated, consider use `prior_callback_on_step_end`", + ) + prior_kwargs["callback_steps"] = kwargs.pop("prior_callback_steps") + + prior_outputs = self.prior_pipe( + prompt=prompt if prompt_embeds is None else None, + height=height, + width=width, + num_inference_steps=prior_num_inference_steps, + timesteps=prior_timesteps, + guidance_scale=prior_guidance_scale, + negative_prompt=negative_prompt if negative_prompt_embeds is None else None, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + num_images_per_prompt=num_images_per_prompt, + generator=generator, + latents=latents, + output_type="pt", + return_dict=False, + callback_on_step_end=prior_callback_on_step_end, + callback_on_step_end_tensor_inputs=prior_callback_on_step_end_tensor_inputs, + **prior_kwargs, + ) + image_embeddings = prior_outputs[0] + + outputs = self.decoder_pipe( + image_embeddings=image_embeddings, + prompt=prompt if prompt is not None else "", + num_inference_steps=num_inference_steps, + timesteps=decoder_timesteps, + guidance_scale=decoder_guidance_scale, + negative_prompt=negative_prompt, + generator=generator, + output_type=output_type, + return_dict=return_dict, + callback_on_step_end=callback_on_step_end, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + **kwargs, + ) + + return outputs diff --git a/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/pipeline_wuerstchen_prior.py b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/pipeline_wuerstchen_prior.py new file mode 100644 index 0000000000000000000000000000000000000000..a32f09204d27a3db16bc2c6c4fee7af701c4536f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/pipelines/wuerstchen/pipeline_wuerstchen_prior.py @@ -0,0 +1,528 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 math import ceil +from typing import Callable, Dict, List, Optional, Union + +import numpy as np +import torch +from transformers import CLIPTextModel, CLIPTokenizer + +from ...loaders import StableDiffusionLoraLoaderMixin +from ...schedulers import DDPMWuerstchenScheduler +from ...utils import BaseOutput, deprecate, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .modeling_wuerstchen_prior import WuerstchenPrior + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +DEFAULT_STAGE_C_TIMESTEPS = list(np.linspace(1.0, 2 / 3, 20)) + list(np.linspace(2 / 3, 0.0, 11))[1:] + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import WuerstchenPriorPipeline + + >>> prior_pipe = WuerstchenPriorPipeline.from_pretrained( + ... "warp-ai/wuerstchen-prior", torch_dtype=torch.float16 + ... ).to("cuda") + + >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet" + >>> prior_output = pipe(prompt) + ``` +""" + + +@dataclass +class WuerstchenPriorPipelineOutput(BaseOutput): + """ + Output class for WuerstchenPriorPipeline. + + Args: + image_embeddings (`torch.Tensor` or `np.ndarray`) + Prior image embeddings for text prompt + + """ + + image_embeddings: Union[torch.Tensor, np.ndarray] + + +class WuerstchenPriorPipeline(DiffusionPipeline, StableDiffusionLoraLoaderMixin): + """ + Pipeline for generating image prior for Wuerstchen. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + prior ([`Prior`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + text_encoder ([`CLIPTextModelWithProjection`]): + Frozen text-encoder. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + scheduler ([`DDPMWuerstchenScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + latent_mean ('float', *optional*, defaults to 42.0): + Mean value for latent diffusers. + latent_std ('float', *optional*, defaults to 1.0): + Standard value for latent diffusers. + resolution_multiple ('float', *optional*, defaults to 42.67): + Default resolution for multiple images generated. + """ + + unet_name = "prior" + text_encoder_name = "text_encoder" + model_cpu_offload_seq = "text_encoder->prior" + _callback_tensor_inputs = ["latents", "text_encoder_hidden_states", "negative_prompt_embeds"] + _lora_loadable_modules = ["prior", "text_encoder"] + + def __init__( + self, + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModel, + prior: WuerstchenPrior, + scheduler: DDPMWuerstchenScheduler, + latent_mean: float = 42.0, + latent_std: float = 1.0, + resolution_multiple: float = 42.67, + ) -> None: + super().__init__() + self.register_modules( + tokenizer=tokenizer, + text_encoder=text_encoder, + prior=prior, + scheduler=scheduler, + ) + self.register_to_config( + latent_mean=latent_mean, latent_std=latent_std, resolution_multiple=resolution_multiple + ) + + # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def encode_prompt( + self, + device, + num_images_per_prompt, + do_classifier_free_guidance, + prompt=None, + negative_prompt=None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + ): + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + attention_mask = text_inputs.attention_mask + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] + attention_mask = attention_mask[:, : self.tokenizer.model_max_length] + + text_encoder_output = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask.to(device) + ) + prompt_embeds = text_encoder_output.last_hidden_state + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + + if negative_prompt_embeds is None and do_classifier_free_guidance: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + negative_prompt_embeds_text_encoder_output = self.text_encoder( + uncond_input.input_ids.to(device), attention_mask=uncond_input.attention_mask.to(device) + ) + + negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.last_hidden_state + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + # done duplicates + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + num_inference_steps, + do_classifier_free_guidance, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if not isinstance(num_inference_steps, int): + raise TypeError( + f"'num_inference_steps' must be of type 'int', but got {type(num_inference_steps)}\ + In Case you want to provide explicit timesteps, please use the 'timesteps' argument." + ) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Optional[Union[str, List[str]]] = None, + height: int = 1024, + width: int = 1024, + num_inference_steps: int = 60, + timesteps: List[float] = None, + guidance_scale: float = 8.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + num_images_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pt", + return_dict: bool = True, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + height (`int`, *optional*, defaults to 1024): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 1024): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 60): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps` + timesteps are used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to 8.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `decoder_guidance_scale` is defined as `w` of + equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by + setting `decoder_guidance_scale > 1`. Higher guidance scale encourages to generate images that are + closely linked to the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `decoder_guidance_scale` is less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.WuerstchenPriorPipelineOutput`] or `tuple` [`~pipelines.WuerstchenPriorPipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated image embeddings. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + # 0. Define commonly used variables + device = self._execution_device + self._guidance_scale = guidance_scale + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 1. Check inputs. Raise error if not correct + if prompt is not None and not isinstance(prompt, list): + if isinstance(prompt, str): + prompt = [prompt] + else: + raise TypeError(f"'prompt' must be of type 'list' or 'str', but got {type(prompt)}.") + + if self.do_classifier_free_guidance: + if negative_prompt is not None and not isinstance(negative_prompt, list): + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] + else: + raise TypeError( + f"'negative_prompt' must be of type 'list' or 'str', but got {type(negative_prompt)}." + ) + + self.check_inputs( + prompt, + negative_prompt, + num_inference_steps, + self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + # 2. Encode caption + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + text_encoder_hidden_states = ( + torch.cat([prompt_embeds, negative_prompt_embeds]) if negative_prompt_embeds is not None else prompt_embeds + ) + + # 3. Determine latent shape of image embeddings + dtype = text_encoder_hidden_states.dtype + latent_height = ceil(height / self.config.resolution_multiple) + latent_width = ceil(width / self.config.resolution_multiple) + num_channels = self.prior.config.c_in + effnet_features_shape = (num_images_per_prompt * batch_size, num_channels, latent_height, latent_width) + + # 4. Prepare and set timesteps + if timesteps is not None: + self.scheduler.set_timesteps(timesteps=timesteps, device=device) + timesteps = self.scheduler.timesteps + num_inference_steps = len(timesteps) + else: + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latents + latents = self.prepare_latents(effnet_features_shape, dtype, device, generator, latents, self.scheduler) + + # 6. Run denoising loop + self._num_timesteps = len(timesteps[:-1]) + for i, t in enumerate(self.progress_bar(timesteps[:-1])): + ratio = t.expand(latents.size(0)).to(dtype) + + # 7. Denoise image embeddings + predicted_image_embedding = self.prior( + torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents, + r=torch.cat([ratio] * 2) if self.do_classifier_free_guidance else ratio, + c=text_encoder_hidden_states, + ) + + # 8. Check for classifier free guidance and apply it + if self.do_classifier_free_guidance: + predicted_image_embedding_text, predicted_image_embedding_uncond = predicted_image_embedding.chunk(2) + predicted_image_embedding = torch.lerp( + predicted_image_embedding_uncond, predicted_image_embedding_text, self.guidance_scale + ) + + # 9. Renoise latents to next timestep + latents = self.scheduler.step( + model_output=predicted_image_embedding, + timestep=ratio, + sample=latents, + generator=generator, + ).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + text_encoder_hidden_states = callback_outputs.pop( + "text_encoder_hidden_states", text_encoder_hidden_states + ) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 10. Denormalize the latents + latents = latents * self.config.latent_mean - self.config.latent_std + + # Offload all models + self.maybe_free_model_hooks() + + if output_type == "np": + latents = latents.cpu().float().numpy() + + if not return_dict: + return (latents,) + + return WuerstchenPriorPipelineOutput(latents) diff --git a/venv/lib/python3.10/site-packages/diffusers/py.typed b/venv/lib/python3.10/site-packages/diffusers/py.typed new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/__init__.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..efd24187532179892108466c386154e1a54e3bde --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/__init__.py @@ -0,0 +1,194 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 inspect +from typing import Dict, List, Optional, Union + +from ..utils import is_transformers_available, logging +from .auto import DiffusersAutoQuantizer +from .base import DiffusersQuantizer +from .quantization_config import QuantizationConfigMixin as DiffQuantConfigMixin + + +try: + from transformers.utils.quantization_config import QuantizationConfigMixin as TransformersQuantConfigMixin +except ImportError: + + class TransformersQuantConfigMixin: + pass + + +logger = logging.get_logger(__name__) + + +class PipelineQuantizationConfig: + """ + Configuration class to be used when applying quantization on-the-fly to [`~DiffusionPipeline.from_pretrained`]. + + Args: + quant_backend (`str`): Quantization backend to be used. When using this option, we assume that the backend + is available to both `diffusers` and `transformers`. + quant_kwargs (`dict`): Params to initialize the quantization backend class. + components_to_quantize (`list`): Components of a pipeline to be quantized. + quant_mapping (`dict`): Mapping defining the quantization specs to be used for the pipeline + components. When using this argument, users are not expected to provide `quant_backend`, `quant_kawargs`, + and `components_to_quantize`. + """ + + def __init__( + self, + quant_backend: str = None, + quant_kwargs: Dict[str, Union[str, float, int, dict]] = None, + components_to_quantize: Optional[List[str]] = None, + quant_mapping: Dict[str, Union[DiffQuantConfigMixin, "TransformersQuantConfigMixin"]] = None, + ): + self.quant_backend = quant_backend + # Initialize kwargs to be {} to set to the defaults. + self.quant_kwargs = quant_kwargs or {} + self.components_to_quantize = components_to_quantize + self.quant_mapping = quant_mapping + + self.post_init() + + def post_init(self): + quant_mapping = self.quant_mapping + self.is_granular = True if quant_mapping is not None else False + + self._validate_init_args() + + def _validate_init_args(self): + if self.quant_backend and self.quant_mapping: + raise ValueError("Both `quant_backend` and `quant_mapping` cannot be specified at the same time.") + + if not self.quant_mapping and not self.quant_backend: + raise ValueError("Must provide a `quant_backend` when not providing a `quant_mapping`.") + + if not self.quant_kwargs and not self.quant_mapping: + raise ValueError("Both `quant_kwargs` and `quant_mapping` cannot be None.") + + if self.quant_backend is not None: + self._validate_init_kwargs_in_backends() + + if self.quant_mapping is not None: + self._validate_quant_mapping_args() + + def _validate_init_kwargs_in_backends(self): + quant_backend = self.quant_backend + + self._check_backend_availability(quant_backend) + + quant_config_mapping_transformers, quant_config_mapping_diffusers = self._get_quant_config_list() + + if quant_config_mapping_transformers is not None: + init_kwargs_transformers = inspect.signature(quant_config_mapping_transformers[quant_backend].__init__) + init_kwargs_transformers = {name for name in init_kwargs_transformers.parameters if name != "self"} + else: + init_kwargs_transformers = None + + init_kwargs_diffusers = inspect.signature(quant_config_mapping_diffusers[quant_backend].__init__) + init_kwargs_diffusers = {name for name in init_kwargs_diffusers.parameters if name != "self"} + + if init_kwargs_transformers != init_kwargs_diffusers: + raise ValueError( + "The signatures of the __init__ methods of the quantization config classes in `diffusers` and `transformers` don't match. " + f"Please provide a `quant_mapping` instead, in the {self.__class__.__name__} class. Refer to [the docs](https://huggingface.co/docs/diffusers/main/en/quantization/overview#pipeline-level-quantization) to learn more about how " + "this mapping would look like." + ) + + def _validate_quant_mapping_args(self): + quant_mapping = self.quant_mapping + transformers_map, diffusers_map = self._get_quant_config_list() + + available_transformers = list(transformers_map.values()) if transformers_map else None + available_diffusers = list(diffusers_map.values()) + + for module_name, config in quant_mapping.items(): + if any(isinstance(config, cfg) for cfg in available_diffusers): + continue + + if available_transformers and any(isinstance(config, cfg) for cfg in available_transformers): + continue + + if available_transformers: + raise ValueError( + f"Provided config for module_name={module_name} could not be found. " + f"Available diffusers configs: {available_diffusers}; " + f"Available transformers configs: {available_transformers}." + ) + else: + raise ValueError( + f"Provided config for module_name={module_name} could not be found. " + f"Available diffusers configs: {available_diffusers}." + ) + + def _check_backend_availability(self, quant_backend: str): + quant_config_mapping_transformers, quant_config_mapping_diffusers = self._get_quant_config_list() + + available_backends_transformers = ( + list(quant_config_mapping_transformers.keys()) if quant_config_mapping_transformers else None + ) + available_backends_diffusers = list(quant_config_mapping_diffusers.keys()) + + if ( + available_backends_transformers and quant_backend not in available_backends_transformers + ) or quant_backend not in quant_config_mapping_diffusers: + error_message = f"Provided quant_backend={quant_backend} was not found." + if available_backends_transformers: + error_message += f"\nAvailable ones (transformers): {available_backends_transformers}." + error_message += f"\nAvailable ones (diffusers): {available_backends_diffusers}." + raise ValueError(error_message) + + def _resolve_quant_config(self, is_diffusers: bool = True, module_name: str = None): + quant_config_mapping_transformers, quant_config_mapping_diffusers = self._get_quant_config_list() + + quant_mapping = self.quant_mapping + components_to_quantize = self.components_to_quantize + + # Granular case + if self.is_granular and module_name in quant_mapping: + logger.debug(f"Initializing quantization config class for {module_name}.") + config = quant_mapping[module_name] + return config + + # Global config case + else: + should_quantize = False + # Only quantize the modules requested for. + if components_to_quantize and module_name in components_to_quantize: + should_quantize = True + # No specification for `components_to_quantize` means all modules should be quantized. + elif not self.is_granular and not components_to_quantize: + should_quantize = True + + if should_quantize: + logger.debug(f"Initializing quantization config class for {module_name}.") + mapping_to_use = quant_config_mapping_diffusers if is_diffusers else quant_config_mapping_transformers + quant_config_cls = mapping_to_use[self.quant_backend] + quant_kwargs = self.quant_kwargs + return quant_config_cls(**quant_kwargs) + + # Fallback: no applicable configuration found. + return None + + def _get_quant_config_list(self): + if is_transformers_available(): + from transformers.quantizers.auto import ( + AUTO_QUANTIZATION_CONFIG_MAPPING as quant_config_mapping_transformers, + ) + else: + quant_config_mapping_transformers = None + + from ..quantizers.auto import AUTO_QUANTIZATION_CONFIG_MAPPING as quant_config_mapping_diffusers + + return quant_config_mapping_transformers, quant_config_mapping_diffusers diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..60ed96c629edfe99f74a4dd382728192520c5969 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/__pycache__/auto.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/__pycache__/auto.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..29bfe7d238c120588f15506192c2c57d3e3e4262 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/__pycache__/auto.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/__pycache__/base.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/__pycache__/base.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f5cc9c9bedc125751a28a8882a7ca38aa9af89f8 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/__pycache__/base.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/__pycache__/quantization_config.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/__pycache__/quantization_config.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f0534a97a115013b7dc9830e688c625e613e1a32 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/__pycache__/quantization_config.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/auto.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/auto.py new file mode 100644 index 0000000000000000000000000000000000000000..ce214ae7bc17da37d542fabe0341b7e541c7e708 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/auto.py @@ -0,0 +1,143 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# 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. +""" +Adapted from +https://github.com/huggingface/transformers/blob/c409cd81777fb27aadc043ed3d8339dbc020fb3b/src/transformers/quantizers/auto.py +""" + +import warnings +from typing import Dict, Optional, Union + +from .bitsandbytes import BnB4BitDiffusersQuantizer, BnB8BitDiffusersQuantizer +from .gguf import GGUFQuantizer +from .quantization_config import ( + BitsAndBytesConfig, + GGUFQuantizationConfig, + QuantizationConfigMixin, + QuantizationMethod, + QuantoConfig, + TorchAoConfig, +) +from .quanto import QuantoQuantizer +from .torchao import TorchAoHfQuantizer + + +AUTO_QUANTIZER_MAPPING = { + "bitsandbytes_4bit": BnB4BitDiffusersQuantizer, + "bitsandbytes_8bit": BnB8BitDiffusersQuantizer, + "gguf": GGUFQuantizer, + "quanto": QuantoQuantizer, + "torchao": TorchAoHfQuantizer, +} + +AUTO_QUANTIZATION_CONFIG_MAPPING = { + "bitsandbytes_4bit": BitsAndBytesConfig, + "bitsandbytes_8bit": BitsAndBytesConfig, + "gguf": GGUFQuantizationConfig, + "quanto": QuantoConfig, + "torchao": TorchAoConfig, +} + + +class DiffusersAutoQuantizer: + """ + The auto diffusers quantizer class that takes care of automatically instantiating to the correct + `DiffusersQuantizer` given the `QuantizationConfig`. + """ + + @classmethod + def from_dict(cls, quantization_config_dict: Dict): + quant_method = quantization_config_dict.get("quant_method", None) + # We need a special care for bnb models to make sure everything is BC .. + if quantization_config_dict.get("load_in_8bit", False) or quantization_config_dict.get("load_in_4bit", False): + suffix = "_4bit" if quantization_config_dict.get("load_in_4bit", False) else "_8bit" + quant_method = QuantizationMethod.BITS_AND_BYTES + suffix + elif quant_method is None: + raise ValueError( + "The model's quantization config from the arguments has no `quant_method` attribute. Make sure that the model has been correctly quantized" + ) + + if quant_method not in AUTO_QUANTIZATION_CONFIG_MAPPING.keys(): + raise ValueError( + f"Unknown quantization type, got {quant_method} - supported types are:" + f" {list(AUTO_QUANTIZER_MAPPING.keys())}" + ) + + target_cls = AUTO_QUANTIZATION_CONFIG_MAPPING[quant_method] + return target_cls.from_dict(quantization_config_dict) + + @classmethod + def from_config(cls, quantization_config: Union[QuantizationConfigMixin, Dict], **kwargs): + # Convert it to a QuantizationConfig if the q_config is a dict + if isinstance(quantization_config, dict): + quantization_config = cls.from_dict(quantization_config) + + quant_method = quantization_config.quant_method + + # Again, we need a special care for bnb as we have a single quantization config + # class for both 4-bit and 8-bit quantization + if quant_method == QuantizationMethod.BITS_AND_BYTES: + if quantization_config.load_in_8bit: + quant_method += "_8bit" + else: + quant_method += "_4bit" + + if quant_method not in AUTO_QUANTIZER_MAPPING.keys(): + raise ValueError( + f"Unknown quantization type, got {quant_method} - supported types are:" + f" {list(AUTO_QUANTIZER_MAPPING.keys())}" + ) + + target_cls = AUTO_QUANTIZER_MAPPING[quant_method] + return target_cls(quantization_config, **kwargs) + + @classmethod + def from_pretrained(cls, pretrained_model_name_or_path, **kwargs): + model_config = cls.load_config(pretrained_model_name_or_path, **kwargs) + if getattr(model_config, "quantization_config", None) is None: + raise ValueError( + f"Did not found a `quantization_config` in {pretrained_model_name_or_path}. Make sure that the model is correctly quantized." + ) + quantization_config_dict = model_config.quantization_config + quantization_config = cls.from_dict(quantization_config_dict) + # Update with potential kwargs that are passed through from_pretrained. + quantization_config.update(kwargs) + + return cls.from_config(quantization_config) + + @classmethod + def merge_quantization_configs( + cls, + quantization_config: Union[dict, QuantizationConfigMixin], + quantization_config_from_args: Optional[QuantizationConfigMixin], + ): + """ + handles situations where both quantization_config from args and quantization_config from model config are + present. + """ + if quantization_config_from_args is not None: + warning_msg = ( + "You passed `quantization_config` or equivalent parameters to `from_pretrained` but the model you're loading" + " already has a `quantization_config` attribute. The `quantization_config` from the model will be used." + ) + else: + warning_msg = "" + + if isinstance(quantization_config, dict): + quantization_config = cls.from_dict(quantization_config) + + if warning_msg != "": + warnings.warn(warning_msg) + + return quantization_config diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/base.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/base.py new file mode 100644 index 0000000000000000000000000000000000000000..357d920d29c45dd28d123ae16d76f83d8a069cc1 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/base.py @@ -0,0 +1,234 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# 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. + +""" +Adapted from +https://github.com/huggingface/transformers/blob/52cb4034ada381fe1ffe8d428a1076e5411a8026/src/transformers/quantizers/base.py +""" + +from abc import ABC, abstractmethod +from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union + +from ..utils import is_torch_available +from .quantization_config import QuantizationConfigMixin + + +if TYPE_CHECKING: + from ..models.modeling_utils import ModelMixin + +if is_torch_available(): + import torch + + +class DiffusersQuantizer(ABC): + """ + Abstract class of the HuggingFace quantizer. Supports for now quantizing HF diffusers models for inference and/or + quantization. This class is used only for diffusers.models.modeling_utils.ModelMixin.from_pretrained and cannot be + easily used outside the scope of that method yet. + + Attributes + quantization_config (`diffusers.quantizers.quantization_config.QuantizationConfigMixin`): + The quantization config that defines the quantization parameters of your model that you want to quantize. + modules_to_not_convert (`List[str]`, *optional*): + The list of module names to not convert when quantizing the model. + required_packages (`List[str]`, *optional*): + The list of required pip packages to install prior to using the quantizer + requires_calibration (`bool`): + Whether the quantization method requires to calibrate the model before using it. + """ + + requires_calibration = False + required_packages = None + + def __init__(self, quantization_config: QuantizationConfigMixin, **kwargs): + self.quantization_config = quantization_config + + # -- Handle extra kwargs below -- + self.modules_to_not_convert = kwargs.pop("modules_to_not_convert", []) + self.pre_quantized = kwargs.pop("pre_quantized", True) + + if not self.pre_quantized and self.requires_calibration: + raise ValueError( + f"The quantization method {quantization_config.quant_method} does require the model to be pre-quantized." + f" You explicitly passed `pre_quantized=False` meaning your model weights are not quantized. Make sure to " + f"pass `pre_quantized=True` while knowing what you are doing." + ) + + def update_torch_dtype(self, torch_dtype: "torch.dtype") -> "torch.dtype": + """ + Some quantization methods require to explicitly set the dtype of the model to a target dtype. You need to + override this method in case you want to make sure that behavior is preserved + + Args: + torch_dtype (`torch.dtype`): + The input dtype that is passed in `from_pretrained` + """ + return torch_dtype + + def update_device_map(self, device_map: Optional[Dict[str, Any]]) -> Optional[Dict[str, Any]]: + """ + Override this method if you want to pass a override the existing device map with a new one. E.g. for + bitsandbytes, since `accelerate` is a hard requirement, if no device_map is passed, the device_map is set to + `"auto"`` + + Args: + device_map (`Union[dict, str]`, *optional*): + The device_map that is passed through the `from_pretrained` method. + """ + return device_map + + def adjust_target_dtype(self, torch_dtype: "torch.dtype") -> "torch.dtype": + """ + Override this method if you want to adjust the `target_dtype` variable used in `from_pretrained` to compute the + device_map in case the device_map is a `str`. E.g. for bitsandbytes we force-set `target_dtype` to `torch.int8` + and for 4-bit we pass a custom enum `accelerate.CustomDtype.int4`. + + Args: + torch_dtype (`torch.dtype`, *optional*): + The torch_dtype that is used to compute the device_map. + """ + return torch_dtype + + def update_missing_keys(self, model, missing_keys: List[str], prefix: str) -> List[str]: + """ + Override this method if you want to adjust the `missing_keys`. + + Args: + missing_keys (`List[str]`, *optional*): + The list of missing keys in the checkpoint compared to the state dict of the model + """ + return missing_keys + + def get_special_dtypes_update(self, model, torch_dtype: "torch.dtype") -> Dict[str, "torch.dtype"]: + """ + returns dtypes for modules that are not quantized - used for the computation of the device_map in case one + passes a str as a device_map. The method will use the `modules_to_not_convert` that is modified in + `_process_model_before_weight_loading`. `diffusers` models don't have any `modules_to_not_convert` attributes + yet but this can change soon in the future. + + Args: + model (`~diffusers.models.modeling_utils.ModelMixin`): + The model to quantize + torch_dtype (`torch.dtype`): + The dtype passed in `from_pretrained` method. + """ + + return { + name: torch_dtype + for name, _ in model.named_parameters() + if any(m in name for m in self.modules_to_not_convert) + } + + def adjust_max_memory(self, max_memory: Dict[str, Union[int, str]]) -> Dict[str, Union[int, str]]: + """adjust max_memory argument for infer_auto_device_map() if extra memory is needed for quantization""" + return max_memory + + def check_if_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + state_dict: Dict[str, Any], + **kwargs, + ) -> bool: + """ + checks if a loaded state_dict component is part of quantized param + some validation; only defined for + quantization methods that require to create a new parameters for quantization. + """ + return False + + def create_quantized_param(self, *args, **kwargs) -> "torch.nn.Parameter": + """ + takes needed components from state_dict and creates quantized param. + """ + return + + def check_quantized_param_shape(self, *args, **kwargs): + """ + checks if the quantized param has expected shape. + """ + return True + + def validate_environment(self, *args, **kwargs): + """ + This method is used to potentially check for potential conflicts with arguments that are passed in + `from_pretrained`. You need to define it for all future quantizers that are integrated with diffusers. If no + explicit check are needed, simply return nothing. + """ + return + + def preprocess_model(self, model: "ModelMixin", **kwargs): + """ + Setting model attributes and/or converting model before weights loading. At this point the model should be + initialized on the meta device so you can freely manipulate the skeleton of the model in order to replace + modules in-place. Make sure to override the abstract method `_process_model_before_weight_loading`. + + Args: + model (`~diffusers.models.modeling_utils.ModelMixin`): + The model to quantize + kwargs (`dict`, *optional*): + The keyword arguments that are passed along `_process_model_before_weight_loading`. + """ + model.is_quantized = True + model.quantization_method = self.quantization_config.quant_method + return self._process_model_before_weight_loading(model, **kwargs) + + def postprocess_model(self, model: "ModelMixin", **kwargs): + """ + Post-process the model post weights loading. Make sure to override the abstract method + `_process_model_after_weight_loading`. + + Args: + model (`~diffusers.models.modeling_utils.ModelMixin`): + The model to quantize + kwargs (`dict`, *optional*): + The keyword arguments that are passed along `_process_model_after_weight_loading`. + """ + return self._process_model_after_weight_loading(model, **kwargs) + + def dequantize(self, model): + """ + Potentially dequantize the model to retrieve the original model, with some loss in accuracy / performance. Note + not all quantization schemes support this. + """ + model = self._dequantize(model) + + # Delete quantizer and quantization config + del model.hf_quantizer + + return model + + def _dequantize(self, model): + raise NotImplementedError( + f"{self.quantization_config.quant_method} has no implementation of `dequantize`, please raise an issue on GitHub." + ) + + @abstractmethod + def _process_model_before_weight_loading(self, model, **kwargs): ... + + @abstractmethod + def _process_model_after_weight_loading(self, model, **kwargs): ... + + @property + @abstractmethod + def is_serializable(self): ... + + @property + @abstractmethod + def is_trainable(self): ... + + @property + def is_compileable(self) -> bool: + """Flag indicating whether the quantized model can be compiled""" + return False diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/__init__.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..9e745bc810faad76608b00ab9c493369466d0970 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/__init__.py @@ -0,0 +1,2 @@ +from .bnb_quantizer import BnB4BitDiffusersQuantizer, BnB8BitDiffusersQuantizer +from .utils import dequantize_and_replace, dequantize_bnb_weight, replace_with_bnb_linear diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..38abd3b99d6d8fb778e076ee0daeaa8d51ff8fee Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/__pycache__/bnb_quantizer.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/__pycache__/bnb_quantizer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..490b438cdc8790bda2e5408e5b88f3a3987491bc Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/__pycache__/bnb_quantizer.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/__pycache__/utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/__pycache__/utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..18b1a46671cb701c1bc18a73762f956dd6e218d0 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/__pycache__/utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/bnb_quantizer.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/bnb_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..0dfdff019b79ef3bfb9b89df6378ab398199beeb --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/bnb_quantizer.py @@ -0,0 +1,577 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# 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. +""" +Adapted from +https://github.com/huggingface/transformers/blob/c409cd81777fb27aadc043ed3d8339dbc020fb3b/src/transformers/quantizers/quantizer_bnb_4bit.py +""" + +from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union + +from ...utils import get_module_from_name +from ..base import DiffusersQuantizer + + +if TYPE_CHECKING: + from ...models.modeling_utils import ModelMixin + +from ...utils import ( + is_accelerate_available, + is_accelerate_version, + is_bitsandbytes_available, + is_bitsandbytes_version, + is_torch_available, + logging, +) + + +if is_torch_available(): + import torch + +logger = logging.get_logger(__name__) + + +class BnB4BitDiffusersQuantizer(DiffusersQuantizer): + """ + 4-bit quantization from bitsandbytes.py quantization method: + before loading: converts transformer layers into Linear4bit during loading: load 16bit weight and pass to the + layer object after: quantizes individual weights in Linear4bit into 4bit at the first .cuda() call saving: + from state dict, as usual; saves weights and `quant_state` components + loading: + need to locate `quant_state` components and pass to Param4bit constructor + """ + + use_keep_in_fp32_modules = True + requires_calibration = False + + def __init__(self, quantization_config, **kwargs): + super().__init__(quantization_config, **kwargs) + + if self.quantization_config.llm_int8_skip_modules is not None: + self.modules_to_not_convert = self.quantization_config.llm_int8_skip_modules + + def validate_environment(self, *args, **kwargs): + if not (torch.cuda.is_available() or torch.xpu.is_available()): + raise RuntimeError("No GPU found. A GPU is needed for quantization.") + if not is_accelerate_available() or is_accelerate_version("<", "0.26.0"): + raise ImportError( + "Using `bitsandbytes` 4-bit quantization requires Accelerate: `pip install 'accelerate>=0.26.0'`" + ) + if not is_bitsandbytes_available() or is_bitsandbytes_version("<", "0.43.3"): + raise ImportError( + "Using `bitsandbytes` 4-bit quantization requires the latest version of bitsandbytes: `pip install -U bitsandbytes`" + ) + + if kwargs.get("from_flax", False): + raise ValueError( + "Converting into 4-bit weights from flax weights is currently not supported, please make" + " sure the weights are in PyTorch format." + ) + + device_map = kwargs.get("device_map", None) + if ( + device_map is not None + and isinstance(device_map, dict) + and not self.quantization_config.llm_int8_enable_fp32_cpu_offload + ): + device_map_without_no_convert = { + key: device_map[key] for key in device_map.keys() if key not in self.modules_to_not_convert + } + if "cpu" in device_map_without_no_convert.values() or "disk" in device_map_without_no_convert.values(): + raise ValueError( + "Some modules are dispatched on the CPU or the disk. Make sure you have enough GPU RAM to fit the " + "quantized model. If you want to dispatch the model on the CPU or the disk while keeping these modules " + "in 32-bit, you need to set `load_in_8bit_fp32_cpu_offload=True` and pass a custom `device_map` to " + "`from_pretrained`. Check " + "https://huggingface.co/docs/transformers/main/en/main_classes/quantization#offload-between-cpu-and-gpu " + "for more details. " + ) + + def adjust_target_dtype(self, target_dtype: "torch.dtype") -> "torch.dtype": + if target_dtype != torch.int8: + from accelerate.utils import CustomDtype + + logger.info("target_dtype {target_dtype} is replaced by `CustomDtype.INT4` for 4-bit BnB quantization") + return CustomDtype.INT4 + else: + raise ValueError(f"Wrong `target_dtype` ({target_dtype}) provided.") + + def check_if_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + state_dict: Dict[str, Any], + **kwargs, + ) -> bool: + import bitsandbytes as bnb + + module, tensor_name = get_module_from_name(model, param_name) + if isinstance(module._parameters.get(tensor_name, None), bnb.nn.Params4bit): + # Add here check for loaded components' dtypes once serialization is implemented + return True + elif isinstance(module, bnb.nn.Linear4bit) and tensor_name == "bias": + # bias could be loaded by regular set_module_tensor_to_device() from accelerate, + # but it would wrongly use uninitialized weight there. + return True + else: + return False + + def create_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + target_device: "torch.device", + state_dict: Dict[str, Any], + unexpected_keys: Optional[List[str]] = None, + **kwargs, + ): + import bitsandbytes as bnb + + module, tensor_name = get_module_from_name(model, param_name) + + if tensor_name not in module._parameters: + raise ValueError(f"{module} does not have a parameter or a buffer named {tensor_name}.") + + old_value = getattr(module, tensor_name) + + if tensor_name == "bias": + if param_value is None: + new_value = old_value.to(target_device) + else: + new_value = param_value.to(target_device) + + new_value = torch.nn.Parameter(new_value, requires_grad=old_value.requires_grad) + module._parameters[tensor_name] = new_value + return + + if not isinstance(module._parameters[tensor_name], bnb.nn.Params4bit): + raise ValueError("this function only loads `Linear4bit components`") + if ( + old_value.device == torch.device("meta") + and target_device not in ["meta", torch.device("meta")] + and param_value is None + ): + raise ValueError(f"{tensor_name} is on the meta device, we need a `value` to put in on {target_device}.") + + # construct `new_value` for the module._parameters[tensor_name]: + if self.pre_quantized: + # 4bit loading. Collecting components for restoring quantized weight + # This can be expanded to make a universal call for any quantized weight loading + + if not self.is_serializable: + raise ValueError( + "Detected int4 weights but the version of bitsandbytes is not compatible with int4 serialization. " + "Make sure to download the latest `bitsandbytes` version. `pip install --upgrade bitsandbytes`." + ) + + if (param_name + ".quant_state.bitsandbytes__fp4" not in state_dict) and ( + param_name + ".quant_state.bitsandbytes__nf4" not in state_dict + ): + raise ValueError( + f"Supplied state dict for {param_name} does not contain `bitsandbytes__*` and possibly other `quantized_stats` components." + ) + + quantized_stats = {} + for k, v in state_dict.items(): + # `startswith` to counter for edge cases where `param_name` + # substring can be present in multiple places in the `state_dict` + if param_name + "." in k and k.startswith(param_name): + quantized_stats[k] = v + if unexpected_keys is not None and k in unexpected_keys: + unexpected_keys.remove(k) + + new_value = bnb.nn.Params4bit.from_prequantized( + data=param_value, + quantized_stats=quantized_stats, + requires_grad=False, + device=target_device, + ) + else: + new_value = param_value.to("cpu") + kwargs = old_value.__dict__ + new_value = bnb.nn.Params4bit(new_value, requires_grad=False, **kwargs).to(target_device) + + module._parameters[tensor_name] = new_value + + def check_quantized_param_shape(self, param_name, current_param, loaded_param): + current_param_shape = current_param.shape + loaded_param_shape = loaded_param.shape + + n = current_param_shape.numel() + inferred_shape = (n,) if "bias" in param_name else ((n + 1) // 2, 1) + if loaded_param_shape != inferred_shape: + raise ValueError( + f"Expected the flattened shape of the current param ({param_name}) to be {loaded_param_shape} but is {inferred_shape}." + ) + else: + return True + + def adjust_max_memory(self, max_memory: Dict[str, Union[int, str]]) -> Dict[str, Union[int, str]]: + # need more space for buffers that are created during quantization + max_memory = {key: val * 0.90 for key, val in max_memory.items()} + return max_memory + + def update_torch_dtype(self, torch_dtype: "torch.dtype") -> "torch.dtype": + if torch_dtype is None: + # We force the `dtype` to be float16, this is a requirement from `bitsandbytes` + logger.info( + "Overriding torch_dtype=%s with `torch_dtype=torch.float16` due to " + "requirements of `bitsandbytes` to enable model loading in 8-bit or 4-bit. " + "Pass your own torch_dtype to specify the dtype of the remaining non-linear layers or pass" + " torch_dtype=torch.float16 to remove this warning.", + torch_dtype, + ) + torch_dtype = torch.float16 + return torch_dtype + + def update_device_map(self, device_map): + if device_map is None: + if torch.xpu.is_available(): + current_device = f"xpu:{torch.xpu.current_device()}" + else: + current_device = f"cuda:{torch.cuda.current_device()}" + device_map = {"": current_device} + logger.info( + "The device_map was not initialized. " + "Setting device_map to {" + ": {current_device}}. " + "If you want to use the model for inference, please set device_map ='auto' " + ) + return device_map + + def _process_model_before_weight_loading( + self, + model: "ModelMixin", + device_map, + keep_in_fp32_modules: List[str] = [], + **kwargs, + ): + from .utils import replace_with_bnb_linear + + load_in_8bit_fp32_cpu_offload = self.quantization_config.llm_int8_enable_fp32_cpu_offload + + # We may keep some modules such as the `proj_out` in their original dtype for numerical stability reasons + self.modules_to_not_convert = self.quantization_config.llm_int8_skip_modules + + if not isinstance(self.modules_to_not_convert, list): + self.modules_to_not_convert = [self.modules_to_not_convert] + + self.modules_to_not_convert.extend(keep_in_fp32_modules) + + # Extend `self.modules_to_not_convert` to keys that are supposed to be offloaded to `cpu` or `disk` + if isinstance(device_map, dict) and len(device_map.keys()) > 1: + keys_on_cpu = [key for key, value in device_map.items() if value in ["disk", "cpu"]] + + if len(keys_on_cpu) > 0 and not load_in_8bit_fp32_cpu_offload: + raise ValueError( + "If you want to offload some keys to `cpu` or `disk`, you need to set " + "`llm_int8_enable_fp32_cpu_offload=True`. Note that these modules will not be " + " converted to 8-bit but kept in 32-bit." + ) + self.modules_to_not_convert.extend(keys_on_cpu) + + # Purge `None`. + # Unlike `transformers`, we don't know if we should always keep certain modules in FP32 + # in case of diffusion transformer models. For language models and others alike, `lm_head` + # and tied modules are usually kept in FP32. + self.modules_to_not_convert = [module for module in self.modules_to_not_convert if module is not None] + + model = replace_with_bnb_linear( + model, modules_to_not_convert=self.modules_to_not_convert, quantization_config=self.quantization_config + ) + model.config.quantization_config = self.quantization_config + model.is_loaded_in_4bit = True + + def _process_model_after_weight_loading(self, model: "ModelMixin", **kwargs): + model.is_4bit_serializable = self.is_serializable + return model + + @property + def is_serializable(self): + # Because we're mandating `bitsandbytes` 0.43.3. + return True + + @property + def is_trainable(self) -> bool: + # Because we're mandating `bitsandbytes` 0.43.3. + return True + + def _dequantize(self, model): + from .utils import dequantize_and_replace + + is_model_on_cpu = model.device.type == "cpu" + if is_model_on_cpu: + logger.info( + "Model was found to be on CPU (could happen as a result of `enable_model_cpu_offload()`). So, moving it to GPU. After dequantization, will move the model back to CPU again to preserve the previous device." + ) + if torch.xpu.is_available(): + model.to(torch.xpu.current_device()) + else: + model.to(torch.cuda.current_device()) + + model = dequantize_and_replace( + model, self.modules_to_not_convert, quantization_config=self.quantization_config + ) + if is_model_on_cpu: + model.to("cpu") + return model + + +class BnB8BitDiffusersQuantizer(DiffusersQuantizer): + """ + 8-bit quantization from bitsandbytes quantization method: + before loading: converts transformer layers into Linear8bitLt during loading: load 16bit weight and pass to the + layer object after: quantizes individual weights in Linear8bitLt into 8bit at fitst .cuda() call + saving: + from state dict, as usual; saves weights and 'SCB' component + loading: + need to locate SCB component and pass to the Linear8bitLt object + """ + + use_keep_in_fp32_modules = True + requires_calibration = False + + def __init__(self, quantization_config, **kwargs): + super().__init__(quantization_config, **kwargs) + + if self.quantization_config.llm_int8_skip_modules is not None: + self.modules_to_not_convert = self.quantization_config.llm_int8_skip_modules + + def validate_environment(self, *args, **kwargs): + if not (torch.cuda.is_available() or torch.xpu.is_available()): + raise RuntimeError("No GPU found. A GPU is needed for quantization.") + if not is_accelerate_available() or is_accelerate_version("<", "0.26.0"): + raise ImportError( + "Using `bitsandbytes` 8-bit quantization requires Accelerate: `pip install 'accelerate>=0.26.0'`" + ) + if not is_bitsandbytes_available() or is_bitsandbytes_version("<", "0.43.3"): + raise ImportError( + "Using `bitsandbytes` 8-bit quantization requires the latest version of bitsandbytes: `pip install -U bitsandbytes`" + ) + + if kwargs.get("from_flax", False): + raise ValueError( + "Converting into 8-bit weights from flax weights is currently not supported, please make" + " sure the weights are in PyTorch format." + ) + + device_map = kwargs.get("device_map", None) + if ( + device_map is not None + and isinstance(device_map, dict) + and not self.quantization_config.llm_int8_enable_fp32_cpu_offload + ): + device_map_without_no_convert = { + key: device_map[key] for key in device_map.keys() if key not in self.modules_to_not_convert + } + if "cpu" in device_map_without_no_convert.values() or "disk" in device_map_without_no_convert.values(): + raise ValueError( + "Some modules are dispatched on the CPU or the disk. Make sure you have enough GPU RAM to fit the " + "quantized model. If you want to dispatch the model on the CPU or the disk while keeping these modules " + "in 32-bit, you need to set `load_in_8bit_fp32_cpu_offload=True` and pass a custom `device_map` to " + "`from_pretrained`. Check " + "https://huggingface.co/docs/transformers/main/en/main_classes/quantization#offload-between-cpu-and-gpu " + "for more details. " + ) + + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer.adjust_max_memory + def adjust_max_memory(self, max_memory: Dict[str, Union[int, str]]) -> Dict[str, Union[int, str]]: + # need more space for buffers that are created during quantization + max_memory = {key: val * 0.90 for key, val in max_memory.items()} + return max_memory + + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer.update_torch_dtype + def update_torch_dtype(self, torch_dtype: "torch.dtype") -> "torch.dtype": + if torch_dtype is None: + # We force the `dtype` to be float16, this is a requirement from `bitsandbytes` + logger.info( + "Overriding torch_dtype=%s with `torch_dtype=torch.float16` due to " + "requirements of `bitsandbytes` to enable model loading in 8-bit or 4-bit. " + "Pass your own torch_dtype to specify the dtype of the remaining non-linear layers or pass" + " torch_dtype=torch.float16 to remove this warning.", + torch_dtype, + ) + torch_dtype = torch.float16 + return torch_dtype + + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer.update_device_map + def update_device_map(self, device_map): + if device_map is None: + if torch.xpu.is_available(): + current_device = f"xpu:{torch.xpu.current_device()}" + else: + current_device = f"cuda:{torch.cuda.current_device()}" + device_map = {"": current_device} + logger.info( + "The device_map was not initialized. " + "Setting device_map to {" + ": {current_device}}. " + "If you want to use the model for inference, please set device_map ='auto' " + ) + return device_map + + def adjust_target_dtype(self, target_dtype: "torch.dtype") -> "torch.dtype": + if target_dtype != torch.int8: + logger.info("target_dtype {target_dtype} is replaced by `torch.int8` for 8-bit BnB quantization") + return torch.int8 + + def check_if_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + state_dict: Dict[str, Any], + **kwargs, + ): + import bitsandbytes as bnb + + module, tensor_name = get_module_from_name(model, param_name) + if isinstance(module._parameters.get(tensor_name, None), bnb.nn.Int8Params): + if self.pre_quantized: + if param_name.replace("weight", "SCB") not in state_dict.keys(): + raise ValueError("Missing quantization component `SCB`") + if param_value.dtype != torch.int8: + raise ValueError( + f"Incompatible dtype `{param_value.dtype}` when loading 8-bit prequantized weight. Expected `torch.int8`." + ) + return True + return False + + def create_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + target_device: "torch.device", + state_dict: Dict[str, Any], + unexpected_keys: Optional[List[str]] = None, + **kwargs, + ): + import bitsandbytes as bnb + + fp16_statistics_key = param_name.replace("weight", "SCB") + fp16_weights_format_key = param_name.replace("weight", "weight_format") + + fp16_statistics = state_dict.get(fp16_statistics_key, None) + fp16_weights_format = state_dict.get(fp16_weights_format_key, None) + + module, tensor_name = get_module_from_name(model, param_name) + if tensor_name not in module._parameters: + raise ValueError(f"{module} does not have a parameter or a buffer named {tensor_name}.") + + old_value = getattr(module, tensor_name) + + if not isinstance(module._parameters[tensor_name], bnb.nn.Int8Params): + raise ValueError(f"Parameter `{tensor_name}` should only be a `bnb.nn.Int8Params` instance.") + if ( + old_value.device == torch.device("meta") + and target_device not in ["meta", torch.device("meta")] + and param_value is None + ): + raise ValueError(f"{tensor_name} is on the meta device, we need a `value` to put in on {target_device}.") + + new_value = param_value.to("cpu") + if self.pre_quantized and not self.is_serializable: + raise ValueError( + "Detected int8 weights but the version of bitsandbytes is not compatible with int8 serialization. " + "Make sure to download the latest `bitsandbytes` version. `pip install --upgrade bitsandbytes`." + ) + + kwargs = old_value.__dict__ + new_value = bnb.nn.Int8Params(new_value, requires_grad=False, **kwargs).to(target_device) + + module._parameters[tensor_name] = new_value + if fp16_statistics is not None: + setattr(module.weight, "SCB", fp16_statistics.to(target_device)) + if unexpected_keys is not None: + unexpected_keys.remove(fp16_statistics_key) + + # We just need to pop the `weight_format` keys from the state dict to remove unneeded + # messages. The correct format is correctly retrieved during the first forward pass. + if fp16_weights_format is not None and unexpected_keys is not None: + unexpected_keys.remove(fp16_weights_format_key) + + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer._process_model_after_weight_loading with 4bit->8bit + def _process_model_after_weight_loading(self, model: "ModelMixin", **kwargs): + model.is_8bit_serializable = self.is_serializable + return model + + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer._process_model_before_weight_loading with 4bit->8bit + def _process_model_before_weight_loading( + self, + model: "ModelMixin", + device_map, + keep_in_fp32_modules: List[str] = [], + **kwargs, + ): + from .utils import replace_with_bnb_linear + + load_in_8bit_fp32_cpu_offload = self.quantization_config.llm_int8_enable_fp32_cpu_offload + + # We may keep some modules such as the `proj_out` in their original dtype for numerical stability reasons + self.modules_to_not_convert = self.quantization_config.llm_int8_skip_modules + + if not isinstance(self.modules_to_not_convert, list): + self.modules_to_not_convert = [self.modules_to_not_convert] + + self.modules_to_not_convert.extend(keep_in_fp32_modules) + + # Extend `self.modules_to_not_convert` to keys that are supposed to be offloaded to `cpu` or `disk` + if isinstance(device_map, dict) and len(device_map.keys()) > 1: + keys_on_cpu = [key for key, value in device_map.items() if value in ["disk", "cpu"]] + + if len(keys_on_cpu) > 0 and not load_in_8bit_fp32_cpu_offload: + raise ValueError( + "If you want to offload some keys to `cpu` or `disk`, you need to set " + "`llm_int8_enable_fp32_cpu_offload=True`. Note that these modules will not be " + " converted to 8-bit but kept in 32-bit." + ) + self.modules_to_not_convert.extend(keys_on_cpu) + + # Purge `None`. + # Unlike `transformers`, we don't know if we should always keep certain modules in FP32 + # in case of diffusion transformer models. For language models and others alike, `lm_head` + # and tied modules are usually kept in FP32. + self.modules_to_not_convert = [module for module in self.modules_to_not_convert if module is not None] + + model = replace_with_bnb_linear( + model, modules_to_not_convert=self.modules_to_not_convert, quantization_config=self.quantization_config + ) + model.config.quantization_config = self.quantization_config + model.is_loaded_in_8bit = True + + @property + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer.is_serializable + def is_serializable(self): + # Because we're mandating `bitsandbytes` 0.43.3. + return True + + @property + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer.is_serializable + def is_trainable(self) -> bool: + # Because we're mandating `bitsandbytes` 0.43.3. + return True + + @property + def is_compileable(self) -> bool: + return True + + def _dequantize(self, model): + from .utils import dequantize_and_replace + + model = dequantize_and_replace( + model, self.modules_to_not_convert, quantization_config=self.quantization_config + ) + return model diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/utils.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..429aabb8fae6de0a2620cde68fba61db9a0557a7 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/bitsandbytes/utils.py @@ -0,0 +1,319 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# 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. +""" +Adapted from +https://github.com/huggingface/transformers/blob/c409cd81777fb27aadc043ed3d8339dbc020fb3b/src/transformers/integrations/bitsandbytes.py +""" + +import inspect +from inspect import signature +from typing import Union + +from ...utils import is_accelerate_available, is_bitsandbytes_available, is_torch_available, logging +from ..quantization_config import QuantizationMethod + + +if is_torch_available(): + import torch + import torch.nn as nn + +if is_bitsandbytes_available(): + import bitsandbytes as bnb + +if is_accelerate_available(): + import accelerate + from accelerate import init_empty_weights + from accelerate.hooks import add_hook_to_module, remove_hook_from_module + +logger = logging.get_logger(__name__) + + +def _replace_with_bnb_linear( + model, + modules_to_not_convert=None, + current_key_name=None, + quantization_config=None, + has_been_replaced=False, +): + """ + Private method that wraps the recursion for module replacement. + + Returns the converted model and a boolean that indicates if the conversion has been successful or not. + """ + for name, module in model.named_children(): + if current_key_name is None: + current_key_name = [] + current_key_name.append(name) + + if isinstance(module, nn.Linear) and name not in modules_to_not_convert: + # Check if the current key is not in the `modules_to_not_convert` + current_key_name_str = ".".join(current_key_name) + if not any( + (key + "." in current_key_name_str) or (key == current_key_name_str) for key in modules_to_not_convert + ): + with init_empty_weights(): + in_features = module.in_features + out_features = module.out_features + + if quantization_config.quantization_method() == "llm_int8": + model._modules[name] = bnb.nn.Linear8bitLt( + in_features, + out_features, + module.bias is not None, + has_fp16_weights=quantization_config.llm_int8_has_fp16_weight, + threshold=quantization_config.llm_int8_threshold, + ) + has_been_replaced = True + else: + if ( + quantization_config.llm_int8_skip_modules is not None + and name in quantization_config.llm_int8_skip_modules + ): + pass + else: + extra_kwargs = ( + {"quant_storage": quantization_config.bnb_4bit_quant_storage} + if "quant_storage" in list(signature(bnb.nn.Linear4bit).parameters) + else {} + ) + model._modules[name] = bnb.nn.Linear4bit( + in_features, + out_features, + module.bias is not None, + quantization_config.bnb_4bit_compute_dtype, + compress_statistics=quantization_config.bnb_4bit_use_double_quant, + quant_type=quantization_config.bnb_4bit_quant_type, + **extra_kwargs, + ) + has_been_replaced = True + # Store the module class in case we need to transpose the weight later + model._modules[name].source_cls = type(module) + # Force requires grad to False to avoid unexpected errors + model._modules[name].requires_grad_(False) + if len(list(module.children())) > 0: + _, has_been_replaced = _replace_with_bnb_linear( + module, + modules_to_not_convert, + current_key_name, + quantization_config, + has_been_replaced=has_been_replaced, + ) + # Remove the last key for recursion + current_key_name.pop(-1) + return model, has_been_replaced + + +def replace_with_bnb_linear(model, modules_to_not_convert=None, current_key_name=None, quantization_config=None): + """ + Helper function to replace the `nn.Linear` layers within `model` with either `bnb.nn.Linear8bit` or + `bnb.nn.Linear4bit` using the `bitsandbytes` library. + + References: + * `bnb.nn.Linear8bit`: [LLM.int8(): 8-bit Matrix Multiplication for Transformers at + Scale](https://huggingface.co/papers/2208.07339) + * `bnb.nn.Linear4bit`: [QLoRA: Efficient Finetuning of Quantized + LLMs](https://huggingface.co/papers/2305.14314) + + Parameters: + model (`torch.nn.Module`): + Input model or `torch.nn.Module` as the function is run recursively. + modules_to_not_convert (`List[`str`]`, *optional*, defaults to `[]`): + Names of the modules to not convert in `Linear8bitLt`. In practice we keep the `modules_to_not_convert` in + full precision for numerical stability reasons. + current_key_name (`List[`str`]`, *optional*): + An array to track the current key of the recursion. This is used to check whether the current key (part of + it) is not in the list of modules to not convert (for instances modules that are offloaded to `cpu` or + `disk`). + quantization_config ('transformers.utils.quantization_config.BitsAndBytesConfig'): + To configure and manage settings related to quantization, a technique used to compress neural network + models by reducing the precision of the weights and activations, thus making models more efficient in terms + of both storage and computation. + """ + model, _ = _replace_with_bnb_linear(model, modules_to_not_convert, current_key_name, quantization_config) + + has_been_replaced = any( + isinstance(replaced_module, (bnb.nn.Linear4bit, bnb.nn.Linear8bitLt)) + for _, replaced_module in model.named_modules() + ) + if not has_been_replaced: + logger.warning( + "You are loading your model in 8bit or 4bit but no linear modules were found in your model." + " Please double check your model architecture, or submit an issue on github if you think this is" + " a bug." + ) + + return model + + +# Adapted from PEFT: https://github.com/huggingface/peft/blob/6d458b300fc2ed82e19f796b53af4c97d03ea604/src/peft/utils/integrations.py#L81 +def dequantize_bnb_weight(weight: "torch.nn.Parameter", state=None, dtype: "torch.dtype" = None): + """ + Helper function to dequantize 4bit or 8bit bnb weights. + + If the weight is not a bnb quantized weight, it will be returned as is. + """ + if not isinstance(weight, torch.nn.Parameter): + raise TypeError(f"Input weight should be of type nn.Parameter, got {type(weight)} instead") + + cls_name = weight.__class__.__name__ + if cls_name not in ("Params4bit", "Int8Params"): + return weight + + if cls_name == "Params4bit": + output_tensor = bnb.functional.dequantize_4bit(weight.data, weight.quant_state) + msg = f"The model is going to be dequantized in {output_tensor.dtype} - if you want to upcast it to another dtype, make sure to pass the desired dtype when quantizing the model through `bnb_4bit_quant_type` argument of `BitsAndBytesConfig`" + if dtype: + msg = f"The model is going to be first dequantized in {output_tensor.dtype} and type-casted to {dtype}" + output_tensor = output_tensor.to(dtype) + logger.warning_once(msg) + return output_tensor + + if state.SCB is None: + state.SCB = weight.SCB + + if hasattr(bnb.functional, "int8_vectorwise_dequant"): + # Use bitsandbytes API if available (requires v0.45.0+) + dequantized = bnb.functional.int8_vectorwise_dequant(weight.data, state.SCB) + else: + # Multiply by (scale/127) to dequantize. + dequantized = weight.data * state.SCB.view(-1, 1) * 7.874015718698502e-3 + + if dtype: + dequantized = dequantized.to(dtype) + return dequantized + + +def _create_accelerate_new_hook(old_hook): + r""" + Creates a new hook based on the old hook. Use it only if you know what you are doing ! This method is a copy of: + https://github.com/huggingface/peft/blob/748f7968f3a31ec06a1c2b0328993319ad9a150a/src/peft/utils/other.py#L245 with + some changes + """ + old_hook_cls = getattr(accelerate.hooks, old_hook.__class__.__name__) + old_hook_attr = old_hook.__dict__ + filtered_old_hook_attr = {} + old_hook_init_signature = inspect.signature(old_hook_cls.__init__) + for k in old_hook_attr.keys(): + if k in old_hook_init_signature.parameters: + filtered_old_hook_attr[k] = old_hook_attr[k] + new_hook = old_hook_cls(**filtered_old_hook_attr) + return new_hook + + +def _dequantize_and_replace( + model, + dtype, + modules_to_not_convert=None, + current_key_name=None, + quantization_config=None, + has_been_replaced=False, +): + """ + Converts a quantized model into its dequantized original version. The newly converted model will have some + performance drop compared to the original model before quantization - use it only for specific usecases such as + QLoRA adapters merging. + + Returns the converted model and a boolean that indicates if the conversion has been successful or not. + """ + quant_method = quantization_config.quantization_method() + + target_cls = bnb.nn.Linear8bitLt if quant_method == "llm_int8" else bnb.nn.Linear4bit + + for name, module in model.named_children(): + if current_key_name is None: + current_key_name = [] + current_key_name.append(name) + + if isinstance(module, target_cls) and name not in modules_to_not_convert: + # Check if the current key is not in the `modules_to_not_convert` + current_key_name_str = ".".join(current_key_name) + + if not any( + (key + "." in current_key_name_str) or (key == current_key_name_str) for key in modules_to_not_convert + ): + bias = getattr(module, "bias", None) + + device = module.weight.device + with init_empty_weights(): + new_module = torch.nn.Linear(module.in_features, module.out_features, bias=bias is not None) + + if quant_method == "llm_int8": + state = module.state + else: + state = None + + new_module.weight = torch.nn.Parameter(dequantize_bnb_weight(module.weight, state, dtype)) + + if bias is not None: + new_module.bias = bias + + # Create a new hook and attach it in case we use accelerate + if hasattr(module, "_hf_hook"): + old_hook = module._hf_hook + new_hook = _create_accelerate_new_hook(old_hook) + + remove_hook_from_module(module) + add_hook_to_module(new_module, new_hook) + + new_module.to(device) + model._modules[name] = new_module + has_been_replaced = True + if len(list(module.children())) > 0: + _, has_been_replaced = _dequantize_and_replace( + module, + dtype=dtype, + modules_to_not_convert=modules_to_not_convert, + current_key_name=current_key_name, + quantization_config=quantization_config, + has_been_replaced=has_been_replaced, + ) + # Remove the last key for recursion + current_key_name.pop(-1) + return model, has_been_replaced + + +def dequantize_and_replace( + model, + modules_to_not_convert=None, + quantization_config=None, +): + model, _ = _dequantize_and_replace( + model, + dtype=model.dtype, + modules_to_not_convert=modules_to_not_convert, + quantization_config=quantization_config, + ) + has_been_replaced = any( + isinstance(replaced_module, torch.nn.Linear) for _, replaced_module in model.named_modules() + ) + if not has_been_replaced: + logger.warning( + "Some linear modules were not dequantized. This could lead to unexpected behaviour. Please check your model." + ) + + return model + + +def _check_bnb_status(module) -> Union[bool, bool]: + is_loaded_in_4bit_bnb = ( + hasattr(module, "is_loaded_in_4bit") + and module.is_loaded_in_4bit + and getattr(module, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES + ) + is_loaded_in_8bit_bnb = ( + hasattr(module, "is_loaded_in_8bit") + and module.is_loaded_in_8bit + and getattr(module, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES + ) + return is_loaded_in_4bit_bnb or is_loaded_in_8bit_bnb, is_loaded_in_4bit_bnb, is_loaded_in_8bit_bnb diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/__init__.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b3d9082ac803c8da97ba7bb37021445307b96f7c --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/__init__.py @@ -0,0 +1 @@ +from .gguf_quantizer import GGUFQuantizer diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1174f604c19982645ae6c956c9d0ee56cb8fc6b3 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/__pycache__/gguf_quantizer.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/__pycache__/gguf_quantizer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ad6622eb1d932bde851668d74bd0ff5ad9819195 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/__pycache__/gguf_quantizer.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/__pycache__/utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/__pycache__/utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4b206f9348b90b1c4e3b3c468f80383cdd50bd6f Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/__pycache__/utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/gguf_quantizer.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/gguf_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..aa5ebf5711a3731c4768da9ede0206b2edc64ade --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/gguf_quantizer.py @@ -0,0 +1,169 @@ +from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union + +from ..base import DiffusersQuantizer + + +if TYPE_CHECKING: + from ...models.modeling_utils import ModelMixin + + +from ...utils import ( + get_module_from_name, + is_accelerate_available, + is_accelerate_version, + is_gguf_available, + is_gguf_version, + is_torch_available, + logging, +) + + +if is_torch_available() and is_gguf_available(): + import torch + + from .utils import ( + GGML_QUANT_SIZES, + GGUFParameter, + _dequantize_gguf_and_restore_linear, + _quant_shape_from_byte_shape, + _replace_with_gguf_linear, + ) + + +logger = logging.get_logger(__name__) + + +class GGUFQuantizer(DiffusersQuantizer): + use_keep_in_fp32_modules = True + + def __init__(self, quantization_config, **kwargs): + super().__init__(quantization_config, **kwargs) + + self.compute_dtype = quantization_config.compute_dtype + self.pre_quantized = quantization_config.pre_quantized + self.modules_to_not_convert = quantization_config.modules_to_not_convert + + if not isinstance(self.modules_to_not_convert, list): + self.modules_to_not_convert = [self.modules_to_not_convert] + + def validate_environment(self, *args, **kwargs): + if not is_accelerate_available() or is_accelerate_version("<", "0.26.0"): + raise ImportError( + "Loading GGUF Parameters requires `accelerate` installed in your environment: `pip install 'accelerate>=0.26.0'`" + ) + if not is_gguf_available() or is_gguf_version("<", "0.10.0"): + raise ImportError( + "To load GGUF format files you must have `gguf` installed in your environment: `pip install gguf>=0.10.0`" + ) + + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer.adjust_max_memory + def adjust_max_memory(self, max_memory: Dict[str, Union[int, str]]) -> Dict[str, Union[int, str]]: + # need more space for buffers that are created during quantization + max_memory = {key: val * 0.90 for key, val in max_memory.items()} + return max_memory + + def adjust_target_dtype(self, target_dtype: "torch.dtype") -> "torch.dtype": + if target_dtype != torch.uint8: + logger.info(f"target_dtype {target_dtype} is replaced by `torch.uint8` for GGUF quantization") + return torch.uint8 + + def update_torch_dtype(self, torch_dtype: "torch.dtype") -> "torch.dtype": + if torch_dtype is None: + torch_dtype = self.compute_dtype + return torch_dtype + + def check_quantized_param_shape(self, param_name, current_param, loaded_param): + loaded_param_shape = loaded_param.shape + current_param_shape = current_param.shape + quant_type = loaded_param.quant_type + + block_size, type_size = GGML_QUANT_SIZES[quant_type] + + inferred_shape = _quant_shape_from_byte_shape(loaded_param_shape, type_size, block_size) + if inferred_shape != current_param_shape: + raise ValueError( + f"{param_name} has an expected quantized shape of: {inferred_shape}, but received shape: {loaded_param_shape}" + ) + + return True + + def check_if_quantized_param( + self, + model: "ModelMixin", + param_value: Union["GGUFParameter", "torch.Tensor"], + param_name: str, + state_dict: Dict[str, Any], + **kwargs, + ) -> bool: + if isinstance(param_value, GGUFParameter): + return True + + return False + + def create_quantized_param( + self, + model: "ModelMixin", + param_value: Union["GGUFParameter", "torch.Tensor"], + param_name: str, + target_device: "torch.device", + state_dict: Optional[Dict[str, Any]] = None, + unexpected_keys: Optional[List[str]] = None, + **kwargs, + ): + module, tensor_name = get_module_from_name(model, param_name) + if tensor_name not in module._parameters and tensor_name not in module._buffers: + raise ValueError(f"{module} does not have a parameter or a buffer named {tensor_name}.") + + if tensor_name in module._parameters: + module._parameters[tensor_name] = param_value.to(target_device) + if tensor_name in module._buffers: + module._buffers[tensor_name] = param_value.to(target_device) + + def _process_model_before_weight_loading( + self, + model: "ModelMixin", + device_map, + keep_in_fp32_modules: List[str] = [], + **kwargs, + ): + state_dict = kwargs.get("state_dict", None) + + self.modules_to_not_convert.extend(keep_in_fp32_modules) + self.modules_to_not_convert = [module for module in self.modules_to_not_convert if module is not None] + + _replace_with_gguf_linear( + model, self.compute_dtype, state_dict, modules_to_not_convert=self.modules_to_not_convert + ) + + def _process_model_after_weight_loading(self, model: "ModelMixin", **kwargs): + return model + + @property + def is_serializable(self): + return False + + @property + def is_trainable(self) -> bool: + return False + + @property + def is_compileable(self) -> bool: + return True + + def _dequantize(self, model): + is_model_on_cpu = model.device.type == "cpu" + if is_model_on_cpu: + logger.info( + "Model was found to be on CPU (could happen as a result of `enable_model_cpu_offload()`). So, moving it to accelerator. After dequantization, will move the model back to CPU again to preserve the previous device." + ) + device = ( + torch.accelerator.current_accelerator() + if hasattr(torch, "accelerator") + else torch.cuda.current_device() + ) + model.to(device) + + model = _dequantize_gguf_and_restore_linear(model, self.modules_to_not_convert) + if is_model_on_cpu: + model.to("cpu") + return model diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/utils.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..41d351712961a1a8ecd1330a97c9341eb896b639 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/gguf/utils.py @@ -0,0 +1,461 @@ +# Copyright 2025 The HuggingFace Team and City96. All rights reserved. +# # +# # 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 inspect +from contextlib import nullcontext + +import gguf +import torch +import torch.nn as nn + +from ...utils import is_accelerate_available + + +if is_accelerate_available(): + import accelerate + from accelerate import init_empty_weights + from accelerate.hooks import add_hook_to_module, remove_hook_from_module + + +# Copied from diffusers.quantizers.bitsandbytes.utils._create_accelerate_new_hook +def _create_accelerate_new_hook(old_hook): + r""" + Creates a new hook based on the old hook. Use it only if you know what you are doing ! This method is a copy of: + https://github.com/huggingface/peft/blob/748f7968f3a31ec06a1c2b0328993319ad9a150a/src/peft/utils/other.py#L245 with + some changes + """ + old_hook_cls = getattr(accelerate.hooks, old_hook.__class__.__name__) + old_hook_attr = old_hook.__dict__ + filtered_old_hook_attr = {} + old_hook_init_signature = inspect.signature(old_hook_cls.__init__) + for k in old_hook_attr.keys(): + if k in old_hook_init_signature.parameters: + filtered_old_hook_attr[k] = old_hook_attr[k] + new_hook = old_hook_cls(**filtered_old_hook_attr) + return new_hook + + +def _replace_with_gguf_linear(model, compute_dtype, state_dict, prefix="", modules_to_not_convert=[]): + def _should_convert_to_gguf(state_dict, prefix): + weight_key = prefix + "weight" + return weight_key in state_dict and isinstance(state_dict[weight_key], GGUFParameter) + + has_children = list(model.children()) + if not has_children: + return + + for name, module in model.named_children(): + module_prefix = prefix + name + "." + _replace_with_gguf_linear(module, compute_dtype, state_dict, module_prefix, modules_to_not_convert) + + if ( + isinstance(module, nn.Linear) + and _should_convert_to_gguf(state_dict, module_prefix) + and name not in modules_to_not_convert + ): + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + model._modules[name] = GGUFLinear( + module.in_features, + module.out_features, + module.bias is not None, + compute_dtype=compute_dtype, + ) + model._modules[name].source_cls = type(module) + # Force requires_grad to False to avoid unexpected errors + model._modules[name].requires_grad_(False) + + return model + + +def _dequantize_gguf_and_restore_linear(model, modules_to_not_convert=[]): + for name, module in model.named_children(): + if isinstance(module, GGUFLinear) and name not in modules_to_not_convert: + device = module.weight.device + bias = getattr(module, "bias", None) + + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + new_module = nn.Linear( + module.in_features, + module.out_features, + module.bias is not None, + device=device, + ) + new_module.weight = nn.Parameter(dequantize_gguf_tensor(module.weight)) + if bias is not None: + new_module.bias = bias + + # Create a new hook and attach it in case we use accelerate + if hasattr(module, "_hf_hook"): + old_hook = module._hf_hook + new_hook = _create_accelerate_new_hook(old_hook) + + remove_hook_from_module(module) + add_hook_to_module(new_module, new_hook) + + new_module.to(device) + model._modules[name] = new_module + + has_children = list(module.children()) + if has_children: + _dequantize_gguf_and_restore_linear(module, modules_to_not_convert) + + return model + + +# dequantize operations based on torch ports of GGUF dequantize_functions +# from City96 +# more info: https://github.com/city96/ComfyUI-GGUF/blob/main/dequant.py + + +QK_K = 256 +K_SCALE_SIZE = 12 + + +def to_uint32(x): + x = x.view(torch.uint8).to(torch.int32) + return (x[:, 0] | x[:, 1] << 8 | x[:, 2] << 16 | x[:, 3] << 24).unsqueeze(1) + + +def split_block_dims(blocks, *args): + n_max = blocks.shape[1] + dims = list(args) + [n_max - sum(args)] + return torch.split(blocks, dims, dim=1) + + +def get_scale_min(scales): + n_blocks = scales.shape[0] + scales = scales.view(torch.uint8) + scales = scales.reshape((n_blocks, 3, 4)) + + d, m, m_d = torch.split(scales, scales.shape[-2] // 3, dim=-2) + + sc = torch.cat([d & 0x3F, (m_d & 0x0F) | ((d >> 2) & 0x30)], dim=-1) + min = torch.cat([m & 0x3F, (m_d >> 4) | ((m >> 2) & 0x30)], dim=-1) + + return (sc.reshape((n_blocks, 8)), min.reshape((n_blocks, 8))) + + +def dequantize_blocks_Q8_0(blocks, block_size, type_size, dtype=None): + d, x = split_block_dims(blocks, 2) + d = d.view(torch.float16).to(dtype) + x = x.view(torch.int8) + return d * x + + +def dequantize_blocks_Q5_1(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + d, m, qh, qs = split_block_dims(blocks, 2, 2, 4) + d = d.view(torch.float16).to(dtype) + m = m.view(torch.float16).to(dtype) + qh = to_uint32(qh) + + qh = qh.reshape((n_blocks, 1)) >> torch.arange(32, device=d.device, dtype=torch.int32).reshape(1, 32) + ql = qs.reshape((n_blocks, -1, 1, block_size // 2)) >> torch.tensor( + [0, 4], device=d.device, dtype=torch.uint8 + ).reshape(1, 1, 2, 1) + qh = (qh & 1).to(torch.uint8) + ql = (ql & 0x0F).reshape((n_blocks, -1)) + + qs = ql | (qh << 4) + return (d * qs) + m + + +def dequantize_blocks_Q5_0(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + d, qh, qs = split_block_dims(blocks, 2, 4) + d = d.view(torch.float16).to(dtype) + qh = to_uint32(qh) + + qh = qh.reshape(n_blocks, 1) >> torch.arange(32, device=d.device, dtype=torch.int32).reshape(1, 32) + ql = qs.reshape(n_blocks, -1, 1, block_size // 2) >> torch.tensor( + [0, 4], device=d.device, dtype=torch.uint8 + ).reshape(1, 1, 2, 1) + + qh = (qh & 1).to(torch.uint8) + ql = (ql & 0x0F).reshape(n_blocks, -1) + + qs = (ql | (qh << 4)).to(torch.int8) - 16 + return d * qs + + +def dequantize_blocks_Q4_1(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + d, m, qs = split_block_dims(blocks, 2, 2) + d = d.view(torch.float16).to(dtype) + m = m.view(torch.float16).to(dtype) + + qs = qs.reshape((n_blocks, -1, 1, block_size // 2)) >> torch.tensor( + [0, 4], device=d.device, dtype=torch.uint8 + ).reshape(1, 1, 2, 1) + qs = (qs & 0x0F).reshape(n_blocks, -1) + + return (d * qs) + m + + +def dequantize_blocks_Q4_0(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + d, qs = split_block_dims(blocks, 2) + d = d.view(torch.float16).to(dtype) + + qs = qs.reshape((n_blocks, -1, 1, block_size // 2)) >> torch.tensor( + [0, 4], device=d.device, dtype=torch.uint8 + ).reshape((1, 1, 2, 1)) + qs = (qs & 0x0F).reshape((n_blocks, -1)).to(torch.int8) - 8 + return d * qs + + +def dequantize_blocks_Q6_K(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + ( + ql, + qh, + scales, + d, + ) = split_block_dims(blocks, QK_K // 2, QK_K // 4, QK_K // 16) + + scales = scales.view(torch.int8).to(dtype) + d = d.view(torch.float16).to(dtype) + d = (d * scales).reshape((n_blocks, QK_K // 16, 1)) + + ql = ql.reshape((n_blocks, -1, 1, 64)) >> torch.tensor([0, 4], device=d.device, dtype=torch.uint8).reshape( + (1, 1, 2, 1) + ) + ql = (ql & 0x0F).reshape((n_blocks, -1, 32)) + qh = qh.reshape((n_blocks, -1, 1, 32)) >> torch.tensor([0, 2, 4, 6], device=d.device, dtype=torch.uint8).reshape( + (1, 1, 4, 1) + ) + qh = (qh & 0x03).reshape((n_blocks, -1, 32)) + q = (ql | (qh << 4)).to(torch.int8) - 32 + q = q.reshape((n_blocks, QK_K // 16, -1)) + + return (d * q).reshape((n_blocks, QK_K)) + + +def dequantize_blocks_Q5_K(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + d, dmin, scales, qh, qs = split_block_dims(blocks, 2, 2, K_SCALE_SIZE, QK_K // 8) + + d = d.view(torch.float16).to(dtype) + dmin = dmin.view(torch.float16).to(dtype) + + sc, m = get_scale_min(scales) + + d = (d * sc).reshape((n_blocks, -1, 1)) + dm = (dmin * m).reshape((n_blocks, -1, 1)) + + ql = qs.reshape((n_blocks, -1, 1, 32)) >> torch.tensor([0, 4], device=d.device, dtype=torch.uint8).reshape( + (1, 1, 2, 1) + ) + qh = qh.reshape((n_blocks, -1, 1, 32)) >> torch.arange(0, 8, device=d.device, dtype=torch.uint8).reshape( + (1, 1, 8, 1) + ) + ql = (ql & 0x0F).reshape((n_blocks, -1, 32)) + qh = (qh & 0x01).reshape((n_blocks, -1, 32)) + q = ql | (qh << 4) + + return (d * q - dm).reshape((n_blocks, QK_K)) + + +def dequantize_blocks_Q4_K(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + d, dmin, scales, qs = split_block_dims(blocks, 2, 2, K_SCALE_SIZE) + d = d.view(torch.float16).to(dtype) + dmin = dmin.view(torch.float16).to(dtype) + + sc, m = get_scale_min(scales) + + d = (d * sc).reshape((n_blocks, -1, 1)) + dm = (dmin * m).reshape((n_blocks, -1, 1)) + + qs = qs.reshape((n_blocks, -1, 1, 32)) >> torch.tensor([0, 4], device=d.device, dtype=torch.uint8).reshape( + (1, 1, 2, 1) + ) + qs = (qs & 0x0F).reshape((n_blocks, -1, 32)) + + return (d * qs - dm).reshape((n_blocks, QK_K)) + + +def dequantize_blocks_Q3_K(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + hmask, qs, scales, d = split_block_dims(blocks, QK_K // 8, QK_K // 4, 12) + d = d.view(torch.float16).to(dtype) + + lscales, hscales = scales[:, :8], scales[:, 8:] + lscales = lscales.reshape((n_blocks, 1, 8)) >> torch.tensor([0, 4], device=d.device, dtype=torch.uint8).reshape( + (1, 2, 1) + ) + lscales = lscales.reshape((n_blocks, 16)) + hscales = hscales.reshape((n_blocks, 1, 4)) >> torch.tensor( + [0, 2, 4, 6], device=d.device, dtype=torch.uint8 + ).reshape((1, 4, 1)) + hscales = hscales.reshape((n_blocks, 16)) + scales = (lscales & 0x0F) | ((hscales & 0x03) << 4) + scales = scales.to(torch.int8) - 32 + + dl = (d * scales).reshape((n_blocks, 16, 1)) + + ql = qs.reshape((n_blocks, -1, 1, 32)) >> torch.tensor([0, 2, 4, 6], device=d.device, dtype=torch.uint8).reshape( + (1, 1, 4, 1) + ) + qh = hmask.reshape(n_blocks, -1, 1, 32) >> torch.arange(0, 8, device=d.device, dtype=torch.uint8).reshape( + (1, 1, 8, 1) + ) + ql = ql.reshape((n_blocks, 16, QK_K // 16)) & 3 + qh = (qh.reshape((n_blocks, 16, QK_K // 16)) & 1) ^ 1 + q = ql.to(torch.int8) - (qh << 2).to(torch.int8) + + return (dl * q).reshape((n_blocks, QK_K)) + + +def dequantize_blocks_Q2_K(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + scales, qs, d, dmin = split_block_dims(blocks, QK_K // 16, QK_K // 4, 2) + d = d.view(torch.float16).to(dtype) + dmin = dmin.view(torch.float16).to(dtype) + + # (n_blocks, 16, 1) + dl = (d * (scales & 0xF)).reshape((n_blocks, QK_K // 16, 1)) + ml = (dmin * (scales >> 4)).reshape((n_blocks, QK_K // 16, 1)) + + shift = torch.tensor([0, 2, 4, 6], device=d.device, dtype=torch.uint8).reshape((1, 1, 4, 1)) + + qs = (qs.reshape((n_blocks, -1, 1, 32)) >> shift) & 3 + qs = qs.reshape((n_blocks, QK_K // 16, 16)) + qs = dl * qs - ml + + return qs.reshape((n_blocks, -1)) + + +def dequantize_blocks_BF16(blocks, block_size, type_size, dtype=None): + return (blocks.view(torch.int16).to(torch.int32) << 16).view(torch.float32) + + +GGML_QUANT_SIZES = gguf.GGML_QUANT_SIZES +dequantize_functions = { + gguf.GGMLQuantizationType.BF16: dequantize_blocks_BF16, + gguf.GGMLQuantizationType.Q8_0: dequantize_blocks_Q8_0, + gguf.GGMLQuantizationType.Q5_1: dequantize_blocks_Q5_1, + gguf.GGMLQuantizationType.Q5_0: dequantize_blocks_Q5_0, + gguf.GGMLQuantizationType.Q4_1: dequantize_blocks_Q4_1, + gguf.GGMLQuantizationType.Q4_0: dequantize_blocks_Q4_0, + gguf.GGMLQuantizationType.Q6_K: dequantize_blocks_Q6_K, + gguf.GGMLQuantizationType.Q5_K: dequantize_blocks_Q5_K, + gguf.GGMLQuantizationType.Q4_K: dequantize_blocks_Q4_K, + gguf.GGMLQuantizationType.Q3_K: dequantize_blocks_Q3_K, + gguf.GGMLQuantizationType.Q2_K: dequantize_blocks_Q2_K, +} +SUPPORTED_GGUF_QUANT_TYPES = list(dequantize_functions.keys()) + + +def _quant_shape_from_byte_shape(shape, type_size, block_size): + return (*shape[:-1], shape[-1] // type_size * block_size) + + +def dequantize_gguf_tensor(tensor): + if not hasattr(tensor, "quant_type"): + return tensor + + quant_type = tensor.quant_type + dequant_fn = dequantize_functions[quant_type] + + block_size, type_size = GGML_QUANT_SIZES[quant_type] + + tensor = tensor.view(torch.uint8) + shape = _quant_shape_from_byte_shape(tensor.shape, type_size, block_size) + + n_blocks = tensor.numel() // type_size + blocks = tensor.reshape((n_blocks, type_size)) + + dequant = dequant_fn(blocks, block_size, type_size) + dequant = dequant.reshape(shape) + + return dequant.as_tensor() + + +class GGUFParameter(torch.nn.Parameter): + def __new__(cls, data, requires_grad=False, quant_type=None): + data = data if data is not None else torch.empty(0) + self = torch.Tensor._make_subclass(cls, data, requires_grad) + self.quant_type = quant_type + block_size, type_size = GGML_QUANT_SIZES[quant_type] + self.quant_shape = _quant_shape_from_byte_shape(self.shape, type_size, block_size) + + return self + + def as_tensor(self): + return torch.Tensor._make_subclass(torch.Tensor, self, self.requires_grad) + + @staticmethod + def _extract_quant_type(args): + # When converting from original format checkpoints we often use splits, cats etc on tensors + # this method ensures that the returned tensor type from those operations remains GGUFParameter + # so that we preserve quant_type information + for arg in args: + if isinstance(arg, list) and isinstance(arg[0], GGUFParameter): + return arg[0].quant_type + if isinstance(arg, GGUFParameter): + return arg.quant_type + return None + + @classmethod + def __torch_function__(cls, func, types, args=(), kwargs=None): + if kwargs is None: + kwargs = {} + + result = super().__torch_function__(func, types, args, kwargs) + + if isinstance(result, torch.Tensor): + quant_type = cls._extract_quant_type(args) + return cls(result, quant_type=quant_type) + # Handle tuples and lists + elif type(result) in (list, tuple): + # Preserve the original type (tuple or list) + quant_type = cls._extract_quant_type(args) + wrapped = [cls(x, quant_type=quant_type) if isinstance(x, torch.Tensor) else x for x in result] + return type(result)(wrapped) + else: + return result + + +class GGUFLinear(nn.Linear): + def __init__( + self, + in_features, + out_features, + bias=False, + compute_dtype=None, + device=None, + ) -> None: + super().__init__(in_features, out_features, bias, device) + self.compute_dtype = compute_dtype + + def forward(self, inputs): + weight = dequantize_gguf_tensor(self.weight) + weight = weight.to(self.compute_dtype) + bias = self.bias.to(self.compute_dtype) if self.bias is not None else None + + output = torch.nn.functional.linear(inputs, weight, bias) + return output diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/quantization_config.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/quantization_config.py new file mode 100644 index 0000000000000000000000000000000000000000..871faf076e5aa361efd45656c31dbcd96af9a316 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/quantization_config.py @@ -0,0 +1,726 @@ +#!/usr/bin/env python +# coding=utf-8 + +# Copyright 2023 The HuggingFace Inc. team. All rights reserved. +# +# 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. + +""" +Adapted from +https://github.com/huggingface/transformers/blob/52cb4034ada381fe1ffe8d428a1076e5411a8026/src/transformers/utils/quantization_config.py +""" + +import copy +import importlib.metadata +import inspect +import json +import os +from dataclasses import dataclass +from enum import Enum +from functools import partial +from typing import Any, Dict, List, Optional, Union + +from packaging import version + +from ..utils import is_torch_available, is_torchao_available, logging + + +if is_torch_available(): + import torch + +logger = logging.get_logger(__name__) + + +class QuantizationMethod(str, Enum): + BITS_AND_BYTES = "bitsandbytes" + GGUF = "gguf" + TORCHAO = "torchao" + QUANTO = "quanto" + + +if is_torchao_available(): + from torchao.quantization.quant_primitives import MappingType + + class TorchAoJSONEncoder(json.JSONEncoder): + def default(self, obj): + if isinstance(obj, MappingType): + return obj.name + return super().default(obj) + + +@dataclass +class QuantizationConfigMixin: + """ + Mixin class for quantization config + """ + + quant_method: QuantizationMethod + _exclude_attributes_at_init = [] + + @classmethod + def from_dict(cls, config_dict, return_unused_kwargs=False, **kwargs): + """ + Instantiates a [`QuantizationConfigMixin`] from a Python dictionary of parameters. + + Args: + config_dict (`Dict[str, Any]`): + Dictionary that will be used to instantiate the configuration object. + return_unused_kwargs (`bool`, *optional*, defaults to `False`): + Whether or not to return a list of unused keyword arguments. Used for `from_pretrained` method in + `PreTrainedModel`. + kwargs (`Dict[str, Any]`): + Additional parameters from which to initialize the configuration object. + + Returns: + [`QuantizationConfigMixin`]: The configuration object instantiated from those parameters. + """ + + config = cls(**config_dict) + + to_remove = [] + for key, value in kwargs.items(): + if hasattr(config, key): + setattr(config, key, value) + to_remove.append(key) + for key in to_remove: + kwargs.pop(key, None) + + if return_unused_kwargs: + return config, kwargs + else: + return config + + def to_json_file(self, json_file_path: Union[str, os.PathLike]): + """ + Save this instance to a JSON file. + + Args: + json_file_path (`str` or `os.PathLike`): + Path to the JSON file in which this configuration instance's parameters will be saved. + use_diff (`bool`, *optional*, defaults to `True`): + If set to `True`, only the difference between the config instance and the default + `QuantizationConfig()` is serialized to JSON file. + """ + with open(json_file_path, "w", encoding="utf-8") as writer: + config_dict = self.to_dict() + json_string = json.dumps(config_dict, indent=2, sort_keys=True) + "\n" + + writer.write(json_string) + + def to_dict(self) -> Dict[str, Any]: + """ + Serializes this instance to a Python dictionary. Returns: + `Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance. + """ + return copy.deepcopy(self.__dict__) + + def __iter__(self): + """allows `dict(obj)` for situations where obj may be a dict or QuantizationConfigMixin""" + for attr, value in copy.deepcopy(self.__dict__).items(): + yield attr, value + + def __repr__(self): + return f"{self.__class__.__name__} {self.to_json_string()}" + + def to_json_string(self, use_diff: bool = True) -> str: + """ + Serializes this instance to a JSON string. + + Args: + use_diff (`bool`, *optional*, defaults to `True`): + If set to `True`, only the difference between the config instance and the default `PretrainedConfig()` + is serialized to JSON string. + + Returns: + `str`: String containing all the attributes that make up this configuration instance in JSON format. + """ + if use_diff is True: + config_dict = self.to_diff_dict() + else: + config_dict = self.to_dict() + return json.dumps(config_dict, indent=2, sort_keys=True) + "\n" + + def update(self, **kwargs): + """ + Updates attributes of this class instance with attributes from `kwargs` if they match existing attributes, + returning all the unused kwargs. + + Args: + kwargs (`Dict[str, Any]`): + Dictionary of attributes to tentatively update this class. + + Returns: + `Dict[str, Any]`: Dictionary containing all the key-value pairs that were not used to update the instance. + """ + to_remove = [] + for key, value in kwargs.items(): + if hasattr(self, key): + setattr(self, key, value) + to_remove.append(key) + + # Remove all the attributes that were updated, without modifying the input dict + unused_kwargs = {key: value for key, value in kwargs.items() if key not in to_remove} + return unused_kwargs + + +@dataclass +class BitsAndBytesConfig(QuantizationConfigMixin): + """ + This is a wrapper class about all possible attributes and features that you can play with a model that has been + loaded using `bitsandbytes`. + + This replaces `load_in_8bit` or `load_in_4bit` therefore both options are mutually exclusive. + + Currently only supports `LLM.int8()`, `FP4`, and `NF4` quantization. If more methods are added to `bitsandbytes`, + then more arguments will be added to this class. + + Args: + load_in_8bit (`bool`, *optional*, defaults to `False`): + This flag is used to enable 8-bit quantization with LLM.int8(). + load_in_4bit (`bool`, *optional*, defaults to `False`): + This flag is used to enable 4-bit quantization by replacing the Linear layers with FP4/NF4 layers from + `bitsandbytes`. + llm_int8_threshold (`float`, *optional*, defaults to 6.0): + This corresponds to the outlier threshold for outlier detection as described in `LLM.int8() : 8-bit Matrix + Multiplication for Transformers at Scale` paper: https://huggingface.co/papers/2208.07339 Any hidden states + value that is above this threshold will be considered an outlier and the operation on those values will be + done in fp16. Values are usually normally distributed, that is, most values are in the range [-3.5, 3.5], + but there are some exceptional systematic outliers that are very differently distributed for large models. + These outliers are often in the interval [-60, -6] or [6, 60]. Int8 quantization works well for values of + magnitude ~5, but beyond that, there is a significant performance penalty. A good default threshold is 6, + but a lower threshold might be needed for more unstable models (small models, fine-tuning). + llm_int8_skip_modules (`List[str]`, *optional*): + An explicit list of the modules that we do not want to convert in 8-bit. This is useful for models such as + Jukebox that has several heads in different places and not necessarily at the last position. For example + for `CausalLM` models, the last `lm_head` is typically kept in its original `dtype`. + llm_int8_enable_fp32_cpu_offload (`bool`, *optional*, defaults to `False`): + This flag is used for advanced use cases and users that are aware of this feature. If you want to split + your model in different parts and run some parts in int8 on GPU and some parts in fp32 on CPU, you can use + this flag. This is useful for offloading large models such as `google/flan-t5-xxl`. Note that the int8 + operations will not be run on CPU. + llm_int8_has_fp16_weight (`bool`, *optional*, defaults to `False`): + This flag runs LLM.int8() with 16-bit main weights. This is useful for fine-tuning as the weights do not + have to be converted back and forth for the backward pass. + bnb_4bit_compute_dtype (`torch.dtype` or str, *optional*, defaults to `torch.float32`): + This sets the computational type which might be different than the input type. For example, inputs might be + fp32, but computation can be set to bf16 for speedups. + bnb_4bit_quant_type (`str`, *optional*, defaults to `"fp4"`): + This sets the quantization data type in the bnb.nn.Linear4Bit layers. Options are FP4 and NF4 data types + which are specified by `fp4` or `nf4`. + bnb_4bit_use_double_quant (`bool`, *optional*, defaults to `False`): + This flag is used for nested quantization where the quantization constants from the first quantization are + quantized again. + bnb_4bit_quant_storage (`torch.dtype` or str, *optional*, defaults to `torch.uint8`): + This sets the storage type to pack the quanitzed 4-bit prarams. + kwargs (`Dict[str, Any]`, *optional*): + Additional parameters from which to initialize the configuration object. + """ + + _exclude_attributes_at_init = ["_load_in_4bit", "_load_in_8bit", "quant_method"] + + def __init__( + self, + load_in_8bit=False, + load_in_4bit=False, + llm_int8_threshold=6.0, + llm_int8_skip_modules=None, + llm_int8_enable_fp32_cpu_offload=False, + llm_int8_has_fp16_weight=False, + bnb_4bit_compute_dtype=None, + bnb_4bit_quant_type="fp4", + bnb_4bit_use_double_quant=False, + bnb_4bit_quant_storage=None, + **kwargs, + ): + self.quant_method = QuantizationMethod.BITS_AND_BYTES + + if load_in_4bit and load_in_8bit: + raise ValueError("load_in_4bit and load_in_8bit are both True, but only one can be used at the same time") + + self._load_in_8bit = load_in_8bit + self._load_in_4bit = load_in_4bit + self.llm_int8_threshold = llm_int8_threshold + self.llm_int8_skip_modules = llm_int8_skip_modules + self.llm_int8_enable_fp32_cpu_offload = llm_int8_enable_fp32_cpu_offload + self.llm_int8_has_fp16_weight = llm_int8_has_fp16_weight + self.bnb_4bit_quant_type = bnb_4bit_quant_type + self.bnb_4bit_use_double_quant = bnb_4bit_use_double_quant + + if bnb_4bit_compute_dtype is None: + self.bnb_4bit_compute_dtype = torch.float32 + elif isinstance(bnb_4bit_compute_dtype, str): + self.bnb_4bit_compute_dtype = getattr(torch, bnb_4bit_compute_dtype) + elif isinstance(bnb_4bit_compute_dtype, torch.dtype): + self.bnb_4bit_compute_dtype = bnb_4bit_compute_dtype + else: + raise ValueError("bnb_4bit_compute_dtype must be a string or a torch.dtype") + + if bnb_4bit_quant_storage is None: + self.bnb_4bit_quant_storage = torch.uint8 + elif isinstance(bnb_4bit_quant_storage, str): + if bnb_4bit_quant_storage not in ["float16", "float32", "int8", "uint8", "float64", "bfloat16"]: + raise ValueError( + "`bnb_4bit_quant_storage` must be a valid string (one of 'float16', 'float32', 'int8', 'uint8', 'float64', 'bfloat16') " + ) + self.bnb_4bit_quant_storage = getattr(torch, bnb_4bit_quant_storage) + elif isinstance(bnb_4bit_quant_storage, torch.dtype): + self.bnb_4bit_quant_storage = bnb_4bit_quant_storage + else: + raise ValueError("bnb_4bit_quant_storage must be a string or a torch.dtype") + + if kwargs and not all(k in self._exclude_attributes_at_init for k in kwargs): + logger.warning(f"Unused kwargs: {list(kwargs.keys())}. These kwargs are not used in {self.__class__}.") + + self.post_init() + + @property + def load_in_4bit(self): + return self._load_in_4bit + + @load_in_4bit.setter + def load_in_4bit(self, value: bool): + if not isinstance(value, bool): + raise TypeError("load_in_4bit must be a boolean") + + if self.load_in_8bit and value: + raise ValueError("load_in_4bit and load_in_8bit are both True, but only one can be used at the same time") + self._load_in_4bit = value + + @property + def load_in_8bit(self): + return self._load_in_8bit + + @load_in_8bit.setter + def load_in_8bit(self, value: bool): + if not isinstance(value, bool): + raise TypeError("load_in_8bit must be a boolean") + + if self.load_in_4bit and value: + raise ValueError("load_in_4bit and load_in_8bit are both True, but only one can be used at the same time") + self._load_in_8bit = value + + def post_init(self): + r""" + Safety checker that arguments are correct - also replaces some NoneType arguments with their default values. + """ + if not isinstance(self.load_in_4bit, bool): + raise TypeError("load_in_4bit must be a boolean") + + if not isinstance(self.load_in_8bit, bool): + raise TypeError("load_in_8bit must be a boolean") + + if not isinstance(self.llm_int8_threshold, float): + raise TypeError("llm_int8_threshold must be a float") + + if self.llm_int8_skip_modules is not None and not isinstance(self.llm_int8_skip_modules, list): + raise TypeError("llm_int8_skip_modules must be a list of strings") + if not isinstance(self.llm_int8_enable_fp32_cpu_offload, bool): + raise TypeError("llm_int8_enable_fp32_cpu_offload must be a boolean") + + if not isinstance(self.llm_int8_has_fp16_weight, bool): + raise TypeError("llm_int8_has_fp16_weight must be a boolean") + + if self.bnb_4bit_compute_dtype is not None and not isinstance(self.bnb_4bit_compute_dtype, torch.dtype): + raise TypeError("bnb_4bit_compute_dtype must be torch.dtype") + + if not isinstance(self.bnb_4bit_quant_type, str): + raise TypeError("bnb_4bit_quant_type must be a string") + + if not isinstance(self.bnb_4bit_use_double_quant, bool): + raise TypeError("bnb_4bit_use_double_quant must be a boolean") + + if self.load_in_4bit and not version.parse(importlib.metadata.version("bitsandbytes")) >= version.parse( + "0.39.0" + ): + raise ValueError( + "4 bit quantization requires bitsandbytes>=0.39.0 - please upgrade your bitsandbytes version" + ) + + def is_quantizable(self): + r""" + Returns `True` if the model is quantizable, `False` otherwise. + """ + return self.load_in_8bit or self.load_in_4bit + + def quantization_method(self): + r""" + This method returns the quantization method used for the model. If the model is not quantizable, it returns + `None`. + """ + if self.load_in_8bit: + return "llm_int8" + elif self.load_in_4bit and self.bnb_4bit_quant_type == "fp4": + return "fp4" + elif self.load_in_4bit and self.bnb_4bit_quant_type == "nf4": + return "nf4" + else: + return None + + def to_dict(self) -> Dict[str, Any]: + """ + Serializes this instance to a Python dictionary. Returns: + `Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance. + """ + output = copy.deepcopy(self.__dict__) + output["bnb_4bit_compute_dtype"] = str(output["bnb_4bit_compute_dtype"]).split(".")[1] + output["bnb_4bit_quant_storage"] = str(output["bnb_4bit_quant_storage"]).split(".")[1] + output["load_in_4bit"] = self.load_in_4bit + output["load_in_8bit"] = self.load_in_8bit + + return output + + def __repr__(self): + config_dict = self.to_dict() + return f"{self.__class__.__name__} {json.dumps(config_dict, indent=2, sort_keys=True)}\n" + + def to_diff_dict(self) -> Dict[str, Any]: + """ + Removes all attributes from config which correspond to the default config attributes for better readability and + serializes to a Python dictionary. + + Returns: + `Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance, + """ + config_dict = self.to_dict() + + # get the default config dict + default_config_dict = BitsAndBytesConfig().to_dict() + + serializable_config_dict = {} + + # only serialize values that differ from the default config + for key, value in config_dict.items(): + if value != default_config_dict[key]: + serializable_config_dict[key] = value + + return serializable_config_dict + + +@dataclass +class GGUFQuantizationConfig(QuantizationConfigMixin): + """This is a config class for GGUF Quantization techniques. + + Args: + compute_dtype: (`torch.dtype`, defaults to `torch.float32`): + This sets the computational type which might be different than the input type. For example, inputs might be + fp32, but computation can be set to bf16 for speedups. + + """ + + def __init__(self, compute_dtype: Optional["torch.dtype"] = None): + self.quant_method = QuantizationMethod.GGUF + self.compute_dtype = compute_dtype + self.pre_quantized = True + + # TODO: (Dhruv) Add this as an init argument when we can support loading unquantized checkpoints. + self.modules_to_not_convert = None + + if self.compute_dtype is None: + self.compute_dtype = torch.float32 + + +@dataclass +class TorchAoConfig(QuantizationConfigMixin): + """This is a config class for torchao quantization/sparsity techniques. + + Args: + quant_type (`str`): + The type of quantization we want to use, currently supporting: + - **Integer quantization:** + - Full function names: `int4_weight_only`, `int8_dynamic_activation_int4_weight`, + `int8_weight_only`, `int8_dynamic_activation_int8_weight` + - Shorthands: `int4wo`, `int4dq`, `int8wo`, `int8dq` + + - **Floating point 8-bit quantization:** + - Full function names: `float8_weight_only`, `float8_dynamic_activation_float8_weight`, + `float8_static_activation_float8_weight` + - Shorthands: `float8wo`, `float8wo_e5m2`, `float8wo_e4m3`, `float8dq`, `float8dq_e4m3`, + `float8_e4m3_tensor`, `float8_e4m3_row`, + + - **Floating point X-bit quantization:** + - Full function names: `fpx_weight_only` + - Shorthands: `fpX_eAwB`, where `X` is the number of bits (between `1` to `7`), `A` is the number + of exponent bits and `B` is the number of mantissa bits. The constraint of `X == A + B + 1` must + be satisfied for a given shorthand notation. + + - **Unsigned Integer quantization:** + - Full function names: `uintx_weight_only` + - Shorthands: `uint1wo`, `uint2wo`, `uint3wo`, `uint4wo`, `uint5wo`, `uint6wo`, `uint7wo` + modules_to_not_convert (`List[str]`, *optional*, default to `None`): + The list of modules to not quantize, useful for quantizing models that explicitly require to have some + modules left in their original precision. + kwargs (`Dict[str, Any]`, *optional*): + The keyword arguments for the chosen type of quantization, for example, int4_weight_only quantization + supports two keyword arguments `group_size` and `inner_k_tiles` currently. More API examples and + documentation of arguments can be found in + https://github.com/pytorch/ao/tree/main/torchao/quantization#other-available-quantization-techniques + + Example: + ```python + from diffusers import FluxTransformer2DModel, TorchAoConfig + + quantization_config = TorchAoConfig("int8wo") + transformer = FluxTransformer2DModel.from_pretrained( + "black-forest-labs/Flux.1-Dev", + subfolder="transformer", + quantization_config=quantization_config, + torch_dtype=torch.bfloat16, + ) + ``` + """ + + def __init__(self, quant_type: str, modules_to_not_convert: Optional[List[str]] = None, **kwargs) -> None: + self.quant_method = QuantizationMethod.TORCHAO + self.quant_type = quant_type + self.modules_to_not_convert = modules_to_not_convert + + # When we load from serialized config, "quant_type_kwargs" will be the key + if "quant_type_kwargs" in kwargs: + self.quant_type_kwargs = kwargs["quant_type_kwargs"] + else: + self.quant_type_kwargs = kwargs + + TORCHAO_QUANT_TYPE_METHODS = self._get_torchao_quant_type_to_method() + if self.quant_type not in TORCHAO_QUANT_TYPE_METHODS.keys(): + is_floating_quant_type = self.quant_type.startswith("float") or self.quant_type.startswith("fp") + if is_floating_quant_type and not self._is_xpu_or_cuda_capability_atleast_8_9(): + raise ValueError( + f"Requested quantization type: {self.quant_type} is not supported on GPUs with CUDA capability <= 8.9. You " + f"can check the CUDA capability of your GPU using `torch.cuda.get_device_capability()`." + ) + + raise ValueError( + f"Requested quantization type: {self.quant_type} is not supported or is an incorrect `quant_type` name. If you think the " + f"provided quantization type should be supported, please open an issue at https://github.com/huggingface/diffusers/issues." + ) + + method = TORCHAO_QUANT_TYPE_METHODS[self.quant_type] + signature = inspect.signature(method) + all_kwargs = { + param.name + for param in signature.parameters.values() + if param.kind in [inspect.Parameter.KEYWORD_ONLY, inspect.Parameter.POSITIONAL_OR_KEYWORD] + } + unsupported_kwargs = list(self.quant_type_kwargs.keys() - all_kwargs) + + if len(unsupported_kwargs) > 0: + raise ValueError( + f'The quantization method "{quant_type}" does not support the following keyword arguments: ' + f"{unsupported_kwargs}. The following keywords arguments are supported: {all_kwargs}." + ) + + @classmethod + def _get_torchao_quant_type_to_method(cls): + r""" + Returns supported torchao quantization types with all commonly used notations. + """ + + if is_torchao_available(): + # TODO(aryan): Support autoquant and sparsify + from torchao.quantization import ( + float8_dynamic_activation_float8_weight, + float8_static_activation_float8_weight, + float8_weight_only, + fpx_weight_only, + int4_weight_only, + int8_dynamic_activation_int4_weight, + int8_dynamic_activation_int8_weight, + int8_weight_only, + uintx_weight_only, + ) + + # TODO(aryan): Add a note on how to use PerAxis and PerGroup observers + from torchao.quantization.observer import PerRow, PerTensor + + def generate_float8dq_types(dtype: torch.dtype): + name = "e5m2" if dtype == torch.float8_e5m2 else "e4m3" + types = {} + + for granularity_cls in [PerTensor, PerRow]: + # Note: Activation and Weights cannot have different granularities + granularity_name = "tensor" if granularity_cls is PerTensor else "row" + types[f"float8dq_{name}_{granularity_name}"] = partial( + float8_dynamic_activation_float8_weight, + activation_dtype=dtype, + weight_dtype=dtype, + granularity=(granularity_cls(), granularity_cls()), + ) + + return types + + def generate_fpx_quantization_types(bits: int): + types = {} + + for ebits in range(1, bits): + mbits = bits - ebits - 1 + types[f"fp{bits}_e{ebits}m{mbits}"] = partial(fpx_weight_only, ebits=ebits, mbits=mbits) + + non_sign_bits = bits - 1 + default_ebits = (non_sign_bits + 1) // 2 + default_mbits = non_sign_bits - default_ebits + types[f"fp{bits}"] = partial(fpx_weight_only, ebits=default_ebits, mbits=default_mbits) + + return types + + INT4_QUANTIZATION_TYPES = { + # int4 weight + bfloat16/float16 activation + "int4wo": int4_weight_only, + "int4_weight_only": int4_weight_only, + # int4 weight + int8 activation + "int4dq": int8_dynamic_activation_int4_weight, + "int8_dynamic_activation_int4_weight": int8_dynamic_activation_int4_weight, + } + + INT8_QUANTIZATION_TYPES = { + # int8 weight + bfloat16/float16 activation + "int8wo": int8_weight_only, + "int8_weight_only": int8_weight_only, + # int8 weight + int8 activation + "int8dq": int8_dynamic_activation_int8_weight, + "int8_dynamic_activation_int8_weight": int8_dynamic_activation_int8_weight, + } + + # TODO(aryan): handle torch 2.2/2.3 + FLOATX_QUANTIZATION_TYPES = { + # float8_e5m2 weight + bfloat16/float16 activation + "float8wo": partial(float8_weight_only, weight_dtype=torch.float8_e5m2), + "float8_weight_only": float8_weight_only, + "float8wo_e5m2": partial(float8_weight_only, weight_dtype=torch.float8_e5m2), + # float8_e4m3 weight + bfloat16/float16 activation + "float8wo_e4m3": partial(float8_weight_only, weight_dtype=torch.float8_e4m3fn), + # float8_e5m2 weight + float8 activation (dynamic) + "float8dq": float8_dynamic_activation_float8_weight, + "float8_dynamic_activation_float8_weight": float8_dynamic_activation_float8_weight, + # ===== Matrix multiplication is not supported in float8_e5m2 so the following errors out. + # However, changing activation_dtype=torch.float8_e4m3 might work here ===== + # "float8dq_e5m2": partial( + # float8_dynamic_activation_float8_weight, + # activation_dtype=torch.float8_e5m2, + # weight_dtype=torch.float8_e5m2, + # ), + # **generate_float8dq_types(torch.float8_e5m2), + # ===== ===== + # float8_e4m3 weight + float8 activation (dynamic) + "float8dq_e4m3": partial( + float8_dynamic_activation_float8_weight, + activation_dtype=torch.float8_e4m3fn, + weight_dtype=torch.float8_e4m3fn, + ), + **generate_float8dq_types(torch.float8_e4m3fn), + # float8 weight + float8 activation (static) + "float8_static_activation_float8_weight": float8_static_activation_float8_weight, + # For fpx, only x <= 8 is supported by default. Other dtypes can be explored by users directly + # fpx weight + bfloat16/float16 activation + **generate_fpx_quantization_types(3), + **generate_fpx_quantization_types(4), + **generate_fpx_quantization_types(5), + **generate_fpx_quantization_types(6), + **generate_fpx_quantization_types(7), + } + + UINTX_QUANTIZATION_DTYPES = { + "uintx_weight_only": uintx_weight_only, + "uint1wo": partial(uintx_weight_only, dtype=torch.uint1), + "uint2wo": partial(uintx_weight_only, dtype=torch.uint2), + "uint3wo": partial(uintx_weight_only, dtype=torch.uint3), + "uint4wo": partial(uintx_weight_only, dtype=torch.uint4), + "uint5wo": partial(uintx_weight_only, dtype=torch.uint5), + "uint6wo": partial(uintx_weight_only, dtype=torch.uint6), + "uint7wo": partial(uintx_weight_only, dtype=torch.uint7), + # "uint8wo": partial(uintx_weight_only, dtype=torch.uint8), # uint8 quantization is not supported + } + + QUANTIZATION_TYPES = {} + QUANTIZATION_TYPES.update(INT4_QUANTIZATION_TYPES) + QUANTIZATION_TYPES.update(INT8_QUANTIZATION_TYPES) + QUANTIZATION_TYPES.update(UINTX_QUANTIZATION_DTYPES) + + if cls._is_xpu_or_cuda_capability_atleast_8_9(): + QUANTIZATION_TYPES.update(FLOATX_QUANTIZATION_TYPES) + + return QUANTIZATION_TYPES + else: + raise ValueError( + "TorchAoConfig requires torchao to be installed, please install with `pip install torchao`" + ) + + @staticmethod + def _is_xpu_or_cuda_capability_atleast_8_9() -> bool: + if torch.cuda.is_available(): + major, minor = torch.cuda.get_device_capability() + if major == 8: + return minor >= 9 + return major >= 9 + elif torch.xpu.is_available(): + return True + else: + raise RuntimeError("TorchAO requires a CUDA compatible GPU or Intel XPU and installation of PyTorch.") + + def get_apply_tensor_subclass(self): + TORCHAO_QUANT_TYPE_METHODS = self._get_torchao_quant_type_to_method() + return TORCHAO_QUANT_TYPE_METHODS[self.quant_type](**self.quant_type_kwargs) + + def __repr__(self): + r""" + Example of how this looks for `TorchAoConfig("uint4wo", group_size=32)`: + + ``` + TorchAoConfig { + "modules_to_not_convert": null, + "quant_method": "torchao", + "quant_type": "uint4wo", + "quant_type_kwargs": { + "group_size": 32 + } + } + ``` + """ + config_dict = self.to_dict() + return ( + f"{self.__class__.__name__} {json.dumps(config_dict, indent=2, sort_keys=True, cls=TorchAoJSONEncoder)}\n" + ) + + +@dataclass +class QuantoConfig(QuantizationConfigMixin): + """ + This is a wrapper class about all possible attributes and features that you can play with a model that has been + loaded using `quanto`. + + Args: + weights_dtype (`str`, *optional*, defaults to `"int8"`): + The target dtype for the weights after quantization. Supported values are ("float8","int8","int4","int2") + modules_to_not_convert (`list`, *optional*, default to `None`): + The list of modules to not quantize, useful for quantizing models that explicitly require to have some + modules left in their original precision (e.g. Whisper encoder, Llava encoder, Mixtral gate layers). + """ + + def __init__( + self, + weights_dtype: str = "int8", + modules_to_not_convert: Optional[List[str]] = None, + **kwargs, + ): + self.quant_method = QuantizationMethod.QUANTO + self.weights_dtype = weights_dtype + self.modules_to_not_convert = modules_to_not_convert + + self.post_init() + + def post_init(self): + r""" + Safety checker that arguments are correct + """ + accepted_weights = ["float8", "int8", "int4", "int2"] + if self.weights_dtype not in accepted_weights: + raise ValueError(f"Only support weights in {accepted_weights} but found {self.weights_dtype}") diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/__init__.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a4e8a1f41a1e4bf3bbb53c5684398277eb671a83 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/__init__.py @@ -0,0 +1 @@ +from .quanto_quantizer import QuantoQuantizer diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a78e8241791383429220fa1fa4c88cf6c6f33128 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/__pycache__/quanto_quantizer.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/__pycache__/quanto_quantizer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5e0acaa6027b9df9c0899f36b5b4ae0436a84e59 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/__pycache__/quanto_quantizer.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/__pycache__/utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/__pycache__/utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a919ef2d5b695809dded5bf58862da1ca99abd67 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/__pycache__/utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/quanto_quantizer.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/quanto_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..c5f71f816fc343e39f38cf69f817322312c182d2 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/quanto_quantizer.py @@ -0,0 +1,181 @@ +from typing import TYPE_CHECKING, Any, Dict, List, Union + +from diffusers.utils.import_utils import is_optimum_quanto_version + +from ...utils import ( + get_module_from_name, + is_accelerate_available, + is_accelerate_version, + is_optimum_quanto_available, + is_torch_available, + logging, +) +from ..base import DiffusersQuantizer + + +if TYPE_CHECKING: + from ...models.modeling_utils import ModelMixin + + +if is_torch_available(): + import torch + +if is_accelerate_available(): + from accelerate.utils import CustomDtype, set_module_tensor_to_device + +if is_optimum_quanto_available(): + from .utils import _replace_with_quanto_layers + +logger = logging.get_logger(__name__) + + +class QuantoQuantizer(DiffusersQuantizer): + r""" + Diffusers Quantizer for Optimum Quanto + """ + + use_keep_in_fp32_modules = True + requires_calibration = False + required_packages = ["quanto", "accelerate"] + + def __init__(self, quantization_config, **kwargs): + super().__init__(quantization_config, **kwargs) + + def validate_environment(self, *args, **kwargs): + if not is_optimum_quanto_available(): + raise ImportError( + "Loading an optimum-quanto quantized model requires optimum-quanto library (`pip install optimum-quanto`)" + ) + if not is_optimum_quanto_version(">=", "0.2.6"): + raise ImportError( + "Loading an optimum-quanto quantized model requires `optimum-quanto>=0.2.6`. " + "Please upgrade your installation with `pip install --upgrade optimum-quanto" + ) + + if not is_accelerate_available(): + raise ImportError( + "Loading an optimum-quanto quantized model requires accelerate library (`pip install accelerate`)" + ) + + device_map = kwargs.get("device_map", None) + if isinstance(device_map, dict) and len(device_map.keys()) > 1: + raise ValueError( + "`device_map` for multi-GPU inference or CPU/disk offload is currently not supported with Diffusers and the Quanto backend" + ) + + def check_if_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + state_dict: Dict[str, Any], + **kwargs, + ): + # Quanto imports diffusers internally. This is here to prevent circular imports + from optimum.quanto import QModuleMixin, QTensor + from optimum.quanto.tensor.packed import PackedTensor + + module, tensor_name = get_module_from_name(model, param_name) + if self.pre_quantized and any(isinstance(module, t) for t in [QTensor, PackedTensor]): + return True + elif isinstance(module, QModuleMixin) and "weight" in tensor_name: + return not module.frozen + + return False + + def create_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + target_device: "torch.device", + *args, + **kwargs, + ): + """ + Create the quantized parameter by calling .freeze() after setting it to the module. + """ + + dtype = kwargs.get("dtype", torch.float32) + module, tensor_name = get_module_from_name(model, param_name) + if self.pre_quantized: + setattr(module, tensor_name, param_value) + else: + set_module_tensor_to_device(model, param_name, target_device, param_value, dtype) + module.freeze() + module.weight.requires_grad = False + + def adjust_max_memory(self, max_memory: Dict[str, Union[int, str]]) -> Dict[str, Union[int, str]]: + max_memory = {key: val * 0.90 for key, val in max_memory.items()} + return max_memory + + def adjust_target_dtype(self, target_dtype: "torch.dtype") -> "torch.dtype": + if is_accelerate_version(">=", "0.27.0"): + mapping = { + "int8": torch.int8, + "float8": CustomDtype.FP8, + "int4": CustomDtype.INT4, + "int2": CustomDtype.INT2, + } + target_dtype = mapping[self.quantization_config.weights_dtype] + + return target_dtype + + def update_torch_dtype(self, torch_dtype: "torch.dtype" = None) -> "torch.dtype": + if torch_dtype is None: + logger.info("You did not specify `torch_dtype` in `from_pretrained`. Setting it to `torch.float32`.") + torch_dtype = torch.float32 + return torch_dtype + + def update_missing_keys(self, model, missing_keys: List[str], prefix: str) -> List[str]: + # Quanto imports diffusers internally. This is here to prevent circular imports + from optimum.quanto import QModuleMixin + + not_missing_keys = [] + for name, module in model.named_modules(): + if isinstance(module, QModuleMixin): + for missing in missing_keys: + if ( + (name in missing or name in f"{prefix}.{missing}") + and not missing.endswith(".weight") + and not missing.endswith(".bias") + ): + not_missing_keys.append(missing) + return [k for k in missing_keys if k not in not_missing_keys] + + def _process_model_before_weight_loading( + self, + model: "ModelMixin", + device_map, + keep_in_fp32_modules: List[str] = [], + **kwargs, + ): + self.modules_to_not_convert = self.quantization_config.modules_to_not_convert + + if not isinstance(self.modules_to_not_convert, list): + self.modules_to_not_convert = [self.modules_to_not_convert] + + self.modules_to_not_convert.extend(keep_in_fp32_modules) + + model = _replace_with_quanto_layers( + model, + modules_to_not_convert=self.modules_to_not_convert, + quantization_config=self.quantization_config, + pre_quantized=self.pre_quantized, + ) + model.config.quantization_config = self.quantization_config + + def _process_model_after_weight_loading(self, model, **kwargs): + return model + + @property + def is_trainable(self): + return True + + @property + def is_serializable(self): + return True + + @property + def is_compileable(self) -> bool: + return True diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/utils.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..6f41fd36b43a7c222de17a297523939d27f34537 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/quanto/utils.py @@ -0,0 +1,60 @@ +import torch.nn as nn + +from ...utils import is_accelerate_available, logging + + +logger = logging.get_logger(__name__) + +if is_accelerate_available(): + from accelerate import init_empty_weights + + +def _replace_with_quanto_layers(model, quantization_config, modules_to_not_convert: list, pre_quantized=False): + # Quanto imports diffusers internally. These are placed here to avoid circular imports + from optimum.quanto import QLinear, freeze, qfloat8, qint2, qint4, qint8 + + def _get_weight_type(dtype: str): + return {"float8": qfloat8, "int8": qint8, "int4": qint4, "int2": qint2}[dtype] + + def _replace_layers(model, quantization_config, modules_to_not_convert): + has_children = list(model.children()) + if not has_children: + return model + + for name, module in model.named_children(): + _replace_layers(module, quantization_config, modules_to_not_convert) + + if name in modules_to_not_convert: + continue + + if isinstance(module, nn.Linear): + with init_empty_weights(): + qlinear = QLinear( + in_features=module.in_features, + out_features=module.out_features, + bias=module.bias is not None, + dtype=module.weight.dtype, + weights=_get_weight_type(quantization_config.weights_dtype), + ) + model._modules[name] = qlinear + model._modules[name].source_cls = type(module) + model._modules[name].requires_grad_(False) + + return model + + model = _replace_layers(model, quantization_config, modules_to_not_convert) + has_been_replaced = any(isinstance(replaced_module, QLinear) for _, replaced_module in model.named_modules()) + + if not has_been_replaced: + logger.warning( + f"{model.__class__.__name__} does not appear to have any `nn.Linear` modules. Quantization will not be applied." + " Please check your model architecture, or submit an issue on Github if you think this is a bug." + " https://github.com/huggingface/diffusers/issues/new" + ) + + # We need to freeze the pre_quantized model in order for the loaded state_dict and model state dict + # to match when trying to load weights with load_model_dict_into_meta + if pre_quantized: + freeze(model) + + return model diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/torchao/__init__.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/torchao/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..c56bf54c25158fb9e7ef9ff3bc1eb651743764fb --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/torchao/__init__.py @@ -0,0 +1,15 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# 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 .torchao_quantizer import TorchAoHfQuantizer diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/torchao/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/torchao/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..76137c06c82c94aafe48447323f50f06cc02a5c5 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/torchao/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/torchao/__pycache__/torchao_quantizer.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/quantizers/torchao/__pycache__/torchao_quantizer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1f885277632913c028c21a97af9b7cdbed6e3b13 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/quantizers/torchao/__pycache__/torchao_quantizer.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/quantizers/torchao/torchao_quantizer.py b/venv/lib/python3.10/site-packages/diffusers/quantizers/torchao/torchao_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..c12513f061da6764d5b4a1bf31707c7b9573ad8b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/quantizers/torchao/torchao_quantizer.py @@ -0,0 +1,341 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# 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. + +""" +Adapted from +https://github.com/huggingface/transformers/blob/3a8eb74668e9c2cc563b2f5c62fac174797063e0/src/transformers/quantizers/quantizer_torchao.py +""" + +import importlib +import types +from typing import TYPE_CHECKING, Any, Dict, List, Union + +from packaging import version + +from ...utils import ( + get_module_from_name, + is_torch_available, + is_torch_version, + is_torchao_available, + is_torchao_version, + logging, +) +from ..base import DiffusersQuantizer + + +if TYPE_CHECKING: + from ...models.modeling_utils import ModelMixin + + +if is_torch_available(): + import torch + import torch.nn as nn + + if is_torch_version(">=", "2.5"): + SUPPORTED_TORCH_DTYPES_FOR_QUANTIZATION = ( + # At the moment, only int8 is supported for integer quantization dtypes. + # In Torch 2.6, int1-int7 will be introduced, so this can be visited in the future + # to support more quantization methods, such as intx_weight_only. + torch.int8, + torch.float8_e4m3fn, + torch.float8_e5m2, + torch.uint1, + torch.uint2, + torch.uint3, + torch.uint4, + torch.uint5, + torch.uint6, + torch.uint7, + ) + else: + SUPPORTED_TORCH_DTYPES_FOR_QUANTIZATION = ( + torch.int8, + torch.float8_e4m3fn, + torch.float8_e5m2, + ) + +if is_torchao_available(): + from torchao.quantization import quantize_ + + +def _update_torch_safe_globals(): + safe_globals = [ + (torch.uint1, "torch.uint1"), + (torch.uint2, "torch.uint2"), + (torch.uint3, "torch.uint3"), + (torch.uint4, "torch.uint4"), + (torch.uint5, "torch.uint5"), + (torch.uint6, "torch.uint6"), + (torch.uint7, "torch.uint7"), + ] + try: + from torchao.dtypes import NF4Tensor + from torchao.dtypes.floatx.float8_layout import Float8AQTTensorImpl + from torchao.dtypes.uintx.uint4_layout import UInt4Tensor + from torchao.dtypes.uintx.uintx_layout import UintxAQTTensorImpl, UintxTensor + + safe_globals.extend([UintxTensor, UInt4Tensor, UintxAQTTensorImpl, Float8AQTTensorImpl, NF4Tensor]) + + except (ImportError, ModuleNotFoundError) as e: + logger.warning( + "Unable to import `torchao` Tensor objects. This may affect loading checkpoints serialized with `torchao`" + ) + logger.debug(e) + + finally: + torch.serialization.add_safe_globals(safe_globals=safe_globals) + + +if ( + is_torch_available() + and is_torch_version(">=", "2.6.0") + and is_torchao_available() + and is_torchao_version(">=", "0.7.0") +): + _update_torch_safe_globals() + + +logger = logging.get_logger(__name__) + + +def _quantization_type(weight): + from torchao.dtypes import AffineQuantizedTensor + from torchao.quantization.linear_activation_quantized_tensor import LinearActivationQuantizedTensor + + if isinstance(weight, AffineQuantizedTensor): + return f"{weight.__class__.__name__}({weight._quantization_type()})" + + if isinstance(weight, LinearActivationQuantizedTensor): + return f"{weight.__class__.__name__}(activation={weight.input_quant_func}, weight={_quantization_type(weight.original_weight_tensor)})" + + +def _linear_extra_repr(self): + weight = _quantization_type(self.weight) + if weight is None: + return f"in_features={self.weight.shape[1]}, out_features={self.weight.shape[0]}, weight=None" + else: + return f"in_features={self.weight.shape[1]}, out_features={self.weight.shape[0]}, weight={weight}" + + +class TorchAoHfQuantizer(DiffusersQuantizer): + r""" + Diffusers Quantizer for TorchAO: https://github.com/pytorch/ao/. + """ + + requires_calibration = False + required_packages = ["torchao"] + + def __init__(self, quantization_config, **kwargs): + super().__init__(quantization_config, **kwargs) + + def validate_environment(self, *args, **kwargs): + if not is_torchao_available(): + raise ImportError( + "Loading a TorchAO quantized model requires the torchao library. Please install with `pip install torchao`" + ) + torchao_version = version.parse(importlib.metadata.version("torch")) + if torchao_version < version.parse("0.7.0"): + raise RuntimeError( + f"The minimum required version of `torchao` is 0.7.0, but the current version is {torchao_version}. Please upgrade with `pip install -U torchao`." + ) + + self.offload = False + + device_map = kwargs.get("device_map", None) + if isinstance(device_map, dict): + if "cpu" in device_map.values() or "disk" in device_map.values(): + if self.pre_quantized: + raise ValueError( + "You are attempting to perform cpu/disk offload with a pre-quantized torchao model " + "This is not supported yet. Please remove the CPU or disk device from the `device_map` argument." + ) + else: + self.offload = True + + if self.pre_quantized: + weights_only = kwargs.get("weights_only", None) + if weights_only: + torch_version = version.parse(importlib.metadata.version("torch")) + if torch_version < version.parse("2.5.0"): + # TODO(aryan): TorchAO is compatible with Pytorch >= 2.2 for certain quantization types. Try to see if we can support it in future + raise RuntimeError( + f"In order to use TorchAO pre-quantized model, you need to have torch>=2.5.0. However, the current version is {torch_version}." + ) + + def update_torch_dtype(self, torch_dtype): + quant_type = self.quantization_config.quant_type + + if quant_type.startswith("int") or quant_type.startswith("uint"): + if torch_dtype is not None and torch_dtype != torch.bfloat16: + logger.warning( + f"You are trying to set torch_dtype to {torch_dtype} for int4/int8/uintx quantization, but " + f"only bfloat16 is supported right now. Please set `torch_dtype=torch.bfloat16`." + ) + + if torch_dtype is None: + # We need to set the torch_dtype, otherwise we have dtype mismatch when performing the quantized linear op + logger.warning( + "Overriding `torch_dtype` with `torch_dtype=torch.bfloat16` due to requirements of `torchao` " + "to enable model loading in different precisions. Pass your own `torch_dtype` to specify the " + "dtype of the remaining non-linear layers, or pass torch_dtype=torch.bfloat16, to remove this warning." + ) + torch_dtype = torch.bfloat16 + + return torch_dtype + + def adjust_target_dtype(self, target_dtype: "torch.dtype") -> "torch.dtype": + quant_type = self.quantization_config.quant_type + + if quant_type.startswith("int8") or quant_type.startswith("int4"): + # Note that int4 weights are created by packing into torch.int8, but since there is no torch.int4, we use torch.int8 + return torch.int8 + elif quant_type == "uintx_weight_only": + return self.quantization_config.quant_type_kwargs.get("dtype", torch.uint8) + elif quant_type.startswith("uint"): + return { + 1: torch.uint1, + 2: torch.uint2, + 3: torch.uint3, + 4: torch.uint4, + 5: torch.uint5, + 6: torch.uint6, + 7: torch.uint7, + }[int(quant_type[4])] + elif quant_type.startswith("float") or quant_type.startswith("fp"): + return torch.bfloat16 + + if isinstance(target_dtype, SUPPORTED_TORCH_DTYPES_FOR_QUANTIZATION): + return target_dtype + + # We need one of the supported dtypes to be selected in order for accelerate to determine + # the total size of modules/parameters for auto device placement. + possible_device_maps = ["auto", "balanced", "balanced_low_0", "sequential"] + raise ValueError( + f"You have set `device_map` as one of {possible_device_maps} on a TorchAO quantized model but a suitable target dtype " + f"could not be inferred. The supported target_dtypes are: {SUPPORTED_TORCH_DTYPES_FOR_QUANTIZATION}. If you think the " + f"dtype you are using should be supported, please open an issue at https://github.com/huggingface/diffusers/issues." + ) + + def adjust_max_memory(self, max_memory: Dict[str, Union[int, str]]) -> Dict[str, Union[int, str]]: + max_memory = {key: val * 0.9 for key, val in max_memory.items()} + return max_memory + + def check_if_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + state_dict: Dict[str, Any], + **kwargs, + ) -> bool: + param_device = kwargs.pop("param_device", None) + # Check if the param_name is not in self.modules_to_not_convert + if any((key + "." in param_name) or (key == param_name) for key in self.modules_to_not_convert): + return False + elif param_device == "cpu" and self.offload: + # We don't quantize weights that we offload + return False + else: + # We only quantize the weight of nn.Linear + module, tensor_name = get_module_from_name(model, param_name) + return isinstance(module, torch.nn.Linear) and (tensor_name == "weight") + + def create_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + target_device: "torch.device", + state_dict: Dict[str, Any], + unexpected_keys: List[str], + **kwargs, + ): + r""" + Each nn.Linear layer that needs to be quantized is processed here. First, we set the value the weight tensor, + then we move it to the target device. Finally, we quantize the module. + """ + module, tensor_name = get_module_from_name(model, param_name) + + if self.pre_quantized: + # If we're loading pre-quantized weights, replace the repr of linear layers for pretty printing info + # about AffineQuantizedTensor + module._parameters[tensor_name] = torch.nn.Parameter(param_value.to(device=target_device)) + if isinstance(module, nn.Linear): + module.extra_repr = types.MethodType(_linear_extra_repr, module) + else: + # As we perform quantization here, the repr of linear layers is that of AQT, so we don't have to do it ourselves + module._parameters[tensor_name] = torch.nn.Parameter(param_value).to(device=target_device) + quantize_(module, self.quantization_config.get_apply_tensor_subclass()) + + def _process_model_before_weight_loading( + self, + model: "ModelMixin", + device_map, + keep_in_fp32_modules: List[str] = [], + **kwargs, + ): + self.modules_to_not_convert = self.quantization_config.modules_to_not_convert + + if not isinstance(self.modules_to_not_convert, list): + self.modules_to_not_convert = [self.modules_to_not_convert] + + self.modules_to_not_convert.extend(keep_in_fp32_modules) + + # Extend `self.modules_to_not_convert` to keys that are supposed to be offloaded to `cpu` or `disk` + if isinstance(device_map, dict) and len(device_map.keys()) > 1: + keys_on_cpu = [key for key, value in device_map.items() if value in ["disk", "cpu"]] + self.modules_to_not_convert.extend(keys_on_cpu) + + # Purge `None`. + # Unlike `transformers`, we don't know if we should always keep certain modules in FP32 + # in case of diffusion transformer models. For language models and others alike, `lm_head` + # and tied modules are usually kept in FP32. + self.modules_to_not_convert = [module for module in self.modules_to_not_convert if module is not None] + + model.config.quantization_config = self.quantization_config + + def _process_model_after_weight_loading(self, model: "ModelMixin"): + return model + + def is_serializable(self, safe_serialization=None): + # TODO(aryan): needs to be tested + if safe_serialization: + logger.warning( + "torchao quantized model does not support safe serialization, please set `safe_serialization` to False." + ) + return False + + _is_torchao_serializable = version.parse(importlib.metadata.version("huggingface_hub")) >= version.parse( + "0.25.0" + ) + + if not _is_torchao_serializable: + logger.warning("torchao quantized model is only serializable after huggingface_hub >= 0.25.0 ") + + if self.offload and self.quantization_config.modules_to_not_convert is None: + logger.warning( + "The model contains offloaded modules and these modules are not quantized. We don't recommend saving the model as we won't be able to reload them." + "If you want to specify modules to not quantize, please specify modules_to_not_convert in the quantization_config." + ) + return False + + return _is_torchao_serializable + + @property + def is_trainable(self): + return self.quantization_config.quant_type.startswith("int8") + + @property + def is_compileable(self) -> bool: + return True diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__init__.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..29052c1ba0cbf404bd446487a39d2eb7ca3abf40 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/__init__.py @@ -0,0 +1,224 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 TYPE_CHECKING + +from ..utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_flax_available, + is_scipy_available, + is_torch_available, + is_torchsde_available, +) + + +_dummy_modules = {} +_import_structure = {} + +try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_pt_objects # noqa F403 + + _dummy_modules.update(get_objects_from_module(dummy_pt_objects)) + +else: + _import_structure["deprecated"] = ["KarrasVeScheduler", "ScoreSdeVpScheduler"] + _import_structure["scheduling_amused"] = ["AmusedScheduler"] + _import_structure["scheduling_consistency_decoder"] = ["ConsistencyDecoderScheduler"] + _import_structure["scheduling_consistency_models"] = ["CMStochasticIterativeScheduler"] + _import_structure["scheduling_ddim"] = ["DDIMScheduler"] + _import_structure["scheduling_ddim_cogvideox"] = ["CogVideoXDDIMScheduler"] + _import_structure["scheduling_ddim_inverse"] = ["DDIMInverseScheduler"] + _import_structure["scheduling_ddim_parallel"] = ["DDIMParallelScheduler"] + _import_structure["scheduling_ddpm"] = ["DDPMScheduler"] + _import_structure["scheduling_ddpm_parallel"] = ["DDPMParallelScheduler"] + _import_structure["scheduling_ddpm_wuerstchen"] = ["DDPMWuerstchenScheduler"] + _import_structure["scheduling_deis_multistep"] = ["DEISMultistepScheduler"] + _import_structure["scheduling_dpm_cogvideox"] = ["CogVideoXDPMScheduler"] + _import_structure["scheduling_dpmsolver_multistep"] = ["DPMSolverMultistepScheduler"] + _import_structure["scheduling_dpmsolver_multistep_inverse"] = ["DPMSolverMultistepInverseScheduler"] + _import_structure["scheduling_dpmsolver_singlestep"] = ["DPMSolverSinglestepScheduler"] + _import_structure["scheduling_edm_dpmsolver_multistep"] = ["EDMDPMSolverMultistepScheduler"] + _import_structure["scheduling_edm_euler"] = ["EDMEulerScheduler"] + _import_structure["scheduling_euler_ancestral_discrete"] = ["EulerAncestralDiscreteScheduler"] + _import_structure["scheduling_euler_discrete"] = ["EulerDiscreteScheduler"] + _import_structure["scheduling_flow_match_euler_discrete"] = ["FlowMatchEulerDiscreteScheduler"] + _import_structure["scheduling_flow_match_heun_discrete"] = ["FlowMatchHeunDiscreteScheduler"] + _import_structure["scheduling_flow_match_lcm"] = ["FlowMatchLCMScheduler"] + _import_structure["scheduling_heun_discrete"] = ["HeunDiscreteScheduler"] + _import_structure["scheduling_ipndm"] = ["IPNDMScheduler"] + _import_structure["scheduling_k_dpm_2_ancestral_discrete"] = ["KDPM2AncestralDiscreteScheduler"] + _import_structure["scheduling_k_dpm_2_discrete"] = ["KDPM2DiscreteScheduler"] + _import_structure["scheduling_lcm"] = ["LCMScheduler"] + _import_structure["scheduling_pndm"] = ["PNDMScheduler"] + _import_structure["scheduling_repaint"] = ["RePaintScheduler"] + _import_structure["scheduling_sasolver"] = ["SASolverScheduler"] + _import_structure["scheduling_scm"] = ["SCMScheduler"] + _import_structure["scheduling_sde_ve"] = ["ScoreSdeVeScheduler"] + _import_structure["scheduling_tcd"] = ["TCDScheduler"] + _import_structure["scheduling_unclip"] = ["UnCLIPScheduler"] + _import_structure["scheduling_unipc_multistep"] = ["UniPCMultistepScheduler"] + _import_structure["scheduling_utils"] = ["AysSchedules", "KarrasDiffusionSchedulers", "SchedulerMixin"] + _import_structure["scheduling_vq_diffusion"] = ["VQDiffusionScheduler"] + +try: + if not is_flax_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_flax_objects # noqa F403 + + _dummy_modules.update(get_objects_from_module(dummy_flax_objects)) + +else: + _import_structure["scheduling_ddim_flax"] = ["FlaxDDIMScheduler"] + _import_structure["scheduling_ddpm_flax"] = ["FlaxDDPMScheduler"] + _import_structure["scheduling_dpmsolver_multistep_flax"] = ["FlaxDPMSolverMultistepScheduler"] + _import_structure["scheduling_euler_discrete_flax"] = ["FlaxEulerDiscreteScheduler"] + _import_structure["scheduling_karras_ve_flax"] = ["FlaxKarrasVeScheduler"] + _import_structure["scheduling_lms_discrete_flax"] = ["FlaxLMSDiscreteScheduler"] + _import_structure["scheduling_pndm_flax"] = ["FlaxPNDMScheduler"] + _import_structure["scheduling_sde_ve_flax"] = ["FlaxScoreSdeVeScheduler"] + _import_structure["scheduling_utils_flax"] = [ + "FlaxKarrasDiffusionSchedulers", + "FlaxSchedulerMixin", + "FlaxSchedulerOutput", + "broadcast_to_shape_from_left", + ] + + +try: + if not (is_torch_available() and is_scipy_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_torch_and_scipy_objects # noqa F403 + + _dummy_modules.update(get_objects_from_module(dummy_torch_and_scipy_objects)) + +else: + _import_structure["scheduling_lms_discrete"] = ["LMSDiscreteScheduler"] + +try: + if not (is_torch_available() and is_torchsde_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_torch_and_torchsde_objects # noqa F403 + + _dummy_modules.update(get_objects_from_module(dummy_torch_and_torchsde_objects)) + +else: + _import_structure["scheduling_cosine_dpmsolver_multistep"] = ["CosineDPMSolverMultistepScheduler"] + _import_structure["scheduling_dpmsolver_sde"] = ["DPMSolverSDEScheduler"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + from ..utils import ( + OptionalDependencyNotAvailable, + is_flax_available, + is_scipy_available, + is_torch_available, + is_torchsde_available, + ) + + try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_pt_objects import * # noqa F403 + else: + from .deprecated import KarrasVeScheduler, ScoreSdeVpScheduler + from .scheduling_amused import AmusedScheduler + from .scheduling_consistency_decoder import ConsistencyDecoderScheduler + from .scheduling_consistency_models import CMStochasticIterativeScheduler + from .scheduling_ddim import DDIMScheduler + from .scheduling_ddim_cogvideox import CogVideoXDDIMScheduler + from .scheduling_ddim_inverse import DDIMInverseScheduler + from .scheduling_ddim_parallel import DDIMParallelScheduler + from .scheduling_ddpm import DDPMScheduler + from .scheduling_ddpm_parallel import DDPMParallelScheduler + from .scheduling_ddpm_wuerstchen import DDPMWuerstchenScheduler + from .scheduling_deis_multistep import DEISMultistepScheduler + from .scheduling_dpm_cogvideox import CogVideoXDPMScheduler + from .scheduling_dpmsolver_multistep import DPMSolverMultistepScheduler + from .scheduling_dpmsolver_multistep_inverse import DPMSolverMultistepInverseScheduler + from .scheduling_dpmsolver_singlestep import DPMSolverSinglestepScheduler + from .scheduling_edm_dpmsolver_multistep import EDMDPMSolverMultistepScheduler + from .scheduling_edm_euler import EDMEulerScheduler + from .scheduling_euler_ancestral_discrete import EulerAncestralDiscreteScheduler + from .scheduling_euler_discrete import EulerDiscreteScheduler + from .scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler + from .scheduling_flow_match_heun_discrete import FlowMatchHeunDiscreteScheduler + from .scheduling_flow_match_lcm import FlowMatchLCMScheduler + from .scheduling_heun_discrete import HeunDiscreteScheduler + from .scheduling_ipndm import IPNDMScheduler + from .scheduling_k_dpm_2_ancestral_discrete import KDPM2AncestralDiscreteScheduler + from .scheduling_k_dpm_2_discrete import KDPM2DiscreteScheduler + from .scheduling_lcm import LCMScheduler + from .scheduling_pndm import PNDMScheduler + from .scheduling_repaint import RePaintScheduler + from .scheduling_sasolver import SASolverScheduler + from .scheduling_scm import SCMScheduler + from .scheduling_sde_ve import ScoreSdeVeScheduler + from .scheduling_tcd import TCDScheduler + from .scheduling_unclip import UnCLIPScheduler + from .scheduling_unipc_multistep import UniPCMultistepScheduler + from .scheduling_utils import AysSchedules, KarrasDiffusionSchedulers, SchedulerMixin + from .scheduling_vq_diffusion import VQDiffusionScheduler + try: + if not is_flax_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_flax_objects import * # noqa F403 + else: + from .scheduling_ddim_flax import FlaxDDIMScheduler + from .scheduling_ddpm_flax import FlaxDDPMScheduler + from .scheduling_dpmsolver_multistep_flax import FlaxDPMSolverMultistepScheduler + from .scheduling_euler_discrete_flax import FlaxEulerDiscreteScheduler + from .scheduling_karras_ve_flax import FlaxKarrasVeScheduler + from .scheduling_lms_discrete_flax import FlaxLMSDiscreteScheduler + from .scheduling_pndm_flax import FlaxPNDMScheduler + from .scheduling_sde_ve_flax import FlaxScoreSdeVeScheduler + from .scheduling_utils_flax import ( + FlaxKarrasDiffusionSchedulers, + FlaxSchedulerMixin, + FlaxSchedulerOutput, + broadcast_to_shape_from_left, + ) + + try: + if not (is_torch_available() and is_scipy_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_torch_and_scipy_objects import * # noqa F403 + else: + from .scheduling_lms_discrete import LMSDiscreteScheduler + + try: + if not (is_torch_available() and is_torchsde_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_torch_and_torchsde_objects import * # noqa F403 + else: + from .scheduling_cosine_dpmsolver_multistep import CosineDPMSolverMultistepScheduler + from .scheduling_dpmsolver_sde import DPMSolverSDEScheduler + +else: + import sys + + sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__) + for name, value in _dummy_modules.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b9b78a39f2ede55c1301beccc88f9a9d060797d2 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_amused.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_amused.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..567350aa83374b1deb7006066fa16f8e4dba07de Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_amused.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_consistency_decoder.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_consistency_decoder.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bfaff3fd8f4f9c7444f1e06c1d5218c76cba9dc0 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_consistency_decoder.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_consistency_models.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_consistency_models.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c3a81a86fc4f3f09b3ec4ebe9741e891f9baeb67 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_consistency_models.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_cosine_dpmsolver_multistep.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_cosine_dpmsolver_multistep.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..16aa5471e5b42b5bb8f7b9c78b860fa405668a41 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_cosine_dpmsolver_multistep.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f69b1b7610f8dce8c43a5094d8c7bebe2fd52e53 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim_cogvideox.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim_cogvideox.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1ca9c8aa5f61158483f58bd729c90ce6c28d13a8 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim_cogvideox.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim_flax.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim_flax.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4d228268a19437683060638f44f69ddd5f7f0ec2 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim_flax.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim_inverse.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim_inverse.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ab6fe086697379d69b0caf830473354574b57843 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim_inverse.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim_parallel.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim_parallel.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3bbbcbba1699f70d90867d25526edf9a895fb4c5 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddim_parallel.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddpm.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddpm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..52ffb39d7f83d7ce1368b0487d36b26980f06475 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddpm.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddpm_flax.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddpm_flax.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..15a7f0405963bf3600ea296e754aa5c548bed114 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddpm_flax.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddpm_parallel.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddpm_parallel.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ad1dc019e9129873fa58886c62ff7ca823432f80 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddpm_parallel.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddpm_wuerstchen.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddpm_wuerstchen.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ec07a576ea3786e515806b7380b6a9755073695f Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ddpm_wuerstchen.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_deis_multistep.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_deis_multistep.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..50881ed21c2769b3cdccfc63d248df73f6f43a0c Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_deis_multistep.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpm_cogvideox.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpm_cogvideox.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cf57dcbf0b4fa92e9a3d0641e82c4423d9cbe340 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpm_cogvideox.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_multistep.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_multistep.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9e59fa9a9b1fc577a0c471dec284426bb4b46e31 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_multistep.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_multistep_flax.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_multistep_flax.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8a1a774b92a0acb20336d36438196d6f69bc424d Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_multistep_flax.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_multistep_inverse.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_multistep_inverse.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e0b035a40556a6ede1d20b00f5865e8091d61703 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_multistep_inverse.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_sde.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_sde.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e5ca3d752b839e57bad7d6b787753c77dfba048b Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_sde.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_singlestep.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_singlestep.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0b53084acb26195168f05a2b470a18f31a8bca3f Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_dpmsolver_singlestep.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_edm_dpmsolver_multistep.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_edm_dpmsolver_multistep.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..986bf7549984ec412c0d61e49b42b0e549820050 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_edm_dpmsolver_multistep.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_edm_euler.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_edm_euler.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..802e80927a04075b990b075a19be50feb5d0e28a Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_edm_euler.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_euler_ancestral_discrete.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_euler_ancestral_discrete.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..07a505ab4d15e81c1bfa65fa263235687a17b8a9 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_euler_ancestral_discrete.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_euler_discrete.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_euler_discrete.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..80901e0af3018d5babfab73123bdeda7ee2fcf61 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_euler_discrete.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_euler_discrete_flax.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_euler_discrete_flax.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0cdfdd96903c0f61a389ecf4f3c3759411e71479 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_euler_discrete_flax.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_flow_match_euler_discrete.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_flow_match_euler_discrete.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..71b13ec8ccc49c11a3928bb0dc077643707fd95d Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_flow_match_euler_discrete.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_flow_match_heun_discrete.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_flow_match_heun_discrete.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bea09eb2cfc848e84b2acf105db4d5b20940a2e2 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_flow_match_heun_discrete.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_flow_match_lcm.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_flow_match_lcm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..41bf0ce8052c617362eb3ef81ddf2a3238045413 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_flow_match_lcm.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_heun_discrete.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_heun_discrete.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4cacb731d7e5172aa60ca6a3504b521fa5e53621 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_heun_discrete.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ipndm.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ipndm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3047dd755069c2b972c9e3c035820e9b00651f8f Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_ipndm.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_k_dpm_2_ancestral_discrete.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_k_dpm_2_ancestral_discrete.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c2745676be1de1bb7f3c5b7a4722768a360f5bc0 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_k_dpm_2_ancestral_discrete.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_k_dpm_2_discrete.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_k_dpm_2_discrete.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..47e52c027458759c25005c66bbb185180c785d46 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_k_dpm_2_discrete.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_karras_ve_flax.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_karras_ve_flax.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e2dc04f2c8a53adfcd7f19ad79c7f475eccb60bb Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_karras_ve_flax.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_lcm.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_lcm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..81266fc5d5ce084f74c8aa8b73d207d9ceac829f Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_lcm.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_lms_discrete.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_lms_discrete.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..46747e69516458b65ff0f5e24fd5bb4bb9c55fa8 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_lms_discrete.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_lms_discrete_flax.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_lms_discrete_flax.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4c60232150092456389b554803182ad6b5488dad Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_lms_discrete_flax.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_pndm.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_pndm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..26c54b005ebe5d9a556e1c559fbf8dcea6337b5e Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_pndm.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_pndm_flax.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_pndm_flax.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d2c3a12e4048cf62e8d083ec51adbaee0c201ca5 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_pndm_flax.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_repaint.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_repaint.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5a046228d256a25be8f0bc1e901f4cbdda6d34f1 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_repaint.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_sasolver.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_sasolver.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..707a1e509d75ea841f3e6bd379c795cde41ce7e2 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_sasolver.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_scm.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_scm.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fccb6cbbc00138769e41f7343ea4c8db5a78f85d Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_scm.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_sde_ve.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_sde_ve.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..dd0d87d197327649ea91300eb55befe37ed26556 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_sde_ve.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_sde_ve_flax.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_sde_ve_flax.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6e5cc8ada8b2ac3ccf505765890109aa9cb8af55 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_sde_ve_flax.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_tcd.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_tcd.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1e8a6fe522e75c18b23199fc5f6ccb6827015dcf Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_tcd.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_unclip.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_unclip.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8a29bb96de4fd08f120cf4636bd537c3cf6c0d1c Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_unclip.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_unipc_multistep.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_unipc_multistep.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0195c0a00bc21a2fb10b03e6ae08f7355c3a89e6 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_unipc_multistep.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9c52305f1037645e2d061ea71f4043b08ce8bd13 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_utils_flax.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_utils_flax.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9ee12cd7cf5038a2ad345861604ee235f84aad33 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_utils_flax.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_vq_diffusion.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_vq_diffusion.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ee2f70aa348f984782dcc204587f14db3166df04 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/__pycache__/scheduling_vq_diffusion.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/__init__.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..479cf9bd568be320b60e87a831bf80835e9943f2 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/__init__.py @@ -0,0 +1,50 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_pt_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_pt_objects)) +else: + _import_structure["scheduling_karras_ve"] = ["KarrasVeScheduler"] + _import_structure["scheduling_sde_vp"] = ["ScoreSdeVpScheduler"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_pt_objects import * # noqa F403 + else: + from .scheduling_karras_ve import KarrasVeScheduler + from .scheduling_sde_vp import ScoreSdeVpScheduler + + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c906eefd486b8f0f6a651343f22d44cd9db3926a Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/__pycache__/scheduling_karras_ve.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/__pycache__/scheduling_karras_ve.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..029273852c36a82b810f3ec4e3d3a038ede03c70 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/__pycache__/scheduling_karras_ve.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/__pycache__/scheduling_sde_vp.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/__pycache__/scheduling_sde_vp.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..118d0b9e1ff884762ea98545b593ff09fe6a41ba Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/__pycache__/scheduling_sde_vp.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/scheduling_karras_ve.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/scheduling_karras_ve.py new file mode 100644 index 0000000000000000000000000000000000000000..6b968e70814556d93ed2d5937a2321342b97f16b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/scheduling_karras_ve.py @@ -0,0 +1,244 @@ +# Copyright 2025 NVIDIA and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +from dataclasses import dataclass +from typing import Optional, Tuple, Union + +import numpy as np +import torch + +from ...configuration_utils import ConfigMixin, register_to_config +from ...utils import BaseOutput +from ...utils.torch_utils import randn_tensor +from ..scheduling_utils import SchedulerMixin + + +@dataclass +class KarrasVeOutput(BaseOutput): + """ + Output class for the scheduler's step function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + derivative (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Derivative of predicted original image sample (x_0). + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample (x_{0}) based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + derivative: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +class KarrasVeScheduler(SchedulerMixin, ConfigMixin): + """ + A stochastic scheduler tailored to variance-expanding models. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + + + For more details on the parameters, see [Appendix E](https://huggingface.co/papers/2206.00364). The grid search + values used to find the optimal `{s_noise, s_churn, s_min, s_max}` for a specific model are described in Table 5 of + the paper. + + + + Args: + sigma_min (`float`, defaults to 0.02): + The minimum noise magnitude. + sigma_max (`float`, defaults to 100): + The maximum noise magnitude. + s_noise (`float`, defaults to 1.007): + The amount of additional noise to counteract loss of detail during sampling. A reasonable range is [1.000, + 1.011]. + s_churn (`float`, defaults to 80): + The parameter controlling the overall amount of stochasticity. A reasonable range is [0, 100]. + s_min (`float`, defaults to 0.05): + The start value of the sigma range to add noise (enable stochasticity). A reasonable range is [0, 10]. + s_max (`float`, defaults to 50): + The end value of the sigma range to add noise. A reasonable range is [0.2, 80]. + """ + + order = 2 + + @register_to_config + def __init__( + self, + sigma_min: float = 0.02, + sigma_max: float = 100, + s_noise: float = 1.007, + s_churn: float = 80, + s_min: float = 0.05, + s_max: float = 50, + ): + # standard deviation of the initial noise distribution + self.init_noise_sigma = sigma_max + + # setable values + self.num_inference_steps: int = None + self.timesteps: np.IntTensor = None + self.schedule: torch.Tensor = None # sigma(t_i) + + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.num_inference_steps = num_inference_steps + timesteps = np.arange(0, self.num_inference_steps)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps).to(device) + schedule = [ + ( + self.config.sigma_max**2 + * (self.config.sigma_min**2 / self.config.sigma_max**2) ** (i / (num_inference_steps - 1)) + ) + for i in self.timesteps + ] + self.schedule = torch.tensor(schedule, dtype=torch.float32, device=device) + + def add_noise_to_input( + self, sample: torch.Tensor, sigma: float, generator: Optional[torch.Generator] = None + ) -> Tuple[torch.Tensor, float]: + """ + Explicit Langevin-like "churn" step of adding noise to the sample according to a `gamma_i ≥ 0` to reach a + higher noise level `sigma_hat = sigma_i + gamma_i*sigma_i`. + + Args: + sample (`torch.Tensor`): + The input sample. + sigma (`float`): + generator (`torch.Generator`, *optional*): + A random number generator. + """ + if self.config.s_min <= sigma <= self.config.s_max: + gamma = min(self.config.s_churn / self.num_inference_steps, 2**0.5 - 1) + else: + gamma = 0 + + # sample eps ~ N(0, S_noise^2 * I) + eps = self.config.s_noise * randn_tensor(sample.shape, generator=generator).to(sample.device) + sigma_hat = sigma + gamma * sigma + sample_hat = sample + ((sigma_hat**2 - sigma**2) ** 0.5 * eps) + + return sample_hat, sigma_hat + + def step( + self, + model_output: torch.Tensor, + sigma_hat: float, + sigma_prev: float, + sample_hat: torch.Tensor, + return_dict: bool = True, + ) -> Union[KarrasVeOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + sigma_hat (`float`): + sigma_prev (`float`): + sample_hat (`torch.Tensor`): + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_karras_ve.KarrasVESchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_karras_ve.KarrasVESchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_karras_ve.KarrasVESchedulerOutput`] is returned, + otherwise a tuple is returned where the first element is the sample tensor. + + """ + + pred_original_sample = sample_hat + sigma_hat * model_output + derivative = (sample_hat - pred_original_sample) / sigma_hat + sample_prev = sample_hat + (sigma_prev - sigma_hat) * derivative + + if not return_dict: + return (sample_prev, derivative) + + return KarrasVeOutput( + prev_sample=sample_prev, derivative=derivative, pred_original_sample=pred_original_sample + ) + + def step_correct( + self, + model_output: torch.Tensor, + sigma_hat: float, + sigma_prev: float, + sample_hat: torch.Tensor, + sample_prev: torch.Tensor, + derivative: torch.Tensor, + return_dict: bool = True, + ) -> Union[KarrasVeOutput, Tuple]: + """ + Corrects the predicted sample based on the `model_output` of the network. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + sigma_hat (`float`): TODO + sigma_prev (`float`): TODO + sample_hat (`torch.Tensor`): TODO + sample_prev (`torch.Tensor`): TODO + derivative (`torch.Tensor`): TODO + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`. + + Returns: + prev_sample (TODO): updated sample in the diffusion chain. derivative (TODO): TODO + + """ + pred_original_sample = sample_prev + sigma_prev * model_output + derivative_corr = (sample_prev - pred_original_sample) / sigma_prev + sample_prev = sample_hat + (sigma_prev - sigma_hat) * (0.5 * derivative + 0.5 * derivative_corr) + + if not return_dict: + return (sample_prev, derivative) + + return KarrasVeOutput( + prev_sample=sample_prev, derivative=derivative, pred_original_sample=pred_original_sample + ) + + def add_noise(self, original_samples, noise, timesteps): + raise NotImplementedError() diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/scheduling_sde_vp.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/scheduling_sde_vp.py new file mode 100644 index 0000000000000000000000000000000000000000..5088bdb4976182468cbe64a3b378974c23654888 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/deprecated/scheduling_sde_vp.py @@ -0,0 +1,109 @@ +# Copyright 2025 Google Brain and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/yang-song/score_sde_pytorch + +import math +from typing import Union + +import torch + +from ...configuration_utils import ConfigMixin, register_to_config +from ...utils.torch_utils import randn_tensor +from ..scheduling_utils import SchedulerMixin + + +class ScoreSdeVpScheduler(SchedulerMixin, ConfigMixin): + """ + `ScoreSdeVpScheduler` is a variance preserving stochastic differential equation (SDE) scheduler. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 2000): + The number of diffusion steps to train the model. + beta_min (`int`, defaults to 0.1): + beta_max (`int`, defaults to 20): + sampling_eps (`int`, defaults to 1e-3): + The end value of sampling where timesteps decrease progressively from 1 to epsilon. + """ + + order = 1 + + @register_to_config + def __init__(self, num_train_timesteps=2000, beta_min=0.1, beta_max=20, sampling_eps=1e-3): + self.sigmas = None + self.discrete_sigmas = None + self.timesteps = None + + def set_timesteps(self, num_inference_steps, device: Union[str, torch.device] = None): + """ + Sets the continuous timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.timesteps = torch.linspace(1, self.config.sampling_eps, num_inference_steps, device=device) + + def step_pred(self, score, x, t, generator=None): + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + score (): + x (): + t (): + generator (`torch.Generator`, *optional*): + A random number generator. + """ + if self.timesteps is None: + raise ValueError( + "`self.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler" + ) + + # TODO(Patrick) better comments + non-PyTorch + # postprocess model score + log_mean_coeff = -0.25 * t**2 * (self.config.beta_max - self.config.beta_min) - 0.5 * t * self.config.beta_min + std = torch.sqrt(1.0 - torch.exp(2.0 * log_mean_coeff)) + std = std.flatten() + while len(std.shape) < len(score.shape): + std = std.unsqueeze(-1) + score = -score / std + + # compute + dt = -1.0 / len(self.timesteps) + + beta_t = self.config.beta_min + t * (self.config.beta_max - self.config.beta_min) + beta_t = beta_t.flatten() + while len(beta_t.shape) < len(x.shape): + beta_t = beta_t.unsqueeze(-1) + drift = -0.5 * beta_t * x + + diffusion = torch.sqrt(beta_t) + drift = drift - diffusion**2 * score + x_mean = x + drift * dt + + # add noise + noise = randn_tensor(x.shape, layout=x.layout, generator=generator, device=x.device, dtype=x.dtype) + x = x_mean + diffusion * math.sqrt(-dt) * noise + + return x, x_mean + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_amused.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_amused.py new file mode 100644 index 0000000000000000000000000000000000000000..238b8d8691715a83105df815c0ebcad748092ec1 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_amused.py @@ -0,0 +1,162 @@ +import math +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from .scheduling_utils import SchedulerMixin + + +def gumbel_noise(t, generator=None): + device = generator.device if generator is not None else t.device + noise = torch.zeros_like(t, device=device).uniform_(0, 1, generator=generator).to(t.device) + return -torch.log((-torch.log(noise.clamp(1e-20))).clamp(1e-20)) + + +def mask_by_random_topk(mask_len, probs, temperature=1.0, generator=None): + confidence = torch.log(probs.clamp(1e-20)) + temperature * gumbel_noise(probs, generator=generator) + sorted_confidence = torch.sort(confidence, dim=-1).values + cut_off = torch.gather(sorted_confidence, 1, mask_len.long()) + masking = confidence < cut_off + return masking + + +@dataclass +class AmusedSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor = None + + +class AmusedScheduler(SchedulerMixin, ConfigMixin): + order = 1 + + temperatures: torch.Tensor + + @register_to_config + def __init__( + self, + mask_token_id: int, + masking_schedule: str = "cosine", + ): + self.temperatures = None + self.timesteps = None + + def set_timesteps( + self, + num_inference_steps: int, + temperature: Union[int, Tuple[int, int], List[int]] = (2, 0), + device: Union[str, torch.device] = None, + ): + self.timesteps = torch.arange(num_inference_steps, device=device).flip(0) + + if isinstance(temperature, (tuple, list)): + self.temperatures = torch.linspace(temperature[0], temperature[1], num_inference_steps, device=device) + else: + self.temperatures = torch.linspace(temperature, 0.01, num_inference_steps, device=device) + + def step( + self, + model_output: torch.Tensor, + timestep: torch.long, + sample: torch.LongTensor, + starting_mask_ratio: int = 1, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[AmusedSchedulerOutput, Tuple]: + two_dim_input = sample.ndim == 3 and model_output.ndim == 4 + + if two_dim_input: + batch_size, codebook_size, height, width = model_output.shape + sample = sample.reshape(batch_size, height * width) + model_output = model_output.reshape(batch_size, codebook_size, height * width).permute(0, 2, 1) + + unknown_map = sample == self.config.mask_token_id + + probs = model_output.softmax(dim=-1) + + device = probs.device + probs_ = probs.to(generator.device) if generator is not None else probs # handles when generator is on CPU + if probs_.device.type == "cpu" and probs_.dtype != torch.float32: + probs_ = probs_.float() # multinomial is not implemented for cpu half precision + probs_ = probs_.reshape(-1, probs.size(-1)) + pred_original_sample = torch.multinomial(probs_, 1, generator=generator).to(device=device) + pred_original_sample = pred_original_sample[:, 0].view(*probs.shape[:-1]) + pred_original_sample = torch.where(unknown_map, pred_original_sample, sample) + + if timestep == 0: + prev_sample = pred_original_sample + else: + seq_len = sample.shape[1] + step_idx = (self.timesteps == timestep).nonzero() + ratio = (step_idx + 1) / len(self.timesteps) + + if self.config.masking_schedule == "cosine": + mask_ratio = torch.cos(ratio * math.pi / 2) + elif self.config.masking_schedule == "linear": + mask_ratio = 1 - ratio + else: + raise ValueError(f"unknown masking schedule {self.config.masking_schedule}") + + mask_ratio = starting_mask_ratio * mask_ratio + + mask_len = (seq_len * mask_ratio).floor() + # do not mask more than amount previously masked + mask_len = torch.min(unknown_map.sum(dim=-1, keepdim=True) - 1, mask_len) + # mask at least one + mask_len = torch.max(torch.tensor([1], device=model_output.device), mask_len) + + selected_probs = torch.gather(probs, -1, pred_original_sample[:, :, None])[:, :, 0] + # Ignores the tokens given in the input by overwriting their confidence. + selected_probs = torch.where(unknown_map, selected_probs, torch.finfo(selected_probs.dtype).max) + + masking = mask_by_random_topk(mask_len, selected_probs, self.temperatures[step_idx], generator) + + # Masks tokens with lower confidence. + prev_sample = torch.where(masking, self.config.mask_token_id, pred_original_sample) + + if two_dim_input: + prev_sample = prev_sample.reshape(batch_size, height, width) + pred_original_sample = pred_original_sample.reshape(batch_size, height, width) + + if not return_dict: + return (prev_sample, pred_original_sample) + + return AmusedSchedulerOutput(prev_sample, pred_original_sample) + + def add_noise(self, sample, timesteps, generator=None): + step_idx = (self.timesteps == timesteps).nonzero() + ratio = (step_idx + 1) / len(self.timesteps) + + if self.config.masking_schedule == "cosine": + mask_ratio = torch.cos(ratio * math.pi / 2) + elif self.config.masking_schedule == "linear": + mask_ratio = 1 - ratio + else: + raise ValueError(f"unknown masking schedule {self.config.masking_schedule}") + + mask_indices = ( + torch.rand( + sample.shape, device=generator.device if generator is not None else sample.device, generator=generator + ).to(sample.device) + < mask_ratio + ) + + masked_sample = sample.clone() + + masked_sample[mask_indices] = self.config.mask_token_id + + return masked_sample diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_consistency_decoder.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_consistency_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..d7af018b284a59878c3cae86cec57e02d344e227 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_consistency_decoder.py @@ -0,0 +1,180 @@ +import math +from dataclasses import dataclass +from typing import Optional, Tuple, Union + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +@dataclass +class ConsistencyDecoderSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.Tensor + + +class ConsistencyDecoderScheduler(SchedulerMixin, ConfigMixin): + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1024, + sigma_data: float = 0.5, + ): + betas = betas_for_alpha_bar(num_train_timesteps) + + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + + self.sqrt_alphas_cumprod = torch.sqrt(alphas_cumprod) + self.sqrt_one_minus_alphas_cumprod = torch.sqrt(1.0 - alphas_cumprod) + + sigmas = torch.sqrt(1.0 / alphas_cumprod - 1) + + sqrt_recip_alphas_cumprod = torch.sqrt(1.0 / alphas_cumprod) + + self.c_skip = sqrt_recip_alphas_cumprod * sigma_data**2 / (sigmas**2 + sigma_data**2) + self.c_out = sigmas * sigma_data / (sigmas**2 + sigma_data**2) ** 0.5 + self.c_in = sqrt_recip_alphas_cumprod / (sigmas**2 + sigma_data**2) ** 0.5 + + def set_timesteps( + self, + num_inference_steps: Optional[int] = None, + device: Union[str, torch.device] = None, + ): + if num_inference_steps != 2: + raise ValueError("Currently more than 2 inference steps are not supported.") + + self.timesteps = torch.tensor([1008, 512], dtype=torch.long, device=device) + self.sqrt_alphas_cumprod = self.sqrt_alphas_cumprod.to(device) + self.sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod.to(device) + self.c_skip = self.c_skip.to(device) + self.c_out = self.c_out.to(device) + self.c_in = self.c_in.to(device) + + @property + def init_noise_sigma(self): + return self.sqrt_one_minus_alphas_cumprod[self.timesteps[0]] + + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample * self.c_in[timestep] + + def step( + self, + model_output: torch.Tensor, + timestep: Union[float, torch.Tensor], + sample: torch.Tensor, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[ConsistencyDecoderSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`float`): + The current timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a + [`~schedulers.scheduling_consistency_models.ConsistencyDecoderSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_consistency_models.ConsistencyDecoderSchedulerOutput`] or `tuple`: + If return_dict is `True`, + [`~schedulers.scheduling_consistency_models.ConsistencyDecoderSchedulerOutput`] is returned, otherwise + a tuple is returned where the first element is the sample tensor. + """ + x_0 = self.c_out[timestep] * model_output + self.c_skip[timestep] * sample + + timestep_idx = torch.where(self.timesteps == timestep)[0] + + if timestep_idx == len(self.timesteps) - 1: + prev_sample = x_0 + else: + noise = randn_tensor(x_0.shape, generator=generator, dtype=x_0.dtype, device=x_0.device) + prev_sample = ( + self.sqrt_alphas_cumprod[self.timesteps[timestep_idx + 1]].to(x_0.dtype) * x_0 + + self.sqrt_one_minus_alphas_cumprod[self.timesteps[timestep_idx + 1]].to(x_0.dtype) * noise + ) + + if not return_dict: + return (prev_sample,) + + return ConsistencyDecoderSchedulerOutput(prev_sample=prev_sample) diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_consistency_models.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_consistency_models.py new file mode 100644 index 0000000000000000000000000000000000000000..0f506225880091a9d3635d35de0dc9aaca1a245b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_consistency_models.py @@ -0,0 +1,445 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class CMStochasticIterativeSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.Tensor + + +class CMStochasticIterativeScheduler(SchedulerMixin, ConfigMixin): + """ + Multistep and onestep sampling for consistency models. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 40): + The number of diffusion steps to train the model. + sigma_min (`float`, defaults to 0.002): + Minimum noise magnitude in the sigma schedule. Defaults to 0.002 from the original implementation. + sigma_max (`float`, defaults to 80.0): + Maximum noise magnitude in the sigma schedule. Defaults to 80.0 from the original implementation. + sigma_data (`float`, defaults to 0.5): + The standard deviation of the data distribution from the EDM + [paper](https://huggingface.co/papers/2206.00364). Defaults to 0.5 from the original implementation. + s_noise (`float`, defaults to 1.0): + The amount of additional noise to counteract loss of detail during sampling. A reasonable range is [1.000, + 1.011]. Defaults to 1.0 from the original implementation. + rho (`float`, defaults to 7.0): + The parameter for calculating the Karras sigma schedule from the EDM + [paper](https://huggingface.co/papers/2206.00364). Defaults to 7.0 from the original implementation. + clip_denoised (`bool`, defaults to `True`): + Whether to clip the denoised outputs to `(-1, 1)`. + timesteps (`List` or `np.ndarray` or `torch.Tensor`, *optional*): + An explicit timestep schedule that can be optionally specified. The timesteps are expected to be in + increasing order. + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 40, + sigma_min: float = 0.002, + sigma_max: float = 80.0, + sigma_data: float = 0.5, + s_noise: float = 1.0, + rho: float = 7.0, + clip_denoised: bool = True, + ): + # standard deviation of the initial noise distribution + self.init_noise_sigma = sigma_max + + ramp = np.linspace(0, 1, num_train_timesteps) + sigmas = self._convert_to_karras(ramp) + timesteps = self.sigma_to_t(sigmas) + + # setable values + self.num_inference_steps = None + self.sigmas = torch.from_numpy(sigmas) + self.timesteps = torch.from_numpy(timesteps) + self.custom_timesteps = False + self.is_scale_input_called = False + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor: + """ + Scales the consistency model input by `(sigma**2 + sigma_data**2) ** 0.5`. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`float` or `torch.Tensor`): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + # Get sigma corresponding to timestep + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + + sample = sample / ((sigma**2 + self.config.sigma_data**2) ** 0.5) + + self.is_scale_input_called = True + return sample + + def sigma_to_t(self, sigmas: Union[float, np.ndarray]): + """ + Gets scaled timesteps from the Karras sigmas for input to the consistency model. + + Args: + sigmas (`float` or `np.ndarray`): + A single Karras sigma or an array of Karras sigmas. + + Returns: + `float` or `np.ndarray`: + A scaled input timestep or scaled input timestep array. + """ + if not isinstance(sigmas, np.ndarray): + sigmas = np.array(sigmas, dtype=np.float64) + + timesteps = 1000 * 0.25 * np.log(sigmas + 1e-44) + + return timesteps + + def set_timesteps( + self, + num_inference_steps: Optional[int] = None, + device: Union[str, torch.device] = None, + timesteps: Optional[List[int]] = None, + ): + """ + Sets the timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of equal spacing between timesteps is used. If `timesteps` is passed, + `num_inference_steps` must be `None`. + """ + if num_inference_steps is None and timesteps is None: + raise ValueError("Exactly one of `num_inference_steps` or `timesteps` must be supplied.") + + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Can only pass one of `num_inference_steps` or `timesteps`.") + + # Follow DDPMScheduler custom timesteps logic + if timesteps is not None: + for i in range(1, len(timesteps)): + if timesteps[i] >= timesteps[i - 1]: + raise ValueError("`timesteps` must be in descending order.") + + if timesteps[0] >= self.config.num_train_timesteps: + raise ValueError( + f"`timesteps` must start before `self.config.train_timesteps`: {self.config.num_train_timesteps}." + ) + + timesteps = np.array(timesteps, dtype=np.int64) + self.custom_timesteps = True + else: + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + self.custom_timesteps = False + + # Map timesteps to Karras sigmas directly for multistep sampling + # See https://github.com/openai/consistency_models/blob/main/cm/karras_diffusion.py#L675 + num_train_timesteps = self.config.num_train_timesteps + ramp = timesteps[::-1].copy() + ramp = ramp / (num_train_timesteps - 1) + sigmas = self._convert_to_karras(ramp) + timesteps = self.sigma_to_t(sigmas) + + sigmas = np.concatenate([sigmas, [self.config.sigma_min]]).astype(np.float32) + self.sigmas = torch.from_numpy(sigmas).to(device=device) + + if str(device).startswith("mps"): + # mps does not support float64 + self.timesteps = torch.from_numpy(timesteps).to(device, dtype=torch.float32) + else: + self.timesteps = torch.from_numpy(timesteps).to(device=device) + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Modified _convert_to_karras implementation that takes in ramp as argument + def _convert_to_karras(self, ramp): + """Constructs the noise schedule of Karras et al. (2022).""" + + sigma_min: float = self.config.sigma_min + sigma_max: float = self.config.sigma_max + + rho = self.config.rho + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + def get_scalings(self, sigma): + sigma_data = self.config.sigma_data + + c_skip = sigma_data**2 / (sigma**2 + sigma_data**2) + c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + return c_skip, c_out + + def get_scalings_for_boundary_condition(self, sigma): + """ + Gets the scalings used in the consistency model parameterization (from Appendix C of the + [paper](https://huggingface.co/papers/2303.01469)) to enforce boundary condition. + + + + `epsilon` in the equations for `c_skip` and `c_out` is set to `sigma_min`. + + + + Args: + sigma (`torch.Tensor`): + The current sigma in the Karras sigma schedule. + + Returns: + `tuple`: + A two-element tuple where `c_skip` (which weights the current sample) is the first element and `c_out` + (which weights the consistency model output) is the second element. + """ + sigma_min = self.config.sigma_min + sigma_data = self.config.sigma_data + + c_skip = sigma_data**2 / ((sigma - sigma_min) ** 2 + sigma_data**2) + c_out = (sigma - sigma_min) * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + return c_skip, c_out + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: Union[float, torch.Tensor], + sample: torch.Tensor, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[CMStochasticIterativeSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`float`): + The current timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a + [`~schedulers.scheduling_consistency_models.CMStochasticIterativeSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_consistency_models.CMStochasticIterativeSchedulerOutput`] or `tuple`: + If return_dict is `True`, + [`~schedulers.scheduling_consistency_models.CMStochasticIterativeSchedulerOutput`] is returned, + otherwise a tuple is returned where the first element is the sample tensor. + """ + + if isinstance(timestep, (int, torch.IntTensor, torch.LongTensor)): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + f" `{self.__class__}.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if not self.is_scale_input_called: + logger.warning( + "The `scale_model_input` function should be called before `step` to ensure correct denoising. " + "See `StableDiffusionPipeline` for a usage example." + ) + + sigma_min = self.config.sigma_min + sigma_max = self.config.sigma_max + + if self.step_index is None: + self._init_step_index(timestep) + + # sigma_next corresponds to next_t in original implementation + sigma = self.sigmas[self.step_index] + if self.step_index + 1 < self.config.num_train_timesteps: + sigma_next = self.sigmas[self.step_index + 1] + else: + # Set sigma_next to sigma_min + sigma_next = self.sigmas[-1] + + # Get scalings for boundary conditions + c_skip, c_out = self.get_scalings_for_boundary_condition(sigma) + + # 1. Denoise model output using boundary conditions + denoised = c_out * model_output + c_skip * sample + if self.config.clip_denoised: + denoised = denoised.clamp(-1, 1) + + # 2. Sample z ~ N(0, s_noise^2 * I) + # Noise is not used for onestep sampling. + if len(self.timesteps) > 1: + noise = randn_tensor( + model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator + ) + else: + noise = torch.zeros_like(model_output) + z = noise * self.config.s_noise + + sigma_hat = sigma_next.clamp(min=sigma_min, max=sigma_max) + + # 3. Return noisy sample + # tau = sigma_hat, eps = sigma_min + prev_sample = denoised + z * (sigma_hat**2 - sigma_min**2) ** 0.5 + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return CMStochasticIterativeSchedulerOutput(prev_sample=prev_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_cosine_dpmsolver_multistep.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_cosine_dpmsolver_multistep.py new file mode 100644 index 0000000000000000000000000000000000000000..66ed296da8ea89d5cfd3f240d698cf640d08a808 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_cosine_dpmsolver_multistep.py @@ -0,0 +1,577 @@ +# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver and https://github.com/NVlabs/edm + +import math +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from .scheduling_dpmsolver_sde import BrownianTreeNoiseSampler +from .scheduling_utils import SchedulerMixin, SchedulerOutput + + +class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin): + """ + Implements a variant of `DPMSolverMultistepScheduler` with cosine schedule, proposed by Nichol and Dhariwal (2021). + This scheduler was used in Stable Audio Open [1]. + + [1] Evans, Parker, et al. "Stable Audio Open" https://huggingface.co/papers/2407.14358 + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + sigma_min (`float`, *optional*, defaults to 0.3): + Minimum noise magnitude in the sigma schedule. This was set to 0.3 in Stable Audio Open [1]. + sigma_max (`float`, *optional*, defaults to 500): + Maximum noise magnitude in the sigma schedule. This was set to 500 in Stable Audio Open [1]. + sigma_data (`float`, *optional*, defaults to 1.0): + The standard deviation of the data distribution. This is set to 1.0 in Stable Audio Open [1]. + sigma_schedule (`str`, *optional*, defaults to `exponential`): + Sigma schedule to compute the `sigmas`. By default, we the schedule introduced in the EDM paper + (https://huggingface.co/papers/2206.00364). Other acceptable value is "exponential". The exponential + schedule was incorporated in this model: https://huggingface.co/stabilityai/cosxl. + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + solver_order (`int`, defaults to 2): + The DPMSolver order which can be `1` or `2`. It is recommended to use `solver_order=2`. + prediction_type (`str`, defaults to `v_prediction`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + solver_type (`str`, defaults to `midpoint`): + Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the + sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers. + lower_order_final (`bool`, defaults to `True`): + Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can + stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10. + euler_at_final (`bool`, defaults to `False`): + Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail + richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference + steps, but sometimes may result in blurring. + final_sigmas_type (`str`, defaults to `"zero"`): + The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final + sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0. + """ + + _compatibles = [] + order = 1 + + @register_to_config + def __init__( + self, + sigma_min: float = 0.3, + sigma_max: float = 500, + sigma_data: float = 1.0, + sigma_schedule: str = "exponential", + num_train_timesteps: int = 1000, + solver_order: int = 2, + prediction_type: str = "v_prediction", + rho: float = 7.0, + solver_type: str = "midpoint", + lower_order_final: bool = True, + euler_at_final: bool = False, + final_sigmas_type: Optional[str] = "zero", # "zero", "sigma_min" + ): + if solver_type not in ["midpoint", "heun"]: + if solver_type in ["logrho", "bh1", "bh2"]: + self.register_to_config(solver_type="midpoint") + else: + raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}") + + ramp = torch.linspace(0, 1, num_train_timesteps) + if sigma_schedule == "karras": + sigmas = self._compute_karras_sigmas(ramp) + elif sigma_schedule == "exponential": + sigmas = self._compute_exponential_sigmas(ramp) + + self.timesteps = self.precondition_noise(sigmas) + + self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)]) + + # setable values + self.num_inference_steps = None + self.model_outputs = [None] * solver_order + self.lower_order_nums = 0 + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self): + # standard deviation of the initial noise distribution + return (self.config.sigma_max**2 + 1) ** 0.5 + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_inputs + def precondition_inputs(self, sample, sigma): + c_in = self._get_conditioning_c_in(sigma) + scaled_sample = sample * c_in + return scaled_sample + + def precondition_noise(self, sigma): + if not isinstance(sigma, torch.Tensor): + sigma = torch.tensor([sigma]) + + return sigma.atan() / math.pi * 2 + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_outputs + def precondition_outputs(self, sample, model_output, sigma): + sigma_data = self.config.sigma_data + c_skip = sigma_data**2 / (sigma**2 + sigma_data**2) + + if self.config.prediction_type == "epsilon": + c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + elif self.config.prediction_type == "v_prediction": + c_out = -sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + else: + raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.") + + denoised = c_skip * sample + c_out * model_output + + return denoised + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.scale_model_input + def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = self.precondition_inputs(sample, sigma) + + self.is_scale_input_called = True + return sample + + def set_timesteps(self, num_inference_steps: int = None, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + + self.num_inference_steps = num_inference_steps + + ramp = torch.linspace(0, 1, self.num_inference_steps) + if self.config.sigma_schedule == "karras": + sigmas = self._compute_karras_sigmas(ramp) + elif self.config.sigma_schedule == "exponential": + sigmas = self._compute_exponential_sigmas(ramp) + + sigmas = sigmas.to(dtype=torch.float32, device=device) + self.timesteps = self.precondition_noise(sigmas) + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = self.config.sigma_min + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + + self.sigmas = torch.cat([sigmas, torch.tensor([sigma_last], dtype=torch.float32, device=device)]) + + self.model_outputs = [ + None, + ] * self.config.solver_order + self.lower_order_nums = 0 + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # if a noise sampler is used, reinitialise it + self.noise_sampler = None + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._compute_karras_sigmas + def _compute_karras_sigmas(self, ramp, sigma_min=None, sigma_max=None) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + sigma_min = sigma_min or self.config.sigma_min + sigma_max = sigma_max or self.config.sigma_max + + rho = self.config.rho + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._compute_exponential_sigmas + def _compute_exponential_sigmas(self, ramp, sigma_min=None, sigma_max=None) -> torch.Tensor: + """Implementation closely follows k-diffusion. + + https://github.com/crowsonkb/k-diffusion/blob/6ab5146d4a5ef63901326489f31f1d8e7dd36b48/k_diffusion/sampling.py#L26 + """ + sigma_min = sigma_min or self.config.sigma_min + sigma_max = sigma_max or self.config.sigma_max + sigmas = torch.linspace(math.log(sigma_min), math.log(sigma_max), len(ramp)).exp().flip(0) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + def _sigma_to_alpha_sigma_t(self, sigma): + alpha_t = torch.tensor(1) # Inputs are pre-scaled before going into unet, so alpha_t = 1 + sigma_t = sigma + + return alpha_t, sigma_t + + def convert_model_output( + self, + model_output: torch.Tensor, + sample: torch.Tensor = None, + ) -> torch.Tensor: + """ + Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is + designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an + integral of the data prediction model. + + + + The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both noise + prediction and data prediction models. + + + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + sigma = self.sigmas[self.step_index] + x0_pred = self.precondition_outputs(sample, model_output, sigma) + + return x0_pred + + def dpm_solver_first_order_update( + self, + model_output: torch.Tensor, + sample: torch.Tensor = None, + noise: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + """ + One step for the first-order DPMSolver (equivalent to DDIM). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s = torch.log(alpha_s) - torch.log(sigma_s) + + h = lambda_t - lambda_s + assert noise is not None + x_t = ( + (sigma_t / sigma_s * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + + return x_t + + def multistep_dpm_solver_second_order_update( + self, + model_output_list: List[torch.Tensor], + sample: torch.Tensor = None, + noise: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + """ + One step for the second-order multistep DPMSolver. + + Args: + model_output_list (`List[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + sigma_t, sigma_s0, sigma_s1 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + + m0, m1 = model_output_list[-1], model_output_list[-2] + + h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1 + r0 = h_0 / h + D0, D1 = m0, (1.0 / r0) * (m0 - m1) + + # sde-dpmsolver++ + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + + return x_t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + index_candidates = (schedule_timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + return step_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index + def _init_step_index(self, timestep): + """ + Initialize the step_index counter for the scheduler. + """ + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: Union[int, torch.Tensor], + sample: torch.Tensor, + generator=None, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the multistep DPMSolver. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Improve numerical stability for small number of steps + lower_order_final = (self.step_index == len(self.timesteps) - 1) and ( + self.config.euler_at_final + or (self.config.lower_order_final and len(self.timesteps) < 15) + or self.config.final_sigmas_type == "zero" + ) + lower_order_second = ( + (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15 + ) + + model_output = self.convert_model_output(model_output, sample=sample) + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.model_outputs[-1] = model_output + + if self.noise_sampler is None: + seed = None + if generator is not None: + seed = ( + [g.initial_seed() for g in generator] if isinstance(generator, list) else generator.initial_seed() + ) + self.noise_sampler = BrownianTreeNoiseSampler( + model_output, sigma_min=self.config.sigma_min, sigma_max=self.config.sigma_max, seed=seed + ) + noise = self.noise_sampler(self.sigmas[self.step_index], self.sigmas[self.step_index + 1]).to( + model_output.device + ) + + if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final: + prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise) + elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second: + prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise) + + if self.lower_order_nums < self.config.solver_order: + self.lower_order_nums += 1 + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._get_conditioning_c_in + def _get_conditioning_c_in(self, sigma): + c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5) + return c_in + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim.py new file mode 100644 index 0000000000000000000000000000000000000000..5ee0d084f060920820b686056ecca4b48fef543e --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim.py @@ -0,0 +1,521 @@ +# Copyright 2025 Stanford University Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +import math +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM +class DDIMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +def rescale_zero_terminal_snr(betas): + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + + Args: + betas (`torch.Tensor`): + the betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class DDIMScheduler(SchedulerMixin, ConfigMixin): + """ + `DDIMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with + non-Markovian guidance. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to 1.0): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, defaults to `True`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + clip_sample: bool = True, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: str = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + clip_sample_range: float = 1.0, + sample_max_value: float = 1.0, + timestep_spacing: str = "leading", + rescale_betas_zero_snr: bool = False, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # At every step in ddim, we are looking into the previous alphas_cumprod + # For the final step, there is no previous alphas_cumprod because we are already at 0 + # `set_alpha_to_one` decides whether we set this parameter simply to one or + # whether we use the final alpha of the "non-previous" one. + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def _get_variance(self, timestep, prev_timestep): + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + """ + + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps) + .round()[::-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + eta: float = 0.0, + use_clipped_model_output: bool = False, + generator=None, + variance_noise: Optional[torch.Tensor] = None, + return_dict: bool = True, + ) -> Union[DDIMSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + eta (`float`): + The weight of noise for added noise in diffusion step. + use_clipped_model_output (`bool`, defaults to `False`): + If `True`, computes "corrected" `model_output` from the clipped predicted original sample. Necessary + because predicted original sample is clipped to [-1, 1] when `self.config.clip_sample` is `True`. If no + clipping has happened, "corrected" `model_output` would coincide with the one provided as input and + `use_clipped_model_output` has no effect. + generator (`torch.Generator`, *optional*): + A random number generator. + variance_noise (`torch.Tensor`): + Alternative to generating noise with `generator` by directly providing the noise for the variance + itself. Useful for methods such as [`CycleDiffusion`]. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502 + # Ideally, read DDIM paper in-detail understanding + + # Notation ( -> + # - pred_noise_t -> e_theta(x_t, t) + # - pred_original_sample -> f_theta(x_t, t) or x_0 + # - std_dev_t -> sigma_t + # - eta -> η + # - pred_sample_direction -> "direction pointing to x_t" + # - pred_prev_sample -> "x_t-1" + + # 1. get previous step value (=t-1) + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + # 2. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + pred_epsilon = model_output + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction`" + ) + + # 4. Clip or threshold "predicted x_0" + if self.config.thresholding: + pred_original_sample = self._threshold_sample(pred_original_sample) + elif self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 5. compute variance: "sigma_t(η)" -> see formula (16) + # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1) + variance = self._get_variance(timestep, prev_timestep) + std_dev_t = eta * variance ** (0.5) + + if use_clipped_model_output: + # the pred_epsilon is always re-derived from the clipped x_0 in Glide + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + + # 6. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon + + # 7. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502 + prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction + + if eta > 0: + if variance_noise is not None and generator is not None: + raise ValueError( + "Cannot pass both generator and variance_noise. Please make sure that either `generator` or" + " `variance_noise` stays `None`." + ) + + if variance_noise is None: + variance_noise = randn_tensor( + model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype + ) + variance = std_dev_t * variance_noise + + prev_sample = prev_sample + variance + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim_cogvideox.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim_cogvideox.py new file mode 100644 index 0000000000000000000000000000000000000000..c19efdc7834dadd2c55d32f5c29e7e05f6d2b27f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim_cogvideox.py @@ -0,0 +1,452 @@ +# Copyright 2025 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team. +# All rights reserved. +# +# 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. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +import math +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM +class DDIMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +def rescale_zero_terminal_snr(alphas_cumprod): + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + + Args: + betas (`torch.Tensor`): + the betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + + return alphas_bar + + +class CogVideoXDDIMScheduler(SchedulerMixin, ConfigMixin): + """ + `DDIMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with + non-Markovian guidance. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to 1.0): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, defaults to `True`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, + beta_end: float = 0.0120, + beta_schedule: str = "scaled_linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + clip_sample: bool = True, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: str = "epsilon", + clip_sample_range: float = 1.0, + sample_max_value: float = 1.0, + timestep_spacing: str = "leading", + rescale_betas_zero_snr: bool = False, + snr_shift_scale: float = 3.0, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float64) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # Modify: SNR shift following SD3 + self.alphas_cumprod = self.alphas_cumprod / (snr_shift_scale + (1 - snr_shift_scale) * self.alphas_cumprod) + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.alphas_cumprod = rescale_zero_terminal_snr(self.alphas_cumprod) + + # At every step in ddim, we are looking into the previous alphas_cumprod + # For the final step, there is no previous alphas_cumprod because we are already at 0 + # `set_alpha_to_one` decides whether we set this parameter simply to one or + # whether we use the final alpha of the "non-previous" one. + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + + def _get_variance(self, timestep, prev_timestep): + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + """ + + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps) + .round()[::-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + eta: float = 0.0, + use_clipped_model_output: bool = False, + generator=None, + variance_noise: Optional[torch.Tensor] = None, + return_dict: bool = True, + ) -> Union[DDIMSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + eta (`float`): + The weight of noise for added noise in diffusion step. + use_clipped_model_output (`bool`, defaults to `False`): + If `True`, computes "corrected" `model_output` from the clipped predicted original sample. Necessary + because predicted original sample is clipped to [-1, 1] when `self.config.clip_sample` is `True`. If no + clipping has happened, "corrected" `model_output` would coincide with the one provided as input and + `use_clipped_model_output` has no effect. + generator (`torch.Generator`, *optional*): + A random number generator. + variance_noise (`torch.Tensor`): + Alternative to generating noise with `generator` by directly providing the noise for the variance + itself. Useful for methods such as [`CycleDiffusion`]. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502 + # Ideally, read DDIM paper in-detail understanding + + # Notation ( -> + # - pred_noise_t -> e_theta(x_t, t) + # - pred_original_sample -> f_theta(x_t, t) or x_0 + # - std_dev_t -> sigma_t + # - eta -> η + # - pred_sample_direction -> "direction pointing to x_t" + # - pred_prev_sample -> "x_t-1" + + # 1. get previous step value (=t-1) + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + # 2. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + # To make style tests pass, commented out `pred_epsilon` as it is an unused variable + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + # pred_epsilon = model_output + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + # pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + # pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction`" + ) + + a_t = ((1 - alpha_prod_t_prev) / (1 - alpha_prod_t)) ** 0.5 + b_t = alpha_prod_t_prev**0.5 - alpha_prod_t**0.5 * a_t + + prev_sample = a_t * sample + b_t * pred_original_sample + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim_flax.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim_flax.py new file mode 100644 index 0000000000000000000000000000000000000000..802d8f79779de52c39ffd1210571518c74e37b02 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim_flax.py @@ -0,0 +1,314 @@ +# Copyright 2025 Stanford University Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +from dataclasses import dataclass +from typing import Optional, Tuple, Union + +import flax +import jax.numpy as jnp + +from ..configuration_utils import ConfigMixin, register_to_config +from .scheduling_utils_flax import ( + CommonSchedulerState, + FlaxKarrasDiffusionSchedulers, + FlaxSchedulerMixin, + FlaxSchedulerOutput, + add_noise_common, + get_velocity_common, +) + + +@flax.struct.dataclass +class DDIMSchedulerState: + common: CommonSchedulerState + final_alpha_cumprod: jnp.ndarray + + # setable values + init_noise_sigma: jnp.ndarray + timesteps: jnp.ndarray + num_inference_steps: Optional[int] = None + + @classmethod + def create( + cls, + common: CommonSchedulerState, + final_alpha_cumprod: jnp.ndarray, + init_noise_sigma: jnp.ndarray, + timesteps: jnp.ndarray, + ): + return cls( + common=common, + final_alpha_cumprod=final_alpha_cumprod, + init_noise_sigma=init_noise_sigma, + timesteps=timesteps, + ) + + +@dataclass +class FlaxDDIMSchedulerOutput(FlaxSchedulerOutput): + state: DDIMSchedulerState + + +class FlaxDDIMScheduler(FlaxSchedulerMixin, ConfigMixin): + """ + Denoising diffusion implicit models is a scheduler that extends the denoising procedure introduced in denoising + diffusion probabilistic models (DDPMs) with non-Markovian guidance. + + [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` + function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. + [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and + [`~SchedulerMixin.from_pretrained`] functions. + + For more details, see the original paper: https://huggingface.co/papers/2010.02502 + + Args: + num_train_timesteps (`int`): number of diffusion steps used to train the model. + beta_start (`float`): the starting `beta` value of inference. + beta_end (`float`): the final `beta` value. + beta_schedule (`str`): + the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`jnp.ndarray`, optional): + option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc. + clip_sample (`bool`, default `True`): + option to clip predicted sample between for numerical stability. The clip range is determined by + `clip_sample_range`. + clip_sample_range (`float`, default `1.0`): + the maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, default `True`): + each diffusion step uses the value of alphas product at that step and at the previous one. For the final + step there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the value of alpha at step 0. + steps_offset (`int`, default `0`): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, default `epsilon`): + indicates whether the model predicts the noise (epsilon), or the samples. One of `epsilon`, `sample`. + `v-prediction` is not supported for this scheduler. + dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`): + the `dtype` used for params and computation. + """ + + _compatibles = [e.name for e in FlaxKarrasDiffusionSchedulers] + + dtype: jnp.dtype + + @property + def has_state(self): + return True + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[jnp.ndarray] = None, + clip_sample: bool = True, + clip_sample_range: float = 1.0, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: str = "epsilon", + dtype: jnp.dtype = jnp.float32, + ): + self.dtype = dtype + + def create_state(self, common: Optional[CommonSchedulerState] = None) -> DDIMSchedulerState: + if common is None: + common = CommonSchedulerState.create(self) + + # At every step in ddim, we are looking into the previous alphas_cumprod + # For the final step, there is no previous alphas_cumprod because we are already at 0 + # `set_alpha_to_one` decides whether we set this parameter simply to one or + # whether we use the final alpha of the "non-previous" one. + final_alpha_cumprod = ( + jnp.array(1.0, dtype=self.dtype) if self.config.set_alpha_to_one else common.alphas_cumprod[0] + ) + + # standard deviation of the initial noise distribution + init_noise_sigma = jnp.array(1.0, dtype=self.dtype) + + timesteps = jnp.arange(0, self.config.num_train_timesteps).round()[::-1] + + return DDIMSchedulerState.create( + common=common, + final_alpha_cumprod=final_alpha_cumprod, + init_noise_sigma=init_noise_sigma, + timesteps=timesteps, + ) + + def scale_model_input( + self, state: DDIMSchedulerState, sample: jnp.ndarray, timestep: Optional[int] = None + ) -> jnp.ndarray: + """ + Args: + state (`PNDMSchedulerState`): the `FlaxPNDMScheduler` state data class instance. + sample (`jnp.ndarray`): input sample + timestep (`int`, optional): current timestep + + Returns: + `jnp.ndarray`: scaled input sample + """ + return sample + + def set_timesteps( + self, state: DDIMSchedulerState, num_inference_steps: int, shape: Tuple = () + ) -> DDIMSchedulerState: + """ + Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference. + + Args: + state (`DDIMSchedulerState`): + the `FlaxDDIMScheduler` state data class instance. + num_inference_steps (`int`): + the number of diffusion steps used when generating samples with a pre-trained model. + """ + step_ratio = self.config.num_train_timesteps // num_inference_steps + # creates integer timesteps by multiplying by ratio + # rounding to avoid issues when num_inference_step is power of 3 + timesteps = (jnp.arange(0, num_inference_steps) * step_ratio).round()[::-1] + self.config.steps_offset + + return state.replace( + num_inference_steps=num_inference_steps, + timesteps=timesteps, + ) + + def _get_variance(self, state: DDIMSchedulerState, timestep, prev_timestep): + alpha_prod_t = state.common.alphas_cumprod[timestep] + alpha_prod_t_prev = jnp.where( + prev_timestep >= 0, state.common.alphas_cumprod[prev_timestep], state.final_alpha_cumprod + ) + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + def step( + self, + state: DDIMSchedulerState, + model_output: jnp.ndarray, + timestep: int, + sample: jnp.ndarray, + eta: float = 0.0, + return_dict: bool = True, + ) -> Union[FlaxDDIMSchedulerOutput, Tuple]: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + state (`DDIMSchedulerState`): the `FlaxDDIMScheduler` state data class instance. + model_output (`jnp.ndarray`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + return_dict (`bool`): option for returning tuple rather than FlaxDDIMSchedulerOutput class + + Returns: + [`FlaxDDIMSchedulerOutput`] or `tuple`: [`FlaxDDIMSchedulerOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is the sample tensor. + + """ + if state.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502 + # Ideally, read DDIM paper in-detail understanding + + # Notation ( -> + # - pred_noise_t -> e_theta(x_t, t) + # - pred_original_sample -> f_theta(x_t, t) or x_0 + # - std_dev_t -> sigma_t + # - eta -> η + # - pred_sample_direction -> "direction pointing to x_t" + # - pred_prev_sample -> "x_t-1" + + # 1. get previous step value (=t-1) + prev_timestep = timestep - self.config.num_train_timesteps // state.num_inference_steps + + alphas_cumprod = state.common.alphas_cumprod + final_alpha_cumprod = state.final_alpha_cumprod + + # 2. compute alphas, betas + alpha_prod_t = alphas_cumprod[timestep] + alpha_prod_t_prev = jnp.where(prev_timestep >= 0, alphas_cumprod[prev_timestep], final_alpha_cumprod) + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + pred_epsilon = model_output + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction`" + ) + + # 4. Clip or threshold "predicted x_0" + if self.config.clip_sample: + pred_original_sample = pred_original_sample.clip( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 4. compute variance: "sigma_t(η)" -> see formula (16) + # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1) + variance = self._get_variance(state, timestep, prev_timestep) + std_dev_t = eta * variance ** (0.5) + + # 5. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon + + # 6. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502 + prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction + + if not return_dict: + return (prev_sample, state) + + return FlaxDDIMSchedulerOutput(prev_sample=prev_sample, state=state) + + def add_noise( + self, + state: DDIMSchedulerState, + original_samples: jnp.ndarray, + noise: jnp.ndarray, + timesteps: jnp.ndarray, + ) -> jnp.ndarray: + return add_noise_common(state.common, original_samples, noise, timesteps) + + def get_velocity( + self, + state: DDIMSchedulerState, + sample: jnp.ndarray, + noise: jnp.ndarray, + timesteps: jnp.ndarray, + ) -> jnp.ndarray: + return get_velocity_common(state.common, sample, noise, timesteps) + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim_inverse.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim_inverse.py new file mode 100644 index 0000000000000000000000000000000000000000..49dba840d08982d18225c05585e407433153327a --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim_inverse.py @@ -0,0 +1,374 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion +import math +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.schedulers.scheduling_utils import SchedulerMixin +from diffusers.utils import BaseOutput, deprecate + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM +class DDIMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas): + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + + Args: + betas (`torch.Tensor`): + the betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class DDIMInverseScheduler(SchedulerMixin, ConfigMixin): + """ + `DDIMInverseScheduler` is the reverse scheduler of [`DDIMScheduler`]. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to 1.0): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, defaults to `True`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to 0, otherwise + it uses the alpha value at step `num_train_timesteps - 1`. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + order = 1 + ignore_for_config = ["kwargs"] + _deprecated_kwargs = ["set_alpha_to_zero"] + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + clip_sample: bool = True, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: str = "epsilon", + clip_sample_range: float = 1.0, + timestep_spacing: str = "leading", + rescale_betas_zero_snr: bool = False, + **kwargs, + ): + if kwargs.get("set_alpha_to_zero", None) is not None: + deprecation_message = ( + "The `set_alpha_to_zero` argument is deprecated. Please use `set_alpha_to_one` instead." + ) + deprecate("set_alpha_to_zero", "1.0.0", deprecation_message, standard_warn=False) + set_alpha_to_one = kwargs["set_alpha_to_zero"] + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # At every step in inverted ddim, we are looking into the next alphas_cumprod + # For the initial step, there is no current alphas_cumprod, and the index is out of bounds + # `set_alpha_to_one` decides whether we set this parameter simply to one + # in this case, self.step() just output the predicted noise + # or whether we use the initial alpha used in training the diffusion model. + self.initial_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps).copy().astype(np.int64)) + + # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler.scale_model_input + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + """ + + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + + # "leading" and "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round().copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)[::-1]).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + return_dict: bool = True, + ) -> Union[DDIMSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + eta (`float`): + The weight of noise for added noise in diffusion step. + use_clipped_model_output (`bool`, defaults to `False`): + If `True`, computes "corrected" `model_output` from the clipped predicted original sample. Necessary + because predicted original sample is clipped to [-1, 1] when `self.config.clip_sample` is `True`. If no + clipping has happened, "corrected" `model_output` would coincide with the one provided as input and + `use_clipped_model_output` has no effect. + variance_noise (`torch.Tensor`): + Alternative to generating noise with `generator` by directly providing the noise for the variance + itself. Useful for methods such as [`CycleDiffusion`]. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_ddim_inverse.DDIMInverseSchedulerOutput`] or + `tuple`. + + Returns: + [`~schedulers.scheduling_ddim_inverse.DDIMInverseSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_ddim_inverse.DDIMInverseSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + + """ + # 1. get previous step value (=t+1) + prev_timestep = timestep + timestep = min( + timestep - self.config.num_train_timesteps // self.num_inference_steps, self.config.num_train_timesteps - 1 + ) + + # 2. compute alphas, betas + # change original implementation to exactly match noise levels for analogous forward process + alpha_prod_t = self.alphas_cumprod[timestep] if timestep >= 0 else self.initial_alpha_cumprod + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + pred_epsilon = model_output + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction`" + ) + + # 4. Clip or threshold "predicted x_0" + if self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 5. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_sample_direction = (1 - alpha_prod_t_prev) ** (0.5) * pred_epsilon + + # 6. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502 + prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction + + if not return_dict: + return (prev_sample, pred_original_sample) + return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim_parallel.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim_parallel.py new file mode 100644 index 0000000000000000000000000000000000000000..7c3f03a8dbe19e76946769ce8915bcbdffa7e0c3 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddim_parallel.py @@ -0,0 +1,646 @@ +# Copyright 2025 ParaDiGMS authors and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +import math +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput +class DDIMParallelSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas): + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + + Args: + betas (`torch.Tensor`): + the betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class DDIMParallelScheduler(SchedulerMixin, ConfigMixin): + """ + Denoising diffusion implicit models is a scheduler that extends the denoising procedure introduced in denoising + diffusion probabilistic models (DDPMs) with non-Markovian guidance. + + [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` + function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. + [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and + [`~SchedulerMixin.from_pretrained`] functions. + + For more details, see the original paper: https://huggingface.co/papers/2010.02502 + + Args: + num_train_timesteps (`int`): number of diffusion steps used to train the model. + beta_start (`float`): the starting `beta` value of inference. + beta_end (`float`): the final `beta` value. + beta_schedule (`str`): + the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, optional): + option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc. + clip_sample (`bool`, default `True`): + option to clip predicted sample for numerical stability. + clip_sample_range (`float`, default `1.0`): + the maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, default `True`): + each diffusion step uses the value of alphas product at that step and at the previous one. For the final + step there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the value of alpha at step 0. + steps_offset (`int`, default `0`): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, default `epsilon`, optional): + prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion + process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4 + https://imagen.research.google/video/paper.pdf) + thresholding (`bool`, default `False`): + whether to use the "dynamic thresholding" method (introduced by Imagen, + https://huggingface.co/papers/2205.11487). Note that the thresholding method is unsuitable for latent-space + diffusion models (such as stable-diffusion). + dynamic_thresholding_ratio (`float`, default `0.995`): + the ratio for the dynamic thresholding method. Default is `0.995`, the same as Imagen + (https://huggingface.co/papers/2205.11487). Valid only when `thresholding=True`. + sample_max_value (`float`, default `1.0`): + the threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`str`, default `"leading"`): + The way the timesteps should be scaled. Refer to Table 2. of [Common Diffusion Noise Schedules and Sample + Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + rescale_betas_zero_snr (`bool`, default `False`): + whether to rescale the betas to have zero terminal SNR (proposed by + https://huggingface.co/papers/2305.08891). This can enable the model to generate very bright and dark + samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + _is_ode_scheduler = True + + @register_to_config + # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler.__init__ + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + clip_sample: bool = True, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: str = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + clip_sample_range: float = 1.0, + sample_max_value: float = 1.0, + timestep_spacing: str = "leading", + rescale_betas_zero_snr: bool = False, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # At every step in ddim, we are looking into the previous alphas_cumprod + # For the final step, there is no previous alphas_cumprod because we are already at 0 + # `set_alpha_to_one` decides whether we set this parameter simply to one or + # whether we use the final alpha of the "non-previous" one. + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + + # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler.scale_model_input + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def _get_variance(self, timestep, prev_timestep=None): + if prev_timestep is None: + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + def _batch_get_variance(self, t, prev_t): + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[torch.clip(prev_t, min=0)] + alpha_prod_t_prev[prev_t < 0] = torch.tensor(1.0) + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler.set_timesteps + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + """ + + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps) + .round()[::-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + eta: float = 0.0, + use_clipped_model_output: bool = False, + generator=None, + variance_noise: Optional[torch.Tensor] = None, + return_dict: bool = True, + ) -> Union[DDIMParallelSchedulerOutput, Tuple]: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + current instance of sample being created by diffusion process. + eta (`float`): weight of noise for added noise in diffusion step. + use_clipped_model_output (`bool`): if `True`, compute "corrected" `model_output` from the clipped + predicted original sample. Necessary because predicted original sample is clipped to [-1, 1] when + `self.config.clip_sample` is `True`. If no clipping has happened, "corrected" `model_output` would + coincide with the one provided as input and `use_clipped_model_output` will have not effect. + generator: random number generator. + variance_noise (`torch.Tensor`): instead of generating noise for the variance using `generator`, we + can directly provide the noise for the variance itself. This is useful for methods such as + CycleDiffusion. (https://huggingface.co/papers/2210.05559) + return_dict (`bool`): option for returning tuple rather than DDIMParallelSchedulerOutput class + + Returns: + [`~schedulers.scheduling_utils.DDIMParallelSchedulerOutput`] or `tuple`: + [`~schedulers.scheduling_utils.DDIMParallelSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. + When returning a tuple, the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502 + # Ideally, read DDIM paper in-detail understanding + + # Notation ( -> + # - pred_noise_t -> e_theta(x_t, t) + # - pred_original_sample -> f_theta(x_t, t) or x_0 + # - std_dev_t -> sigma_t + # - eta -> η + # - pred_sample_direction -> "direction pointing to x_t" + # - pred_prev_sample -> "x_t-1" + + # 1. get previous step value (=t-1) + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + # 2. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + pred_epsilon = model_output + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction`" + ) + + # 4. Clip or threshold "predicted x_0" + if self.config.thresholding: + pred_original_sample = self._threshold_sample(pred_original_sample) + elif self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 5. compute variance: "sigma_t(η)" -> see formula (16) + # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1) + variance = self._get_variance(timestep, prev_timestep) + std_dev_t = eta * variance ** (0.5) + + if use_clipped_model_output: + # the pred_epsilon is always re-derived from the clipped x_0 in Glide + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + + # 6. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon + + # 7. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502 + prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction + + if eta > 0: + if variance_noise is not None and generator is not None: + raise ValueError( + "Cannot pass both generator and variance_noise. Please make sure that either `generator` or" + " `variance_noise` stays `None`." + ) + + if variance_noise is None: + variance_noise = randn_tensor( + model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype + ) + variance = std_dev_t * variance_noise + + prev_sample = prev_sample + variance + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return DDIMParallelSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + def batch_step_no_noise( + self, + model_output: torch.Tensor, + timesteps: List[int], + sample: torch.Tensor, + eta: float = 0.0, + use_clipped_model_output: bool = False, + ) -> torch.Tensor: + """ + Batched version of the `step` function, to be able to reverse the SDE for multiple samples/timesteps at once. + Also, does not add any noise to the predicted sample, which is necessary for parallel sampling where the noise + is pre-sampled by the pipeline. + + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): direct output from learned diffusion model. + timesteps (`List[int]`): + current discrete timesteps in the diffusion chain. This is now a list of integers. + sample (`torch.Tensor`): + current instance of sample being created by diffusion process. + eta (`float`): weight of noise for added noise in diffusion step. + use_clipped_model_output (`bool`): if `True`, compute "corrected" `model_output` from the clipped + predicted original sample. Necessary because predicted original sample is clipped to [-1, 1] when + `self.config.clip_sample` is `True`. If no clipping has happened, "corrected" `model_output` would + coincide with the one provided as input and `use_clipped_model_output` will have not effect. + + Returns: + `torch.Tensor`: sample tensor at previous timestep. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + assert eta == 0.0 + + # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502 + # Ideally, read DDIM paper in-detail understanding + + # Notation ( -> + # - pred_noise_t -> e_theta(x_t, t) + # - pred_original_sample -> f_theta(x_t, t) or x_0 + # - std_dev_t -> sigma_t + # - eta -> η + # - pred_sample_direction -> "direction pointing to x_t" + # - pred_prev_sample -> "x_t-1" + + # 1. get previous step value (=t-1) + t = timesteps + prev_t = t - self.config.num_train_timesteps // self.num_inference_steps + + t = t.view(-1, *([1] * (model_output.ndim - 1))) + prev_t = prev_t.view(-1, *([1] * (model_output.ndim - 1))) + + # 1. compute alphas, betas + self.alphas_cumprod = self.alphas_cumprod.to(model_output.device) + self.final_alpha_cumprod = self.final_alpha_cumprod.to(model_output.device) + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[torch.clip(prev_t, min=0)] + alpha_prod_t_prev[prev_t < 0] = torch.tensor(1.0) + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + pred_epsilon = model_output + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction`" + ) + + # 4. Clip or threshold "predicted x_0" + if self.config.thresholding: + pred_original_sample = self._threshold_sample(pred_original_sample) + elif self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 5. compute variance: "sigma_t(η)" -> see formula (16) + # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1) + variance = self._batch_get_variance(t, prev_t).to(model_output.device).view(*alpha_prod_t_prev.shape) + std_dev_t = eta * variance ** (0.5) + + if use_clipped_model_output: + # the pred_epsilon is always re-derived from the clipped x_0 in Glide + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + + # 6. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon + + # 7. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502 + prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction + + return prev_sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddpm.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddpm.py new file mode 100644 index 0000000000000000000000000000000000000000..0fab6d910a822e688cd8f2c9ac162b2bf1d0d269 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddpm.py @@ -0,0 +1,558 @@ +# Copyright 2025 UC Berkeley Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim + +import math +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +@dataclass +class DDPMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas): + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + + Args: + betas (`torch.Tensor`): + the betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class DDPMScheduler(SchedulerMixin, ConfigMixin): + """ + `DDPMScheduler` explores the connections between denoising score matching and Langevin dynamics sampling. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, `squaredcos_cap_v2`, or `sigmoid`. + trained_betas (`np.ndarray`, *optional*): + An array of betas to pass directly to the constructor without using `beta_start` and `beta_end`. + variance_type (`str`, defaults to `"fixed_small"`): + Clip the variance when adding noise to the denoised sample. Choose from `fixed_small`, `fixed_small_log`, + `fixed_large`, `fixed_large_log`, `learned` or `learned_range`. + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to 1.0): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + variance_type: str = "fixed_small", + clip_sample: bool = True, + prediction_type: str = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + clip_sample_range: float = 1.0, + sample_max_value: float = 1.0, + timestep_spacing: str = "leading", + steps_offset: int = 0, + rescale_betas_zero_snr: bool = False, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + elif beta_schedule == "sigmoid": + # GeoDiff sigmoid schedule + betas = torch.linspace(-6, 6, num_train_timesteps) + self.betas = torch.sigmoid(betas) * (beta_end - beta_start) + beta_start + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + self.one = torch.tensor(1.0) + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.custom_timesteps = False + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy()) + + self.variance_type = variance_type + + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps( + self, + num_inference_steps: Optional[int] = None, + device: Union[str, torch.device] = None, + timesteps: Optional[List[int]] = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, + `timesteps` must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of equal spacing between timesteps is used. If `timesteps` is passed, + `num_inference_steps` must be `None`. + + """ + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.") + + if timesteps is not None: + for i in range(1, len(timesteps)): + if timesteps[i] >= timesteps[i - 1]: + raise ValueError("`custom_timesteps` must be in descending order.") + + if timesteps[0] >= self.config.num_train_timesteps: + raise ValueError( + f"`timesteps` must start before `self.config.train_timesteps`: {self.config.num_train_timesteps}." + ) + + timesteps = np.array(timesteps, dtype=np.int64) + self.custom_timesteps = True + else: + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + self.custom_timesteps = False + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps) + .round()[::-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + def _get_variance(self, t, predicted_variance=None, variance_type=None): + prev_t = self.previous_timestep(t) + + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one + current_beta_t = 1 - alpha_prod_t / alpha_prod_t_prev + + # For t > 0, compute predicted variance βt (see formula (6) and (7) from https://huggingface.co/papers/2006.11239) + # and sample from it to get previous sample + # x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample + variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * current_beta_t + + # we always take the log of variance, so clamp it to ensure it's not 0 + variance = torch.clamp(variance, min=1e-20) + + if variance_type is None: + variance_type = self.config.variance_type + + # hacks - were probably added for training stability + if variance_type == "fixed_small": + variance = variance + # for rl-diffuser https://huggingface.co/papers/2205.09991 + elif variance_type == "fixed_small_log": + variance = torch.log(variance) + variance = torch.exp(0.5 * variance) + elif variance_type == "fixed_large": + variance = current_beta_t + elif variance_type == "fixed_large_log": + # Glide max_log + variance = torch.log(current_beta_t) + elif variance_type == "learned": + return predicted_variance + elif variance_type == "learned_range": + min_log = torch.log(variance) + max_log = torch.log(current_beta_t) + frac = (predicted_variance + 1) / 2 + variance = frac * max_log + (1 - frac) * min_log + + return variance + + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + generator=None, + return_dict: bool = True, + ) -> Union[DDPMSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + t = timestep + + prev_t = self.previous_timestep(t) + + if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type in ["learned", "learned_range"]: + model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1) + else: + predicted_variance = None + + # 1. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + current_alpha_t = alpha_prod_t / alpha_prod_t_prev + current_beta_t = 1 - current_alpha_t + + # 2. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (15) from https://huggingface.co/papers/2006.11239 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or" + " `v_prediction` for the DDPMScheduler." + ) + + # 3. Clip or threshold "predicted x_0" + if self.config.thresholding: + pred_original_sample = self._threshold_sample(pred_original_sample) + elif self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * current_beta_t) / beta_prod_t + current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t + + # 5. Compute predicted previous sample µ_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample + + # 6. Add noise + variance = 0 + if t > 0: + device = model_output.device + variance_noise = randn_tensor( + model_output.shape, generator=generator, device=device, dtype=model_output.dtype + ) + if self.variance_type == "fixed_small_log": + variance = self._get_variance(t, predicted_variance=predicted_variance) * variance_noise + elif self.variance_type == "learned_range": + variance = self._get_variance(t, predicted_variance=predicted_variance) + variance = torch.exp(0.5 * variance) * variance_noise + else: + variance = (self._get_variance(t, predicted_variance=predicted_variance) ** 0.5) * variance_noise + + pred_prev_sample = pred_prev_sample + variance + + if not return_dict: + return ( + pred_prev_sample, + pred_original_sample, + ) + + return DDPMSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample) + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self): + return self.config.num_train_timesteps + + def previous_timestep(self, timestep): + if self.custom_timesteps or self.num_inference_steps: + index = (self.timesteps == timestep).nonzero(as_tuple=True)[0][0] + if index == self.timesteps.shape[0] - 1: + prev_t = torch.tensor(-1) + else: + prev_t = self.timesteps[index + 1] + else: + prev_t = timestep - 1 + return prev_t diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddpm_flax.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddpm_flax.py new file mode 100644 index 0000000000000000000000000000000000000000..a3264f54f572f60c1d44898bd220bf512684f1cb --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddpm_flax.py @@ -0,0 +1,303 @@ +# Copyright 2025 UC Berkeley Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim + +from dataclasses import dataclass +from typing import Optional, Tuple, Union + +import flax +import jax +import jax.numpy as jnp + +from ..configuration_utils import ConfigMixin, register_to_config +from .scheduling_utils_flax import ( + CommonSchedulerState, + FlaxKarrasDiffusionSchedulers, + FlaxSchedulerMixin, + FlaxSchedulerOutput, + add_noise_common, + get_velocity_common, +) + + +@flax.struct.dataclass +class DDPMSchedulerState: + common: CommonSchedulerState + + # setable values + init_noise_sigma: jnp.ndarray + timesteps: jnp.ndarray + num_inference_steps: Optional[int] = None + + @classmethod + def create(cls, common: CommonSchedulerState, init_noise_sigma: jnp.ndarray, timesteps: jnp.ndarray): + return cls(common=common, init_noise_sigma=init_noise_sigma, timesteps=timesteps) + + +@dataclass +class FlaxDDPMSchedulerOutput(FlaxSchedulerOutput): + state: DDPMSchedulerState + + +class FlaxDDPMScheduler(FlaxSchedulerMixin, ConfigMixin): + """ + Denoising diffusion probabilistic models (DDPMs) explores the connections between denoising score matching and + Langevin dynamics sampling. + + [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` + function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. + [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and + [`~SchedulerMixin.from_pretrained`] functions. + + For more details, see the original paper: https://huggingface.co/papers/2006.11239 + + Args: + num_train_timesteps (`int`): number of diffusion steps used to train the model. + beta_start (`float`): the starting `beta` value of inference. + beta_end (`float`): the final `beta` value. + beta_schedule (`str`): + the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, optional): + option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc. + variance_type (`str`): + options to clip the variance used when adding noise to the denoised sample. Choose from `fixed_small`, + `fixed_small_log`, `fixed_large`, `fixed_large_log`, `learned` or `learned_range`. + clip_sample (`bool`, default `True`): + option to clip predicted sample between -1 and 1 for numerical stability. + prediction_type (`str`, default `epsilon`): + indicates whether the model predicts the noise (epsilon), or the samples. One of `epsilon`, `sample`. + `v-prediction` is not supported for this scheduler. + dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`): + the `dtype` used for params and computation. + """ + + _compatibles = [e.name for e in FlaxKarrasDiffusionSchedulers] + + dtype: jnp.dtype + + @property + def has_state(self): + return True + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[jnp.ndarray] = None, + variance_type: str = "fixed_small", + clip_sample: bool = True, + prediction_type: str = "epsilon", + dtype: jnp.dtype = jnp.float32, + ): + self.dtype = dtype + + def create_state(self, common: Optional[CommonSchedulerState] = None) -> DDPMSchedulerState: + if common is None: + common = CommonSchedulerState.create(self) + + # standard deviation of the initial noise distribution + init_noise_sigma = jnp.array(1.0, dtype=self.dtype) + + timesteps = jnp.arange(0, self.config.num_train_timesteps).round()[::-1] + + return DDPMSchedulerState.create( + common=common, + init_noise_sigma=init_noise_sigma, + timesteps=timesteps, + ) + + def scale_model_input( + self, state: DDPMSchedulerState, sample: jnp.ndarray, timestep: Optional[int] = None + ) -> jnp.ndarray: + """ + Args: + state (`PNDMSchedulerState`): the `FlaxPNDMScheduler` state data class instance. + sample (`jnp.ndarray`): input sample + timestep (`int`, optional): current timestep + + Returns: + `jnp.ndarray`: scaled input sample + """ + return sample + + def set_timesteps( + self, state: DDPMSchedulerState, num_inference_steps: int, shape: Tuple = () + ) -> DDPMSchedulerState: + """ + Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference. + + Args: + state (`DDIMSchedulerState`): + the `FlaxDDPMScheduler` state data class instance. + num_inference_steps (`int`): + the number of diffusion steps used when generating samples with a pre-trained model. + """ + + step_ratio = self.config.num_train_timesteps // num_inference_steps + # creates integer timesteps by multiplying by ratio + # rounding to avoid issues when num_inference_step is power of 3 + timesteps = (jnp.arange(0, num_inference_steps) * step_ratio).round()[::-1] + + return state.replace( + num_inference_steps=num_inference_steps, + timesteps=timesteps, + ) + + def _get_variance(self, state: DDPMSchedulerState, t, predicted_variance=None, variance_type=None): + alpha_prod_t = state.common.alphas_cumprod[t] + alpha_prod_t_prev = jnp.where(t > 0, state.common.alphas_cumprod[t - 1], jnp.array(1.0, dtype=self.dtype)) + + # For t > 0, compute predicted variance βt (see formula (6) and (7) from https://huggingface.co/papers/2006.11239) + # and sample from it to get previous sample + # x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample + variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * state.common.betas[t] + + if variance_type is None: + variance_type = self.config.variance_type + + # hacks - were probably added for training stability + if variance_type == "fixed_small": + variance = jnp.clip(variance, a_min=1e-20) + # for rl-diffuser https://huggingface.co/papers/2205.09991 + elif variance_type == "fixed_small_log": + variance = jnp.log(jnp.clip(variance, a_min=1e-20)) + elif variance_type == "fixed_large": + variance = state.common.betas[t] + elif variance_type == "fixed_large_log": + # Glide max_log + variance = jnp.log(state.common.betas[t]) + elif variance_type == "learned": + return predicted_variance + elif variance_type == "learned_range": + min_log = variance + max_log = state.common.betas[t] + frac = (predicted_variance + 1) / 2 + variance = frac * max_log + (1 - frac) * min_log + + return variance + + def step( + self, + state: DDPMSchedulerState, + model_output: jnp.ndarray, + timestep: int, + sample: jnp.ndarray, + key: Optional[jax.Array] = None, + return_dict: bool = True, + ) -> Union[FlaxDDPMSchedulerOutput, Tuple]: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + state (`DDPMSchedulerState`): the `FlaxDDPMScheduler` state data class instance. + model_output (`jnp.ndarray`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + key (`jax.Array`): a PRNG key. + return_dict (`bool`): option for returning tuple rather than FlaxDDPMSchedulerOutput class + + Returns: + [`FlaxDDPMSchedulerOutput`] or `tuple`: [`FlaxDDPMSchedulerOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is the sample tensor. + + """ + t = timestep + + if key is None: + key = jax.random.key(0) + + if ( + len(model_output.shape) > 1 + and model_output.shape[1] == sample.shape[1] * 2 + and self.config.variance_type in ["learned", "learned_range"] + ): + model_output, predicted_variance = jnp.split(model_output, sample.shape[1], axis=1) + else: + predicted_variance = None + + # 1. compute alphas, betas + alpha_prod_t = state.common.alphas_cumprod[t] + alpha_prod_t_prev = jnp.where(t > 0, state.common.alphas_cumprod[t - 1], jnp.array(1.0, dtype=self.dtype)) + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + # 2. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (15) from https://huggingface.co/papers/2006.11239 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` " + " for the FlaxDDPMScheduler." + ) + + # 3. Clip "predicted x_0" + if self.config.clip_sample: + pred_original_sample = jnp.clip(pred_original_sample, -1, 1) + + # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * state.common.betas[t]) / beta_prod_t + current_sample_coeff = state.common.alphas[t] ** (0.5) * beta_prod_t_prev / beta_prod_t + + # 5. Compute predicted previous sample µ_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample + + # 6. Add noise + def random_variance(): + split_key = jax.random.split(key, num=1)[0] + noise = jax.random.normal(split_key, shape=model_output.shape, dtype=self.dtype) + return (self._get_variance(state, t, predicted_variance=predicted_variance) ** 0.5) * noise + + variance = jnp.where(t > 0, random_variance(), jnp.zeros(model_output.shape, dtype=self.dtype)) + + pred_prev_sample = pred_prev_sample + variance + + if not return_dict: + return (pred_prev_sample, state) + + return FlaxDDPMSchedulerOutput(prev_sample=pred_prev_sample, state=state) + + def add_noise( + self, + state: DDPMSchedulerState, + original_samples: jnp.ndarray, + noise: jnp.ndarray, + timesteps: jnp.ndarray, + ) -> jnp.ndarray: + return add_noise_common(state.common, original_samples, noise, timesteps) + + def get_velocity( + self, + state: DDPMSchedulerState, + sample: jnp.ndarray, + noise: jnp.ndarray, + timesteps: jnp.ndarray, + ) -> jnp.ndarray: + return get_velocity_common(state.common, sample, noise, timesteps) + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddpm_parallel.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddpm_parallel.py new file mode 100644 index 0000000000000000000000000000000000000000..ec741f9ecb7dd1f28fde6b53a9c7f5c9fa43b5c9 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddpm_parallel.py @@ -0,0 +1,649 @@ +# Copyright 2025 ParaDiGMS authors and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim + +import math +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput +class DDPMParallelSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas): + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + + Args: + betas (`torch.Tensor`): + the betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class DDPMParallelScheduler(SchedulerMixin, ConfigMixin): + """ + Denoising diffusion probabilistic models (DDPMs) explores the connections between denoising score matching and + Langevin dynamics sampling. + + [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` + function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. + [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and + [`~SchedulerMixin.from_pretrained`] functions. + + For more details, see the original paper: https://huggingface.co/papers/2006.11239 + + Args: + num_train_timesteps (`int`): number of diffusion steps used to train the model. + beta_start (`float`): the starting `beta` value of inference. + beta_end (`float`): the final `beta` value. + beta_schedule (`str`): + the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, `squaredcos_cap_v2` or `sigmoid`. + trained_betas (`np.ndarray`, optional): + option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc. + variance_type (`str`): + options to clip the variance used when adding noise to the denoised sample. Choose from `fixed_small`, + `fixed_small_log`, `fixed_large`, `fixed_large_log`, `learned` or `learned_range`. + clip_sample (`bool`, default `True`): + option to clip predicted sample for numerical stability. + clip_sample_range (`float`, default `1.0`): + the maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + prediction_type (`str`, default `epsilon`, optional): + prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion + process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4 + https://imagen.research.google/video/paper.pdf) + thresholding (`bool`, default `False`): + whether to use the "dynamic thresholding" method (introduced by Imagen, + https://huggingface.co/papers/2205.11487). Note that the thresholding method is unsuitable for latent-space + diffusion models (such as stable-diffusion). + dynamic_thresholding_ratio (`float`, default `0.995`): + the ratio for the dynamic thresholding method. Default is `0.995`, the same as Imagen + (https://huggingface.co/papers/2205.11487). Valid only when `thresholding=True`. + sample_max_value (`float`, default `1.0`): + the threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`str`, default `"leading"`): + The way the timesteps should be scaled. Refer to Table 2. of [Common Diffusion Noise Schedules and Sample + Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, default `0`): + An offset added to the inference steps, as required by some model families. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + _is_ode_scheduler = False + + @register_to_config + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.__init__ + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + variance_type: str = "fixed_small", + clip_sample: bool = True, + prediction_type: str = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + clip_sample_range: float = 1.0, + sample_max_value: float = 1.0, + timestep_spacing: str = "leading", + steps_offset: int = 0, + rescale_betas_zero_snr: bool = False, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + elif beta_schedule == "sigmoid": + # GeoDiff sigmoid schedule + betas = torch.linspace(-6, 6, num_train_timesteps) + self.betas = torch.sigmoid(betas) * (beta_end - beta_start) + beta_start + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + self.one = torch.tensor(1.0) + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.custom_timesteps = False + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy()) + + self.variance_type = variance_type + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.scale_model_input + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.set_timesteps + def set_timesteps( + self, + num_inference_steps: Optional[int] = None, + device: Union[str, torch.device] = None, + timesteps: Optional[List[int]] = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, + `timesteps` must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of equal spacing between timesteps is used. If `timesteps` is passed, + `num_inference_steps` must be `None`. + + """ + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.") + + if timesteps is not None: + for i in range(1, len(timesteps)): + if timesteps[i] >= timesteps[i - 1]: + raise ValueError("`custom_timesteps` must be in descending order.") + + if timesteps[0] >= self.config.num_train_timesteps: + raise ValueError( + f"`timesteps` must start before `self.config.train_timesteps`: {self.config.num_train_timesteps}." + ) + + timesteps = np.array(timesteps, dtype=np.int64) + self.custom_timesteps = True + else: + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + self.custom_timesteps = False + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps) + .round()[::-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._get_variance + def _get_variance(self, t, predicted_variance=None, variance_type=None): + prev_t = self.previous_timestep(t) + + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one + current_beta_t = 1 - alpha_prod_t / alpha_prod_t_prev + + # For t > 0, compute predicted variance βt (see formula (6) and (7) from https://huggingface.co/papers/2006.11239) + # and sample from it to get previous sample + # x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample + variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * current_beta_t + + # we always take the log of variance, so clamp it to ensure it's not 0 + variance = torch.clamp(variance, min=1e-20) + + if variance_type is None: + variance_type = self.config.variance_type + + # hacks - were probably added for training stability + if variance_type == "fixed_small": + variance = variance + # for rl-diffuser https://huggingface.co/papers/2205.09991 + elif variance_type == "fixed_small_log": + variance = torch.log(variance) + variance = torch.exp(0.5 * variance) + elif variance_type == "fixed_large": + variance = current_beta_t + elif variance_type == "fixed_large_log": + # Glide max_log + variance = torch.log(current_beta_t) + elif variance_type == "learned": + return predicted_variance + elif variance_type == "learned_range": + min_log = torch.log(variance) + max_log = torch.log(current_beta_t) + frac = (predicted_variance + 1) / 2 + variance = frac * max_log + (1 - frac) * min_log + + return variance + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + generator=None, + return_dict: bool = True, + ) -> Union[DDPMParallelSchedulerOutput, Tuple]: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + current instance of sample being created by diffusion process. + generator: random number generator. + return_dict (`bool`): option for returning tuple rather than DDPMParallelSchedulerOutput class + + Returns: + [`~schedulers.scheduling_utils.DDPMParallelSchedulerOutput`] or `tuple`: + [`~schedulers.scheduling_utils.DDPMParallelSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. + When returning a tuple, the first element is the sample tensor. + + """ + t = timestep + + prev_t = self.previous_timestep(t) + + if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type in ["learned", "learned_range"]: + model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1) + else: + predicted_variance = None + + # 1. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + current_alpha_t = alpha_prod_t / alpha_prod_t_prev + current_beta_t = 1 - current_alpha_t + + # 2. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (15) from https://huggingface.co/papers/2006.11239 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or" + " `v_prediction` for the DDPMScheduler." + ) + + # 3. Clip or threshold "predicted x_0" + if self.config.thresholding: + pred_original_sample = self._threshold_sample(pred_original_sample) + elif self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * current_beta_t) / beta_prod_t + current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t + + # 5. Compute predicted previous sample µ_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample + + # 6. Add noise + variance = 0 + if t > 0: + device = model_output.device + variance_noise = randn_tensor( + model_output.shape, generator=generator, device=device, dtype=model_output.dtype + ) + if self.variance_type == "fixed_small_log": + variance = self._get_variance(t, predicted_variance=predicted_variance) * variance_noise + elif self.variance_type == "learned_range": + variance = self._get_variance(t, predicted_variance=predicted_variance) + variance = torch.exp(0.5 * variance) * variance_noise + else: + variance = (self._get_variance(t, predicted_variance=predicted_variance) ** 0.5) * variance_noise + + pred_prev_sample = pred_prev_sample + variance + + if not return_dict: + return ( + pred_prev_sample, + pred_original_sample, + ) + + return DDPMParallelSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample) + + def batch_step_no_noise( + self, + model_output: torch.Tensor, + timesteps: List[int], + sample: torch.Tensor, + ) -> torch.Tensor: + """ + Batched version of the `step` function, to be able to reverse the SDE for multiple samples/timesteps at once. + Also, does not add any noise to the predicted sample, which is necessary for parallel sampling where the noise + is pre-sampled by the pipeline. + + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): direct output from learned diffusion model. + timesteps (`List[int]`): + current discrete timesteps in the diffusion chain. This is now a list of integers. + sample (`torch.Tensor`): + current instance of sample being created by diffusion process. + + Returns: + `torch.Tensor`: sample tensor at previous timestep. + """ + t = timesteps + num_inference_steps = self.num_inference_steps if self.num_inference_steps else self.config.num_train_timesteps + prev_t = t - self.config.num_train_timesteps // num_inference_steps + + t = t.view(-1, *([1] * (model_output.ndim - 1))) + prev_t = prev_t.view(-1, *([1] * (model_output.ndim - 1))) + + if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type in ["learned", "learned_range"]: + model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1) + else: + pass + + # 1. compute alphas, betas + self.alphas_cumprod = self.alphas_cumprod.to(model_output.device) + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[torch.clip(prev_t, min=0)] + alpha_prod_t_prev[prev_t < 0] = torch.tensor(1.0) + + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + current_alpha_t = alpha_prod_t / alpha_prod_t_prev + current_beta_t = 1 - current_alpha_t + + # 2. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (15) from https://huggingface.co/papers/2006.11239 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or" + " `v_prediction` for the DDPMParallelScheduler." + ) + + # 3. Clip or threshold "predicted x_0" + if self.config.thresholding: + pred_original_sample = self._threshold_sample(pred_original_sample) + elif self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * current_beta_t) / beta_prod_t + current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t + + # 5. Compute predicted previous sample µ_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample + + return pred_prev_sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self): + return self.config.num_train_timesteps + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.previous_timestep + def previous_timestep(self, timestep): + if self.custom_timesteps or self.num_inference_steps: + index = (self.timesteps == timestep).nonzero(as_tuple=True)[0][0] + if index == self.timesteps.shape[0] - 1: + prev_t = torch.tensor(-1) + else: + prev_t = self.timesteps[index + 1] + else: + prev_t = timestep - 1 + return prev_t diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddpm_wuerstchen.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddpm_wuerstchen.py new file mode 100644 index 0000000000000000000000000000000000000000..71f08277ebd71fc363dcddc4c8dd5fb6fa250e2b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ddpm_wuerstchen.py @@ -0,0 +1,230 @@ +# Copyright (c) 2022 Pablo Pernías MIT License +# Copyright 2025 UC Berkeley Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim + +import math +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +@dataclass +class DDPMWuerstchenSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's step function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.Tensor + + +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class DDPMWuerstchenScheduler(SchedulerMixin, ConfigMixin): + """ + Denoising diffusion probabilistic models (DDPMs) explores the connections between denoising score matching and + Langevin dynamics sampling. + + [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` + function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. + [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and + [`~SchedulerMixin.from_pretrained`] functions. + + For more details, see the original paper: https://huggingface.co/papers/2006.11239 + + Args: + scaler (`float`): .... + s (`float`): .... + """ + + @register_to_config + def __init__( + self, + scaler: float = 1.0, + s: float = 0.008, + ): + self.scaler = scaler + self.s = torch.tensor([s]) + self._init_alpha_cumprod = torch.cos(self.s / (1 + self.s) * torch.pi * 0.5) ** 2 + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + def _alpha_cumprod(self, t, device): + if self.scaler > 1: + t = 1 - (1 - t) ** self.scaler + elif self.scaler < 1: + t = t**self.scaler + alpha_cumprod = torch.cos( + (t + self.s.to(device)) / (1 + self.s.to(device)) * torch.pi * 0.5 + ) ** 2 / self._init_alpha_cumprod.to(device) + return alpha_cumprod.clamp(0.0001, 0.9999) + + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): input sample + timestep (`int`, optional): current timestep + + Returns: + `torch.Tensor`: scaled input sample + """ + return sample + + def set_timesteps( + self, + num_inference_steps: int = None, + timesteps: Optional[List[int]] = None, + device: Union[str, torch.device] = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference. + + Args: + num_inference_steps (`Dict[float, int]`): + the number of diffusion steps used when generating samples with a pre-trained model. If passed, then + `timesteps` must be `None`. + device (`str` or `torch.device`, optional): + the device to which the timesteps are moved to. {2 / 3: 20, 0.0: 10} + """ + if timesteps is None: + timesteps = torch.linspace(1.0, 0.0, num_inference_steps + 1, device=device) + if not isinstance(timesteps, torch.Tensor): + timesteps = torch.Tensor(timesteps).to(device) + self.timesteps = timesteps + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + generator=None, + return_dict: bool = True, + ) -> Union[DDPMWuerstchenSchedulerOutput, Tuple]: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + current instance of sample being created by diffusion process. + generator: random number generator. + return_dict (`bool`): option for returning tuple rather than DDPMWuerstchenSchedulerOutput class + + Returns: + [`DDPMWuerstchenSchedulerOutput`] or `tuple`: [`DDPMWuerstchenSchedulerOutput`] if `return_dict` is True, + otherwise a `tuple`. When returning a tuple, the first element is the sample tensor. + + """ + dtype = model_output.dtype + device = model_output.device + t = timestep + + prev_t = self.previous_timestep(t) + + alpha_cumprod = self._alpha_cumprod(t, device).view(t.size(0), *[1 for _ in sample.shape[1:]]) + alpha_cumprod_prev = self._alpha_cumprod(prev_t, device).view(prev_t.size(0), *[1 for _ in sample.shape[1:]]) + alpha = alpha_cumprod / alpha_cumprod_prev + + mu = (1.0 / alpha).sqrt() * (sample - (1 - alpha) * model_output / (1 - alpha_cumprod).sqrt()) + + std_noise = randn_tensor(mu.shape, generator=generator, device=model_output.device, dtype=model_output.dtype) + std = ((1 - alpha) * (1.0 - alpha_cumprod_prev) / (1.0 - alpha_cumprod)).sqrt() * std_noise + pred = mu + std * (prev_t != 0).float().view(prev_t.size(0), *[1 for _ in sample.shape[1:]]) + + if not return_dict: + return (pred.to(dtype),) + + return DDPMWuerstchenSchedulerOutput(prev_sample=pred.to(dtype)) + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + device = original_samples.device + dtype = original_samples.dtype + alpha_cumprod = self._alpha_cumprod(timesteps, device=device).view( + timesteps.size(0), *[1 for _ in original_samples.shape[1:]] + ) + noisy_samples = alpha_cumprod.sqrt() * original_samples + (1 - alpha_cumprod).sqrt() * noise + return noisy_samples.to(dtype=dtype) + + def __len__(self): + return self.config.num_train_timesteps + + def previous_timestep(self, timestep): + index = (self.timesteps - timestep[0]).abs().argmin().item() + prev_t = self.timesteps[index + 1][None].expand(timestep.shape[0]) + return prev_t diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_deis_multistep.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_deis_multistep.py new file mode 100644 index 0000000000000000000000000000000000000000..748a7e39c0b89414dcdce6c37e63fb4816201f98 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_deis_multistep.py @@ -0,0 +1,886 @@ +# Copyright 2025 FLAIR Lab and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: check https://huggingface.co/papers/2204.13902 and https://github.com/qsh-zh/deis for more info +# The codebase is modified based on https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_dpmsolver_multistep.py + +import math +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import deprecate, is_scipy_available +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +if is_scipy_available(): + import scipy.stats + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class DEISMultistepScheduler(SchedulerMixin, ConfigMixin): + """ + `DEISMultistepScheduler` is a fast high order solver for diffusion ordinary differential equations (ODEs). + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + solver_order (`int`, defaults to 2): + The DEIS order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided + sampling, and `solver_order=3` for unconditional sampling. + prediction_type (`str`, defaults to `epsilon`): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + algorithm_type (`str`, defaults to `deis`): + The algorithm type for the solver. + lower_order_final (`bool`, defaults to `True`): + Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[np.ndarray] = None, + solver_order: int = 2, + prediction_type: str = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + algorithm_type: str = "deis", + solver_type: str = "logrho", + lower_order_final: bool = True, + use_karras_sigmas: Optional[bool] = False, + use_exponential_sigmas: Optional[bool] = False, + use_beta_sigmas: Optional[bool] = False, + use_flow_sigmas: Optional[bool] = False, + flow_shift: Optional[float] = 1.0, + timestep_spacing: str = "linspace", + steps_offset: int = 0, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + # Currently we only support VP-type noise schedule + self.alpha_t = torch.sqrt(self.alphas_cumprod) + self.sigma_t = torch.sqrt(1 - self.alphas_cumprod) + self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t) + self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5 + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # settings for DEIS + if algorithm_type not in ["deis"]: + if algorithm_type in ["dpmsolver", "dpmsolver++"]: + self.register_to_config(algorithm_type="deis") + else: + raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}") + + if solver_type not in ["logrho"]: + if solver_type in ["midpoint", "heun", "bh1", "bh2"]: + self.register_to_config(solver_type="logrho") + else: + raise NotImplementedError(f"solver type {solver_type} is not implemented for {self.__class__}") + + # setable values + self.num_inference_steps = None + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps) + self.model_outputs = [None] * solver_order + self.lower_order_nums = 0 + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps + 1) + .round()[::-1][:-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // (num_inference_steps + 1) + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.arange(self.config.num_train_timesteps, 0, -step_ratio).round().copy().astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + if self.config.use_karras_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round() + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_exponential_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_beta_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_flow_sigmas: + alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1) + sigmas = 1.0 - alphas + sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy() + timesteps = (sigmas * self.config.num_train_timesteps).copy() + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + else: + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5 + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + + self.sigmas = torch.from_numpy(sigmas) + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64) + + self.num_inference_steps = len(timesteps) + + self.model_outputs = [ + None, + ] * self.config.solver_order + self.lower_order_nums = 0 + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t + def _sigma_to_alpha_sigma_t(self, sigma): + if self.config.use_flow_sigmas: + alpha_t = 1 - sigma + sigma_t = sigma + else: + alpha_t = 1 / ((sigma**2 + 1) ** 0.5) + sigma_t = sigma * alpha_t + + return alpha_t, sigma_t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """Constructs an exponential noise schedule.""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def convert_model_output( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + """ + Convert the model output to the corresponding type the DEIS algorithm needs. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + if self.config.prediction_type == "epsilon": + x0_pred = (sample - sigma_t * model_output) / alpha_t + elif self.config.prediction_type == "sample": + x0_pred = model_output + elif self.config.prediction_type == "v_prediction": + x0_pred = alpha_t * sample - sigma_t * model_output + elif self.config.prediction_type == "flow_prediction": + sigma_t = self.sigmas[self.step_index] + x0_pred = sample - sigma_t * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + "`v_prediction`, or `flow_prediction` for the DEISMultistepScheduler." + ) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + if self.config.algorithm_type == "deis": + return (sample - alpha_t * x0_pred) / sigma_t + else: + raise NotImplementedError("only support log-rho multistep deis now") + + def deis_first_order_update( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the first-order DEIS (equivalent to DDIM). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + prev_timestep (`int`): + The previous discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s = torch.log(alpha_s) - torch.log(sigma_s) + + h = lambda_t - lambda_s + if self.config.algorithm_type == "deis": + x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output + else: + raise NotImplementedError("only support log-rho multistep deis now") + return x_t + + def multistep_deis_second_order_update( + self, + model_output_list: List[torch.Tensor], + *args, + sample: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the second-order multistep DEIS. + + Args: + model_output_list (`List[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s0, sigma_s1 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + + m0, m1 = model_output_list[-1], model_output_list[-2] + + rho_t, rho_s0, rho_s1 = sigma_t / alpha_t, sigma_s0 / alpha_s0, sigma_s1 / alpha_s1 + + if self.config.algorithm_type == "deis": + + def ind_fn(t, b, c): + # Integrate[(log(t) - log(c)) / (log(b) - log(c)), {t}] + return t * (-np.log(c) + np.log(t) - 1) / (np.log(b) - np.log(c)) + + coef1 = ind_fn(rho_t, rho_s0, rho_s1) - ind_fn(rho_s0, rho_s0, rho_s1) + coef2 = ind_fn(rho_t, rho_s1, rho_s0) - ind_fn(rho_s0, rho_s1, rho_s0) + + x_t = alpha_t * (sample / alpha_s0 + coef1 * m0 + coef2 * m1) + return x_t + else: + raise NotImplementedError("only support log-rho multistep deis now") + + def multistep_deis_third_order_update( + self, + model_output_list: List[torch.Tensor], + *args, + sample: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the third-order multistep DEIS. + + Args: + model_output_list (`List[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`): + A current instance of a sample created by diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s0, sigma_s1, sigma_s2 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + self.sigmas[self.step_index - 2], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2) + + m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3] + + rho_t, rho_s0, rho_s1, rho_s2 = ( + sigma_t / alpha_t, + sigma_s0 / alpha_s0, + sigma_s1 / alpha_s1, + sigma_s2 / alpha_s2, + ) + + if self.config.algorithm_type == "deis": + + def ind_fn(t, b, c, d): + # Integrate[(log(t) - log(c))(log(t) - log(d)) / (log(b) - log(c))(log(b) - log(d)), {t}] + numerator = t * ( + np.log(c) * (np.log(d) - np.log(t) + 1) + - np.log(d) * np.log(t) + + np.log(d) + + np.log(t) ** 2 + - 2 * np.log(t) + + 2 + ) + denominator = (np.log(b) - np.log(c)) * (np.log(b) - np.log(d)) + return numerator / denominator + + coef1 = ind_fn(rho_t, rho_s0, rho_s1, rho_s2) - ind_fn(rho_s0, rho_s0, rho_s1, rho_s2) + coef2 = ind_fn(rho_t, rho_s1, rho_s2, rho_s0) - ind_fn(rho_s0, rho_s1, rho_s2, rho_s0) + coef3 = ind_fn(rho_t, rho_s2, rho_s0, rho_s1) - ind_fn(rho_s0, rho_s2, rho_s0, rho_s1) + + x_t = alpha_t * (sample / alpha_s0 + coef1 * m0 + coef2 * m1 + coef3 * m2) + + return x_t + else: + raise NotImplementedError("only support log-rho multistep deis now") + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + index_candidates = (schedule_timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + return step_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index + def _init_step_index(self, timestep): + """ + Initialize the step_index counter for the scheduler. + """ + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: Union[int, torch.Tensor], + sample: torch.Tensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the multistep DEIS. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + lower_order_final = ( + (self.step_index == len(self.timesteps) - 1) and self.config.lower_order_final and len(self.timesteps) < 15 + ) + lower_order_second = ( + (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15 + ) + + model_output = self.convert_model_output(model_output, sample=sample) + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.model_outputs[-1] = model_output + + if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final: + prev_sample = self.deis_first_order_update(model_output, sample=sample) + elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second: + prev_sample = self.multistep_deis_second_order_update(self.model_outputs, sample=sample) + else: + prev_sample = self.multistep_deis_third_order_update(self.model_outputs, sample=sample) + + if self.lower_order_nums < self.config.solver_order: + self.lower_order_nums += 1 + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # begin_index is None when the scheduler is used for training or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + noisy_samples = alpha_t * original_samples + sigma_t * noise + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpm_cogvideox.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpm_cogvideox.py new file mode 100644 index 0000000000000000000000000000000000000000..f7b63720e1071890fc7bb6bfe54b9ad808276093 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpm_cogvideox.py @@ -0,0 +1,489 @@ +# Copyright 2025 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team. +# All rights reserved. +# +# 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. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +import math +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM +class DDIMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +def rescale_zero_terminal_snr(alphas_cumprod): + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + + Args: + betas (`torch.Tensor`): + the betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + + return alphas_bar + + +class CogVideoXDPMScheduler(SchedulerMixin, ConfigMixin): + """ + `DDIMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with + non-Markovian guidance. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to 1.0): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, defaults to `True`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, + beta_end: float = 0.0120, + beta_schedule: str = "scaled_linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + clip_sample: bool = True, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: str = "epsilon", + clip_sample_range: float = 1.0, + sample_max_value: float = 1.0, + timestep_spacing: str = "leading", + rescale_betas_zero_snr: bool = False, + snr_shift_scale: float = 3.0, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float64) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # Modify: SNR shift following SD3 + self.alphas_cumprod = self.alphas_cumprod / (snr_shift_scale + (1 - snr_shift_scale) * self.alphas_cumprod) + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.alphas_cumprod = rescale_zero_terminal_snr(self.alphas_cumprod) + + # At every step in ddim, we are looking into the previous alphas_cumprod + # For the final step, there is no previous alphas_cumprod because we are already at 0 + # `set_alpha_to_one` decides whether we set this parameter simply to one or + # whether we use the final alpha of the "non-previous" one. + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + + def _get_variance(self, timestep, prev_timestep): + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + """ + + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps) + .round()[::-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + def get_variables(self, alpha_prod_t, alpha_prod_t_prev, alpha_prod_t_back=None): + lamb = ((alpha_prod_t / (1 - alpha_prod_t)) ** 0.5).log() + lamb_next = ((alpha_prod_t_prev / (1 - alpha_prod_t_prev)) ** 0.5).log() + h = lamb_next - lamb + + if alpha_prod_t_back is not None: + lamb_previous = ((alpha_prod_t_back / (1 - alpha_prod_t_back)) ** 0.5).log() + h_last = lamb - lamb_previous + r = h_last / h + return h, r, lamb, lamb_next + else: + return h, None, lamb, lamb_next + + def get_mult(self, h, r, alpha_prod_t, alpha_prod_t_prev, alpha_prod_t_back): + mult1 = ((1 - alpha_prod_t_prev) / (1 - alpha_prod_t)) ** 0.5 * (-h).exp() + mult2 = (-2 * h).expm1() * alpha_prod_t_prev**0.5 + + if alpha_prod_t_back is not None: + mult3 = 1 + 1 / (2 * r) + mult4 = 1 / (2 * r) + return mult1, mult2, mult3, mult4 + else: + return mult1, mult2 + + def step( + self, + model_output: torch.Tensor, + old_pred_original_sample: torch.Tensor, + timestep: int, + timestep_back: int, + sample: torch.Tensor, + eta: float = 0.0, + use_clipped_model_output: bool = False, + generator=None, + variance_noise: Optional[torch.Tensor] = None, + return_dict: bool = False, + ) -> Union[DDIMSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + eta (`float`): + The weight of noise for added noise in diffusion step. + use_clipped_model_output (`bool`, defaults to `False`): + If `True`, computes "corrected" `model_output` from the clipped predicted original sample. Necessary + because predicted original sample is clipped to [-1, 1] when `self.config.clip_sample` is `True`. If no + clipping has happened, "corrected" `model_output` would coincide with the one provided as input and + `use_clipped_model_output` has no effect. + generator (`torch.Generator`, *optional*): + A random number generator. + variance_noise (`torch.Tensor`): + Alternative to generating noise with `generator` by directly providing the noise for the variance + itself. Useful for methods such as [`CycleDiffusion`]. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502 + # Ideally, read DDIM paper in-detail understanding + + # Notation ( -> + # - pred_noise_t -> e_theta(x_t, t) + # - pred_original_sample -> f_theta(x_t, t) or x_0 + # - std_dev_t -> sigma_t + # - eta -> η + # - pred_sample_direction -> "direction pointing to x_t" + # - pred_prev_sample -> "x_t-1" + + # 1. get previous step value (=t-1) + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + # 2. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + alpha_prod_t_back = self.alphas_cumprod[timestep_back] if timestep_back is not None else None + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + # To make style tests pass, commented out `pred_epsilon` as it is an unused variable + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + # pred_epsilon = model_output + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + # pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + # pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction`" + ) + + h, r, lamb, lamb_next = self.get_variables(alpha_prod_t, alpha_prod_t_prev, alpha_prod_t_back) + mult = list(self.get_mult(h, r, alpha_prod_t, alpha_prod_t_prev, alpha_prod_t_back)) + mult_noise = (1 - alpha_prod_t_prev) ** 0.5 * (1 - (-2 * h).exp()) ** 0.5 + + noise = randn_tensor(sample.shape, generator=generator, device=sample.device, dtype=sample.dtype) + prev_sample = mult[0] * sample - mult[1] * pred_original_sample + mult_noise * noise + + if old_pred_original_sample is None or prev_timestep < 0: + # Save a network evaluation if all noise levels are 0 or on the first step + return prev_sample, pred_original_sample + else: + denoised_d = mult[2] * pred_original_sample - mult[3] * old_pred_original_sample + noise = randn_tensor(sample.shape, generator=generator, device=sample.device, dtype=sample.dtype) + x_advanced = mult[0] * sample - mult[1] * denoised_d + mult_noise * noise + + prev_sample = x_advanced + + if not return_dict: + return (prev_sample, pred_original_sample) + + return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_multistep.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_multistep.py new file mode 100644 index 0000000000000000000000000000000000000000..1a648af5a008fb01fc3c14a9f8f72bf588f50a69 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_multistep.py @@ -0,0 +1,1174 @@ +# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver + +import math +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import deprecate, is_scipy_available +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +if is_scipy_available(): + import scipy.stats + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas): + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + + Args: + betas (`torch.Tensor`): + the betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin): + """ + `DPMSolverMultistepScheduler` is a fast dedicated high-order solver for diffusion ODEs. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + solver_order (`int`, defaults to 2): + The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided + sampling, and `solver_order=3` for unconditional sampling. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample), `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper), or `flow_prediction`. + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True` and + `algorithm_type="dpmsolver++"`. + algorithm_type (`str`, defaults to `dpmsolver++`): + Algorithm type for the solver; can be `dpmsolver`, `dpmsolver++`, `sde-dpmsolver` or `sde-dpmsolver++`. The + `dpmsolver` type implements the algorithms in the [DPMSolver](https://huggingface.co/papers/2206.00927) + paper, and the `dpmsolver++` type implements the algorithms in the + [DPMSolver++](https://huggingface.co/papers/2211.01095) paper. It is recommended to use `dpmsolver++` or + `sde-dpmsolver++` with `solver_order=2` for guided sampling like in Stable Diffusion. + solver_type (`str`, defaults to `midpoint`): + Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the + sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers. + lower_order_final (`bool`, defaults to `True`): + Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can + stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10. + euler_at_final (`bool`, defaults to `False`): + Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail + richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference + steps, but sometimes may result in blurring. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + use_lu_lambdas (`bool`, *optional*, defaults to `False`): + Whether to use the uniform-logSNR for step sizes proposed by Lu's DPM-Solver in the noise schedule during + the sampling process. If `True`, the sigmas and time steps are determined according to a sequence of + `lambda(t)`. + use_flow_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use flow sigmas for step sizes in the noise schedule during the sampling process. + flow_shift (`float`, *optional*, defaults to 1.0): + The shift value for the timestep schedule for flow matching. + final_sigmas_type (`str`, defaults to `"zero"`): + The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final + sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0. + lambda_min_clipped (`float`, defaults to `-inf`): + Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the + cosine (`squaredcos_cap_v2`) noise schedule. + variance_type (`str`, *optional*): + Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output + contains the predicted Gaussian variance. + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + solver_order: int = 2, + prediction_type: str = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + algorithm_type: str = "dpmsolver++", + solver_type: str = "midpoint", + lower_order_final: bool = True, + euler_at_final: bool = False, + use_karras_sigmas: Optional[bool] = False, + use_exponential_sigmas: Optional[bool] = False, + use_beta_sigmas: Optional[bool] = False, + use_lu_lambdas: Optional[bool] = False, + use_flow_sigmas: Optional[bool] = False, + flow_shift: Optional[float] = 1.0, + final_sigmas_type: Optional[str] = "zero", # "zero", "sigma_min" + lambda_min_clipped: float = -float("inf"), + variance_type: Optional[str] = None, + timestep_spacing: str = "linspace", + steps_offset: int = 0, + rescale_betas_zero_snr: bool = False, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if algorithm_type in ["dpmsolver", "sde-dpmsolver"]: + deprecation_message = f"algorithm_type {algorithm_type} is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead" + deprecate("algorithm_types dpmsolver and sde-dpmsolver", "1.0.0", deprecation_message) + + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + if rescale_betas_zero_snr: + # Close to 0 without being 0 so first sigma is not inf + # FP16 smallest positive subnormal works well here + self.alphas_cumprod[-1] = 2**-24 + + # Currently we only support VP-type noise schedule + self.alpha_t = torch.sqrt(self.alphas_cumprod) + self.sigma_t = torch.sqrt(1 - self.alphas_cumprod) + self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t) + self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5 + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # settings for DPM-Solver + if algorithm_type not in ["dpmsolver", "dpmsolver++", "sde-dpmsolver", "sde-dpmsolver++"]: + if algorithm_type == "deis": + self.register_to_config(algorithm_type="dpmsolver++") + else: + raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}") + + if solver_type not in ["midpoint", "heun"]: + if solver_type in ["logrho", "bh1", "bh2"]: + self.register_to_config(solver_type="midpoint") + else: + raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}") + + if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"] and final_sigmas_type == "zero": + raise ValueError( + f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please choose `sigma_min` instead." + ) + + # setable values + self.num_inference_steps = None + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps) + self.model_outputs = [None] * solver_order + self.lower_order_nums = 0 + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_timesteps( + self, + num_inference_steps: int = None, + device: Union[str, torch.device] = None, + timesteps: Optional[List[int]] = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to support arbitrary timesteps schedule. If `None`, timesteps will be generated + based on the `timestep_spacing` attribute. If `timesteps` is passed, `num_inference_steps` and `sigmas` + must be `None`, and `timestep_spacing` attribute will be ignored. + """ + if num_inference_steps is None and timesteps is None: + raise ValueError("Must pass exactly one of `num_inference_steps` or `timesteps`.") + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.") + if timesteps is not None and self.config.use_karras_sigmas: + raise ValueError("Cannot use `timesteps` with `config.use_karras_sigmas = True`") + if timesteps is not None and self.config.use_lu_lambdas: + raise ValueError("Cannot use `timesteps` with `config.use_lu_lambdas = True`") + if timesteps is not None and self.config.use_exponential_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_exponential_sigmas = True`.") + if timesteps is not None and self.config.use_beta_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_beta_sigmas = True`.") + + if timesteps is not None: + timesteps = np.array(timesteps).astype(np.int64) + else: + # Clipping the minimum of all lambda(t) for numerical stability. + # This is critical for cosine (squaredcos_cap_v2) noise schedule. + clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped) + last_timestep = ((self.config.num_train_timesteps - clipped_idx).numpy()).item() + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, last_timestep - 1, num_inference_steps + 1) + .round()[::-1][:-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = last_timestep // (num_inference_steps + 1) + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = ( + (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64) + ) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.arange(last_timestep, 0, -step_ratio).round().copy().astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + + if self.config.use_karras_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + if self.config.beta_schedule != "squaredcos_cap_v2": + timesteps = timesteps.round() + elif self.config.use_lu_lambdas: + lambdas = np.flip(log_sigmas.copy()) + lambdas = self._convert_to_lu(in_lambdas=lambdas, num_inference_steps=num_inference_steps) + sigmas = np.exp(lambdas) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + if self.config.beta_schedule != "squaredcos_cap_v2": + timesteps = timesteps.round() + elif self.config.use_exponential_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_beta_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_flow_sigmas: + alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1) + sigmas = 1.0 - alphas + sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy() + timesteps = (sigmas * self.config.num_train_timesteps).copy() + else: + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5 + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + + self.sigmas = torch.from_numpy(sigmas) + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64) + + self.num_inference_steps = len(timesteps) + + self.model_outputs = [ + None, + ] * self.config.solver_order + self.lower_order_nums = 0 + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + def _sigma_to_alpha_sigma_t(self, sigma): + if self.config.use_flow_sigmas: + alpha_t = 1 - sigma + sigma_t = sigma + else: + alpha_t = 1 / ((sigma**2 + 1) ** 0.5) + sigma_t = sigma * alpha_t + + return alpha_t, sigma_t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + def _convert_to_lu(self, in_lambdas: torch.Tensor, num_inference_steps) -> torch.Tensor: + """Constructs the noise schedule of Lu et al. (2022).""" + + lambda_min: float = in_lambdas[-1].item() + lambda_max: float = in_lambdas[0].item() + + rho = 1.0 # 1.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = lambda_min ** (1 / rho) + max_inv_rho = lambda_max ** (1 / rho) + lambdas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return lambdas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """Constructs an exponential noise schedule.""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def convert_model_output( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + """ + Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is + designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an + integral of the data prediction model. + + + + The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both noise + prediction and data prediction models. + + + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + # DPM-Solver++ needs to solve an integral of the data prediction model. + if self.config.algorithm_type in ["dpmsolver++", "sde-dpmsolver++"]: + if self.config.prediction_type == "epsilon": + # DPM-Solver and DPM-Solver++ only need the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + model_output = model_output[:, :3] + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = (sample - sigma_t * model_output) / alpha_t + elif self.config.prediction_type == "sample": + x0_pred = model_output + elif self.config.prediction_type == "v_prediction": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = alpha_t * sample - sigma_t * model_output + elif self.config.prediction_type == "flow_prediction": + sigma_t = self.sigmas[self.step_index] + x0_pred = sample - sigma_t * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + "`v_prediction`, or `flow_prediction` for the DPMSolverMultistepScheduler." + ) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + return x0_pred + + # DPM-Solver needs to solve an integral of the noise prediction model. + elif self.config.algorithm_type in ["dpmsolver", "sde-dpmsolver"]: + if self.config.prediction_type == "epsilon": + # DPM-Solver and DPM-Solver++ only need the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + epsilon = model_output[:, :3] + else: + epsilon = model_output + elif self.config.prediction_type == "sample": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + epsilon = (sample - alpha_t * model_output) / sigma_t + elif self.config.prediction_type == "v_prediction": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + epsilon = alpha_t * model_output + sigma_t * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction` for the DPMSolverMultistepScheduler." + ) + + if self.config.thresholding: + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = (sample - sigma_t * epsilon) / alpha_t + x0_pred = self._threshold_sample(x0_pred) + epsilon = (sample - alpha_t * x0_pred) / sigma_t + + return epsilon + + def dpm_solver_first_order_update( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + noise: Optional[torch.Tensor] = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the first-order DPMSolver (equivalent to DDIM). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s = torch.log(alpha_s) - torch.log(sigma_s) + + h = lambda_t - lambda_s + if self.config.algorithm_type == "dpmsolver++": + x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output + elif self.config.algorithm_type == "dpmsolver": + x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + x_t = ( + (sigma_t / sigma_s * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.algorithm_type == "sde-dpmsolver": + assert noise is not None + x_t = ( + (alpha_t / alpha_s) * sample + - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * model_output + + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise + ) + return x_t + + def multistep_dpm_solver_second_order_update( + self, + model_output_list: List[torch.Tensor], + *args, + sample: torch.Tensor = None, + noise: Optional[torch.Tensor] = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the second-order multistep DPMSolver. + + Args: + model_output_list (`List[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s0, sigma_s1 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + + m0, m1 = model_output_list[-1], model_output_list[-2] + + h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1 + r0 = h_0 / h + D0, D1 = m0, (1.0 / r0) * (m0 - m1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2211.01095 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + ) + elif self.config.algorithm_type == "dpmsolver": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + ) + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.algorithm_type == "sde-dpmsolver": + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (alpha_t / alpha_s0) * sample + - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * (torch.exp(h) - 1.0)) * D1 + + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (alpha_t / alpha_s0) * sample + - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0 + - 2.0 * (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise + ) + return x_t + + def multistep_dpm_solver_third_order_update( + self, + model_output_list: List[torch.Tensor], + *args, + sample: torch.Tensor = None, + noise: Optional[torch.Tensor] = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the third-order multistep DPMSolver. + + Args: + model_output_list (`List[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`): + A current instance of a sample created by diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s0, sigma_s1, sigma_s2 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + self.sigmas[self.step_index - 2], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2) + + m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3] + + h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2 + r0, r1 = h_0 / h, h_1 / h + D0 = m0 + D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2) + D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1) + D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2 + ) + elif self.config.algorithm_type == "dpmsolver": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + - (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2 + ) + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h) - 2.0 * h) / (2.0 * h) ** 2 - 0.5)) * D2 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + return x_t + + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + index_candidates = (schedule_timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + return step_index + + def _init_step_index(self, timestep): + """ + Initialize the step_index counter for the scheduler. + """ + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: Union[int, torch.Tensor], + sample: torch.Tensor, + generator=None, + variance_noise: Optional[torch.Tensor] = None, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the multistep DPMSolver. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + variance_noise (`torch.Tensor`): + Alternative to generating noise with `generator` by directly providing the noise for the variance + itself. Useful for methods such as [`LEdits++`]. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Improve numerical stability for small number of steps + lower_order_final = (self.step_index == len(self.timesteps) - 1) and ( + self.config.euler_at_final + or (self.config.lower_order_final and len(self.timesteps) < 15) + or self.config.final_sigmas_type == "zero" + ) + lower_order_second = ( + (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15 + ) + + model_output = self.convert_model_output(model_output, sample=sample) + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.model_outputs[-1] = model_output + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + if self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"] and variance_noise is None: + noise = randn_tensor( + model_output.shape, generator=generator, device=model_output.device, dtype=torch.float32 + ) + elif self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"]: + noise = variance_noise.to(device=model_output.device, dtype=torch.float32) + else: + noise = None + + if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final: + prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise) + elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second: + prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise) + else: + prev_sample = self.multistep_dpm_solver_third_order_update(self.model_outputs, sample=sample, noise=noise) + + if self.lower_order_nums < self.config.solver_order: + self.lower_order_nums += 1 + + # Cast sample back to expected dtype + prev_sample = prev_sample.to(model_output.dtype) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # begin_index is None when the scheduler is used for training or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + noisy_samples = alpha_t * original_samples + sigma_t * noise + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_multistep_flax.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_multistep_flax.py new file mode 100644 index 0000000000000000000000000000000000000000..71b9960bf2ff21d116caffce05c986d8b3a898e4 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_multistep_flax.py @@ -0,0 +1,645 @@ +# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver + +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import flax +import jax +import jax.numpy as jnp + +from ..configuration_utils import ConfigMixin, register_to_config +from .scheduling_utils_flax import ( + CommonSchedulerState, + FlaxKarrasDiffusionSchedulers, + FlaxSchedulerMixin, + FlaxSchedulerOutput, + add_noise_common, +) + + +@flax.struct.dataclass +class DPMSolverMultistepSchedulerState: + common: CommonSchedulerState + alpha_t: jnp.ndarray + sigma_t: jnp.ndarray + lambda_t: jnp.ndarray + + # setable values + init_noise_sigma: jnp.ndarray + timesteps: jnp.ndarray + num_inference_steps: Optional[int] = None + + # running values + model_outputs: Optional[jnp.ndarray] = None + lower_order_nums: Optional[jnp.int32] = None + prev_timestep: Optional[jnp.int32] = None + cur_sample: Optional[jnp.ndarray] = None + + @classmethod + def create( + cls, + common: CommonSchedulerState, + alpha_t: jnp.ndarray, + sigma_t: jnp.ndarray, + lambda_t: jnp.ndarray, + init_noise_sigma: jnp.ndarray, + timesteps: jnp.ndarray, + ): + return cls( + common=common, + alpha_t=alpha_t, + sigma_t=sigma_t, + lambda_t=lambda_t, + init_noise_sigma=init_noise_sigma, + timesteps=timesteps, + ) + + +@dataclass +class FlaxDPMSolverMultistepSchedulerOutput(FlaxSchedulerOutput): + state: DPMSolverMultistepSchedulerState + + +class FlaxDPMSolverMultistepScheduler(FlaxSchedulerMixin, ConfigMixin): + """ + DPM-Solver (and the improved version DPM-Solver++) is a fast dedicated high-order solver for diffusion ODEs with + the convergence order guarantee. Empirically, sampling by DPM-Solver with only 20 steps can generate high-quality + samples, and it can generate quite good samples even in only 10 steps. + + For more details, see the original paper: https://huggingface.co/papers/2206.00927 and + https://huggingface.co/papers/2211.01095 + + Currently, we support the multistep DPM-Solver for both noise prediction models and data prediction models. We + recommend to use `solver_order=2` for guided sampling, and `solver_order=3` for unconditional sampling. + + We also support the "dynamic thresholding" method in Imagen (https://huggingface.co/papers/2205.11487). For + pixel-space diffusion models, you can set both `algorithm_type="dpmsolver++"` and `thresholding=True` to use the + dynamic thresholding. Note that the thresholding method is unsuitable for latent-space diffusion models (such as + stable-diffusion). + + [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` + function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. + [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and + [`~SchedulerMixin.from_pretrained`] functions. + + For more details, see the original paper: https://huggingface.co/papers/2206.00927 and + https://huggingface.co/papers/2211.01095 + + Args: + num_train_timesteps (`int`): number of diffusion steps used to train the model. + beta_start (`float`): the starting `beta` value of inference. + beta_end (`float`): the final `beta` value. + beta_schedule (`str`): + the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, optional): + option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc. + solver_order (`int`, default `2`): + the order of DPM-Solver; can be `1` or `2` or `3`. We recommend to use `solver_order=2` for guided + sampling, and `solver_order=3` for unconditional sampling. + prediction_type (`str`, default `epsilon`): + indicates whether the model predicts the noise (epsilon), or the data / `x0`. One of `epsilon`, `sample`, + or `v-prediction`. + thresholding (`bool`, default `False`): + whether to use the "dynamic thresholding" method (introduced by Imagen, + https://huggingface.co/papers/2205.11487). For pixel-space diffusion models, you can set both + `algorithm_type=dpmsolver++` and `thresholding=True` to use the dynamic thresholding. Note that the + thresholding method is unsuitable for latent-space diffusion models (such as stable-diffusion). + dynamic_thresholding_ratio (`float`, default `0.995`): + the ratio for the dynamic thresholding method. Default is `0.995`, the same as Imagen + (https://huggingface.co/papers/2205.11487). + sample_max_value (`float`, default `1.0`): + the threshold value for dynamic thresholding. Valid only when `thresholding=True` and + `algorithm_type="dpmsolver++`. + algorithm_type (`str`, default `dpmsolver++`): + the algorithm type for the solver. Either `dpmsolver` or `dpmsolver++`. The `dpmsolver` type implements the + algorithms in https://huggingface.co/papers/2206.00927, and the `dpmsolver++` type implements the + algorithms in https://huggingface.co/papers/2211.01095. We recommend to use `dpmsolver++` with + `solver_order=2` for guided sampling (e.g. stable-diffusion). + solver_type (`str`, default `midpoint`): + the solver type for the second-order solver. Either `midpoint` or `heun`. The solver type slightly affects + the sample quality, especially for small number of steps. We empirically find that `midpoint` solvers are + slightly better, so we recommend to use the `midpoint` type. + lower_order_final (`bool`, default `True`): + whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. We empirically + find this trick can stabilize the sampling of DPM-Solver for steps < 15, especially for steps <= 10. + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`): + the `dtype` used for params and computation. + """ + + _compatibles = [e.name for e in FlaxKarrasDiffusionSchedulers] + + dtype: jnp.dtype + + @property + def has_state(self): + return True + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[jnp.ndarray] = None, + solver_order: int = 2, + prediction_type: str = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + algorithm_type: str = "dpmsolver++", + solver_type: str = "midpoint", + lower_order_final: bool = True, + timestep_spacing: str = "linspace", + dtype: jnp.dtype = jnp.float32, + ): + self.dtype = dtype + + def create_state(self, common: Optional[CommonSchedulerState] = None) -> DPMSolverMultistepSchedulerState: + if common is None: + common = CommonSchedulerState.create(self) + + # Currently we only support VP-type noise schedule + alpha_t = jnp.sqrt(common.alphas_cumprod) + sigma_t = jnp.sqrt(1 - common.alphas_cumprod) + lambda_t = jnp.log(alpha_t) - jnp.log(sigma_t) + + # settings for DPM-Solver + if self.config.algorithm_type not in ["dpmsolver", "dpmsolver++"]: + raise NotImplementedError(f"{self.config.algorithm_type} is not implemented for {self.__class__}") + if self.config.solver_type not in ["midpoint", "heun"]: + raise NotImplementedError(f"{self.config.solver_type} is not implemented for {self.__class__}") + + # standard deviation of the initial noise distribution + init_noise_sigma = jnp.array(1.0, dtype=self.dtype) + + timesteps = jnp.arange(0, self.config.num_train_timesteps).round()[::-1] + + return DPMSolverMultistepSchedulerState.create( + common=common, + alpha_t=alpha_t, + sigma_t=sigma_t, + lambda_t=lambda_t, + init_noise_sigma=init_noise_sigma, + timesteps=timesteps, + ) + + def set_timesteps( + self, state: DPMSolverMultistepSchedulerState, num_inference_steps: int, shape: Tuple + ) -> DPMSolverMultistepSchedulerState: + """ + Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference. + + Args: + state (`DPMSolverMultistepSchedulerState`): + the `FlaxDPMSolverMultistepScheduler` state data class instance. + num_inference_steps (`int`): + the number of diffusion steps used when generating samples with a pre-trained model. + shape (`Tuple`): + the shape of the samples to be generated. + """ + last_timestep = self.config.num_train_timesteps + if self.config.timestep_spacing == "linspace": + timesteps = ( + jnp.linspace(0, last_timestep - 1, num_inference_steps + 1).round()[::-1][:-1].astype(jnp.int32) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = last_timestep // (num_inference_steps + 1) + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = ( + (jnp.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(jnp.int32) + ) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = jnp.arange(last_timestep, 0, -step_ratio).round().copy().astype(jnp.int32) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + # initial running values + + model_outputs = jnp.zeros((self.config.solver_order,) + shape, dtype=self.dtype) + lower_order_nums = jnp.int32(0) + prev_timestep = jnp.int32(-1) + cur_sample = jnp.zeros(shape, dtype=self.dtype) + + return state.replace( + num_inference_steps=num_inference_steps, + timesteps=timesteps, + model_outputs=model_outputs, + lower_order_nums=lower_order_nums, + prev_timestep=prev_timestep, + cur_sample=cur_sample, + ) + + def convert_model_output( + self, + state: DPMSolverMultistepSchedulerState, + model_output: jnp.ndarray, + timestep: int, + sample: jnp.ndarray, + ) -> jnp.ndarray: + """ + Convert the model output to the corresponding type that the algorithm (DPM-Solver / DPM-Solver++) needs. + + DPM-Solver is designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to + discretize an integral of the data prediction model. So we need to first convert the model output to the + corresponding type to match the algorithm. + + Note that the algorithm type and the model type is decoupled. That is to say, we can use either DPM-Solver or + DPM-Solver++ for both noise prediction model and data prediction model. + + Args: + model_output (`jnp.ndarray`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + + Returns: + `jnp.ndarray`: the converted model output. + """ + # DPM-Solver++ needs to solve an integral of the data prediction model. + if self.config.algorithm_type == "dpmsolver++": + if self.config.prediction_type == "epsilon": + alpha_t, sigma_t = state.alpha_t[timestep], state.sigma_t[timestep] + x0_pred = (sample - sigma_t * model_output) / alpha_t + elif self.config.prediction_type == "sample": + x0_pred = model_output + elif self.config.prediction_type == "v_prediction": + alpha_t, sigma_t = state.alpha_t[timestep], state.sigma_t[timestep] + x0_pred = alpha_t * sample - sigma_t * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + " or `v_prediction` for the FlaxDPMSolverMultistepScheduler." + ) + + if self.config.thresholding: + # Dynamic thresholding in https://huggingface.co/papers/2205.11487 + dynamic_max_val = jnp.percentile( + jnp.abs(x0_pred), self.config.dynamic_thresholding_ratio, axis=tuple(range(1, x0_pred.ndim)) + ) + dynamic_max_val = jnp.maximum( + dynamic_max_val, self.config.sample_max_value * jnp.ones_like(dynamic_max_val) + ) + x0_pred = jnp.clip(x0_pred, -dynamic_max_val, dynamic_max_val) / dynamic_max_val + return x0_pred + # DPM-Solver needs to solve an integral of the noise prediction model. + elif self.config.algorithm_type == "dpmsolver": + if self.config.prediction_type == "epsilon": + return model_output + elif self.config.prediction_type == "sample": + alpha_t, sigma_t = state.alpha_t[timestep], state.sigma_t[timestep] + epsilon = (sample - alpha_t * model_output) / sigma_t + return epsilon + elif self.config.prediction_type == "v_prediction": + alpha_t, sigma_t = state.alpha_t[timestep], state.sigma_t[timestep] + epsilon = alpha_t * model_output + sigma_t * sample + return epsilon + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + " or `v_prediction` for the FlaxDPMSolverMultistepScheduler." + ) + + def dpm_solver_first_order_update( + self, + state: DPMSolverMultistepSchedulerState, + model_output: jnp.ndarray, + timestep: int, + prev_timestep: int, + sample: jnp.ndarray, + ) -> jnp.ndarray: + """ + One step for the first-order DPM-Solver (equivalent to DDIM). + + See https://huggingface.co/papers/2206.00927 for the detailed derivation. + + Args: + model_output (`jnp.ndarray`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + prev_timestep (`int`): previous discrete timestep in the diffusion chain. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + + Returns: + `jnp.ndarray`: the sample tensor at the previous timestep. + """ + t, s0 = prev_timestep, timestep + m0 = model_output + lambda_t, lambda_s = state.lambda_t[t], state.lambda_t[s0] + alpha_t, alpha_s = state.alpha_t[t], state.alpha_t[s0] + sigma_t, sigma_s = state.sigma_t[t], state.sigma_t[s0] + h = lambda_t - lambda_s + if self.config.algorithm_type == "dpmsolver++": + x_t = (sigma_t / sigma_s) * sample - (alpha_t * (jnp.exp(-h) - 1.0)) * m0 + elif self.config.algorithm_type == "dpmsolver": + x_t = (alpha_t / alpha_s) * sample - (sigma_t * (jnp.exp(h) - 1.0)) * m0 + return x_t + + def multistep_dpm_solver_second_order_update( + self, + state: DPMSolverMultistepSchedulerState, + model_output_list: jnp.ndarray, + timestep_list: List[int], + prev_timestep: int, + sample: jnp.ndarray, + ) -> jnp.ndarray: + """ + One step for the second-order multistep DPM-Solver. + + Args: + model_output_list (`List[jnp.ndarray]`): + direct outputs from learned diffusion model at current and latter timesteps. + timestep (`int`): current and latter discrete timestep in the diffusion chain. + prev_timestep (`int`): previous discrete timestep in the diffusion chain. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + + Returns: + `jnp.ndarray`: the sample tensor at the previous timestep. + """ + t, s0, s1 = prev_timestep, timestep_list[-1], timestep_list[-2] + m0, m1 = model_output_list[-1], model_output_list[-2] + lambda_t, lambda_s0, lambda_s1 = state.lambda_t[t], state.lambda_t[s0], state.lambda_t[s1] + alpha_t, alpha_s0 = state.alpha_t[t], state.alpha_t[s0] + sigma_t, sigma_s0 = state.sigma_t[t], state.sigma_t[s0] + h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1 + r0 = h_0 / h + D0, D1 = m0, (1.0 / r0) * (m0 - m1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2211.01095 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (jnp.exp(-h) - 1.0)) * D0 + - 0.5 * (alpha_t * (jnp.exp(-h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (jnp.exp(-h) - 1.0)) * D0 + + (alpha_t * ((jnp.exp(-h) - 1.0) / h + 1.0)) * D1 + ) + elif self.config.algorithm_type == "dpmsolver": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (jnp.exp(h) - 1.0)) * D0 + - 0.5 * (sigma_t * (jnp.exp(h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (jnp.exp(h) - 1.0)) * D0 + - (sigma_t * ((jnp.exp(h) - 1.0) / h - 1.0)) * D1 + ) + return x_t + + def multistep_dpm_solver_third_order_update( + self, + state: DPMSolverMultistepSchedulerState, + model_output_list: jnp.ndarray, + timestep_list: List[int], + prev_timestep: int, + sample: jnp.ndarray, + ) -> jnp.ndarray: + """ + One step for the third-order multistep DPM-Solver. + + Args: + model_output_list (`List[jnp.ndarray]`): + direct outputs from learned diffusion model at current and latter timesteps. + timestep (`int`): current and latter discrete timestep in the diffusion chain. + prev_timestep (`int`): previous discrete timestep in the diffusion chain. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + + Returns: + `jnp.ndarray`: the sample tensor at the previous timestep. + """ + t, s0, s1, s2 = prev_timestep, timestep_list[-1], timestep_list[-2], timestep_list[-3] + m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3] + lambda_t, lambda_s0, lambda_s1, lambda_s2 = ( + state.lambda_t[t], + state.lambda_t[s0], + state.lambda_t[s1], + state.lambda_t[s2], + ) + alpha_t, alpha_s0 = state.alpha_t[t], state.alpha_t[s0] + sigma_t, sigma_s0 = state.sigma_t[t], state.sigma_t[s0] + h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2 + r0, r1 = h_0 / h, h_1 / h + D0 = m0 + D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2) + D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1) + D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (jnp.exp(-h) - 1.0)) * D0 + + (alpha_t * ((jnp.exp(-h) - 1.0) / h + 1.0)) * D1 + - (alpha_t * ((jnp.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2 + ) + elif self.config.algorithm_type == "dpmsolver": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (jnp.exp(h) - 1.0)) * D0 + - (sigma_t * ((jnp.exp(h) - 1.0) / h - 1.0)) * D1 + - (sigma_t * ((jnp.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2 + ) + return x_t + + def step( + self, + state: DPMSolverMultistepSchedulerState, + model_output: jnp.ndarray, + timestep: int, + sample: jnp.ndarray, + return_dict: bool = True, + ) -> Union[FlaxDPMSolverMultistepSchedulerOutput, Tuple]: + """ + Predict the sample at the previous timestep by DPM-Solver. Core function to propagate the diffusion process + from the learned model outputs (most often the predicted noise). + + Args: + state (`DPMSolverMultistepSchedulerState`): + the `FlaxDPMSolverMultistepScheduler` state data class instance. + model_output (`jnp.ndarray`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + return_dict (`bool`): option for returning tuple rather than FlaxDPMSolverMultistepSchedulerOutput class + + Returns: + [`FlaxDPMSolverMultistepSchedulerOutput`] or `tuple`: [`FlaxDPMSolverMultistepSchedulerOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample tensor. + + """ + if state.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + (step_index,) = jnp.where(state.timesteps == timestep, size=1) + step_index = step_index[0] + + prev_timestep = jax.lax.select(step_index == len(state.timesteps) - 1, 0, state.timesteps[step_index + 1]) + + model_output = self.convert_model_output(state, model_output, timestep, sample) + + model_outputs_new = jnp.roll(state.model_outputs, -1, axis=0) + model_outputs_new = model_outputs_new.at[-1].set(model_output) + state = state.replace( + model_outputs=model_outputs_new, + prev_timestep=prev_timestep, + cur_sample=sample, + ) + + def step_1(state: DPMSolverMultistepSchedulerState) -> jnp.ndarray: + return self.dpm_solver_first_order_update( + state, + state.model_outputs[-1], + state.timesteps[step_index], + state.prev_timestep, + state.cur_sample, + ) + + def step_23(state: DPMSolverMultistepSchedulerState) -> jnp.ndarray: + def step_2(state: DPMSolverMultistepSchedulerState) -> jnp.ndarray: + timestep_list = jnp.array([state.timesteps[step_index - 1], state.timesteps[step_index]]) + return self.multistep_dpm_solver_second_order_update( + state, + state.model_outputs, + timestep_list, + state.prev_timestep, + state.cur_sample, + ) + + def step_3(state: DPMSolverMultistepSchedulerState) -> jnp.ndarray: + timestep_list = jnp.array( + [ + state.timesteps[step_index - 2], + state.timesteps[step_index - 1], + state.timesteps[step_index], + ] + ) + return self.multistep_dpm_solver_third_order_update( + state, + state.model_outputs, + timestep_list, + state.prev_timestep, + state.cur_sample, + ) + + step_2_output = step_2(state) + step_3_output = step_3(state) + + if self.config.solver_order == 2: + return step_2_output + elif self.config.lower_order_final and len(state.timesteps) < 15: + return jax.lax.select( + state.lower_order_nums < 2, + step_2_output, + jax.lax.select( + step_index == len(state.timesteps) - 2, + step_2_output, + step_3_output, + ), + ) + else: + return jax.lax.select( + state.lower_order_nums < 2, + step_2_output, + step_3_output, + ) + + step_1_output = step_1(state) + step_23_output = step_23(state) + + if self.config.solver_order == 1: + prev_sample = step_1_output + + elif self.config.lower_order_final and len(state.timesteps) < 15: + prev_sample = jax.lax.select( + state.lower_order_nums < 1, + step_1_output, + jax.lax.select( + step_index == len(state.timesteps) - 1, + step_1_output, + step_23_output, + ), + ) + + else: + prev_sample = jax.lax.select( + state.lower_order_nums < 1, + step_1_output, + step_23_output, + ) + + state = state.replace( + lower_order_nums=jnp.minimum(state.lower_order_nums + 1, self.config.solver_order), + ) + + if not return_dict: + return (prev_sample, state) + + return FlaxDPMSolverMultistepSchedulerOutput(prev_sample=prev_sample, state=state) + + def scale_model_input( + self, state: DPMSolverMultistepSchedulerState, sample: jnp.ndarray, timestep: Optional[int] = None + ) -> jnp.ndarray: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + state (`DPMSolverMultistepSchedulerState`): + the `FlaxDPMSolverMultistepScheduler` state data class instance. + sample (`jnp.ndarray`): input sample + timestep (`int`, optional): current timestep + + Returns: + `jnp.ndarray`: scaled input sample + """ + return sample + + def add_noise( + self, + state: DPMSolverMultistepSchedulerState, + original_samples: jnp.ndarray, + noise: jnp.ndarray, + timesteps: jnp.ndarray, + ) -> jnp.ndarray: + return add_noise_common(state.common, original_samples, noise, timesteps) + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_multistep_inverse.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_multistep_inverse.py new file mode 100644 index 0000000000000000000000000000000000000000..9ec95885111172344cb953d96a70eee0cd12776c --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_multistep_inverse.py @@ -0,0 +1,1027 @@ +# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver + +import math +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import deprecate, is_scipy_available +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +if is_scipy_available(): + import scipy.stats + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class DPMSolverMultistepInverseScheduler(SchedulerMixin, ConfigMixin): + """ + `DPMSolverMultistepInverseScheduler` is the reverse scheduler of [`DPMSolverMultistepScheduler`]. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + solver_order (`int`, defaults to 2): + The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided + sampling, and `solver_order=3` for unconditional sampling. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True` and + `algorithm_type="dpmsolver++"`. + algorithm_type (`str`, defaults to `dpmsolver++`): + Algorithm type for the solver; can be `dpmsolver`, `dpmsolver++`, `sde-dpmsolver` or `sde-dpmsolver++`. The + `dpmsolver` type implements the algorithms in the [DPMSolver](https://huggingface.co/papers/2206.00927) + paper, and the `dpmsolver++` type implements the algorithms in the + [DPMSolver++](https://huggingface.co/papers/2211.01095) paper. It is recommended to use `dpmsolver++` or + `sde-dpmsolver++` with `solver_order=2` for guided sampling like in Stable Diffusion. + solver_type (`str`, defaults to `midpoint`): + Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the + sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers. + lower_order_final (`bool`, defaults to `True`): + Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can + stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10. + euler_at_final (`bool`, defaults to `False`): + Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail + richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference + steps, but sometimes may result in blurring. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + lambda_min_clipped (`float`, defaults to `-inf`): + Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the + cosine (`squaredcos_cap_v2`) noise schedule. + variance_type (`str`, *optional*): + Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output + contains the predicted Gaussian variance. + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + solver_order: int = 2, + prediction_type: str = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + algorithm_type: str = "dpmsolver++", + solver_type: str = "midpoint", + lower_order_final: bool = True, + euler_at_final: bool = False, + use_karras_sigmas: Optional[bool] = False, + use_exponential_sigmas: Optional[bool] = False, + use_beta_sigmas: Optional[bool] = False, + use_flow_sigmas: Optional[bool] = False, + flow_shift: Optional[float] = 1.0, + lambda_min_clipped: float = -float("inf"), + variance_type: Optional[str] = None, + timestep_spacing: str = "linspace", + steps_offset: int = 0, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if algorithm_type in ["dpmsolver", "sde-dpmsolver"]: + deprecation_message = f"algorithm_type {algorithm_type} is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead" + deprecate("algorithm_types dpmsolver and sde-dpmsolver", "1.0.0", deprecation_message) + + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + # Currently we only support VP-type noise schedule + self.alpha_t = torch.sqrt(self.alphas_cumprod) + self.sigma_t = torch.sqrt(1 - self.alphas_cumprod) + self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t) + self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5 + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # settings for DPM-Solver + if algorithm_type not in ["dpmsolver", "dpmsolver++", "sde-dpmsolver", "sde-dpmsolver++"]: + if algorithm_type == "deis": + self.register_to_config(algorithm_type="dpmsolver++") + else: + raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}") + + if solver_type not in ["midpoint", "heun"]: + if solver_type in ["logrho", "bh1", "bh2"]: + self.register_to_config(solver_type="midpoint") + else: + raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}") + + # setable values + self.num_inference_steps = None + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32).copy() + self.timesteps = torch.from_numpy(timesteps) + self.model_outputs = [None] * solver_order + self.lower_order_nums = 0 + self._step_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + self.use_karras_sigmas = use_karras_sigmas + self.use_exponential_sigmas = use_exponential_sigmas + self.use_beta_sigmas = use_beta_sigmas + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + def set_timesteps(self, num_inference_steps: int = None, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + # Clipping the minimum of all lambda(t) for numerical stability. + # This is critical for cosine (squaredcos_cap_v2) noise schedule. + clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped).item() + self.noisiest_timestep = self.config.num_train_timesteps - 1 - clipped_idx + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.noisiest_timestep, num_inference_steps + 1).round()[:-1].copy().astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = (self.noisiest_timestep + 1) // (num_inference_steps + 1) + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[:-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.arange(self.noisiest_timestep + 1, 0, -step_ratio).round()[::-1].copy().astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', " + "'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round() + timesteps = timesteps.copy().astype(np.int64) + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_flow_sigmas: + alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1) + sigmas = 1.0 - alphas + sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy() + timesteps = (sigmas * self.config.num_train_timesteps).copy() + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + else: + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + sigma_max = ( + (1 - self.alphas_cumprod[self.noisiest_timestep]) / self.alphas_cumprod[self.noisiest_timestep] + ) ** 0.5 + sigmas = np.concatenate([sigmas, [sigma_max]]).astype(np.float32) + + self.sigmas = torch.from_numpy(sigmas) + + # when num_inference_steps == num_train_timesteps, we can end up with + # duplicates in timesteps. + _, unique_indices = np.unique(timesteps, return_index=True) + timesteps = timesteps[np.sort(unique_indices)] + + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64) + + self.num_inference_steps = len(timesteps) + + self.model_outputs = [ + None, + ] * self.config.solver_order + self.lower_order_nums = 0 + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t + def _sigma_to_alpha_sigma_t(self, sigma): + if self.config.use_flow_sigmas: + alpha_t = 1 - sigma + sigma_t = sigma + else: + alpha_t = 1 / ((sigma**2 + 1) ** 0.5) + sigma_t = sigma * alpha_t + + return alpha_t, sigma_t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """Constructs an exponential noise schedule.""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.convert_model_output + def convert_model_output( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + """ + Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is + designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an + integral of the data prediction model. + + + + The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both noise + prediction and data prediction models. + + + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + # DPM-Solver++ needs to solve an integral of the data prediction model. + if self.config.algorithm_type in ["dpmsolver++", "sde-dpmsolver++"]: + if self.config.prediction_type == "epsilon": + # DPM-Solver and DPM-Solver++ only need the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + model_output = model_output[:, :3] + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = (sample - sigma_t * model_output) / alpha_t + elif self.config.prediction_type == "sample": + x0_pred = model_output + elif self.config.prediction_type == "v_prediction": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = alpha_t * sample - sigma_t * model_output + elif self.config.prediction_type == "flow_prediction": + sigma_t = self.sigmas[self.step_index] + x0_pred = sample - sigma_t * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + "`v_prediction`, or `flow_prediction` for the DPMSolverMultistepScheduler." + ) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + return x0_pred + + # DPM-Solver needs to solve an integral of the noise prediction model. + elif self.config.algorithm_type in ["dpmsolver", "sde-dpmsolver"]: + if self.config.prediction_type == "epsilon": + # DPM-Solver and DPM-Solver++ only need the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + epsilon = model_output[:, :3] + else: + epsilon = model_output + elif self.config.prediction_type == "sample": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + epsilon = (sample - alpha_t * model_output) / sigma_t + elif self.config.prediction_type == "v_prediction": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + epsilon = alpha_t * model_output + sigma_t * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction` for the DPMSolverMultistepScheduler." + ) + + if self.config.thresholding: + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = (sample - sigma_t * epsilon) / alpha_t + x0_pred = self._threshold_sample(x0_pred) + epsilon = (sample - alpha_t * x0_pred) / sigma_t + + return epsilon + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.dpm_solver_first_order_update + def dpm_solver_first_order_update( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + noise: Optional[torch.Tensor] = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the first-order DPMSolver (equivalent to DDIM). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s = torch.log(alpha_s) - torch.log(sigma_s) + + h = lambda_t - lambda_s + if self.config.algorithm_type == "dpmsolver++": + x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output + elif self.config.algorithm_type == "dpmsolver": + x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + x_t = ( + (sigma_t / sigma_s * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.algorithm_type == "sde-dpmsolver": + assert noise is not None + x_t = ( + (alpha_t / alpha_s) * sample + - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * model_output + + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise + ) + return x_t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.multistep_dpm_solver_second_order_update + def multistep_dpm_solver_second_order_update( + self, + model_output_list: List[torch.Tensor], + *args, + sample: torch.Tensor = None, + noise: Optional[torch.Tensor] = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the second-order multistep DPMSolver. + + Args: + model_output_list (`List[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s0, sigma_s1 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + + m0, m1 = model_output_list[-1], model_output_list[-2] + + h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1 + r0 = h_0 / h + D0, D1 = m0, (1.0 / r0) * (m0 - m1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2211.01095 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + ) + elif self.config.algorithm_type == "dpmsolver": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + ) + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.algorithm_type == "sde-dpmsolver": + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (alpha_t / alpha_s0) * sample + - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * (torch.exp(h) - 1.0)) * D1 + + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (alpha_t / alpha_s0) * sample + - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0 + - 2.0 * (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise + ) + return x_t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.multistep_dpm_solver_third_order_update + def multistep_dpm_solver_third_order_update( + self, + model_output_list: List[torch.Tensor], + *args, + sample: torch.Tensor = None, + noise: Optional[torch.Tensor] = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the third-order multistep DPMSolver. + + Args: + model_output_list (`List[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`): + A current instance of a sample created by diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s0, sigma_s1, sigma_s2 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + self.sigmas[self.step_index - 2], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2) + + m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3] + + h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2 + r0, r1 = h_0 / h, h_1 / h + D0 = m0 + D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2) + D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1) + D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2 + ) + elif self.config.algorithm_type == "dpmsolver": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + - (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2 + ) + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h) - 2.0 * h) / (2.0 * h) ** 2 - 0.5)) * D2 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + return x_t + + def _init_step_index(self, timestep): + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + + index_candidates = (self.timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + self._step_index = step_index + + def step( + self, + model_output: torch.Tensor, + timestep: Union[int, torch.Tensor], + sample: torch.Tensor, + generator=None, + variance_noise: Optional[torch.Tensor] = None, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the multistep DPMSolver. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + variance_noise (`torch.Tensor`): + Alternative to generating noise with `generator` by directly providing the noise for the variance + itself. Useful for methods such as [`CycleDiffusion`]. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Improve numerical stability for small number of steps + lower_order_final = (self.step_index == len(self.timesteps) - 1) and ( + self.config.euler_at_final or (self.config.lower_order_final and len(self.timesteps) < 15) + ) + lower_order_second = ( + (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15 + ) + + model_output = self.convert_model_output(model_output, sample=sample) + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.model_outputs[-1] = model_output + + if self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"] and variance_noise is None: + noise = randn_tensor( + model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype + ) + elif self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"]: + noise = variance_noise + else: + noise = None + + if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final: + prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise) + elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second: + prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise) + else: + prev_sample = self.multistep_dpm_solver_third_order_update(self.model_outputs, sample=sample) + + if self.lower_order_nums < self.config.solver_order: + self.lower_order_nums += 1 + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.scale_model_input + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + step_indices = [] + for timestep in timesteps: + index_candidates = (schedule_timesteps == timestep).nonzero() + if len(index_candidates) == 0: + step_index = len(schedule_timesteps) - 1 + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + step_indices.append(step_index) + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + noisy_samples = alpha_t * original_samples + sigma_t * noise + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_sde.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_sde.py new file mode 100644 index 0000000000000000000000000000000000000000..eeb06773d97756a6eafe4f0bcee8a669925b692e --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_sde.py @@ -0,0 +1,676 @@ +# Copyright 2025 Katherine Crowson, The HuggingFace Team and hlky. All rights reserved. +# +# 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 dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch +import torchsde + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DPMSolverSDE +class DPMSolverSDESchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +class BatchedBrownianTree: + """A wrapper around torchsde.BrownianTree that enables batches of entropy.""" + + def __init__(self, x, t0, t1, seed=None, **kwargs): + t0, t1, self.sign = self.sort(t0, t1) + w0 = kwargs.get("w0", torch.zeros_like(x)) + if seed is None: + seed = torch.randint(0, 2**63 - 1, []).item() + self.batched = True + try: + assert len(seed) == x.shape[0] + w0 = w0[0] + except TypeError: + seed = [seed] + self.batched = False + self.trees = [ + torchsde.BrownianInterval( + t0=t0, + t1=t1, + size=w0.shape, + dtype=w0.dtype, + device=w0.device, + entropy=s, + tol=1e-6, + pool_size=24, + halfway_tree=True, + ) + for s in seed + ] + + @staticmethod + def sort(a, b): + return (a, b, 1) if a < b else (b, a, -1) + + def __call__(self, t0, t1): + t0, t1, sign = self.sort(t0, t1) + w = torch.stack([tree(t0, t1) for tree in self.trees]) * (self.sign * sign) + return w if self.batched else w[0] + + +class BrownianTreeNoiseSampler: + """A noise sampler backed by a torchsde.BrownianTree. + + Args: + x (Tensor): The tensor whose shape, device and dtype to use to generate + random samples. + sigma_min (float): The low end of the valid interval. + sigma_max (float): The high end of the valid interval. + seed (int or List[int]): The random seed. If a list of seeds is + supplied instead of a single integer, then the noise sampler will use one BrownianTree per batch item, each + with its own seed. + transform (callable): A function that maps sigma to the sampler's + internal timestep. + """ + + def __init__(self, x, sigma_min, sigma_max, seed=None, transform=lambda x: x): + self.transform = transform + t0, t1 = self.transform(torch.as_tensor(sigma_min)), self.transform(torch.as_tensor(sigma_max)) + self.tree = BatchedBrownianTree(x, t0, t1, seed) + + def __call__(self, sigma, sigma_next): + t0, t1 = self.transform(torch.as_tensor(sigma)), self.transform(torch.as_tensor(sigma_next)) + return self.tree(t0, t1) / (t1 - t0).abs().sqrt() + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin): + """ + DPMSolverSDEScheduler implements the stochastic sampler from the [Elucidating the Design Space of Diffusion-Based + Generative Models](https://huggingface.co/papers/2206.00364) paper. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.00085): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.012): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear` or `scaled_linear`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + noise_sampler_seed (`int`, *optional*, defaults to `None`): + The random seed to use for the noise sampler. If `None`, a random seed is generated. + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 2 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, # sensible defaults + beta_end: float = 0.012, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + prediction_type: str = "epsilon", + use_karras_sigmas: Optional[bool] = False, + use_exponential_sigmas: Optional[bool] = False, + use_beta_sigmas: Optional[bool] = False, + noise_sampler_seed: Optional[int] = None, + timestep_spacing: str = "linspace", + steps_offset: int = 0, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # set all values + self.set_timesteps(num_train_timesteps, None, num_train_timesteps) + self.use_karras_sigmas = use_karras_sigmas + self.noise_sampler = None + self.noise_sampler_seed = noise_sampler_seed + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + @property + def init_noise_sigma(self): + # standard deviation of the initial noise distribution + if self.config.timestep_spacing in ["linspace", "trailing"]: + return self.sigmas.max() + + return (self.sigmas.max() ** 2 + 1) ** 0.5 + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input( + self, + sample: torch.Tensor, + timestep: Union[float, torch.Tensor], + ) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sigma_input = sigma if self.state_in_first_order else self.mid_point_sigma + sample = sample / ((sigma_input**2 + 1) ** 0.5) + return sample + + def set_timesteps( + self, + num_inference_steps: int, + device: Union[str, torch.device] = None, + num_train_timesteps: Optional[int] = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.num_inference_steps = num_inference_steps + + num_train_timesteps = num_train_timesteps or self.config.num_train_timesteps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = np.linspace(0, num_train_timesteps - 1, num_inference_steps, dtype=float)[::-1].copy() + elif self.config.timestep_spacing == "leading": + step_ratio = num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(float) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(num_train_timesteps, 0, -step_ratio)).round().copy().astype(float) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + second_order_timesteps = self._second_order_timesteps(sigmas, log_sigmas) + + sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32) + sigmas = torch.from_numpy(sigmas).to(device=device) + self.sigmas = torch.cat([sigmas[:1], sigmas[1:-1].repeat_interleave(2), sigmas[-1:]]) + + timesteps = torch.from_numpy(timesteps) + second_order_timesteps = torch.from_numpy(second_order_timesteps) + timesteps = torch.cat([timesteps[:1], timesteps[1:].repeat_interleave(2)]) + timesteps[1::2] = second_order_timesteps + + if str(device).startswith("mps"): + # mps does not support float64 + self.timesteps = timesteps.to(device, dtype=torch.float32) + else: + self.timesteps = timesteps.to(device=device) + + # empty first order variables + self.sample = None + self.mid_point_sigma = None + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + self.noise_sampler = None + + def _second_order_timesteps(self, sigmas, log_sigmas): + def sigma_fn(_t): + return np.exp(-_t) + + def t_fn(_sigma): + return -np.log(_sigma) + + midpoint_ratio = 0.5 + t = t_fn(sigmas) + delta_time = np.diff(t) + t_proposed = t[:-1] + delta_time * midpoint_ratio + sig_proposed = sigma_fn(t_proposed) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sig_proposed]) + return timesteps + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + + sigma_min: float = in_sigmas[-1].item() + sigma_max: float = in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, self.num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """Constructs an exponential noise schedule.""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + @property + def state_in_first_order(self): + return self.sample is None + + def step( + self, + model_output: Union[torch.Tensor, np.ndarray], + timestep: Union[float, torch.Tensor], + sample: Union[torch.Tensor, np.ndarray], + return_dict: bool = True, + s_noise: float = 1.0, + ) -> Union[DPMSolverSDESchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor` or `np.ndarray`): + The direct output from learned diffusion model. + timestep (`float` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor` or `np.ndarray`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_dpmsolver_sde.DPMSolverSDESchedulerOutput`] or + tuple. + s_noise (`float`, *optional*, defaults to 1.0): + Scaling factor for noise added to the sample. + + Returns: + [`~schedulers.scheduling_dpmsolver_sde.DPMSolverSDESchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_dpmsolver_sde.DPMSolverSDESchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + """ + if self.step_index is None: + self._init_step_index(timestep) + + # Create a noise sampler if it hasn't been created yet + if self.noise_sampler is None: + min_sigma, max_sigma = self.sigmas[self.sigmas > 0].min(), self.sigmas.max() + self.noise_sampler = BrownianTreeNoiseSampler(sample, min_sigma, max_sigma, self.noise_sampler_seed) + + # Define functions to compute sigma and t from each other + def sigma_fn(_t: torch.Tensor) -> torch.Tensor: + return _t.neg().exp() + + def t_fn(_sigma: torch.Tensor) -> torch.Tensor: + return _sigma.log().neg() + + if self.state_in_first_order: + sigma = self.sigmas[self.step_index] + sigma_next = self.sigmas[self.step_index + 1] + else: + # 2nd order + sigma = self.sigmas[self.step_index - 1] + sigma_next = self.sigmas[self.step_index] + + # Set the midpoint and step size for the current step + midpoint_ratio = 0.5 + t, t_next = t_fn(sigma), t_fn(sigma_next) + delta_time = t_next - t + t_proposed = t + delta_time * midpoint_ratio + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + sigma_input = sigma if self.state_in_first_order else sigma_fn(t_proposed) + pred_original_sample = sample - sigma_input * model_output + elif self.config.prediction_type == "v_prediction": + sigma_input = sigma if self.state_in_first_order else sigma_fn(t_proposed) + pred_original_sample = model_output * (-sigma_input / (sigma_input**2 + 1) ** 0.5) + ( + sample / (sigma_input**2 + 1) + ) + elif self.config.prediction_type == "sample": + raise NotImplementedError("prediction_type not implemented yet: sample") + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + if sigma_next == 0: + derivative = (sample - pred_original_sample) / sigma + dt = sigma_next - sigma + prev_sample = sample + derivative * dt + else: + if self.state_in_first_order: + t_next = t_proposed + else: + sample = self.sample + + sigma_from = sigma_fn(t) + sigma_to = sigma_fn(t_next) + sigma_up = min(sigma_to, (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5) + sigma_down = (sigma_to**2 - sigma_up**2) ** 0.5 + ancestral_t = t_fn(sigma_down) + prev_sample = (sigma_fn(ancestral_t) / sigma_fn(t)) * sample - ( + t - ancestral_t + ).expm1() * pred_original_sample + prev_sample = prev_sample + self.noise_sampler(sigma_fn(t), sigma_fn(t_next)) * s_noise * sigma_up + + if self.state_in_first_order: + # store for 2nd order step + self.sample = sample + self.mid_point_sigma = sigma_fn(t_next) + else: + # free for "first order mode" + self.sample = None + self.mid_point_sigma = None + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return DPMSolverSDESchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_singlestep.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_singlestep.py new file mode 100644 index 0000000000000000000000000000000000000000..9e3e830039bb141bc3917d64f3bafa539403f1af --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_dpmsolver_singlestep.py @@ -0,0 +1,1168 @@ +# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver + +import math +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import deprecate, is_scipy_available, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +if is_scipy_available(): + import scipy.stats + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin): + """ + `DPMSolverSinglestepScheduler` is a fast dedicated high-order solver for diffusion ODEs. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + solver_order (`int`, defaults to 2): + The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided + sampling, and `solver_order=3` for unconditional sampling. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True` and + `algorithm_type="dpmsolver++"`. + algorithm_type (`str`, defaults to `dpmsolver++`): + Algorithm type for the solver; can be `dpmsolver` or `dpmsolver++` or `sde-dpmsolver++`. The `dpmsolver` + type implements the algorithms in the [DPMSolver](https://huggingface.co/papers/2206.00927) paper, and the + `dpmsolver++` type implements the algorithms in the [DPMSolver++](https://huggingface.co/papers/2211.01095) + paper. It is recommended to use `dpmsolver++` or `sde-dpmsolver++` with `solver_order=2` for guided + sampling like in Stable Diffusion. + solver_type (`str`, defaults to `midpoint`): + Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the + sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers. + lower_order_final (`bool`, defaults to `True`): + Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can + stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + final_sigmas_type (`str`, *optional*, defaults to `"zero"`): + The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final + sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0. + lambda_min_clipped (`float`, defaults to `-inf`): + Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the + cosine (`squaredcos_cap_v2`) noise schedule. + variance_type (`str`, *optional*): + Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output + contains the predicted Gaussian variance. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[np.ndarray] = None, + solver_order: int = 2, + prediction_type: str = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + algorithm_type: str = "dpmsolver++", + solver_type: str = "midpoint", + lower_order_final: bool = False, + use_karras_sigmas: Optional[bool] = False, + use_exponential_sigmas: Optional[bool] = False, + use_beta_sigmas: Optional[bool] = False, + use_flow_sigmas: Optional[bool] = False, + flow_shift: Optional[float] = 1.0, + final_sigmas_type: Optional[str] = "zero", # "zero", "sigma_min" + lambda_min_clipped: float = -float("inf"), + variance_type: Optional[str] = None, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if algorithm_type == "dpmsolver": + deprecation_message = "algorithm_type `dpmsolver` is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead" + deprecate("algorithm_types=dpmsolver", "1.0.0", deprecation_message) + + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + # Currently we only support VP-type noise schedule + self.alpha_t = torch.sqrt(self.alphas_cumprod) + self.sigma_t = torch.sqrt(1 - self.alphas_cumprod) + self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t) + self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5 + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # settings for DPM-Solver + if algorithm_type not in ["dpmsolver", "dpmsolver++", "sde-dpmsolver++"]: + if algorithm_type == "deis": + self.register_to_config(algorithm_type="dpmsolver++") + else: + raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}") + if solver_type not in ["midpoint", "heun"]: + if solver_type in ["logrho", "bh1", "bh2"]: + self.register_to_config(solver_type="midpoint") + else: + raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}") + + if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"] and final_sigmas_type == "zero": + raise ValueError( + f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please choose `sigma_min` instead." + ) + + # setable values + self.num_inference_steps = None + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps) + self.model_outputs = [None] * solver_order + self.sample = None + self.order_list = self.get_order_list(num_train_timesteps) + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + def get_order_list(self, num_inference_steps: int) -> List[int]: + """ + Computes the solver order at each time step. + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + """ + steps = num_inference_steps + order = self.config.solver_order + if order > 3: + raise ValueError("Order > 3 is not supported by this scheduler") + if self.config.lower_order_final: + if order == 3: + if steps % 3 == 0: + orders = [1, 2, 3] * (steps // 3 - 1) + [1, 2] + [1] + elif steps % 3 == 1: + orders = [1, 2, 3] * (steps // 3) + [1] + else: + orders = [1, 2, 3] * (steps // 3) + [1, 2] + elif order == 2: + if steps % 2 == 0: + orders = [1, 2] * (steps // 2 - 1) + [1, 1] + else: + orders = [1, 2] * (steps // 2) + [1] + elif order == 1: + orders = [1] * steps + else: + if order == 3: + orders = [1, 2, 3] * (steps // 3) + elif order == 2: + orders = [1, 2] * (steps // 2) + elif order == 1: + orders = [1] * steps + + if self.config.final_sigmas_type == "zero": + orders[-1] = 1 + + return orders + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_timesteps( + self, + num_inference_steps: int = None, + device: Union[str, torch.device] = None, + timesteps: Optional[List[int]] = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of equal spacing between timesteps schedule is used. If `timesteps` is + passed, `num_inference_steps` must be `None`. + """ + if num_inference_steps is None and timesteps is None: + raise ValueError("Must pass exactly one of `num_inference_steps` or `timesteps`.") + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Must pass exactly one of `num_inference_steps` or `timesteps`.") + if timesteps is not None and self.config.use_karras_sigmas: + raise ValueError("Cannot use `timesteps` when `config.use_karras_sigmas=True`.") + if timesteps is not None and self.config.use_exponential_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_exponential_sigmas = True`.") + if timesteps is not None and self.config.use_beta_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_beta_sigmas = True`.") + + num_inference_steps = num_inference_steps or len(timesteps) + self.num_inference_steps = num_inference_steps + + if timesteps is not None: + timesteps = np.array(timesteps).astype(np.int64) + else: + # Clipping the minimum of all lambda(t) for numerical stability. + # This is critical for cosine (squaredcos_cap_v2) noise schedule. + clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped) + clipped_idx = clipped_idx.item() + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1 - clipped_idx, num_inference_steps + 1) + .round()[::-1][:-1] + .copy() + .astype(np.int64) + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + if self.config.use_karras_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round() + elif self.config.use_exponential_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_beta_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_flow_sigmas: + alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1) + sigmas = 1.0 - alphas + sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy() + timesteps = (sigmas * self.config.num_train_timesteps).copy() + else: + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5 + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f" `final_sigmas_type` must be one of `sigma_min` or `zero`, but got {self.config.final_sigmas_type}" + ) + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + + self.sigmas = torch.from_numpy(sigmas).to(device=device) + + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64) + self.model_outputs = [None] * self.config.solver_order + self.sample = None + + if not self.config.lower_order_final and num_inference_steps % self.config.solver_order != 0: + logger.warning( + "Changing scheduler {self.config} to have `lower_order_final` set to True to handle uneven amount of inference steps. Please make sure to always use an even number of `num_inference steps when using `lower_order_final=False`." + ) + self.register_to_config(lower_order_final=True) + + if not self.config.lower_order_final and self.config.final_sigmas_type == "zero": + logger.warning( + " `last_sigmas_type='zero'` is not supported for `lower_order_final=False`. Changing scheduler {self.config} to have `lower_order_final` set to True." + ) + self.register_to_config(lower_order_final=True) + + self.order_list = self.get_order_list(num_inference_steps) + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t + def _sigma_to_alpha_sigma_t(self, sigma): + if self.config.use_flow_sigmas: + alpha_t = 1 - sigma + sigma_t = sigma + else: + alpha_t = 1 / ((sigma**2 + 1) ** 0.5) + sigma_t = sigma * alpha_t + + return alpha_t, sigma_t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """Constructs an exponential noise schedule.""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def convert_model_output( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + """ + Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is + designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an + integral of the data prediction model. + + + + The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both noise + prediction and data prediction models. + + + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + # DPM-Solver++ needs to solve an integral of the data prediction model. + if self.config.algorithm_type in ["dpmsolver++", "sde-dpmsolver++"]: + if self.config.prediction_type == "epsilon": + # DPM-Solver and DPM-Solver++ only need the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + model_output = model_output[:, :3] + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = (sample - sigma_t * model_output) / alpha_t + elif self.config.prediction_type == "sample": + x0_pred = model_output + elif self.config.prediction_type == "v_prediction": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = alpha_t * sample - sigma_t * model_output + elif self.config.prediction_type == "flow_prediction": + sigma_t = self.sigmas[self.step_index] + x0_pred = sample - sigma_t * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + "`v_prediction`, or `flow_prediction` for the DPMSolverSinglestepScheduler." + ) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + return x0_pred + + # DPM-Solver needs to solve an integral of the noise prediction model. + elif self.config.algorithm_type == "dpmsolver": + if self.config.prediction_type == "epsilon": + # DPM-Solver and DPM-Solver++ only need the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + epsilon = model_output[:, :3] + else: + epsilon = model_output + elif self.config.prediction_type == "sample": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + epsilon = (sample - alpha_t * model_output) / sigma_t + elif self.config.prediction_type == "v_prediction": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + epsilon = alpha_t * model_output + sigma_t * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction` for the DPMSolverSinglestepScheduler." + ) + + if self.config.thresholding: + alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep] + x0_pred = (sample - sigma_t * epsilon) / alpha_t + x0_pred = self._threshold_sample(x0_pred) + epsilon = (sample - alpha_t * x0_pred) / sigma_t + + return epsilon + + def dpm_solver_first_order_update( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + noise: Optional[torch.Tensor] = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the first-order DPMSolver (equivalent to DDIM). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + prev_timestep (`int`): + The previous discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s = torch.log(alpha_s) - torch.log(sigma_s) + h = lambda_t - lambda_s + if self.config.algorithm_type == "dpmsolver++": + x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output + elif self.config.algorithm_type == "dpmsolver": + x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + x_t = ( + (sigma_t / sigma_s * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + return x_t + + def singlestep_dpm_solver_second_order_update( + self, + model_output_list: List[torch.Tensor], + *args, + sample: torch.Tensor = None, + noise: Optional[torch.Tensor] = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the second-order singlestep DPMSolver that computes the solution at time `prev_timestep` from the + time `timestep_list[-2]`. + + Args: + model_output_list (`List[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + timestep (`int`): + The current and latter discrete timestep in the diffusion chain. + prev_timestep (`int`): + The previous discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + sigma_t, sigma_s0, sigma_s1 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + + m0, m1 = model_output_list[-1], model_output_list[-2] + + h, h_0 = lambda_t - lambda_s1, lambda_s0 - lambda_s1 + r0 = h_0 / h + D0, D1 = m1, (1.0 / r0) * (m0 - m1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2211.01095 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s1) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s1) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + ) + elif self.config.algorithm_type == "dpmsolver": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (alpha_t / alpha_s1) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (alpha_t / alpha_s1) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + ) + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s1 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s1 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + return x_t + + def singlestep_dpm_solver_third_order_update( + self, + model_output_list: List[torch.Tensor], + *args, + sample: torch.Tensor = None, + noise: Optional[torch.Tensor] = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the third-order singlestep DPMSolver that computes the solution at time `prev_timestep` from the + time `timestep_list[-3]`. + + Args: + model_output_list (`List[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + timestep (`int`): + The current and latter discrete timestep in the diffusion chain. + prev_timestep (`int`): + The previous discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s0, sigma_s1, sigma_s2 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + self.sigmas[self.step_index - 2], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2) + + m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3] + + h, h_0, h_1 = lambda_t - lambda_s2, lambda_s0 - lambda_s2, lambda_s1 - lambda_s2 + r0, r1 = h_0 / h, h_1 / h + D0 = m2 + D1_0, D1_1 = (1.0 / r1) * (m1 - m2), (1.0 / r0) * (m0 - m2) + D1 = (r0 * D1_0 - r1 * D1_1) / (r0 - r1) + D2 = 2.0 * (D1_1 - D1_0) / (r0 - r1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s2) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1_1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s2) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2 + ) + elif self.config.algorithm_type == "dpmsolver": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (alpha_t / alpha_s2) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1_1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (alpha_t / alpha_s2) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + - (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2 + ) + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s2 * torch.exp(-h)) * sample + + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1_1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s2 * torch.exp(-h)) * sample + + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h) + (-2.0 * h)) / (-2.0 * h) ** 2 - 0.5)) * D2 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + return x_t + + def singlestep_dpm_solver_update( + self, + model_output_list: List[torch.Tensor], + *args, + sample: torch.Tensor = None, + order: int = None, + noise: Optional[torch.Tensor] = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the singlestep DPMSolver. + + Args: + model_output_list (`List[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + timestep (`int`): + The current and latter discrete timestep in the diffusion chain. + prev_timestep (`int`): + The previous discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by diffusion process. + order (`int`): + The solver order at this step. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if order is None: + if len(args) > 3: + order = args[3] + else: + raise ValueError("missing `order` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if order == 1: + return self.dpm_solver_first_order_update(model_output_list[-1], sample=sample, noise=noise) + elif order == 2: + return self.singlestep_dpm_solver_second_order_update(model_output_list, sample=sample, noise=noise) + elif order == 3: + return self.singlestep_dpm_solver_third_order_update(model_output_list, sample=sample, noise=noise) + else: + raise ValueError(f"Order must be 1, 2, 3, got {order}") + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + index_candidates = (schedule_timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + return step_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index + def _init_step_index(self, timestep): + """ + Initialize the step_index counter for the scheduler. + """ + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: Union[int, torch.Tensor], + sample: torch.Tensor, + generator=None, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the singlestep DPMSolver. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + model_output = self.convert_model_output(model_output, sample=sample) + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.model_outputs[-1] = model_output + + if self.config.algorithm_type == "sde-dpmsolver++": + noise = randn_tensor( + model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype + ) + else: + noise = None + + order = self.order_list[self.step_index] + + # For img2img denoising might start with order>1 which is not possible + # In this case make sure that the first two steps are both order=1 + while self.model_outputs[-order] is None: + order -= 1 + + # For single-step solvers, we use the initial value at each time with order = 1. + if order == 1: + self.sample = sample + + prev_sample = self.singlestep_dpm_solver_update( + self.model_outputs, sample=self.sample, order=order, noise=noise + ) + + # upon completion increase step index by one, noise=noise + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # begin_index is None when the scheduler is used for training or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + noisy_samples = alpha_t * original_samples + sigma_t * noise + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_edm_dpmsolver_multistep.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_edm_dpmsolver_multistep.py new file mode 100644 index 0000000000000000000000000000000000000000..f1b38aaff56c5c212a0312196d411e4e550d9da4 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_edm_dpmsolver_multistep.py @@ -0,0 +1,712 @@ +# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver and https://github.com/NVlabs/edm + +import math +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin, SchedulerOutput + + +class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin): + """ + Implements DPMSolverMultistepScheduler in EDM formulation as presented in Karras et al. 2022 [1]. + `EDMDPMSolverMultistepScheduler` is a fast dedicated high-order solver for diffusion ODEs. + + [1] Karras, Tero, et al. "Elucidating the Design Space of Diffusion-Based Generative Models." + https://huggingface.co/papers/2206.00364 + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + sigma_min (`float`, *optional*, defaults to 0.002): + Minimum noise magnitude in the sigma schedule. This was set to 0.002 in the EDM paper [1]; a reasonable + range is [0, 10]. + sigma_max (`float`, *optional*, defaults to 80.0): + Maximum noise magnitude in the sigma schedule. This was set to 80.0 in the EDM paper [1]; a reasonable + range is [0.2, 80.0]. + sigma_data (`float`, *optional*, defaults to 0.5): + The standard deviation of the data distribution. This is set to 0.5 in the EDM paper [1]. + sigma_schedule (`str`, *optional*, defaults to `karras`): + Sigma schedule to compute the `sigmas`. By default, we the schedule introduced in the EDM paper + (https://huggingface.co/papers/2206.00364). Other acceptable value is "exponential". The exponential + schedule was incorporated in this model: https://huggingface.co/stabilityai/cosxl. + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + solver_order (`int`, defaults to 2): + The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided + sampling, and `solver_order=3` for unconditional sampling. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True` and + `algorithm_type="dpmsolver++"`. + algorithm_type (`str`, defaults to `dpmsolver++`): + Algorithm type for the solver; can be `dpmsolver++` or `sde-dpmsolver++`. The `dpmsolver++` type implements + the algorithms in the [DPMSolver++](https://huggingface.co/papers/2211.01095) paper. It is recommended to + use `dpmsolver++` or `sde-dpmsolver++` with `solver_order=2` for guided sampling like in Stable Diffusion. + solver_type (`str`, defaults to `midpoint`): + Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the + sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers. + lower_order_final (`bool`, defaults to `True`): + Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can + stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10. + euler_at_final (`bool`, defaults to `False`): + Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail + richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference + steps, but sometimes may result in blurring. + final_sigmas_type (`str`, defaults to `"zero"`): + The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final + sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0. + """ + + _compatibles = [] + order = 1 + + @register_to_config + def __init__( + self, + sigma_min: float = 0.002, + sigma_max: float = 80.0, + sigma_data: float = 0.5, + sigma_schedule: str = "karras", + num_train_timesteps: int = 1000, + prediction_type: str = "epsilon", + rho: float = 7.0, + solver_order: int = 2, + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + algorithm_type: str = "dpmsolver++", + solver_type: str = "midpoint", + lower_order_final: bool = True, + euler_at_final: bool = False, + final_sigmas_type: Optional[str] = "zero", # "zero", "sigma_min" + ): + # settings for DPM-Solver + if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"]: + if algorithm_type == "deis": + self.register_to_config(algorithm_type="dpmsolver++") + else: + raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}") + + if solver_type not in ["midpoint", "heun"]: + if solver_type in ["logrho", "bh1", "bh2"]: + self.register_to_config(solver_type="midpoint") + else: + raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}") + + if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"] and final_sigmas_type == "zero": + raise ValueError( + f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please choose `sigma_min` instead." + ) + + ramp = torch.linspace(0, 1, num_train_timesteps) + if sigma_schedule == "karras": + sigmas = self._compute_karras_sigmas(ramp) + elif sigma_schedule == "exponential": + sigmas = self._compute_exponential_sigmas(ramp) + + self.timesteps = self.precondition_noise(sigmas) + + self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)]) + + # setable values + self.num_inference_steps = None + self.model_outputs = [None] * solver_order + self.lower_order_nums = 0 + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self): + # standard deviation of the initial noise distribution + return (self.config.sigma_max**2 + 1) ** 0.5 + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_inputs + def precondition_inputs(self, sample, sigma): + c_in = self._get_conditioning_c_in(sigma) + scaled_sample = sample * c_in + return scaled_sample + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_noise + def precondition_noise(self, sigma): + if not isinstance(sigma, torch.Tensor): + sigma = torch.tensor([sigma]) + + c_noise = 0.25 * torch.log(sigma) + + return c_noise + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_outputs + def precondition_outputs(self, sample, model_output, sigma): + sigma_data = self.config.sigma_data + c_skip = sigma_data**2 / (sigma**2 + sigma_data**2) + + if self.config.prediction_type == "epsilon": + c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + elif self.config.prediction_type == "v_prediction": + c_out = -sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + else: + raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.") + + denoised = c_skip * sample + c_out * model_output + + return denoised + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.scale_model_input + def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = self.precondition_inputs(sample, sigma) + + self.is_scale_input_called = True + return sample + + def set_timesteps(self, num_inference_steps: int = None, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + + self.num_inference_steps = num_inference_steps + + ramp = torch.linspace(0, 1, self.num_inference_steps) + if self.config.sigma_schedule == "karras": + sigmas = self._compute_karras_sigmas(ramp) + elif self.config.sigma_schedule == "exponential": + sigmas = self._compute_exponential_sigmas(ramp) + + sigmas = sigmas.to(dtype=torch.float32, device=device) + self.timesteps = self.precondition_noise(sigmas) + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = self.config.sigma_min + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + + self.sigmas = torch.cat([sigmas, torch.tensor([sigma_last], dtype=torch.float32, device=device)]) + + self.model_outputs = [ + None, + ] * self.config.solver_order + self.lower_order_nums = 0 + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._compute_karras_sigmas + def _compute_karras_sigmas(self, ramp, sigma_min=None, sigma_max=None) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + sigma_min = sigma_min or self.config.sigma_min + sigma_max = sigma_max or self.config.sigma_max + + rho = self.config.rho + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._compute_exponential_sigmas + def _compute_exponential_sigmas(self, ramp, sigma_min=None, sigma_max=None) -> torch.Tensor: + """Implementation closely follows k-diffusion. + + https://github.com/crowsonkb/k-diffusion/blob/6ab5146d4a5ef63901326489f31f1d8e7dd36b48/k_diffusion/sampling.py#L26 + """ + sigma_min = sigma_min or self.config.sigma_min + sigma_max = sigma_max or self.config.sigma_max + sigmas = torch.linspace(math.log(sigma_min), math.log(sigma_max), len(ramp)).exp().flip(0) + return sigmas + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + def _sigma_to_alpha_sigma_t(self, sigma): + alpha_t = torch.tensor(1) # Inputs are pre-scaled before going into unet, so alpha_t = 1 + sigma_t = sigma + + return alpha_t, sigma_t + + def convert_model_output( + self, + model_output: torch.Tensor, + sample: torch.Tensor = None, + ) -> torch.Tensor: + """ + Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is + designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an + integral of the data prediction model. + + + + The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both noise + prediction and data prediction models. + + + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + sigma = self.sigmas[self.step_index] + x0_pred = self.precondition_outputs(sample, model_output, sigma) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + return x0_pred + + def dpm_solver_first_order_update( + self, + model_output: torch.Tensor, + sample: torch.Tensor = None, + noise: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + """ + One step for the first-order DPMSolver (equivalent to DDIM). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s = torch.log(alpha_s) - torch.log(sigma_s) + + h = lambda_t - lambda_s + if self.config.algorithm_type == "dpmsolver++": + x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + x_t = ( + (sigma_t / sigma_s * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + + return x_t + + def multistep_dpm_solver_second_order_update( + self, + model_output_list: List[torch.Tensor], + sample: torch.Tensor = None, + noise: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + """ + One step for the second-order multistep DPMSolver. + + Args: + model_output_list (`List[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + sigma_t, sigma_s0, sigma_s1 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + + m0, m1 = model_output_list[-1], model_output_list[-2] + + h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1 + r0 = h_0 / h + D0, D1 = m0, (1.0 / r0) * (m0 - m1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2211.01095 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + ) + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + + return x_t + + def multistep_dpm_solver_third_order_update( + self, + model_output_list: List[torch.Tensor], + sample: torch.Tensor = None, + ) -> torch.Tensor: + """ + One step for the third-order multistep DPMSolver. + + Args: + model_output_list (`List[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`): + A current instance of a sample created by diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + sigma_t, sigma_s0, sigma_s1, sigma_s2 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + self.sigmas[self.step_index - 2], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2) + + m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3] + + h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2 + r0, r1 = h_0 / h, h_1 / h + D0 = m0 + D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2) + D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1) + D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2 + ) + + return x_t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + index_candidates = (schedule_timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + return step_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index + def _init_step_index(self, timestep): + """ + Initialize the step_index counter for the scheduler. + """ + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: Union[int, torch.Tensor], + sample: torch.Tensor, + generator=None, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the multistep DPMSolver. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Improve numerical stability for small number of steps + lower_order_final = (self.step_index == len(self.timesteps) - 1) and ( + self.config.euler_at_final + or (self.config.lower_order_final and len(self.timesteps) < 15) + or self.config.final_sigmas_type == "zero" + ) + lower_order_second = ( + (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15 + ) + + model_output = self.convert_model_output(model_output, sample=sample) + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.model_outputs[-1] = model_output + + if self.config.algorithm_type == "sde-dpmsolver++": + noise = randn_tensor( + model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype + ) + else: + noise = None + + if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final: + prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise) + elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second: + prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise) + else: + prev_sample = self.multistep_dpm_solver_third_order_update(self.model_outputs, sample=sample) + + if self.lower_order_nums < self.config.solver_order: + self.lower_order_nums += 1 + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._get_conditioning_c_in + def _get_conditioning_c_in(self, sigma): + c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5) + return c_in + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_edm_euler.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_edm_euler.py new file mode 100644 index 0000000000000000000000000000000000000000..dbeff3de56527c3591a44aaa523d449d7b60c374 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_edm_euler.py @@ -0,0 +1,448 @@ +# Copyright 2025 Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# 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 dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->EulerDiscrete +class EDMEulerSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +class EDMEulerScheduler(SchedulerMixin, ConfigMixin): + """ + Implements the Euler scheduler in EDM formulation as presented in Karras et al. 2022 [1]. + + [1] Karras, Tero, et al. "Elucidating the Design Space of Diffusion-Based Generative Models." + https://huggingface.co/papers/2206.00364 + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + sigma_min (`float`, *optional*, defaults to 0.002): + Minimum noise magnitude in the sigma schedule. This was set to 0.002 in the EDM paper [1]; a reasonable + range is [0, 10]. + sigma_max (`float`, *optional*, defaults to 80.0): + Maximum noise magnitude in the sigma schedule. This was set to 80.0 in the EDM paper [1]; a reasonable + range is [0.2, 80.0]. + sigma_data (`float`, *optional*, defaults to 0.5): + The standard deviation of the data distribution. This is set to 0.5 in the EDM paper [1]. + sigma_schedule (`str`, *optional*, defaults to `karras`): + Sigma schedule to compute the `sigmas`. By default, we the schedule introduced in the EDM paper + (https://huggingface.co/papers/2206.00364). Other acceptable value is "exponential". The exponential + schedule was incorporated in this model: https://huggingface.co/stabilityai/cosxl. + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + rho (`float`, *optional*, defaults to 7.0): + The rho parameter used for calculating the Karras sigma schedule, which is set to 7.0 in the EDM paper [1]. + final_sigmas_type (`str`, defaults to `"zero"`): + The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final + sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0. + """ + + _compatibles = [] + order = 1 + + @register_to_config + def __init__( + self, + sigma_min: float = 0.002, + sigma_max: float = 80.0, + sigma_data: float = 0.5, + sigma_schedule: str = "karras", + num_train_timesteps: int = 1000, + prediction_type: str = "epsilon", + rho: float = 7.0, + final_sigmas_type: str = "zero", # can be "zero" or "sigma_min" + ): + if sigma_schedule not in ["karras", "exponential"]: + raise ValueError(f"Wrong value for provided for `{sigma_schedule=}`.`") + + # setable values + self.num_inference_steps = None + + sigmas_dtype = torch.float32 if torch.backends.mps.is_available() else torch.float64 + sigmas = torch.arange(num_train_timesteps + 1, dtype=sigmas_dtype) / num_train_timesteps + if sigma_schedule == "karras": + sigmas = self._compute_karras_sigmas(sigmas) + elif sigma_schedule == "exponential": + sigmas = self._compute_exponential_sigmas(sigmas) + sigmas = sigmas.to(torch.float32) + + self.timesteps = self.precondition_noise(sigmas) + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = sigmas[-1] + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + + self.sigmas = torch.cat([sigmas, torch.full((1,), fill_value=sigma_last, device=sigmas.device)]) + + self.is_scale_input_called = False + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self): + # standard deviation of the initial noise distribution + return (self.config.sigma_max**2 + 1) ** 0.5 + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def precondition_inputs(self, sample, sigma): + c_in = self._get_conditioning_c_in(sigma) + scaled_sample = sample * c_in + return scaled_sample + + def precondition_noise(self, sigma): + if not isinstance(sigma, torch.Tensor): + sigma = torch.tensor([sigma]) + + c_noise = 0.25 * torch.log(sigma) + + return c_noise + + def precondition_outputs(self, sample, model_output, sigma): + sigma_data = self.config.sigma_data + c_skip = sigma_data**2 / (sigma**2 + sigma_data**2) + + if self.config.prediction_type == "epsilon": + c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + elif self.config.prediction_type == "v_prediction": + c_out = -sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + else: + raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.") + + denoised = c_skip * sample + c_out * model_output + + return denoised + + def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = self.precondition_inputs(sample, sigma) + + self.is_scale_input_called = True + return sample + + def set_timesteps( + self, + num_inference_steps: int = None, + device: Union[str, torch.device] = None, + sigmas: Optional[Union[torch.Tensor, List[float]]] = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + sigmas (`Union[torch.Tensor, List[float]]`, *optional*): + Custom sigmas to use for the denoising process. If not defined, the default behavior when + `num_inference_steps` is passed will be used. + """ + self.num_inference_steps = num_inference_steps + + sigmas_dtype = torch.float32 if torch.backends.mps.is_available() else torch.float64 + if sigmas is None: + sigmas = torch.linspace(0, 1, self.num_inference_steps, dtype=sigmas_dtype) + elif isinstance(sigmas, float): + sigmas = torch.tensor(sigmas, dtype=sigmas_dtype) + else: + sigmas = sigmas.to(sigmas_dtype) + if self.config.sigma_schedule == "karras": + sigmas = self._compute_karras_sigmas(sigmas) + elif self.config.sigma_schedule == "exponential": + sigmas = self._compute_exponential_sigmas(sigmas) + sigmas = sigmas.to(dtype=torch.float32, device=device) + + self.timesteps = self.precondition_noise(sigmas) + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = sigmas[-1] + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + + self.sigmas = torch.cat([sigmas, torch.full((1,), fill_value=sigma_last, device=sigmas.device)]) + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Taken from https://github.com/crowsonkb/k-diffusion/blob/686dbad0f39640ea25c8a8c6a6e56bb40eacefa2/k_diffusion/sampling.py#L17 + def _compute_karras_sigmas(self, ramp, sigma_min=None, sigma_max=None) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + sigma_min = sigma_min or self.config.sigma_min + sigma_max = sigma_max or self.config.sigma_max + + rho = self.config.rho + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + def _compute_exponential_sigmas(self, ramp, sigma_min=None, sigma_max=None) -> torch.Tensor: + """Implementation closely follows k-diffusion. + + https://github.com/crowsonkb/k-diffusion/blob/6ab5146d4a5ef63901326489f31f1d8e7dd36b48/k_diffusion/sampling.py#L26 + """ + sigma_min = sigma_min or self.config.sigma_min + sigma_max = sigma_max or self.config.sigma_max + sigmas = torch.linspace(math.log(sigma_min), math.log(sigma_max), len(ramp)).exp().flip(0) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: Union[float, torch.Tensor], + sample: torch.Tensor, + s_churn: float = 0.0, + s_tmin: float = 0.0, + s_tmax: float = float("inf"), + s_noise: float = 1.0, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + pred_original_sample: Optional[torch.Tensor] = None, + ) -> Union[EDMEulerSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + s_churn (`float`): + s_tmin (`float`): + s_tmax (`float`): + s_noise (`float`, defaults to 1.0): + Scaling factor for noise added to the sample. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_euler_discrete.EDMEulerSchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_euler_discrete.EDMEulerSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_euler_discrete.EDMEulerSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + """ + + if isinstance(timestep, (int, torch.IntTensor, torch.LongTensor)): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `EDMEulerScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if not self.is_scale_input_called: + logger.warning( + "The `scale_model_input` function should be called before `step` to ensure correct denoising. " + "See `StableDiffusionPipeline` for a usage example." + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + + sigma = self.sigmas[self.step_index] + + gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0 + + sigma_hat = sigma * (gamma + 1) + + if gamma > 0: + noise = randn_tensor( + model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator + ) + eps = noise * s_noise + sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5 + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if pred_original_sample is None: + pred_original_sample = self.precondition_outputs(sample, model_output, sigma_hat) + + # 2. Convert to an ODE derivative + derivative = (sample - pred_original_sample) / sigma_hat + + dt = self.sigmas[self.step_index + 1] - sigma_hat + + prev_sample = sample + derivative * dt + + # Cast sample back to model compatible dtype + prev_sample = prev_sample.to(model_output.dtype) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return EDMEulerSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def _get_conditioning_c_in(self, sigma): + c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5) + return c_in + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_euler_ancestral_discrete.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_euler_ancestral_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..9cdaa2c5e10116b5b000c976e10f06e4cd8f0138 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_euler_ancestral_discrete.py @@ -0,0 +1,482 @@ +# Copyright 2025 Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# 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 dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->EulerAncestralDiscrete +class EulerAncestralDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas): + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + + Args: + betas (`torch.Tensor`): + the betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + Ancestral sampling with Euler method steps. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear` or `scaled_linear`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + prediction_type: str = "epsilon", + timestep_spacing: str = "linspace", + steps_offset: int = 0, + rescale_betas_zero_snr: bool = False, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + if rescale_betas_zero_snr: + # Close to 0 without being 0 so first sigma is not inf + # FP16 smallest positive subnormal works well here + self.alphas_cumprod[-1] = 2**-24 + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + sigmas = np.concatenate([sigmas[::-1], [0.0]]).astype(np.float32) + self.sigmas = torch.from_numpy(sigmas) + + # setable values + self.num_inference_steps = None + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=float)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps) + self.is_scale_input_called = False + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self): + # standard deviation of the initial noise distribution + if self.config.timestep_spacing in ["linspace", "trailing"]: + return self.sigmas.max() + + return (self.sigmas.max() ** 2 + 1) ** 0.5 + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = sample / ((sigma**2 + 1) ** 0.5) + self.is_scale_input_called = True + return sample + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.num_inference_steps = num_inference_steps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[ + ::-1 + ].copy() + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(self.config.num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32) + self.sigmas = torch.from_numpy(sigmas).to(device=device) + + self.timesteps = torch.from_numpy(timesteps).to(device=device) + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: Union[float, torch.Tensor], + sample: torch.Tensor, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[EulerAncestralDiscreteSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`): + Whether or not to return a + [`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] or `tuple`: + If return_dict is `True`, + [`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] is returned, + otherwise a tuple is returned where the first element is the sample tensor. + + """ + + if isinstance(timestep, (int, torch.IntTensor, torch.LongTensor)): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if not self.is_scale_input_called: + logger.warning( + "The `scale_model_input` function should be called before `step` to ensure correct denoising. " + "See `StableDiffusionPipeline` for a usage example." + ) + + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + pred_original_sample = sample - sigma * model_output + elif self.config.prediction_type == "v_prediction": + # * c_out + input * c_skip + pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1)) + elif self.config.prediction_type == "sample": + raise NotImplementedError("prediction_type not implemented yet: sample") + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + sigma_from = self.sigmas[self.step_index] + sigma_to = self.sigmas[self.step_index + 1] + sigma_up = (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5 + sigma_down = (sigma_to**2 - sigma_up**2) ** 0.5 + + # 2. Convert to an ODE derivative + derivative = (sample - pred_original_sample) / sigma + + dt = sigma_down - sigma + + prev_sample = sample + derivative * dt + + device = model_output.device + noise = randn_tensor(model_output.shape, dtype=model_output.dtype, device=device, generator=generator) + + prev_sample = prev_sample + noise * sigma_up + + # Cast sample back to model compatible dtype + prev_sample = prev_sample.to(model_output.dtype) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return EulerAncestralDiscreteSchedulerOutput( + prev_sample=prev_sample, pred_original_sample=pred_original_sample + ) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_euler_discrete.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_euler_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..f58d918dbfbe54babe068eb0b2d60189bf92e61e --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_euler_discrete.py @@ -0,0 +1,757 @@ +# Copyright 2025 Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# 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 dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->EulerDiscrete +class EulerDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas): + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + + Args: + betas (`torch.Tensor`): + the betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + Euler scheduler. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear` or `scaled_linear`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + interpolation_type(`str`, defaults to `"linear"`, *optional*): + The interpolation type to compute intermediate sigmas for the scheduler denoising steps. Should be on of + `"linear"` or `"log_linear"`. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + final_sigmas_type (`str`, defaults to `"zero"`): + The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final + sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + prediction_type: str = "epsilon", + interpolation_type: str = "linear", + use_karras_sigmas: Optional[bool] = False, + use_exponential_sigmas: Optional[bool] = False, + use_beta_sigmas: Optional[bool] = False, + sigma_min: Optional[float] = None, + sigma_max: Optional[float] = None, + timestep_spacing: str = "linspace", + timestep_type: str = "discrete", # can be "discrete" or "continuous" + steps_offset: int = 0, + rescale_betas_zero_snr: bool = False, + final_sigmas_type: str = "zero", # can be "zero" or "sigma_min" + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + if rescale_betas_zero_snr: + # Close to 0 without being 0 so first sigma is not inf + # FP16 smallest positive subnormal works well here + self.alphas_cumprod[-1] = 2**-24 + + sigmas = (((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5).flip(0) + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=float)[::-1].copy() + timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32) + + # setable values + self.num_inference_steps = None + + # TODO: Support the full EDM scalings for all prediction types and timestep types + if timestep_type == "continuous" and prediction_type == "v_prediction": + self.timesteps = torch.Tensor([0.25 * sigma.log() for sigma in sigmas]) + else: + self.timesteps = timesteps + + self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)]) + + self.is_scale_input_called = False + self.use_karras_sigmas = use_karras_sigmas + self.use_exponential_sigmas = use_exponential_sigmas + self.use_beta_sigmas = use_beta_sigmas + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self): + # standard deviation of the initial noise distribution + max_sigma = max(self.sigmas) if isinstance(self.sigmas, list) else self.sigmas.max() + if self.config.timestep_spacing in ["linspace", "trailing"]: + return max_sigma + + return (max_sigma**2 + 1) ** 0.5 + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = sample / ((sigma**2 + 1) ** 0.5) + + self.is_scale_input_called = True + return sample + + def set_timesteps( + self, + num_inference_steps: int = None, + device: Union[str, torch.device] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to support arbitrary timesteps schedule. If `None`, timesteps will be generated + based on the `timestep_spacing` attribute. If `timesteps` is passed, `num_inference_steps` and `sigmas` + must be `None`, and `timestep_spacing` attribute will be ignored. + sigmas (`List[float]`, *optional*): + Custom sigmas used to support arbitrary timesteps schedule schedule. If `None`, timesteps and sigmas + will be generated based on the relevant scheduler attributes. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`, and the timesteps will be generated based on the + custom sigmas schedule. + """ + + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` should be set.") + if num_inference_steps is None and timesteps is None and sigmas is None: + raise ValueError("Must pass exactly one of `num_inference_steps` or `timesteps` or `sigmas.") + if num_inference_steps is not None and (timesteps is not None or sigmas is not None): + raise ValueError("Can only pass one of `num_inference_steps` or `timesteps` or `sigmas`.") + if timesteps is not None and self.config.use_karras_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_karras_sigmas = True`.") + if timesteps is not None and self.config.use_exponential_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_exponential_sigmas = True`.") + if timesteps is not None and self.config.use_beta_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_beta_sigmas = True`.") + if ( + timesteps is not None + and self.config.timestep_type == "continuous" + and self.config.prediction_type == "v_prediction" + ): + raise ValueError( + "Cannot set `timesteps` with `config.timestep_type = 'continuous'` and `config.prediction_type = 'v_prediction'`." + ) + + if num_inference_steps is None: + num_inference_steps = len(timesteps) if timesteps is not None else len(sigmas) - 1 + self.num_inference_steps = num_inference_steps + + if sigmas is not None: + log_sigmas = np.log(np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)) + sigmas = np.array(sigmas).astype(np.float32) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas[:-1]]) + + else: + if timesteps is not None: + timesteps = np.array(timesteps).astype(np.float32) + else: + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = np.linspace( + 0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=np.float32 + )[::-1].copy() + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = ( + (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32) + ) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = ( + (np.arange(self.config.num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32) + ) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + if self.config.interpolation_type == "linear": + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + elif self.config.interpolation_type == "log_linear": + sigmas = torch.linspace(np.log(sigmas[-1]), np.log(sigmas[0]), num_inference_steps + 1).exp().numpy() + else: + raise ValueError( + f"{self.config.interpolation_type} is not implemented. Please specify interpolation_type to either" + " 'linear' or 'log_linear'" + ) + + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=self.num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5 + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + + sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device) + + # TODO: Support the full EDM scalings for all prediction types and timestep types + if self.config.timestep_type == "continuous" and self.config.prediction_type == "v_prediction": + self.timesteps = torch.Tensor([0.25 * sigma.log() for sigma in sigmas[:-1]]).to(device=device) + else: + self.timesteps = torch.from_numpy(timesteps.astype(np.float32)).to(device=device) + + self._step_index = None + self._begin_index = None + self.sigmas = sigmas.to("cpu") # to avoid too much CPU/GPU communication + + def _sigma_to_t(self, sigma, log_sigmas): + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from https://github.com/crowsonkb/k-diffusion/blob/686dbad0f39640ea25c8a8c6a6e56bb40eacefa2/k_diffusion/sampling.py#L17 + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from https://github.com/crowsonkb/k-diffusion/blob/686dbad0f39640ea25c8a8c6a6e56bb40eacefa2/k_diffusion/sampling.py#L26 + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """Constructs an exponential noise schedule.""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: Union[float, torch.Tensor], + sample: torch.Tensor, + s_churn: float = 0.0, + s_tmin: float = 0.0, + s_tmax: float = float("inf"), + s_noise: float = 1.0, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[EulerDiscreteSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + s_churn (`float`): + s_tmin (`float`): + s_tmax (`float`): + s_noise (`float`, defaults to 1.0): + Scaling factor for noise added to the sample. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or + tuple. + + Returns: + [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + """ + + if isinstance(timestep, (int, torch.IntTensor, torch.LongTensor)): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if not self.is_scale_input_called: + logger.warning( + "The `scale_model_input` function should be called before `step` to ensure correct denoising. " + "See `StableDiffusionPipeline` for a usage example." + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + + sigma = self.sigmas[self.step_index] + + gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0 + + sigma_hat = sigma * (gamma + 1) + + if gamma > 0: + noise = randn_tensor( + model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator + ) + eps = noise * s_noise + sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5 + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + # NOTE: "original_sample" should not be an expected prediction_type but is left in for + # backwards compatibility + if self.config.prediction_type == "original_sample" or self.config.prediction_type == "sample": + pred_original_sample = model_output + elif self.config.prediction_type == "epsilon": + pred_original_sample = sample - sigma_hat * model_output + elif self.config.prediction_type == "v_prediction": + # denoised = model_output * c_out + input * c_skip + pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1)) + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + # 2. Convert to an ODE derivative + derivative = (sample - pred_original_sample) / sigma_hat + + dt = self.sigmas[self.step_index + 1] - sigma_hat + + prev_sample = sample + derivative * dt + + # Cast sample back to model compatible dtype + prev_sample = prev_sample.to(model_output.dtype) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return EulerDiscreteSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.Tensor) -> torch.Tensor: + if ( + isinstance(timesteps, int) + or isinstance(timesteps, torch.IntTensor) + or isinstance(timesteps, torch.LongTensor) + ): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `EulerDiscreteScheduler.get_velocity()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if sample.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(sample.device, dtype=torch.float32) + timesteps = timesteps.to(sample.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(sample.device) + timesteps = timesteps.to(sample.device) + + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + alphas_cumprod = self.alphas_cumprod.to(sample) + sqrt_alpha_prod = alphas_cumprod[step_indices] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[step_indices]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_euler_discrete_flax.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_euler_discrete_flax.py new file mode 100644 index 0000000000000000000000000000000000000000..dae01302ac6a440f7903cf8b1e22da250b9a3434 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_euler_discrete_flax.py @@ -0,0 +1,265 @@ +# Copyright 2025 Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass +from typing import Optional, Tuple, Union + +import flax +import jax.numpy as jnp + +from ..configuration_utils import ConfigMixin, register_to_config +from .scheduling_utils_flax import ( + CommonSchedulerState, + FlaxKarrasDiffusionSchedulers, + FlaxSchedulerMixin, + FlaxSchedulerOutput, + broadcast_to_shape_from_left, +) + + +@flax.struct.dataclass +class EulerDiscreteSchedulerState: + common: CommonSchedulerState + + # setable values + init_noise_sigma: jnp.ndarray + timesteps: jnp.ndarray + sigmas: jnp.ndarray + num_inference_steps: Optional[int] = None + + @classmethod + def create( + cls, common: CommonSchedulerState, init_noise_sigma: jnp.ndarray, timesteps: jnp.ndarray, sigmas: jnp.ndarray + ): + return cls(common=common, init_noise_sigma=init_noise_sigma, timesteps=timesteps, sigmas=sigmas) + + +@dataclass +class FlaxEulerDiscreteSchedulerOutput(FlaxSchedulerOutput): + state: EulerDiscreteSchedulerState + + +class FlaxEulerDiscreteScheduler(FlaxSchedulerMixin, ConfigMixin): + """ + Euler scheduler (Algorithm 2) from Karras et al. (2022) https://huggingface.co/papers/2206.00364. . Based on the + original k-diffusion implementation by Katherine Crowson: + https://github.com/crowsonkb/k-diffusion/blob/481677d114f6ea445aa009cf5bd7a9cdee909e47/k_diffusion/sampling.py#L51 + + + [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` + function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. + [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and + [`~SchedulerMixin.from_pretrained`] functions. + + Args: + num_train_timesteps (`int`): number of diffusion steps used to train the model. + beta_start (`float`): the starting `beta` value of inference. + beta_end (`float`): the final `beta` value. + beta_schedule (`str`): + the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear` or `scaled_linear`. + trained_betas (`jnp.ndarray`, optional): + option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc. + prediction_type (`str`, default `epsilon`, optional): + prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion + process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4 + https://imagen.research.google/video/paper.pdf) + dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`): + the `dtype` used for params and computation. + """ + + _compatibles = [e.name for e in FlaxKarrasDiffusionSchedulers] + + dtype: jnp.dtype + + @property + def has_state(self): + return True + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[jnp.ndarray] = None, + prediction_type: str = "epsilon", + timestep_spacing: str = "linspace", + dtype: jnp.dtype = jnp.float32, + ): + self.dtype = dtype + + def create_state(self, common: Optional[CommonSchedulerState] = None) -> EulerDiscreteSchedulerState: + if common is None: + common = CommonSchedulerState.create(self) + + timesteps = jnp.arange(0, self.config.num_train_timesteps).round()[::-1] + sigmas = ((1 - common.alphas_cumprod) / common.alphas_cumprod) ** 0.5 + sigmas = jnp.interp(timesteps, jnp.arange(0, len(sigmas)), sigmas) + sigmas = jnp.concatenate([sigmas, jnp.array([0.0], dtype=self.dtype)]) + + # standard deviation of the initial noise distribution + if self.config.timestep_spacing in ["linspace", "trailing"]: + init_noise_sigma = sigmas.max() + else: + init_noise_sigma = (sigmas.max() ** 2 + 1) ** 0.5 + + return EulerDiscreteSchedulerState.create( + common=common, + init_noise_sigma=init_noise_sigma, + timesteps=timesteps, + sigmas=sigmas, + ) + + def scale_model_input(self, state: EulerDiscreteSchedulerState, sample: jnp.ndarray, timestep: int) -> jnp.ndarray: + """ + Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm. + + Args: + state (`EulerDiscreteSchedulerState`): + the `FlaxEulerDiscreteScheduler` state data class instance. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + timestep (`int`): + current discrete timestep in the diffusion chain. + + Returns: + `jnp.ndarray`: scaled input sample + """ + (step_index,) = jnp.where(state.timesteps == timestep, size=1) + step_index = step_index[0] + + sigma = state.sigmas[step_index] + sample = sample / ((sigma**2 + 1) ** 0.5) + return sample + + def set_timesteps( + self, state: EulerDiscreteSchedulerState, num_inference_steps: int, shape: Tuple = () + ) -> EulerDiscreteSchedulerState: + """ + Sets the timesteps used for the diffusion chain. Supporting function to be run before inference. + + Args: + state (`EulerDiscreteSchedulerState`): + the `FlaxEulerDiscreteScheduler` state data class instance. + num_inference_steps (`int`): + the number of diffusion steps used when generating samples with a pre-trained model. + """ + + if self.config.timestep_spacing == "linspace": + timesteps = jnp.linspace(self.config.num_train_timesteps - 1, 0, num_inference_steps, dtype=self.dtype) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // num_inference_steps + timesteps = (jnp.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(float) + timesteps += 1 + else: + raise ValueError( + f"timestep_spacing must be one of ['linspace', 'leading'], got {self.config.timestep_spacing}" + ) + + sigmas = ((1 - state.common.alphas_cumprod) / state.common.alphas_cumprod) ** 0.5 + sigmas = jnp.interp(timesteps, jnp.arange(0, len(sigmas)), sigmas) + sigmas = jnp.concatenate([sigmas, jnp.array([0.0], dtype=self.dtype)]) + + # standard deviation of the initial noise distribution + if self.config.timestep_spacing in ["linspace", "trailing"]: + init_noise_sigma = sigmas.max() + else: + init_noise_sigma = (sigmas.max() ** 2 + 1) ** 0.5 + + return state.replace( + timesteps=timesteps, + sigmas=sigmas, + num_inference_steps=num_inference_steps, + init_noise_sigma=init_noise_sigma, + ) + + def step( + self, + state: EulerDiscreteSchedulerState, + model_output: jnp.ndarray, + timestep: int, + sample: jnp.ndarray, + return_dict: bool = True, + ) -> Union[FlaxEulerDiscreteSchedulerOutput, Tuple]: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + state (`EulerDiscreteSchedulerState`): + the `FlaxEulerDiscreteScheduler` state data class instance. + model_output (`jnp.ndarray`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + order: coefficient for multi-step inference. + return_dict (`bool`): option for returning tuple rather than FlaxEulerDiscreteScheduler class + + Returns: + [`FlaxEulerDiscreteScheduler`] or `tuple`: [`FlaxEulerDiscreteScheduler`] if `return_dict` is True, + otherwise a `tuple`. When returning a tuple, the first element is the sample tensor. + + """ + if state.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + (step_index,) = jnp.where(state.timesteps == timestep, size=1) + step_index = step_index[0] + + sigma = state.sigmas[step_index] + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + pred_original_sample = sample - sigma * model_output + elif self.config.prediction_type == "v_prediction": + # * c_out + input * c_skip + pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1)) + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + # 2. Convert to an ODE derivative + derivative = (sample - pred_original_sample) / sigma + + # dt = sigma_down - sigma + dt = state.sigmas[step_index + 1] - sigma + + prev_sample = sample + derivative * dt + + if not return_dict: + return (prev_sample, state) + + return FlaxEulerDiscreteSchedulerOutput(prev_sample=prev_sample, state=state) + + def add_noise( + self, + state: EulerDiscreteSchedulerState, + original_samples: jnp.ndarray, + noise: jnp.ndarray, + timesteps: jnp.ndarray, + ) -> jnp.ndarray: + sigma = state.sigmas[timesteps].flatten() + sigma = broadcast_to_shape_from_left(sigma, noise.shape) + + noisy_samples = original_samples + noise * sigma + + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_flow_match_euler_discrete.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_flow_match_euler_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..1a4f12ddfa539bb671e09e1ed4ab1af76892b406 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_flow_match_euler_discrete.py @@ -0,0 +1,561 @@ +# Copyright 2025 Stability AI, Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# 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 dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available, logging +from .scheduling_utils import SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class FlowMatchEulerDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.FloatTensor + + +class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + Euler scheduler. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + shift (`float`, defaults to 1.0): + The shift value for the timestep schedule. + use_dynamic_shifting (`bool`, defaults to False): + Whether to apply timestep shifting on-the-fly based on the image resolution. + base_shift (`float`, defaults to 0.5): + Value to stabilize image generation. Increasing `base_shift` reduces variation and image is more consistent + with desired output. + max_shift (`float`, defaults to 1.15): + Value change allowed to latent vectors. Increasing `max_shift` encourages more variation and image may be + more exaggerated or stylized. + base_image_seq_len (`int`, defaults to 256): + The base image sequence length. + max_image_seq_len (`int`, defaults to 4096): + The maximum image sequence length. + invert_sigmas (`bool`, defaults to False): + Whether to invert the sigmas. + shift_terminal (`float`, defaults to None): + The end value of the shifted timestep schedule. + use_karras_sigmas (`bool`, defaults to False): + Whether to use Karras sigmas for step sizes in the noise schedule during sampling. + use_exponential_sigmas (`bool`, defaults to False): + Whether to use exponential sigmas for step sizes in the noise schedule during sampling. + use_beta_sigmas (`bool`, defaults to False): + Whether to use beta sigmas for step sizes in the noise schedule during sampling. + time_shift_type (`str`, defaults to "exponential"): + The type of dynamic resolution-dependent timestep shifting to apply. Either "exponential" or "linear". + stochastic_sampling (`bool`, defaults to False): + Whether to use stochastic sampling. + """ + + _compatibles = [] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + shift: float = 1.0, + use_dynamic_shifting: bool = False, + base_shift: Optional[float] = 0.5, + max_shift: Optional[float] = 1.15, + base_image_seq_len: Optional[int] = 256, + max_image_seq_len: Optional[int] = 4096, + invert_sigmas: bool = False, + shift_terminal: Optional[float] = None, + use_karras_sigmas: Optional[bool] = False, + use_exponential_sigmas: Optional[bool] = False, + use_beta_sigmas: Optional[bool] = False, + time_shift_type: str = "exponential", + stochastic_sampling: bool = False, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if time_shift_type not in {"exponential", "linear"}: + raise ValueError("`time_shift_type` must either be 'exponential' or 'linear'.") + + timesteps = np.linspace(1, num_train_timesteps, num_train_timesteps, dtype=np.float32)[::-1].copy() + timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32) + + sigmas = timesteps / num_train_timesteps + if not use_dynamic_shifting: + # when use_dynamic_shifting is True, we apply the timestep shifting on the fly based on the image resolution + sigmas = shift * sigmas / (1 + (shift - 1) * sigmas) + + self.timesteps = sigmas * num_train_timesteps + + self._step_index = None + self._begin_index = None + + self._shift = shift + + self.sigmas = sigmas.to("cpu") # to avoid too much CPU/GPU communication + self.sigma_min = self.sigmas[-1].item() + self.sigma_max = self.sigmas[0].item() + + @property + def shift(self): + """ + The value used for shifting. + """ + return self._shift + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_shift(self, shift: float): + self._shift = shift + + def scale_noise( + self, + sample: torch.FloatTensor, + timestep: Union[float, torch.FloatTensor], + noise: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + """ + Forward process in flow-matching + + Args: + sample (`torch.FloatTensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.FloatTensor`: + A scaled input sample. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=sample.device, dtype=sample.dtype) + + if sample.device.type == "mps" and torch.is_floating_point(timestep): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(sample.device, dtype=torch.float32) + timestep = timestep.to(sample.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(sample.device) + timestep = timestep.to(sample.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timestep] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timestep.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timestep.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(sample.shape): + sigma = sigma.unsqueeze(-1) + + sample = sigma * noise + (1.0 - sigma) * sample + + return sample + + def _sigma_to_t(self, sigma): + return sigma * self.config.num_train_timesteps + + def time_shift(self, mu: float, sigma: float, t: torch.Tensor): + if self.config.time_shift_type == "exponential": + return self._time_shift_exponential(mu, sigma, t) + elif self.config.time_shift_type == "linear": + return self._time_shift_linear(mu, sigma, t) + + def stretch_shift_to_terminal(self, t: torch.Tensor) -> torch.Tensor: + r""" + Stretches and shifts the timestep schedule to ensure it terminates at the configured `shift_terminal` config + value. + + Reference: + https://github.com/Lightricks/LTX-Video/blob/a01a171f8fe3d99dce2728d60a73fecf4d4238ae/ltx_video/schedulers/rf.py#L51 + + Args: + t (`torch.Tensor`): + A tensor of timesteps to be stretched and shifted. + + Returns: + `torch.Tensor`: + A tensor of adjusted timesteps such that the final value equals `self.config.shift_terminal`. + """ + one_minus_z = 1 - t + scale_factor = one_minus_z[-1] / (1 - self.config.shift_terminal) + stretched_t = 1 - (one_minus_z / scale_factor) + return stretched_t + + def set_timesteps( + self, + num_inference_steps: Optional[int] = None, + device: Union[str, torch.device] = None, + sigmas: Optional[List[float]] = None, + mu: Optional[float] = None, + timesteps: Optional[List[float]] = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + sigmas (`List[float]`, *optional*): + Custom values for sigmas to be used for each diffusion step. If `None`, the sigmas are computed + automatically. + mu (`float`, *optional*): + Determines the amount of shifting applied to sigmas when performing resolution-dependent timestep + shifting. + timesteps (`List[float]`, *optional*): + Custom values for timesteps to be used for each diffusion step. If `None`, the timesteps are computed + automatically. + """ + if self.config.use_dynamic_shifting and mu is None: + raise ValueError("`mu` must be passed when `use_dynamic_shifting` is set to be `True`") + + if sigmas is not None and timesteps is not None: + if len(sigmas) != len(timesteps): + raise ValueError("`sigmas` and `timesteps` should have the same length") + + if num_inference_steps is not None: + if (sigmas is not None and len(sigmas) != num_inference_steps) or ( + timesteps is not None and len(timesteps) != num_inference_steps + ): + raise ValueError( + "`sigmas` and `timesteps` should have the same length as num_inference_steps, if `num_inference_steps` is provided" + ) + else: + num_inference_steps = len(sigmas) if sigmas is not None else len(timesteps) + + self.num_inference_steps = num_inference_steps + + # 1. Prepare default sigmas + is_timesteps_provided = timesteps is not None + + if is_timesteps_provided: + timesteps = np.array(timesteps).astype(np.float32) + + if sigmas is None: + if timesteps is None: + timesteps = np.linspace( + self._sigma_to_t(self.sigma_max), self._sigma_to_t(self.sigma_min), num_inference_steps + ) + sigmas = timesteps / self.config.num_train_timesteps + else: + sigmas = np.array(sigmas).astype(np.float32) + num_inference_steps = len(sigmas) + + # 2. Perform timestep shifting. Either no shifting is applied, or resolution-dependent shifting of + # "exponential" or "linear" type is applied + if self.config.use_dynamic_shifting: + sigmas = self.time_shift(mu, 1.0, sigmas) + else: + sigmas = self.shift * sigmas / (1 + (self.shift - 1) * sigmas) + + # 3. If required, stretch the sigmas schedule to terminate at the configured `shift_terminal` value + if self.config.shift_terminal: + sigmas = self.stretch_shift_to_terminal(sigmas) + + # 4. If required, convert sigmas to one of karras, exponential, or beta sigma schedules + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + + # 5. Convert sigmas and timesteps to tensors and move to specified device + sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device) + if not is_timesteps_provided: + timesteps = sigmas * self.config.num_train_timesteps + else: + timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32, device=device) + + # 6. Append the terminal sigma value. + # If a model requires inverted sigma schedule for denoising but timesteps without inversion, the + # `invert_sigmas` flag can be set to `True`. This case is only required in Mochi + if self.config.invert_sigmas: + sigmas = 1.0 - sigmas + timesteps = sigmas * self.config.num_train_timesteps + sigmas = torch.cat([sigmas, torch.ones(1, device=sigmas.device)]) + else: + sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)]) + + self.timesteps = timesteps + self.sigmas = sigmas + self._step_index = None + self._begin_index = None + + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.FloatTensor, + timestep: Union[float, torch.FloatTensor], + sample: torch.FloatTensor, + s_churn: float = 0.0, + s_tmin: float = 0.0, + s_tmax: float = float("inf"), + s_noise: float = 1.0, + generator: Optional[torch.Generator] = None, + per_token_timesteps: Optional[torch.Tensor] = None, + return_dict: bool = True, + ) -> Union[FlowMatchEulerDiscreteSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + s_churn (`float`): + s_tmin (`float`): + s_tmax (`float`): + s_noise (`float`, defaults to 1.0): + Scaling factor for noise added to the sample. + generator (`torch.Generator`, *optional*): + A random number generator. + per_token_timesteps (`torch.Tensor`, *optional*): + The timesteps for each token in the sample. + return_dict (`bool`): + Whether or not to return a + [`~schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteSchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteSchedulerOutput`] or `tuple`: + If return_dict is `True`, + [`~schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteSchedulerOutput`] is returned, + otherwise a tuple is returned where the first element is the sample tensor. + """ + + if ( + isinstance(timestep, int) + or isinstance(timestep, torch.IntTensor) + or isinstance(timestep, torch.LongTensor) + ): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `FlowMatchEulerDiscreteScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + + if per_token_timesteps is not None: + per_token_sigmas = per_token_timesteps / self.config.num_train_timesteps + + sigmas = self.sigmas[:, None, None] + lower_mask = sigmas < per_token_sigmas[None] - 1e-6 + lower_sigmas = lower_mask * sigmas + lower_sigmas, _ = lower_sigmas.max(dim=0) + + current_sigma = per_token_sigmas[..., None] + next_sigma = lower_sigmas[..., None] + dt = current_sigma - next_sigma + else: + sigma_idx = self.step_index + sigma = self.sigmas[sigma_idx] + sigma_next = self.sigmas[sigma_idx + 1] + + current_sigma = sigma + next_sigma = sigma_next + dt = sigma_next - sigma + + if self.config.stochastic_sampling: + x0 = sample - current_sigma * model_output + noise = torch.randn_like(sample) + prev_sample = (1.0 - next_sigma) * x0 + next_sigma * noise + else: + prev_sample = sample + dt * model_output + + # upon completion increase step index by one + self._step_index += 1 + if per_token_timesteps is None: + # Cast sample back to model compatible dtype + prev_sample = prev_sample.to(model_output.dtype) + + if not return_dict: + return (prev_sample,) + + return FlowMatchEulerDiscreteSchedulerOutput(prev_sample=prev_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """Constructs an exponential noise schedule.""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def _time_shift_exponential(self, mu, sigma, t): + return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma) + + def _time_shift_linear(self, mu, sigma, t): + return mu / (mu + (1 / t - 1) ** sigma) + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_flow_match_heun_discrete.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_flow_match_heun_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..38e5f1ba77a8511544a3f10d9ff7382ccda5c83b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_flow_match_heun_discrete.py @@ -0,0 +1,321 @@ +# Copyright 2025 Stability AI, Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass +from typing import Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class FlowMatchHeunDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.FloatTensor + + +class FlowMatchHeunDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + Heun scheduler. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + shift (`float`, defaults to 1.0): + The shift value for the timestep schedule. + """ + + _compatibles = [] + order = 2 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + shift: float = 1.0, + ): + timesteps = np.linspace(1, num_train_timesteps, num_train_timesteps, dtype=np.float32)[::-1].copy() + timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32) + + sigmas = timesteps / num_train_timesteps + sigmas = shift * sigmas / (1 + (shift - 1) * sigmas) + + self.timesteps = sigmas * num_train_timesteps + + self._step_index = None + self._begin_index = None + + self.sigmas = sigmas.to("cpu") # to avoid too much CPU/GPU communication + self.sigma_min = self.sigmas[-1].item() + self.sigma_max = self.sigmas[0].item() + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_noise( + self, + sample: torch.FloatTensor, + timestep: Union[float, torch.FloatTensor], + noise: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + """ + Forward process in flow-matching + + Args: + sample (`torch.FloatTensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.FloatTensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = sigma * noise + (1.0 - sigma) * sample + + return sample + + def _sigma_to_t(self, sigma): + return sigma * self.config.num_train_timesteps + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.num_inference_steps = num_inference_steps + + timesteps = np.linspace( + self._sigma_to_t(self.sigma_max), self._sigma_to_t(self.sigma_min), num_inference_steps + ) + + sigmas = timesteps / self.config.num_train_timesteps + sigmas = self.config.shift * sigmas / (1 + (self.config.shift - 1) * sigmas) + sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device) + + timesteps = sigmas * self.config.num_train_timesteps + timesteps = torch.cat([timesteps[:1], timesteps[1:].repeat_interleave(2)]) + self.timesteps = timesteps.to(device=device) + + sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)]) + self.sigmas = torch.cat([sigmas[:1], sigmas[1:-1].repeat_interleave(2), sigmas[-1:]]) + + # empty dt and derivative + self.prev_derivative = None + self.dt = None + + self._step_index = None + self._begin_index = None + + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + @property + def state_in_first_order(self): + return self.dt is None + + def step( + self, + model_output: torch.FloatTensor, + timestep: Union[float, torch.FloatTensor], + sample: torch.FloatTensor, + s_churn: float = 0.0, + s_tmin: float = 0.0, + s_tmax: float = float("inf"), + s_noise: float = 1.0, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[FlowMatchHeunDiscreteSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + s_churn (`float`): + s_tmin (`float`): + s_tmax (`float`): + s_noise (`float`, defaults to 1.0): + Scaling factor for noise added to the sample. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`): + Whether or not to return a + [`~schedulers.scheduling_flow_match_heun_discrete.FlowMatchHeunDiscreteSchedulerOutput`] tuple. + + Returns: + [`~schedulers.scheduling_flow_match_heun_discrete.FlowMatchHeunDiscreteSchedulerOutput`] or `tuple`: + If return_dict is `True`, + [`~schedulers.scheduling_flow_match_heun_discrete.FlowMatchHeunDiscreteSchedulerOutput`] is returned, + otherwise a tuple is returned where the first element is the sample tensor. + """ + + if ( + isinstance(timestep, int) + or isinstance(timestep, torch.IntTensor) + or isinstance(timestep, torch.LongTensor) + ): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `FlowMatchHeunDiscreteScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + + if self.state_in_first_order: + sigma = self.sigmas[self.step_index] + sigma_next = self.sigmas[self.step_index + 1] + else: + # 2nd order / Heun's method + sigma = self.sigmas[self.step_index - 1] + sigma_next = self.sigmas[self.step_index] + + gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0 + + sigma_hat = sigma * (gamma + 1) + + if gamma > 0: + noise = randn_tensor( + model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator + ) + eps = noise * s_noise + sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5 + + if self.state_in_first_order: + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + denoised = sample - model_output * sigma + # 2. convert to an ODE derivative for 1st order + derivative = (sample - denoised) / sigma_hat + # 3. Delta timestep + dt = sigma_next - sigma_hat + + # store for 2nd order step + self.prev_derivative = derivative + self.dt = dt + self.sample = sample + else: + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + denoised = sample - model_output * sigma_next + # 2. 2nd order / Heun's method + derivative = (sample - denoised) / sigma_next + derivative = 0.5 * (self.prev_derivative + derivative) + + # 3. take prev timestep & sample + dt = self.dt + sample = self.sample + + # free dt and derivative + # Note, this puts the scheduler in "first order mode" + self.prev_derivative = None + self.dt = None + self.sample = None + + prev_sample = sample + derivative * dt + # Cast sample back to model compatible dtype + prev_sample = prev_sample.to(model_output.dtype) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return FlowMatchHeunDiscreteSchedulerOutput(prev_sample=prev_sample) + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_flow_match_lcm.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_flow_match_lcm.py new file mode 100644 index 0000000000000000000000000000000000000000..933bb1cf8e3d05d58ca6216ada6c40e241d59fa2 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_flow_match_lcm.py @@ -0,0 +1,561 @@ +# Copyright 2025 Stability AI, Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# 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 dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class FlowMatchLCMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.FloatTensor + + +class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin): + """ + LCM scheduler for Flow Matching. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + shift (`float`, defaults to 1.0): + The shift value for the timestep schedule. + use_dynamic_shifting (`bool`, defaults to False): + Whether to apply timestep shifting on-the-fly based on the image resolution. + base_shift (`float`, defaults to 0.5): + Value to stabilize image generation. Increasing `base_shift` reduces variation and image is more consistent + with desired output. + max_shift (`float`, defaults to 1.15): + Value change allowed to latent vectors. Increasing `max_shift` encourages more variation and image may be + more exaggerated or stylized. + base_image_seq_len (`int`, defaults to 256): + The base image sequence length. + max_image_seq_len (`int`, defaults to 4096): + The maximum image sequence length. + invert_sigmas (`bool`, defaults to False): + Whether to invert the sigmas. + shift_terminal (`float`, defaults to None): + The end value of the shifted timestep schedule. + use_karras_sigmas (`bool`, defaults to False): + Whether to use Karras sigmas for step sizes in the noise schedule during sampling. + use_exponential_sigmas (`bool`, defaults to False): + Whether to use exponential sigmas for step sizes in the noise schedule during sampling. + use_beta_sigmas (`bool`, defaults to False): + Whether to use beta sigmas for step sizes in the noise schedule during sampling. + time_shift_type (`str`, defaults to "exponential"): + The type of dynamic resolution-dependent timestep shifting to apply. Either "exponential" or "linear". + scale_factors ('list', defaults to None) + It defines how to scale the latents at which predictions are made. + upscale_mode ('str', defaults to 'bicubic') + Upscaling method, applied if scale-wise generation is considered + """ + + _compatibles = [] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + shift: float = 1.0, + use_dynamic_shifting: bool = False, + base_shift: Optional[float] = 0.5, + max_shift: Optional[float] = 1.15, + base_image_seq_len: Optional[int] = 256, + max_image_seq_len: Optional[int] = 4096, + invert_sigmas: bool = False, + shift_terminal: Optional[float] = None, + use_karras_sigmas: Optional[bool] = False, + use_exponential_sigmas: Optional[bool] = False, + use_beta_sigmas: Optional[bool] = False, + time_shift_type: str = "exponential", + scale_factors: Optional[List[float]] = None, + upscale_mode: Optional[str] = "bicubic", + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if time_shift_type not in {"exponential", "linear"}: + raise ValueError("`time_shift_type` must either be 'exponential' or 'linear'.") + + timesteps = np.linspace(1, num_train_timesteps, num_train_timesteps, dtype=np.float32)[::-1].copy() + timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32) + + sigmas = timesteps / num_train_timesteps + if not use_dynamic_shifting: + # when use_dynamic_shifting is True, we apply the timestep shifting on the fly based on the image resolution + sigmas = shift * sigmas / (1 + (shift - 1) * sigmas) + + self.timesteps = sigmas * num_train_timesteps + + self._step_index = None + self._begin_index = None + + self._shift = shift + + self._init_size = None + self._scale_factors = scale_factors + self._upscale_mode = upscale_mode + + self.sigmas = sigmas.to("cpu") # to avoid too much CPU/GPU communication + self.sigma_min = self.sigmas[-1].item() + self.sigma_max = self.sigmas[0].item() + + @property + def shift(self): + """ + The value used for shifting. + """ + return self._shift + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_shift(self, shift: float): + self._shift = shift + + def set_scale_factors(self, scale_factors: list, upscale_mode): + """ + Sets scale factors for a scale-wise generation regime. + + Args: + scale_factors (`list`): + The scale factors for each step + upscale_mode (`str`): + Upscaling method + """ + self._scale_factors = scale_factors + self._upscale_mode = upscale_mode + + def scale_noise( + self, + sample: torch.FloatTensor, + timestep: Union[float, torch.FloatTensor], + noise: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + """ + Forward process in flow-matching + + Args: + sample (`torch.FloatTensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.FloatTensor`: + A scaled input sample. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=sample.device, dtype=sample.dtype) + + if sample.device.type == "mps" and torch.is_floating_point(timestep): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(sample.device, dtype=torch.float32) + timestep = timestep.to(sample.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(sample.device) + timestep = timestep.to(sample.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timestep] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timestep.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timestep.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(sample.shape): + sigma = sigma.unsqueeze(-1) + + sample = sigma * noise + (1.0 - sigma) * sample + + return sample + + def _sigma_to_t(self, sigma): + return sigma * self.config.num_train_timesteps + + def time_shift(self, mu: float, sigma: float, t: torch.Tensor): + if self.config.time_shift_type == "exponential": + return self._time_shift_exponential(mu, sigma, t) + elif self.config.time_shift_type == "linear": + return self._time_shift_linear(mu, sigma, t) + + def stretch_shift_to_terminal(self, t: torch.Tensor) -> torch.Tensor: + r""" + Stretches and shifts the timestep schedule to ensure it terminates at the configured `shift_terminal` config + value. + + Reference: + https://github.com/Lightricks/LTX-Video/blob/a01a171f8fe3d99dce2728d60a73fecf4d4238ae/ltx_video/schedulers/rf.py#L51 + + Args: + t (`torch.Tensor`): + A tensor of timesteps to be stretched and shifted. + + Returns: + `torch.Tensor`: + A tensor of adjusted timesteps such that the final value equals `self.config.shift_terminal`. + """ + one_minus_z = 1 - t + scale_factor = one_minus_z[-1] / (1 - self.config.shift_terminal) + stretched_t = 1 - (one_minus_z / scale_factor) + return stretched_t + + def set_timesteps( + self, + num_inference_steps: Optional[int] = None, + device: Union[str, torch.device] = None, + sigmas: Optional[List[float]] = None, + mu: Optional[float] = None, + timesteps: Optional[List[float]] = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + sigmas (`List[float]`, *optional*): + Custom values for sigmas to be used for each diffusion step. If `None`, the sigmas are computed + automatically. + mu (`float`, *optional*): + Determines the amount of shifting applied to sigmas when performing resolution-dependent timestep + shifting. + timesteps (`List[float]`, *optional*): + Custom values for timesteps to be used for each diffusion step. If `None`, the timesteps are computed + automatically. + """ + if self.config.use_dynamic_shifting and mu is None: + raise ValueError("`mu` must be passed when `use_dynamic_shifting` is set to be `True`") + + if sigmas is not None and timesteps is not None: + if len(sigmas) != len(timesteps): + raise ValueError("`sigmas` and `timesteps` should have the same length") + + if num_inference_steps is not None: + if (sigmas is not None and len(sigmas) != num_inference_steps) or ( + timesteps is not None and len(timesteps) != num_inference_steps + ): + raise ValueError( + "`sigmas` and `timesteps` should have the same length as num_inference_steps, if `num_inference_steps` is provided" + ) + else: + num_inference_steps = len(sigmas) if sigmas is not None else len(timesteps) + + self.num_inference_steps = num_inference_steps + + # 1. Prepare default sigmas + is_timesteps_provided = timesteps is not None + + if is_timesteps_provided: + timesteps = np.array(timesteps).astype(np.float32) + + if sigmas is None: + if timesteps is None: + timesteps = np.linspace( + self._sigma_to_t(self.sigma_max), self._sigma_to_t(self.sigma_min), num_inference_steps + ) + sigmas = timesteps / self.config.num_train_timesteps + else: + sigmas = np.array(sigmas).astype(np.float32) + num_inference_steps = len(sigmas) + + # 2. Perform timestep shifting. Either no shifting is applied, or resolution-dependent shifting of + # "exponential" or "linear" type is applied + if self.config.use_dynamic_shifting: + sigmas = self.time_shift(mu, 1.0, sigmas) + else: + sigmas = self.shift * sigmas / (1 + (self.shift - 1) * sigmas) + + # 3. If required, stretch the sigmas schedule to terminate at the configured `shift_terminal` value + if self.config.shift_terminal: + sigmas = self.stretch_shift_to_terminal(sigmas) + + # 4. If required, convert sigmas to one of karras, exponential, or beta sigma schedules + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + + # 5. Convert sigmas and timesteps to tensors and move to specified device + sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device) + if not is_timesteps_provided: + timesteps = sigmas * self.config.num_train_timesteps + else: + timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32, device=device) + + # 6. Append the terminal sigma value. + # If a model requires inverted sigma schedule for denoising but timesteps without inversion, the + # `invert_sigmas` flag can be set to `True`. This case is only required in Mochi + if self.config.invert_sigmas: + sigmas = 1.0 - sigmas + timesteps = sigmas * self.config.num_train_timesteps + sigmas = torch.cat([sigmas, torch.ones(1, device=sigmas.device)]) + else: + sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)]) + + self.timesteps = timesteps + self.sigmas = sigmas + self._step_index = None + self._begin_index = None + + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.FloatTensor, + timestep: Union[float, torch.FloatTensor], + sample: torch.FloatTensor, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[FlowMatchLCMSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_flow_match_lcm.FlowMatchLCMSchedulerOutput`] or + tuple. + + Returns: + [`~schedulers.scheduling_flow_match_lcm.FlowMatchLCMSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_flow_match_lcm.FlowMatchLCMSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + """ + + if ( + isinstance(timestep, int) + or isinstance(timestep, torch.IntTensor) + or isinstance(timestep, torch.LongTensor) + ): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `FlowMatchLCMScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if self._scale_factors and self._upscale_mode and len(self.timesteps) != len(self._scale_factors) + 1: + raise ValueError( + "`_scale_factors` should have the same length as `timesteps` - 1, if `_scale_factors` are set." + ) + + if self._init_size is None or self.step_index is None: + self._init_size = model_output.size()[2:] + + if self.step_index is None: + self._init_step_index(timestep) + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + + sigma = self.sigmas[self.step_index] + sigma_next = self.sigmas[self.step_index + 1] + x0_pred = sample - sigma * model_output + + if self._scale_factors and self._upscale_mode: + if self._step_index < len(self._scale_factors): + size = [round(self._scale_factors[self._step_index] * size) for size in self._init_size] + x0_pred = torch.nn.functional.interpolate(x0_pred, size=size, mode=self._upscale_mode) + + noise = randn_tensor(x0_pred.shape, generator=generator, device=x0_pred.device, dtype=x0_pred.dtype) + prev_sample = (1 - sigma_next) * x0_pred + sigma_next * noise + + # upon completion increase step index by one + self._step_index += 1 + # Cast sample back to model compatible dtype + prev_sample = prev_sample.to(model_output.dtype) + + if not return_dict: + return (prev_sample,) + + return FlowMatchLCMSchedulerOutput(prev_sample=prev_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """Constructs an exponential noise schedule.""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def _time_shift_exponential(self, mu, sigma, t): + return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma) + + def _time_shift_linear(self, mu, sigma, t): + return mu / (mu + (1 / t - 1) ** sigma) + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_heun_discrete.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_heun_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..bd1239cfaec7fb71e07583f940e32ad700da1b04 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_heun_discrete.py @@ -0,0 +1,610 @@ +# Copyright 2025 Katherine Crowson, The HuggingFace Team and hlky. All rights reserved. +# +# 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 dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->HeunDiscrete +class HeunDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class HeunDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + Scheduler with Heun steps for discrete beta schedules. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear` or `scaled_linear`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to 1.0): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 2 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, # sensible defaults + beta_end: float = 0.012, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + prediction_type: str = "epsilon", + use_karras_sigmas: Optional[bool] = False, + use_exponential_sigmas: Optional[bool] = False, + use_beta_sigmas: Optional[bool] = False, + clip_sample: Optional[bool] = False, + clip_sample_range: float = 1.0, + timestep_spacing: str = "linspace", + steps_offset: int = 0, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps, alpha_transform_type="cosine") + elif beta_schedule == "exp": + self.betas = betas_for_alpha_bar(num_train_timesteps, alpha_transform_type="exp") + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # set all values + self.set_timesteps(num_train_timesteps, None, num_train_timesteps) + self.use_karras_sigmas = use_karras_sigmas + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + @property + def init_noise_sigma(self): + # standard deviation of the initial noise distribution + if self.config.timestep_spacing in ["linspace", "trailing"]: + return self.sigmas.max() + + return (self.sigmas.max() ** 2 + 1) ** 0.5 + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input( + self, + sample: torch.Tensor, + timestep: Union[float, torch.Tensor], + ) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = sample / ((sigma**2 + 1) ** 0.5) + return sample + + def set_timesteps( + self, + num_inference_steps: Optional[int] = None, + device: Union[str, torch.device] = None, + num_train_timesteps: Optional[int] = None, + timesteps: Optional[List[int]] = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + num_train_timesteps (`int`, *optional*): + The number of diffusion steps used when training the model. If `None`, the default + `num_train_timesteps` attribute is used. + timesteps (`List[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, timesteps will be + generated based on the `timestep_spacing` attribute. If `timesteps` is passed, `num_inference_steps` + must be `None`, and `timestep_spacing` attribute will be ignored. + """ + if num_inference_steps is None and timesteps is None: + raise ValueError("Must pass exactly one of `num_inference_steps` or `custom_timesteps`.") + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.") + if timesteps is not None and self.config.use_karras_sigmas: + raise ValueError("Cannot use `timesteps` with `config.use_karras_sigmas = True`") + if timesteps is not None and self.config.use_exponential_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_exponential_sigmas = True`.") + if timesteps is not None and self.config.use_beta_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_beta_sigmas = True`.") + + num_inference_steps = num_inference_steps or len(timesteps) + self.num_inference_steps = num_inference_steps + num_train_timesteps = num_train_timesteps or self.config.num_train_timesteps + + if timesteps is not None: + timesteps = np.array(timesteps, dtype=np.float32) + else: + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = np.linspace(0, num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[::-1].copy() + elif self.config.timestep_spacing == "leading": + step_ratio = num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=self.num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32) + sigmas = torch.from_numpy(sigmas).to(device=device) + self.sigmas = torch.cat([sigmas[:1], sigmas[1:-1].repeat_interleave(2), sigmas[-1:]]) + + timesteps = torch.from_numpy(timesteps) + timesteps = torch.cat([timesteps[:1], timesteps[1:].repeat_interleave(2)]) + + self.timesteps = timesteps.to(device=device, dtype=torch.float32) + + # empty dt and derivative + self.prev_derivative = None + self.dt = None + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """Constructs an exponential noise schedule.""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + @property + def state_in_first_order(self): + return self.dt is None + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: Union[torch.Tensor, np.ndarray], + timestep: Union[float, torch.Tensor], + sample: Union[torch.Tensor, np.ndarray], + return_dict: bool = True, + ) -> Union[HeunDiscreteSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_heun_discrete.HeunDiscreteSchedulerOutput`] or + tuple. + + Returns: + [`~schedulers.scheduling_heun_discrete.HeunDiscreteSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_heun_discrete.HeunDiscreteSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + """ + if self.step_index is None: + self._init_step_index(timestep) + + if self.state_in_first_order: + sigma = self.sigmas[self.step_index] + sigma_next = self.sigmas[self.step_index + 1] + else: + # 2nd order / Heun's method + sigma = self.sigmas[self.step_index - 1] + sigma_next = self.sigmas[self.step_index] + + # currently only gamma=0 is supported. This usually works best anyways. + # We can support gamma in the future but then need to scale the timestep before + # passing it to the model which requires a change in API + gamma = 0 + sigma_hat = sigma * (gamma + 1) # Note: sigma_hat == sigma for now + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + sigma_input = sigma_hat if self.state_in_first_order else sigma_next + pred_original_sample = sample - sigma_input * model_output + elif self.config.prediction_type == "v_prediction": + sigma_input = sigma_hat if self.state_in_first_order else sigma_next + pred_original_sample = model_output * (-sigma_input / (sigma_input**2 + 1) ** 0.5) + ( + sample / (sigma_input**2 + 1) + ) + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + if self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + if self.state_in_first_order: + # 2. Convert to an ODE derivative for 1st order + derivative = (sample - pred_original_sample) / sigma_hat + # 3. delta timestep + dt = sigma_next - sigma_hat + + # store for 2nd order step + self.prev_derivative = derivative + self.dt = dt + self.sample = sample + else: + # 2. 2nd order / Heun's method + derivative = (sample - pred_original_sample) / sigma_next + derivative = (self.prev_derivative + derivative) / 2 + + # 3. take prev timestep & sample + dt = self.dt + sample = self.sample + + # free dt and derivative + # Note, this puts the scheduler in "first order mode" + self.prev_derivative = None + self.dt = None + self.sample = None + + prev_sample = sample + derivative * dt + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return HeunDiscreteSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ipndm.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ipndm.py new file mode 100644 index 0000000000000000000000000000000000000000..23bc21f10ca426c8a3bb864d540f3a6b7688da6f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_ipndm.py @@ -0,0 +1,224 @@ +# Copyright 2025 Zhejiang University Team and The HuggingFace Team. All rights reserved. +# +# 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 typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from .scheduling_utils import SchedulerMixin, SchedulerOutput + + +class IPNDMScheduler(SchedulerMixin, ConfigMixin): + """ + A fourth-order Improved Pseudo Linear Multistep scheduler. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + """ + + order = 1 + + @register_to_config + def __init__( + self, num_train_timesteps: int = 1000, trained_betas: Optional[Union[np.ndarray, List[float]]] = None + ): + # set `betas`, `alphas`, `timesteps` + self.set_timesteps(num_train_timesteps) + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # For now we only support F-PNDM, i.e. the runge-kutta method + # For more information on the algorithm please take a look at the paper: https://huggingface.co/papers/2202.09778 + # mainly at formula (9), (12), (13) and the Algorithm 2. + self.pndm_order = 4 + + # running values + self.ets = [] + self._step_index = None + self._begin_index = None + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.num_inference_steps = num_inference_steps + steps = torch.linspace(1, 0, num_inference_steps + 1)[:-1] + steps = torch.cat([steps, torch.tensor([0.0])]) + + if self.config.trained_betas is not None: + self.betas = torch.tensor(self.config.trained_betas, dtype=torch.float32) + else: + self.betas = torch.sin(steps * math.pi / 2) ** 2 + + self.alphas = (1.0 - self.betas**2) ** 0.5 + + timesteps = (torch.atan2(self.betas, self.alphas) / math.pi * 2)[:-1] + self.timesteps = timesteps.to(device) + + self.ets = [] + self._step_index = None + self._begin_index = None + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: Union[int, torch.Tensor], + sample: torch.Tensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the linear multistep method. It performs one forward pass multiple times to approximate the solution. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + if self.step_index is None: + self._init_step_index(timestep) + + timestep_index = self.step_index + prev_timestep_index = self.step_index + 1 + + ets = sample * self.betas[timestep_index] + model_output * self.alphas[timestep_index] + self.ets.append(ets) + + if len(self.ets) == 1: + ets = self.ets[-1] + elif len(self.ets) == 2: + ets = (3 * self.ets[-1] - self.ets[-2]) / 2 + elif len(self.ets) == 3: + ets = (23 * self.ets[-1] - 16 * self.ets[-2] + 5 * self.ets[-3]) / 12 + else: + ets = (1 / 24) * (55 * self.ets[-1] - 59 * self.ets[-2] + 37 * self.ets[-3] - 9 * self.ets[-4]) + + prev_sample = self._get_prev_sample(sample, timestep_index, prev_timestep_index, ets) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def _get_prev_sample(self, sample, timestep_index, prev_timestep_index, ets): + alpha = self.alphas[timestep_index] + sigma = self.betas[timestep_index] + + next_alpha = self.alphas[prev_timestep_index] + next_sigma = self.betas[prev_timestep_index] + + pred = (sample - sigma * ets) / max(alpha, 1e-8) + prev_sample = next_alpha * pred + ets * next_sigma + + return prev_sample + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_k_dpm_2_ancestral_discrete.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_k_dpm_2_ancestral_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..6588464073a12a759ba1d22675ee98dde240378f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_k_dpm_2_ancestral_discrete.py @@ -0,0 +1,617 @@ +# Copyright 2025 Katherine Crowson, The HuggingFace Team and hlky. All rights reserved. +# +# 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 dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->KDPM2AncestralDiscrete +class KDPM2AncestralDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class KDPM2AncestralDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + KDPM2DiscreteScheduler with ancestral sampling is inspired by the DPMSolver2 and Algorithm 2 from the [Elucidating + the Design Space of Diffusion-Based Generative Models](https://huggingface.co/papers/2206.00364) paper. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.00085): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.012): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear` or `scaled_linear`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 2 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, # sensible defaults + beta_end: float = 0.012, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + use_karras_sigmas: Optional[bool] = False, + use_exponential_sigmas: Optional[bool] = False, + use_beta_sigmas: Optional[bool] = False, + prediction_type: str = "epsilon", + timestep_spacing: str = "linspace", + steps_offset: int = 0, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # set all values + self.set_timesteps(num_train_timesteps, None, num_train_timesteps) + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self): + # standard deviation of the initial noise distribution + if self.config.timestep_spacing in ["linspace", "trailing"]: + return self.sigmas.max() + + return (self.sigmas.max() ** 2 + 1) ** 0.5 + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input( + self, + sample: torch.Tensor, + timestep: Union[float, torch.Tensor], + ) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + if self.state_in_first_order: + sigma = self.sigmas[self.step_index] + else: + sigma = self.sigmas_interpol[self.step_index - 1] + + sample = sample / ((sigma**2 + 1) ** 0.5) + return sample + + def set_timesteps( + self, + num_inference_steps: int, + device: Union[str, torch.device] = None, + num_train_timesteps: Optional[int] = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.num_inference_steps = num_inference_steps + + num_train_timesteps = num_train_timesteps or self.config.num_train_timesteps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = np.linspace(0, num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[::-1].copy() + elif self.config.timestep_spacing == "leading": + step_ratio = num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round() + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + self.log_sigmas = torch.from_numpy(log_sigmas).to(device) + sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32) + sigmas = torch.from_numpy(sigmas).to(device=device) + + # compute up and down sigmas + sigmas_next = sigmas.roll(-1) + sigmas_next[-1] = 0.0 + sigmas_up = (sigmas_next**2 * (sigmas**2 - sigmas_next**2) / sigmas**2) ** 0.5 + sigmas_down = (sigmas_next**2 - sigmas_up**2) ** 0.5 + sigmas_down[-1] = 0.0 + + # compute interpolated sigmas + sigmas_interpol = sigmas.log().lerp(sigmas_down.log(), 0.5).exp() + sigmas_interpol[-2:] = 0.0 + + # set sigmas + self.sigmas = torch.cat([sigmas[:1], sigmas[1:].repeat_interleave(2), sigmas[-1:]]) + self.sigmas_interpol = torch.cat( + [sigmas_interpol[:1], sigmas_interpol[1:].repeat_interleave(2), sigmas_interpol[-1:]] + ) + self.sigmas_up = torch.cat([sigmas_up[:1], sigmas_up[1:].repeat_interleave(2), sigmas_up[-1:]]) + self.sigmas_down = torch.cat([sigmas_down[:1], sigmas_down[1:].repeat_interleave(2), sigmas_down[-1:]]) + + if str(device).startswith("mps"): + timesteps = torch.from_numpy(timesteps).to(device, dtype=torch.float32) + else: + timesteps = torch.from_numpy(timesteps).to(device) + + sigmas_interpol = sigmas_interpol.cpu() + log_sigmas = self.log_sigmas.cpu() + timesteps_interpol = np.array( + [self._sigma_to_t(sigma_interpol, log_sigmas) for sigma_interpol in sigmas_interpol] + ) + + timesteps_interpol = torch.from_numpy(timesteps_interpol).to(device, dtype=timesteps.dtype) + interleaved_timesteps = torch.stack((timesteps_interpol[:-2, None], timesteps[1:, None]), dim=-1).flatten() + + self.timesteps = torch.cat([timesteps[:1], interleaved_timesteps]) + + self.sample = None + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """Constructs an exponential noise schedule.""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + @property + def state_in_first_order(self): + return self.sample is None + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: Union[torch.Tensor, np.ndarray], + timestep: Union[float, torch.Tensor], + sample: Union[torch.Tensor, np.ndarray], + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[KDPM2AncestralDiscreteSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`): + Whether or not to return a + [`~schedulers.scheduling_k_dpm_2_ancestral_discrete.KDPM2AncestralDiscreteSchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_k_dpm_2_ancestral_discrete.KDPM2AncestralDiscreteSchedulerOutput`] or `tuple`: + If return_dict is `True`, + [`~schedulers.scheduling_k_dpm_2_ancestral_discrete.KDPM2AncestralDiscreteSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + """ + if self.step_index is None: + self._init_step_index(timestep) + + if self.state_in_first_order: + sigma = self.sigmas[self.step_index] + sigma_interpol = self.sigmas_interpol[self.step_index] + sigma_up = self.sigmas_up[self.step_index] + sigma_down = self.sigmas_down[self.step_index - 1] + else: + # 2nd order / KPDM2's method + sigma = self.sigmas[self.step_index - 1] + sigma_interpol = self.sigmas_interpol[self.step_index - 1] + sigma_up = self.sigmas_up[self.step_index - 1] + sigma_down = self.sigmas_down[self.step_index - 1] + + # currently only gamma=0 is supported. This usually works best anyways. + # We can support gamma in the future but then need to scale the timestep before + # passing it to the model which requires a change in API + gamma = 0 + sigma_hat = sigma * (gamma + 1) # Note: sigma_hat == sigma for now + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + sigma_input = sigma_hat if self.state_in_first_order else sigma_interpol + pred_original_sample = sample - sigma_input * model_output + elif self.config.prediction_type == "v_prediction": + sigma_input = sigma_hat if self.state_in_first_order else sigma_interpol + pred_original_sample = model_output * (-sigma_input / (sigma_input**2 + 1) ** 0.5) + ( + sample / (sigma_input**2 + 1) + ) + elif self.config.prediction_type == "sample": + raise NotImplementedError("prediction_type not implemented yet: sample") + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + if self.state_in_first_order: + # 2. Convert to an ODE derivative for 1st order + derivative = (sample - pred_original_sample) / sigma_hat + # 3. delta timestep + dt = sigma_interpol - sigma_hat + + # store for 2nd order step + self.sample = sample + self.dt = dt + prev_sample = sample + derivative * dt + else: + # DPM-Solver-2 + # 2. Convert to an ODE derivative for 2nd order + derivative = (sample - pred_original_sample) / sigma_interpol + # 3. delta timestep + dt = sigma_down - sigma_hat + + sample = self.sample + self.sample = None + + prev_sample = sample + derivative * dt + noise = randn_tensor( + model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator + ) + prev_sample = prev_sample + noise * sigma_up + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return KDPM2AncestralDiscreteSchedulerOutput( + prev_sample=prev_sample, pred_original_sample=pred_original_sample + ) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_k_dpm_2_discrete.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_k_dpm_2_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..9b4cd4e204d6220596a86055156f6c1d36b0f7a8 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_k_dpm_2_discrete.py @@ -0,0 +1,589 @@ +# Copyright 2025 Katherine Crowson, The HuggingFace Team and hlky. All rights reserved. +# +# 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 dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->KDPM2Discrete +class KDPM2DiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class KDPM2DiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + KDPM2DiscreteScheduler is inspired by the DPMSolver2 and Algorithm 2 from the [Elucidating the Design Space of + Diffusion-Based Generative Models](https://huggingface.co/papers/2206.00364) paper. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.00085): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.012): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear` or `scaled_linear`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 2 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, # sensible defaults + beta_end: float = 0.012, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + use_karras_sigmas: Optional[bool] = False, + use_exponential_sigmas: Optional[bool] = False, + use_beta_sigmas: Optional[bool] = False, + prediction_type: str = "epsilon", + timestep_spacing: str = "linspace", + steps_offset: int = 0, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # set all values + self.set_timesteps(num_train_timesteps, None, num_train_timesteps) + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self): + # standard deviation of the initial noise distribution + if self.config.timestep_spacing in ["linspace", "trailing"]: + return self.sigmas.max() + + return (self.sigmas.max() ** 2 + 1) ** 0.5 + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input( + self, + sample: torch.Tensor, + timestep: Union[float, torch.Tensor], + ) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + if self.state_in_first_order: + sigma = self.sigmas[self.step_index] + else: + sigma = self.sigmas_interpol[self.step_index] + + sample = sample / ((sigma**2 + 1) ** 0.5) + return sample + + def set_timesteps( + self, + num_inference_steps: int, + device: Union[str, torch.device] = None, + num_train_timesteps: Optional[int] = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.num_inference_steps = num_inference_steps + + num_train_timesteps = num_train_timesteps or self.config.num_train_timesteps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = np.linspace(0, num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[::-1].copy() + elif self.config.timestep_spacing == "leading": + step_ratio = num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round() + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + self.log_sigmas = torch.from_numpy(log_sigmas).to(device=device) + sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32) + sigmas = torch.from_numpy(sigmas).to(device=device) + + # interpolate sigmas + sigmas_interpol = sigmas.log().lerp(sigmas.roll(1).log(), 0.5).exp() + + self.sigmas = torch.cat([sigmas[:1], sigmas[1:].repeat_interleave(2), sigmas[-1:]]) + self.sigmas_interpol = torch.cat( + [sigmas_interpol[:1], sigmas_interpol[1:].repeat_interleave(2), sigmas_interpol[-1:]] + ) + + timesteps = torch.from_numpy(timesteps).to(device) + + # interpolate timesteps + sigmas_interpol = sigmas_interpol.cpu() + log_sigmas = self.log_sigmas.cpu() + timesteps_interpol = np.array( + [self._sigma_to_t(sigma_interpol, log_sigmas) for sigma_interpol in sigmas_interpol] + ) + timesteps_interpol = torch.from_numpy(timesteps_interpol).to(device, dtype=timesteps.dtype) + interleaved_timesteps = torch.stack((timesteps_interpol[1:-1, None], timesteps[1:, None]), dim=-1).flatten() + + self.timesteps = torch.cat([timesteps[:1], interleaved_timesteps]) + + self.sample = None + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def state_in_first_order(self): + return self.sample is None + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """Constructs an exponential noise schedule.""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def step( + self, + model_output: Union[torch.Tensor, np.ndarray], + timestep: Union[float, torch.Tensor], + sample: Union[torch.Tensor, np.ndarray], + return_dict: bool = True, + ) -> Union[KDPM2DiscreteSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_k_dpm_2_discrete.KDPM2DiscreteSchedulerOutput`] or + tuple. + + Returns: + [`~schedulers.scheduling_k_dpm_2_discrete.KDPM2DiscreteSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_k_dpm_2_discrete.KDPM2DiscreteSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + """ + if self.step_index is None: + self._init_step_index(timestep) + + if self.state_in_first_order: + sigma = self.sigmas[self.step_index] + sigma_interpol = self.sigmas_interpol[self.step_index + 1] + sigma_next = self.sigmas[self.step_index + 1] + else: + # 2nd order / KDPM2's method + sigma = self.sigmas[self.step_index - 1] + sigma_interpol = self.sigmas_interpol[self.step_index] + sigma_next = self.sigmas[self.step_index] + + # currently only gamma=0 is supported. This usually works best anyways. + # We can support gamma in the future but then need to scale the timestep before + # passing it to the model which requires a change in API + gamma = 0 + sigma_hat = sigma * (gamma + 1) # Note: sigma_hat == sigma for now + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + sigma_input = sigma_hat if self.state_in_first_order else sigma_interpol + pred_original_sample = sample - sigma_input * model_output + elif self.config.prediction_type == "v_prediction": + sigma_input = sigma_hat if self.state_in_first_order else sigma_interpol + pred_original_sample = model_output * (-sigma_input / (sigma_input**2 + 1) ** 0.5) + ( + sample / (sigma_input**2 + 1) + ) + elif self.config.prediction_type == "sample": + raise NotImplementedError("prediction_type not implemented yet: sample") + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + if self.state_in_first_order: + # 2. Convert to an ODE derivative for 1st order + derivative = (sample - pred_original_sample) / sigma_hat + # 3. delta timestep + dt = sigma_interpol - sigma_hat + + # store for 2nd order step + self.sample = sample + else: + # DPM-Solver-2 + # 2. Convert to an ODE derivative for 2nd order + derivative = (sample - pred_original_sample) / sigma_interpol + + # 3. delta timestep + dt = sigma_next - sigma_hat + + sample = self.sample + self.sample = None + + # upon completion increase step index by one + self._step_index += 1 + + prev_sample = sample + derivative * dt + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return KDPM2DiscreteSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_karras_ve_flax.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_karras_ve_flax.py new file mode 100644 index 0000000000000000000000000000000000000000..bacfbd61006d68bb39ce781103cd382ace6d772a --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_karras_ve_flax.py @@ -0,0 +1,238 @@ +# Copyright 2025 NVIDIA and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +from dataclasses import dataclass +from typing import Optional, Tuple, Union + +import flax +import jax +import jax.numpy as jnp +from jax import random + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from .scheduling_utils_flax import FlaxSchedulerMixin + + +@flax.struct.dataclass +class KarrasVeSchedulerState: + # setable values + num_inference_steps: Optional[int] = None + timesteps: Optional[jnp.ndarray] = None + schedule: Optional[jnp.ndarray] = None # sigma(t_i) + + @classmethod + def create(cls): + return cls() + + +@dataclass +class FlaxKarrasVeOutput(BaseOutput): + """ + Output class for the scheduler's step function output. + + Args: + prev_sample (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + derivative (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)` for images): + Derivative of predicted original image sample (x_0). + state (`KarrasVeSchedulerState`): the `FlaxKarrasVeScheduler` state data class. + """ + + prev_sample: jnp.ndarray + derivative: jnp.ndarray + state: KarrasVeSchedulerState + + +class FlaxKarrasVeScheduler(FlaxSchedulerMixin, ConfigMixin): + """ + Stochastic sampling from Karras et al. [1] tailored to the Variance-Expanding (VE) models [2]. Use Algorithm 2 and + the VE column of Table 1 from [1] for reference. + + [1] Karras, Tero, et al. "Elucidating the Design Space of Diffusion-Based Generative Models." + https://huggingface.co/papers/2206.00364 [2] Song, Yang, et al. "Score-based generative modeling through stochastic + differential equations." https://huggingface.co/papers/2011.13456 + + [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` + function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. + [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and + [`~SchedulerMixin.from_pretrained`] functions. + + For more details on the parameters, see the original paper's Appendix E.: "Elucidating the Design Space of + Diffusion-Based Generative Models." https://huggingface.co/papers/2206.00364. The grid search values used to find + the optimal {s_noise, s_churn, s_min, s_max} for a specific model are described in Table 5 of the paper. + + Args: + sigma_min (`float`): minimum noise magnitude + sigma_max (`float`): maximum noise magnitude + s_noise (`float`): the amount of additional noise to counteract loss of detail during sampling. + A reasonable range is [1.000, 1.011]. + s_churn (`float`): the parameter controlling the overall amount of stochasticity. + A reasonable range is [0, 100]. + s_min (`float`): the start value of the sigma range where we add noise (enable stochasticity). + A reasonable range is [0, 10]. + s_max (`float`): the end value of the sigma range where we add noise. + A reasonable range is [0.2, 80]. + """ + + @property + def has_state(self): + return True + + @register_to_config + def __init__( + self, + sigma_min: float = 0.02, + sigma_max: float = 100, + s_noise: float = 1.007, + s_churn: float = 80, + s_min: float = 0.05, + s_max: float = 50, + ): + pass + + def create_state(self): + return KarrasVeSchedulerState.create() + + def set_timesteps( + self, state: KarrasVeSchedulerState, num_inference_steps: int, shape: Tuple = () + ) -> KarrasVeSchedulerState: + """ + Sets the continuous timesteps used for the diffusion chain. Supporting function to be run before inference. + + Args: + state (`KarrasVeSchedulerState`): + the `FlaxKarrasVeScheduler` state data class. + num_inference_steps (`int`): + the number of diffusion steps used when generating samples with a pre-trained model. + + """ + timesteps = jnp.arange(0, num_inference_steps)[::-1].copy() + schedule = [ + ( + self.config.sigma_max**2 + * (self.config.sigma_min**2 / self.config.sigma_max**2) ** (i / (num_inference_steps - 1)) + ) + for i in timesteps + ] + + return state.replace( + num_inference_steps=num_inference_steps, + schedule=jnp.array(schedule, dtype=jnp.float32), + timesteps=timesteps, + ) + + def add_noise_to_input( + self, + state: KarrasVeSchedulerState, + sample: jnp.ndarray, + sigma: float, + key: jax.Array, + ) -> Tuple[jnp.ndarray, float]: + """ + Explicit Langevin-like "churn" step of adding noise to the sample according to a factor gamma_i ≥ 0 to reach a + higher noise level sigma_hat = sigma_i + gamma_i*sigma_i. + + TODO Args: + """ + if self.config.s_min <= sigma <= self.config.s_max: + gamma = min(self.config.s_churn / state.num_inference_steps, 2**0.5 - 1) + else: + gamma = 0 + + # sample eps ~ N(0, S_noise^2 * I) + key = random.split(key, num=1) + eps = self.config.s_noise * random.normal(key=key, shape=sample.shape) + sigma_hat = sigma + gamma * sigma + sample_hat = sample + ((sigma_hat**2 - sigma**2) ** 0.5 * eps) + + return sample_hat, sigma_hat + + def step( + self, + state: KarrasVeSchedulerState, + model_output: jnp.ndarray, + sigma_hat: float, + sigma_prev: float, + sample_hat: jnp.ndarray, + return_dict: bool = True, + ) -> Union[FlaxKarrasVeOutput, Tuple]: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + state (`KarrasVeSchedulerState`): the `FlaxKarrasVeScheduler` state data class. + model_output (`torch.Tensor` or `np.ndarray`): direct output from learned diffusion model. + sigma_hat (`float`): TODO + sigma_prev (`float`): TODO + sample_hat (`torch.Tensor` or `np.ndarray`): TODO + return_dict (`bool`): option for returning tuple rather than FlaxKarrasVeOutput class + + Returns: + [`~schedulers.scheduling_karras_ve_flax.FlaxKarrasVeOutput`] or `tuple`: Updated sample in the diffusion + chain and derivative. [`~schedulers.scheduling_karras_ve_flax.FlaxKarrasVeOutput`] if `return_dict` is + True, otherwise a `tuple`. When returning a tuple, the first element is the sample tensor. + """ + + pred_original_sample = sample_hat + sigma_hat * model_output + derivative = (sample_hat - pred_original_sample) / sigma_hat + sample_prev = sample_hat + (sigma_prev - sigma_hat) * derivative + + if not return_dict: + return (sample_prev, derivative, state) + + return FlaxKarrasVeOutput(prev_sample=sample_prev, derivative=derivative, state=state) + + def step_correct( + self, + state: KarrasVeSchedulerState, + model_output: jnp.ndarray, + sigma_hat: float, + sigma_prev: float, + sample_hat: jnp.ndarray, + sample_prev: jnp.ndarray, + derivative: jnp.ndarray, + return_dict: bool = True, + ) -> Union[FlaxKarrasVeOutput, Tuple]: + """ + Correct the predicted sample based on the output model_output of the network. TODO complete description + + Args: + state (`KarrasVeSchedulerState`): the `FlaxKarrasVeScheduler` state data class. + model_output (`torch.Tensor` or `np.ndarray`): direct output from learned diffusion model. + sigma_hat (`float`): TODO + sigma_prev (`float`): TODO + sample_hat (`torch.Tensor` or `np.ndarray`): TODO + sample_prev (`torch.Tensor` or `np.ndarray`): TODO + derivative (`torch.Tensor` or `np.ndarray`): TODO + return_dict (`bool`): option for returning tuple rather than FlaxKarrasVeOutput class + + Returns: + prev_sample (TODO): updated sample in the diffusion chain. derivative (TODO): TODO + + """ + pred_original_sample = sample_prev + sigma_prev * model_output + derivative_corr = (sample_prev - pred_original_sample) / sigma_prev + sample_prev = sample_hat + (sigma_prev - sigma_hat) * (0.5 * derivative + 0.5 * derivative_corr) + + if not return_dict: + return (sample_prev, derivative, state) + + return FlaxKarrasVeOutput(prev_sample=sample_prev, derivative=derivative, state=state) + + def add_noise(self, state: KarrasVeSchedulerState, original_samples, noise, timesteps): + raise NotImplementedError() diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_lcm.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_lcm.py new file mode 100644 index 0000000000000000000000000000000000000000..cd7a29fe675f98eb7998b60b86da41170db9641f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_lcm.py @@ -0,0 +1,653 @@ +# Copyright 2025 Stanford University Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +import math +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class LCMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + denoised: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas: torch.Tensor) -> torch.Tensor: + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + + Args: + betas (`torch.Tensor`): + the betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class LCMScheduler(SchedulerMixin, ConfigMixin): + """ + `LCMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with + non-Markovian guidance. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. [`~ConfigMixin`] takes care of storing all config + attributes that are passed in the scheduler's `__init__` function, such as `num_train_timesteps`. They can be + accessed via `scheduler.config.num_train_timesteps`. [`SchedulerMixin`] provides general loading and saving + functionality via the [`SchedulerMixin.save_pretrained`] and [`~SchedulerMixin.from_pretrained`] functions. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + original_inference_steps (`int`, *optional*, defaults to 50): + The default number of inference steps used to generate a linearly-spaced timestep schedule, from which we + will ultimately take `num_inference_steps` evenly spaced timesteps to form the final timestep schedule. + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to 1.0): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, defaults to `True`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + timestep_scaling (`float`, defaults to 10.0): + The factor the timesteps will be multiplied by when calculating the consistency model boundary conditions + `c_skip` and `c_out`. Increasing this will decrease the approximation error (although the approximation + error at the default of `10.0` is already pretty small). + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, + beta_end: float = 0.012, + beta_schedule: str = "scaled_linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + original_inference_steps: int = 50, + clip_sample: bool = False, + clip_sample_range: float = 1.0, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: str = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + timestep_spacing: str = "leading", + timestep_scaling: float = 10.0, + rescale_betas_zero_snr: bool = False, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # At every step in ddim, we are looking into the previous alphas_cumprod + # For the final step, there is no previous alphas_cumprod because we are already at 0 + # `set_alpha_to_one` decides whether we set this parameter simply to one or + # whether we use the final alpha of the "non-previous" one. + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + self.custom_timesteps = False + + self._step_index = None + self._begin_index = None + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + @property + def step_index(self): + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + def set_timesteps( + self, + num_inference_steps: Optional[int] = None, + device: Union[str, torch.device] = None, + original_inference_steps: Optional[int] = None, + timesteps: Optional[List[int]] = None, + strength: int = 1.0, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. If used, + `timesteps` must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + original_inference_steps (`int`, *optional*): + The original number of inference steps, which will be used to generate a linearly-spaced timestep + schedule (which is different from the standard `diffusers` implementation). We will then take + `num_inference_steps` timesteps from this schedule, evenly spaced in terms of indices, and use that as + our final timestep schedule. If not set, this will default to the `original_inference_steps` attribute. + timesteps (`List[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of equal spacing between timesteps on the training/distillation timestep + schedule is used. If `timesteps` is passed, `num_inference_steps` must be `None`. + """ + # 0. Check inputs + if num_inference_steps is None and timesteps is None: + raise ValueError("Must pass exactly one of `num_inference_steps` or `custom_timesteps`.") + + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.") + + # 1. Calculate the LCM original training/distillation timestep schedule. + original_steps = ( + original_inference_steps if original_inference_steps is not None else self.config.original_inference_steps + ) + + if original_steps > self.config.num_train_timesteps: + raise ValueError( + f"`original_steps`: {original_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + # LCM Timesteps Setting + # The skipping step parameter k from the paper. + k = self.config.num_train_timesteps // original_steps + # LCM Training/Distillation Steps Schedule + # Currently, only a linearly-spaced schedule is supported (same as in the LCM distillation scripts). + lcm_origin_timesteps = np.asarray(list(range(1, int(original_steps * strength) + 1))) * k - 1 + + # 2. Calculate the LCM inference timestep schedule. + if timesteps is not None: + # 2.1 Handle custom timestep schedules. + train_timesteps = set(lcm_origin_timesteps) + non_train_timesteps = [] + for i in range(1, len(timesteps)): + if timesteps[i] >= timesteps[i - 1]: + raise ValueError("`custom_timesteps` must be in descending order.") + + if timesteps[i] not in train_timesteps: + non_train_timesteps.append(timesteps[i]) + + if timesteps[0] >= self.config.num_train_timesteps: + raise ValueError( + f"`timesteps` must start before `self.config.train_timesteps`: {self.config.num_train_timesteps}." + ) + + # Raise warning if timestep schedule does not start with self.config.num_train_timesteps - 1 + if strength == 1.0 and timesteps[0] != self.config.num_train_timesteps - 1: + logger.warning( + f"The first timestep on the custom timestep schedule is {timesteps[0]}, not" + f" `self.config.num_train_timesteps - 1`: {self.config.num_train_timesteps - 1}. You may get" + f" unexpected results when using this timestep schedule." + ) + + # Raise warning if custom timestep schedule contains timesteps not on original timestep schedule + if non_train_timesteps: + logger.warning( + f"The custom timestep schedule contains the following timesteps which are not on the original" + f" training/distillation timestep schedule: {non_train_timesteps}. You may get unexpected results" + f" when using this timestep schedule." + ) + + # Raise warning if custom timestep schedule is longer than original_steps + if len(timesteps) > original_steps: + logger.warning( + f"The number of timesteps in the custom timestep schedule is {len(timesteps)}, which exceeds the" + f" the length of the timestep schedule used for training: {original_steps}. You may get some" + f" unexpected results when using this timestep schedule." + ) + + timesteps = np.array(timesteps, dtype=np.int64) + self.num_inference_steps = len(timesteps) + self.custom_timesteps = True + + # Apply strength (e.g. for img2img pipelines) (see StableDiffusionImg2ImgPipeline.get_timesteps) + init_timestep = min(int(self.num_inference_steps * strength), self.num_inference_steps) + t_start = max(self.num_inference_steps - init_timestep, 0) + timesteps = timesteps[t_start * self.order :] + # TODO: also reset self.num_inference_steps? + else: + # 2.2 Create the "standard" LCM inference timestep schedule. + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + skipping_step = len(lcm_origin_timesteps) // num_inference_steps + + if skipping_step < 1: + raise ValueError( + f"The combination of `original_steps x strength`: {original_steps} x {strength} is smaller than `num_inference_steps`: {num_inference_steps}. Make sure to either reduce `num_inference_steps` to a value smaller than {int(original_steps * strength)} or increase `strength` to a value higher than {float(num_inference_steps / original_steps)}." + ) + + self.num_inference_steps = num_inference_steps + + if num_inference_steps > original_steps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `original_inference_steps`:" + f" {original_steps} because the final timestep schedule will be a subset of the" + f" `original_inference_steps`-sized initial timestep schedule." + ) + + # LCM Inference Steps Schedule + lcm_origin_timesteps = lcm_origin_timesteps[::-1].copy() + # Select (approximately) evenly spaced indices from lcm_origin_timesteps. + inference_indices = np.linspace(0, len(lcm_origin_timesteps), num=num_inference_steps, endpoint=False) + inference_indices = np.floor(inference_indices).astype(np.int64) + timesteps = lcm_origin_timesteps[inference_indices] + + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.long) + + self._step_index = None + self._begin_index = None + + def get_scalings_for_boundary_condition_discrete(self, timestep): + self.sigma_data = 0.5 # Default: 0.5 + scaled_timestep = timestep * self.config.timestep_scaling + + c_skip = self.sigma_data**2 / (scaled_timestep**2 + self.sigma_data**2) + c_out = scaled_timestep / (scaled_timestep**2 + self.sigma_data**2) ** 0.5 + return c_skip, c_out + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[LCMSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_lcm.LCMSchedulerOutput`] or `tuple`. + Returns: + [`~schedulers.scheduling_utils.LCMSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_lcm.LCMSchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # 1. get previous step value + prev_step_index = self.step_index + 1 + if prev_step_index < len(self.timesteps): + prev_timestep = self.timesteps[prev_step_index] + else: + prev_timestep = timestep + + # 2. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + # 3. Get scalings for boundary conditions + c_skip, c_out = self.get_scalings_for_boundary_condition_discrete(timestep) + + # 4. Compute the predicted original sample x_0 based on the model parameterization + if self.config.prediction_type == "epsilon": # noise-prediction + predicted_original_sample = (sample - beta_prod_t.sqrt() * model_output) / alpha_prod_t.sqrt() + elif self.config.prediction_type == "sample": # x-prediction + predicted_original_sample = model_output + elif self.config.prediction_type == "v_prediction": # v-prediction + predicted_original_sample = alpha_prod_t.sqrt() * sample - beta_prod_t.sqrt() * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or" + " `v_prediction` for `LCMScheduler`." + ) + + # 5. Clip or threshold "predicted x_0" + if self.config.thresholding: + predicted_original_sample = self._threshold_sample(predicted_original_sample) + elif self.config.clip_sample: + predicted_original_sample = predicted_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 6. Denoise model output using boundary conditions + denoised = c_out * predicted_original_sample + c_skip * sample + + # 7. Sample and inject noise z ~ N(0, I) for MultiStep Inference + # Noise is not used on the final timestep of the timestep schedule. + # This also means that noise is not used for one-step sampling. + if self.step_index != self.num_inference_steps - 1: + noise = randn_tensor( + model_output.shape, generator=generator, device=model_output.device, dtype=denoised.dtype + ) + prev_sample = alpha_prod_t_prev.sqrt() * denoised + beta_prod_t_prev.sqrt() * noise + else: + prev_sample = denoised + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample, denoised) + + return LCMSchedulerOutput(prev_sample=prev_sample, denoised=denoised) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self): + return self.config.num_train_timesteps + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.previous_timestep + def previous_timestep(self, timestep): + if self.custom_timesteps or self.num_inference_steps: + index = (self.timesteps == timestep).nonzero(as_tuple=True)[0][0] + if index == self.timesteps.shape[0] - 1: + prev_t = torch.tensor(-1) + else: + prev_t = self.timesteps[index + 1] + else: + prev_t = timestep - 1 + return prev_t diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_lms_discrete.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_lms_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..c2450204aa8f94719a8eeb8b43ce9f62f37583cf --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_lms_discrete.py @@ -0,0 +1,552 @@ +# Copyright 2025 Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# 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 warnings +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import scipy.stats +import torch +from scipy import integrate + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->LMSDiscrete +class LMSDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class LMSDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + A linear multistep scheduler for discrete beta schedules. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear` or `scaled_linear`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + use_karras_sigmas: Optional[bool] = False, + use_exponential_sigmas: Optional[bool] = False, + use_beta_sigmas: Optional[bool] = False, + prediction_type: str = "epsilon", + timestep_spacing: str = "linspace", + steps_offset: int = 0, + ): + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + sigmas = np.concatenate([sigmas[::-1], [0.0]]).astype(np.float32) + self.sigmas = torch.from_numpy(sigmas) + + # setable values + self.num_inference_steps = None + self.use_karras_sigmas = use_karras_sigmas + self.set_timesteps(num_train_timesteps, None) + self.derivatives = [] + self.is_scale_input_called = False + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self): + # standard deviation of the initial noise distribution + if self.config.timestep_spacing in ["linspace", "trailing"]: + return self.sigmas.max() + + return (self.sigmas.max() ** 2 + 1) ** 0.5 + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`float` or `torch.Tensor`): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = sample / ((sigma**2 + 1) ** 0.5) + self.is_scale_input_called = True + return sample + + def get_lms_coefficient(self, order, t, current_order): + """ + Compute the linear multistep coefficient. + + Args: + order (): + t (): + current_order (): + """ + + def lms_derivative(tau): + prod = 1.0 + for k in range(order): + if current_order == k: + continue + prod *= (tau - self.sigmas[t - k]) / (self.sigmas[t - current_order] - self.sigmas[t - k]) + return prod + + integrated_coeff = integrate.quad(lms_derivative, self.sigmas[t], self.sigmas[t + 1], epsrel=1e-4)[0] + + return integrated_coeff + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.num_inference_steps = num_inference_steps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[ + ::-1 + ].copy() + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(self.config.num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32) + + self.sigmas = torch.from_numpy(sigmas).to(device=device) + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.float32) + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + self.derivatives = [] + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + + sigma_min: float = in_sigmas[-1].item() + sigma_max: float = in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, self.num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """Constructs an exponential noise schedule.""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def step( + self, + model_output: torch.Tensor, + timestep: Union[float, torch.Tensor], + sample: torch.Tensor, + order: int = 4, + return_dict: bool = True, + ) -> Union[LMSDiscreteSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + order (`int`, defaults to 4): + The order of the linear multistep method. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if not self.is_scale_input_called: + warnings.warn( + "The `scale_model_input` function should be called before `step` to ensure correct denoising. " + "See `StableDiffusionPipeline` for a usage example." + ) + + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + pred_original_sample = sample - sigma * model_output + elif self.config.prediction_type == "v_prediction": + # * c_out + input * c_skip + pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1)) + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + # 2. Convert to an ODE derivative + derivative = (sample - pred_original_sample) / sigma + self.derivatives.append(derivative) + if len(self.derivatives) > order: + self.derivatives.pop(0) + + # 3. Compute linear multistep coefficients + order = min(self.step_index + 1, order) + lms_coeffs = [self.get_lms_coefficient(order, self.step_index, curr_order) for curr_order in range(order)] + + # 4. Compute previous sample based on the derivatives path + prev_sample = sample + sum( + coeff * derivative for coeff, derivative in zip(lms_coeffs, reversed(self.derivatives)) + ) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return LMSDiscreteSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_lms_discrete_flax.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_lms_discrete_flax.py new file mode 100644 index 0000000000000000000000000000000000000000..b8e08ff9e134f8a77b428179057df589d736a187 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_lms_discrete_flax.py @@ -0,0 +1,283 @@ +# Copyright 2025 Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass +from typing import Optional, Tuple, Union + +import flax +import jax.numpy as jnp +from scipy import integrate + +from ..configuration_utils import ConfigMixin, register_to_config +from .scheduling_utils_flax import ( + CommonSchedulerState, + FlaxKarrasDiffusionSchedulers, + FlaxSchedulerMixin, + FlaxSchedulerOutput, + broadcast_to_shape_from_left, +) + + +@flax.struct.dataclass +class LMSDiscreteSchedulerState: + common: CommonSchedulerState + + # setable values + init_noise_sigma: jnp.ndarray + timesteps: jnp.ndarray + sigmas: jnp.ndarray + num_inference_steps: Optional[int] = None + + # running values + derivatives: Optional[jnp.ndarray] = None + + @classmethod + def create( + cls, common: CommonSchedulerState, init_noise_sigma: jnp.ndarray, timesteps: jnp.ndarray, sigmas: jnp.ndarray + ): + return cls(common=common, init_noise_sigma=init_noise_sigma, timesteps=timesteps, sigmas=sigmas) + + +@dataclass +class FlaxLMSSchedulerOutput(FlaxSchedulerOutput): + state: LMSDiscreteSchedulerState + + +class FlaxLMSDiscreteScheduler(FlaxSchedulerMixin, ConfigMixin): + """ + Linear Multistep Scheduler for discrete beta schedules. Based on the original k-diffusion implementation by + Katherine Crowson: + https://github.com/crowsonkb/k-diffusion/blob/481677d114f6ea445aa009cf5bd7a9cdee909e47/k_diffusion/sampling.py#L181 + + [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` + function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. + [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and + [`~SchedulerMixin.from_pretrained`] functions. + + Args: + num_train_timesteps (`int`): number of diffusion steps used to train the model. + beta_start (`float`): the starting `beta` value of inference. + beta_end (`float`): the final `beta` value. + beta_schedule (`str`): + the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear` or `scaled_linear`. + trained_betas (`jnp.ndarray`, optional): + option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc. + prediction_type (`str`, default `epsilon`, optional): + prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion + process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4 + https://imagen.research.google/video/paper.pdf) + dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`): + the `dtype` used for params and computation. + """ + + _compatibles = [e.name for e in FlaxKarrasDiffusionSchedulers] + + dtype: jnp.dtype + + @property + def has_state(self): + return True + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[jnp.ndarray] = None, + prediction_type: str = "epsilon", + dtype: jnp.dtype = jnp.float32, + ): + self.dtype = dtype + + def create_state(self, common: Optional[CommonSchedulerState] = None) -> LMSDiscreteSchedulerState: + if common is None: + common = CommonSchedulerState.create(self) + + timesteps = jnp.arange(0, self.config.num_train_timesteps).round()[::-1] + sigmas = ((1 - common.alphas_cumprod) / common.alphas_cumprod) ** 0.5 + + # standard deviation of the initial noise distribution + init_noise_sigma = sigmas.max() + + return LMSDiscreteSchedulerState.create( + common=common, + init_noise_sigma=init_noise_sigma, + timesteps=timesteps, + sigmas=sigmas, + ) + + def scale_model_input(self, state: LMSDiscreteSchedulerState, sample: jnp.ndarray, timestep: int) -> jnp.ndarray: + """ + Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the K-LMS algorithm. + + Args: + state (`LMSDiscreteSchedulerState`): + the `FlaxLMSDiscreteScheduler` state data class instance. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + timestep (`int`): + current discrete timestep in the diffusion chain. + + Returns: + `jnp.ndarray`: scaled input sample + """ + (step_index,) = jnp.where(state.timesteps == timestep, size=1) + step_index = step_index[0] + + sigma = state.sigmas[step_index] + sample = sample / ((sigma**2 + 1) ** 0.5) + return sample + + def get_lms_coefficient(self, state: LMSDiscreteSchedulerState, order, t, current_order): + """ + Compute a linear multistep coefficient. + + Args: + order (TODO): + t (TODO): + current_order (TODO): + """ + + def lms_derivative(tau): + prod = 1.0 + for k in range(order): + if current_order == k: + continue + prod *= (tau - state.sigmas[t - k]) / (state.sigmas[t - current_order] - state.sigmas[t - k]) + return prod + + integrated_coeff = integrate.quad(lms_derivative, state.sigmas[t], state.sigmas[t + 1], epsrel=1e-4)[0] + + return integrated_coeff + + def set_timesteps( + self, state: LMSDiscreteSchedulerState, num_inference_steps: int, shape: Tuple = () + ) -> LMSDiscreteSchedulerState: + """ + Sets the timesteps used for the diffusion chain. Supporting function to be run before inference. + + Args: + state (`LMSDiscreteSchedulerState`): + the `FlaxLMSDiscreteScheduler` state data class instance. + num_inference_steps (`int`): + the number of diffusion steps used when generating samples with a pre-trained model. + """ + + timesteps = jnp.linspace(self.config.num_train_timesteps - 1, 0, num_inference_steps, dtype=self.dtype) + + low_idx = jnp.floor(timesteps).astype(jnp.int32) + high_idx = jnp.ceil(timesteps).astype(jnp.int32) + + frac = jnp.mod(timesteps, 1.0) + + sigmas = ((1 - state.common.alphas_cumprod) / state.common.alphas_cumprod) ** 0.5 + sigmas = (1 - frac) * sigmas[low_idx] + frac * sigmas[high_idx] + sigmas = jnp.concatenate([sigmas, jnp.array([0.0], dtype=self.dtype)]) + + timesteps = timesteps.astype(jnp.int32) + + # initial running values + derivatives = jnp.zeros((0,) + shape, dtype=self.dtype) + + return state.replace( + timesteps=timesteps, + sigmas=sigmas, + num_inference_steps=num_inference_steps, + derivatives=derivatives, + ) + + def step( + self, + state: LMSDiscreteSchedulerState, + model_output: jnp.ndarray, + timestep: int, + sample: jnp.ndarray, + order: int = 4, + return_dict: bool = True, + ) -> Union[FlaxLMSSchedulerOutput, Tuple]: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + state (`LMSDiscreteSchedulerState`): the `FlaxLMSDiscreteScheduler` state data class instance. + model_output (`jnp.ndarray`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + order: coefficient for multi-step inference. + return_dict (`bool`): option for returning tuple rather than FlaxLMSSchedulerOutput class + + Returns: + [`FlaxLMSSchedulerOutput`] or `tuple`: [`FlaxLMSSchedulerOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is the sample tensor. + + """ + if state.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + sigma = state.sigmas[timestep] + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + pred_original_sample = sample - sigma * model_output + elif self.config.prediction_type == "v_prediction": + # * c_out + input * c_skip + pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1)) + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + # 2. Convert to an ODE derivative + derivative = (sample - pred_original_sample) / sigma + state = state.replace(derivatives=jnp.append(state.derivatives, derivative)) + if len(state.derivatives) > order: + state = state.replace(derivatives=jnp.delete(state.derivatives, 0)) + + # 3. Compute linear multistep coefficients + order = min(timestep + 1, order) + lms_coeffs = [self.get_lms_coefficient(state, order, timestep, curr_order) for curr_order in range(order)] + + # 4. Compute previous sample based on the derivatives path + prev_sample = sample + sum( + coeff * derivative for coeff, derivative in zip(lms_coeffs, reversed(state.derivatives)) + ) + + if not return_dict: + return (prev_sample, state) + + return FlaxLMSSchedulerOutput(prev_sample=prev_sample, state=state) + + def add_noise( + self, + state: LMSDiscreteSchedulerState, + original_samples: jnp.ndarray, + noise: jnp.ndarray, + timesteps: jnp.ndarray, + ) -> jnp.ndarray: + sigma = state.sigmas[timesteps].flatten() + sigma = broadcast_to_shape_from_left(sigma, noise.shape) + + noisy_samples = original_samples + noise * sigma + + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_pndm.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_pndm.py new file mode 100644 index 0000000000000000000000000000000000000000..c07621179e2b7321f346a42def88cde7cad0bbe8 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_pndm.py @@ -0,0 +1,476 @@ +# Copyright 2025 Zhejiang University Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim + +import math +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class PNDMScheduler(SchedulerMixin, ConfigMixin): + """ + `PNDMScheduler` uses pseudo numerical methods for diffusion models such as the Runge-Kutta and linear multi-step + method. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + skip_prk_steps (`bool`, defaults to `False`): + Allows the scheduler to skip the Runge-Kutta steps defined in the original paper as being required before + PLMS steps. + set_alpha_to_one (`bool`, defaults to `False`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process) + or `v_prediction` (see section 2.4 of [Imagen Video](https://imagen.research.google/video/paper.pdf) + paper). + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + skip_prk_steps: bool = False, + set_alpha_to_one: bool = False, + prediction_type: str = "epsilon", + timestep_spacing: str = "leading", + steps_offset: int = 0, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # For now we only support F-PNDM, i.e. the runge-kutta method + # For more information on the algorithm please take a look at the paper: https://huggingface.co/papers/2202.09778 + # mainly at formula (9), (12), (13) and the Algorithm 2. + self.pndm_order = 4 + + # running values + self.cur_model_output = 0 + self.counter = 0 + self.cur_sample = None + self.ets = [] + + # setable values + self.num_inference_steps = None + self._timesteps = np.arange(0, num_train_timesteps)[::-1].copy() + self.prk_timesteps = None + self.plms_timesteps = None + self.timesteps = None + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + + self.num_inference_steps = num_inference_steps + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + self._timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps).round().astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + self._timesteps = (np.arange(0, num_inference_steps) * step_ratio).round() + self._timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + self._timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio))[::-1].astype( + np.int64 + ) + self._timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + if self.config.skip_prk_steps: + # for some models like stable diffusion the prk steps can/should be skipped to + # produce better results. When using PNDM with `self.config.skip_prk_steps` the implementation + # is based on crowsonkb's PLMS sampler implementation: https://github.com/CompVis/latent-diffusion/pull/51 + self.prk_timesteps = np.array([]) + self.plms_timesteps = np.concatenate([self._timesteps[:-1], self._timesteps[-2:-1], self._timesteps[-1:]])[ + ::-1 + ].copy() + else: + prk_timesteps = np.array(self._timesteps[-self.pndm_order :]).repeat(2) + np.tile( + np.array([0, self.config.num_train_timesteps // num_inference_steps // 2]), self.pndm_order + ) + self.prk_timesteps = (prk_timesteps[:-1].repeat(2)[1:-1])[::-1].copy() + self.plms_timesteps = self._timesteps[:-3][ + ::-1 + ].copy() # we copy to avoid having negative strides which are not supported by torch.from_numpy + + timesteps = np.concatenate([self.prk_timesteps, self.plms_timesteps]).astype(np.int64) + self.timesteps = torch.from_numpy(timesteps).to(device) + + self.ets = [] + self.counter = 0 + self.cur_model_output = 0 + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise), and calls [`~PNDMScheduler.step_prk`] + or [`~PNDMScheduler.step_plms`] depending on the internal variable `counter`. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.counter < len(self.prk_timesteps) and not self.config.skip_prk_steps: + return self.step_prk(model_output=model_output, timestep=timestep, sample=sample, return_dict=return_dict) + else: + return self.step_plms(model_output=model_output, timestep=timestep, sample=sample, return_dict=return_dict) + + def step_prk( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the Runge-Kutta method. It performs four forward passes to approximate the solution to the differential + equation. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + diff_to_prev = 0 if self.counter % 2 else self.config.num_train_timesteps // self.num_inference_steps // 2 + prev_timestep = timestep - diff_to_prev + timestep = self.prk_timesteps[self.counter // 4 * 4] + + if self.counter % 4 == 0: + self.cur_model_output += 1 / 6 * model_output + self.ets.append(model_output) + self.cur_sample = sample + elif (self.counter - 1) % 4 == 0: + self.cur_model_output += 1 / 3 * model_output + elif (self.counter - 2) % 4 == 0: + self.cur_model_output += 1 / 3 * model_output + elif (self.counter - 3) % 4 == 0: + model_output = self.cur_model_output + 1 / 6 * model_output + self.cur_model_output = 0 + + # cur_sample should not be `None` + cur_sample = self.cur_sample if self.cur_sample is not None else sample + + prev_sample = self._get_prev_sample(cur_sample, timestep, prev_timestep, model_output) + self.counter += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def step_plms( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the linear multistep method. It performs one forward pass multiple times to approximate the solution. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if not self.config.skip_prk_steps and len(self.ets) < 3: + raise ValueError( + f"{self.__class__} can only be run AFTER scheduler has been run " + "in 'prk' mode for at least 12 iterations " + "See: https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/pipeline_pndm.py " + "for more information." + ) + + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + if self.counter != 1: + self.ets = self.ets[-3:] + self.ets.append(model_output) + else: + prev_timestep = timestep + timestep = timestep + self.config.num_train_timesteps // self.num_inference_steps + + if len(self.ets) == 1 and self.counter == 0: + model_output = model_output + self.cur_sample = sample + elif len(self.ets) == 1 and self.counter == 1: + model_output = (model_output + self.ets[-1]) / 2 + sample = self.cur_sample + self.cur_sample = None + elif len(self.ets) == 2: + model_output = (3 * self.ets[-1] - self.ets[-2]) / 2 + elif len(self.ets) == 3: + model_output = (23 * self.ets[-1] - 16 * self.ets[-2] + 5 * self.ets[-3]) / 12 + else: + model_output = (1 / 24) * (55 * self.ets[-1] - 59 * self.ets[-2] + 37 * self.ets[-3] - 9 * self.ets[-4]) + + prev_sample = self._get_prev_sample(sample, timestep, prev_timestep, model_output) + self.counter += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def _get_prev_sample(self, sample, timestep, prev_timestep, model_output): + # See formula (9) of PNDM paper https://huggingface.co/papers/2202.09778 + # this function computes x_(t−δ) using the formula of (9) + # Note that x_t needs to be added to both sides of the equation + + # Notation ( -> + # alpha_prod_t -> α_t + # alpha_prod_t_prev -> α_(t−δ) + # beta_prod_t -> (1 - α_t) + # beta_prod_t_prev -> (1 - α_(t−δ)) + # sample -> x_t + # model_output -> e_θ(x_t, t) + # prev_sample -> x_(t−δ) + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + if self.config.prediction_type == "v_prediction": + model_output = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + elif self.config.prediction_type != "epsilon": + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon` or `v_prediction`" + ) + + # corresponds to (α_(t−δ) - α_t) divided by + # denominator of x_t in formula (9) and plus 1 + # Note: (α_(t−δ) - α_t) / (sqrt(α_t) * (sqrt(α_(t−δ)) + sqr(α_t))) = + # sqrt(α_(t−δ)) / sqrt(α_t)) + sample_coeff = (alpha_prod_t_prev / alpha_prod_t) ** (0.5) + + # corresponds to denominator of e_θ(x_t, t) in formula (9) + model_output_denom_coeff = alpha_prod_t * beta_prod_t_prev ** (0.5) + ( + alpha_prod_t * beta_prod_t * alpha_prod_t_prev + ) ** (0.5) + + # full formula (9) + prev_sample = ( + sample_coeff * sample - (alpha_prod_t_prev - alpha_prod_t) * model_output / model_output_denom_coeff + ) + + return prev_sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_pndm_flax.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_pndm_flax.py new file mode 100644 index 0000000000000000000000000000000000000000..12e22005afaf700f8d145433a9c35cc029500cae --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_pndm_flax.py @@ -0,0 +1,509 @@ +# Copyright 2025 Zhejiang University Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim + +from dataclasses import dataclass +from typing import Optional, Tuple, Union + +import flax +import jax +import jax.numpy as jnp + +from ..configuration_utils import ConfigMixin, register_to_config +from .scheduling_utils_flax import ( + CommonSchedulerState, + FlaxKarrasDiffusionSchedulers, + FlaxSchedulerMixin, + FlaxSchedulerOutput, + add_noise_common, +) + + +@flax.struct.dataclass +class PNDMSchedulerState: + common: CommonSchedulerState + final_alpha_cumprod: jnp.ndarray + + # setable values + init_noise_sigma: jnp.ndarray + timesteps: jnp.ndarray + num_inference_steps: Optional[int] = None + prk_timesteps: Optional[jnp.ndarray] = None + plms_timesteps: Optional[jnp.ndarray] = None + + # running values + cur_model_output: Optional[jnp.ndarray] = None + counter: Optional[jnp.int32] = None + cur_sample: Optional[jnp.ndarray] = None + ets: Optional[jnp.ndarray] = None + + @classmethod + def create( + cls, + common: CommonSchedulerState, + final_alpha_cumprod: jnp.ndarray, + init_noise_sigma: jnp.ndarray, + timesteps: jnp.ndarray, + ): + return cls( + common=common, + final_alpha_cumprod=final_alpha_cumprod, + init_noise_sigma=init_noise_sigma, + timesteps=timesteps, + ) + + +@dataclass +class FlaxPNDMSchedulerOutput(FlaxSchedulerOutput): + state: PNDMSchedulerState + + +class FlaxPNDMScheduler(FlaxSchedulerMixin, ConfigMixin): + """ + Pseudo numerical methods for diffusion models (PNDM) proposes using more advanced ODE integration techniques, + namely Runge-Kutta method and a linear multi-step method. + + [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` + function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. + [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and + [`~SchedulerMixin.from_pretrained`] functions. + + For more details, see the original paper: https://huggingface.co/papers/2202.09778 + + Args: + num_train_timesteps (`int`): number of diffusion steps used to train the model. + beta_start (`float`): the starting `beta` value of inference. + beta_end (`float`): the final `beta` value. + beta_schedule (`str`): + the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`jnp.ndarray`, optional): + option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc. + skip_prk_steps (`bool`): + allows the scheduler to skip the Runge-Kutta steps that are defined in the original paper as being required + before plms steps; defaults to `False`. + set_alpha_to_one (`bool`, default `False`): + each diffusion step uses the value of alphas product at that step and at the previous one. For the final + step there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the value of alpha at step 0. + steps_offset (`int`, default `0`): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, default `epsilon`, optional): + prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion + process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4 + https://imagen.research.google/video/paper.pdf) + dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`): + the `dtype` used for params and computation. + """ + + _compatibles = [e.name for e in FlaxKarrasDiffusionSchedulers] + + dtype: jnp.dtype + pndm_order: int + + @property + def has_state(self): + return True + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[jnp.ndarray] = None, + skip_prk_steps: bool = False, + set_alpha_to_one: bool = False, + steps_offset: int = 0, + prediction_type: str = "epsilon", + dtype: jnp.dtype = jnp.float32, + ): + self.dtype = dtype + + # For now we only support F-PNDM, i.e. the runge-kutta method + # For more information on the algorithm please take a look at the paper: https://huggingface.co/papers/2202.09778 + # mainly at formula (9), (12), (13) and the Algorithm 2. + self.pndm_order = 4 + + def create_state(self, common: Optional[CommonSchedulerState] = None) -> PNDMSchedulerState: + if common is None: + common = CommonSchedulerState.create(self) + + # At every step in ddim, we are looking into the previous alphas_cumprod + # For the final step, there is no previous alphas_cumprod because we are already at 0 + # `set_alpha_to_one` decides whether we set this parameter simply to one or + # whether we use the final alpha of the "non-previous" one. + final_alpha_cumprod = ( + jnp.array(1.0, dtype=self.dtype) if self.config.set_alpha_to_one else common.alphas_cumprod[0] + ) + + # standard deviation of the initial noise distribution + init_noise_sigma = jnp.array(1.0, dtype=self.dtype) + + timesteps = jnp.arange(0, self.config.num_train_timesteps).round()[::-1] + + return PNDMSchedulerState.create( + common=common, + final_alpha_cumprod=final_alpha_cumprod, + init_noise_sigma=init_noise_sigma, + timesteps=timesteps, + ) + + def set_timesteps(self, state: PNDMSchedulerState, num_inference_steps: int, shape: Tuple) -> PNDMSchedulerState: + """ + Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference. + + Args: + state (`PNDMSchedulerState`): + the `FlaxPNDMScheduler` state data class instance. + num_inference_steps (`int`): + the number of diffusion steps used when generating samples with a pre-trained model. + shape (`Tuple`): + the shape of the samples to be generated. + """ + + step_ratio = self.config.num_train_timesteps // num_inference_steps + # creates integer timesteps by multiplying by ratio + # rounding to avoid issues when num_inference_step is power of 3 + _timesteps = (jnp.arange(0, num_inference_steps) * step_ratio).round() + self.config.steps_offset + + if self.config.skip_prk_steps: + # for some models like stable diffusion the prk steps can/should be skipped to + # produce better results. When using PNDM with `self.config.skip_prk_steps` the implementation + # is based on crowsonkb's PLMS sampler implementation: https://github.com/CompVis/latent-diffusion/pull/51 + + prk_timesteps = jnp.array([], dtype=jnp.int32) + plms_timesteps = jnp.concatenate([_timesteps[:-1], _timesteps[-2:-1], _timesteps[-1:]])[::-1] + + else: + prk_timesteps = _timesteps[-self.pndm_order :].repeat(2) + jnp.tile( + jnp.array([0, self.config.num_train_timesteps // num_inference_steps // 2], dtype=jnp.int32), + self.pndm_order, + ) + + prk_timesteps = (prk_timesteps[:-1].repeat(2)[1:-1])[::-1] + plms_timesteps = _timesteps[:-3][::-1] + + timesteps = jnp.concatenate([prk_timesteps, plms_timesteps]) + + # initial running values + + cur_model_output = jnp.zeros(shape, dtype=self.dtype) + counter = jnp.int32(0) + cur_sample = jnp.zeros(shape, dtype=self.dtype) + ets = jnp.zeros((4,) + shape, dtype=self.dtype) + + return state.replace( + timesteps=timesteps, + num_inference_steps=num_inference_steps, + prk_timesteps=prk_timesteps, + plms_timesteps=plms_timesteps, + cur_model_output=cur_model_output, + counter=counter, + cur_sample=cur_sample, + ets=ets, + ) + + def scale_model_input( + self, state: PNDMSchedulerState, sample: jnp.ndarray, timestep: Optional[int] = None + ) -> jnp.ndarray: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + state (`PNDMSchedulerState`): the `FlaxPNDMScheduler` state data class instance. + sample (`jnp.ndarray`): input sample + timestep (`int`, optional): current timestep + + Returns: + `jnp.ndarray`: scaled input sample + """ + return sample + + def step( + self, + state: PNDMSchedulerState, + model_output: jnp.ndarray, + timestep: int, + sample: jnp.ndarray, + return_dict: bool = True, + ) -> Union[FlaxPNDMSchedulerOutput, Tuple]: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + This function calls `step_prk()` or `step_plms()` depending on the internal variable `counter`. + + Args: + state (`PNDMSchedulerState`): the `FlaxPNDMScheduler` state data class instance. + model_output (`jnp.ndarray`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + return_dict (`bool`): option for returning tuple rather than FlaxPNDMSchedulerOutput class + + Returns: + [`FlaxPNDMSchedulerOutput`] or `tuple`: [`FlaxPNDMSchedulerOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is the sample tensor. + + """ + + if state.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.config.skip_prk_steps: + prev_sample, state = self.step_plms(state, model_output, timestep, sample) + else: + prk_prev_sample, prk_state = self.step_prk(state, model_output, timestep, sample) + plms_prev_sample, plms_state = self.step_plms(state, model_output, timestep, sample) + + cond = state.counter < len(state.prk_timesteps) + + prev_sample = jax.lax.select(cond, prk_prev_sample, plms_prev_sample) + + state = state.replace( + cur_model_output=jax.lax.select(cond, prk_state.cur_model_output, plms_state.cur_model_output), + ets=jax.lax.select(cond, prk_state.ets, plms_state.ets), + cur_sample=jax.lax.select(cond, prk_state.cur_sample, plms_state.cur_sample), + counter=jax.lax.select(cond, prk_state.counter, plms_state.counter), + ) + + if not return_dict: + return (prev_sample, state) + + return FlaxPNDMSchedulerOutput(prev_sample=prev_sample, state=state) + + def step_prk( + self, + state: PNDMSchedulerState, + model_output: jnp.ndarray, + timestep: int, + sample: jnp.ndarray, + ) -> Union[FlaxPNDMSchedulerOutput, Tuple]: + """ + Step function propagating the sample with the Runge-Kutta method. RK takes 4 forward passes to approximate the + solution to the differential equation. + + Args: + state (`PNDMSchedulerState`): the `FlaxPNDMScheduler` state data class instance. + model_output (`jnp.ndarray`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + return_dict (`bool`): option for returning tuple rather than FlaxPNDMSchedulerOutput class + + Returns: + [`FlaxPNDMSchedulerOutput`] or `tuple`: [`FlaxPNDMSchedulerOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is the sample tensor. + + """ + + if state.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + diff_to_prev = jnp.where( + state.counter % 2, 0, self.config.num_train_timesteps // state.num_inference_steps // 2 + ) + prev_timestep = timestep - diff_to_prev + timestep = state.prk_timesteps[state.counter // 4 * 4] + + model_output = jax.lax.select( + (state.counter % 4) != 3, + model_output, # remainder 0, 1, 2 + state.cur_model_output + 1 / 6 * model_output, # remainder 3 + ) + + state = state.replace( + cur_model_output=jax.lax.select_n( + state.counter % 4, + state.cur_model_output + 1 / 6 * model_output, # remainder 0 + state.cur_model_output + 1 / 3 * model_output, # remainder 1 + state.cur_model_output + 1 / 3 * model_output, # remainder 2 + jnp.zeros_like(state.cur_model_output), # remainder 3 + ), + ets=jax.lax.select( + (state.counter % 4) == 0, + state.ets.at[0:3].set(state.ets[1:4]).at[3].set(model_output), # remainder 0 + state.ets, # remainder 1, 2, 3 + ), + cur_sample=jax.lax.select( + (state.counter % 4) == 0, + sample, # remainder 0 + state.cur_sample, # remainder 1, 2, 3 + ), + ) + + cur_sample = state.cur_sample + prev_sample = self._get_prev_sample(state, cur_sample, timestep, prev_timestep, model_output) + state = state.replace(counter=state.counter + 1) + + return (prev_sample, state) + + def step_plms( + self, + state: PNDMSchedulerState, + model_output: jnp.ndarray, + timestep: int, + sample: jnp.ndarray, + ) -> Union[FlaxPNDMSchedulerOutput, Tuple]: + """ + Step function propagating the sample with the linear multi-step method. This has one forward pass with multiple + times to approximate the solution. + + Args: + state (`PNDMSchedulerState`): the `FlaxPNDMScheduler` state data class instance. + model_output (`jnp.ndarray`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + return_dict (`bool`): option for returning tuple rather than FlaxPNDMSchedulerOutput class + + Returns: + [`FlaxPNDMSchedulerOutput`] or `tuple`: [`FlaxPNDMSchedulerOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is the sample tensor. + + """ + + if state.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + # NOTE: There is no way to check in the jitted runtime if the prk mode was ran before + + prev_timestep = timestep - self.config.num_train_timesteps // state.num_inference_steps + prev_timestep = jnp.where(prev_timestep > 0, prev_timestep, 0) + + # Reference: + # if state.counter != 1: + # state.ets.append(model_output) + # else: + # prev_timestep = timestep + # timestep = timestep + self.config.num_train_timesteps // state.num_inference_steps + + prev_timestep = jnp.where(state.counter == 1, timestep, prev_timestep) + timestep = jnp.where( + state.counter == 1, timestep + self.config.num_train_timesteps // state.num_inference_steps, timestep + ) + + # Reference: + # if len(state.ets) == 1 and state.counter == 0: + # model_output = model_output + # state.cur_sample = sample + # elif len(state.ets) == 1 and state.counter == 1: + # model_output = (model_output + state.ets[-1]) / 2 + # sample = state.cur_sample + # state.cur_sample = None + # elif len(state.ets) == 2: + # model_output = (3 * state.ets[-1] - state.ets[-2]) / 2 + # elif len(state.ets) == 3: + # model_output = (23 * state.ets[-1] - 16 * state.ets[-2] + 5 * state.ets[-3]) / 12 + # else: + # model_output = (1 / 24) * (55 * state.ets[-1] - 59 * state.ets[-2] + 37 * state.ets[-3] - 9 * state.ets[-4]) + + state = state.replace( + ets=jax.lax.select( + state.counter != 1, + state.ets.at[0:3].set(state.ets[1:4]).at[3].set(model_output), # counter != 1 + state.ets, # counter 1 + ), + cur_sample=jax.lax.select( + state.counter != 1, + sample, # counter != 1 + state.cur_sample, # counter 1 + ), + ) + + state = state.replace( + cur_model_output=jax.lax.select_n( + jnp.clip(state.counter, 0, 4), + model_output, # counter 0 + (model_output + state.ets[-1]) / 2, # counter 1 + (3 * state.ets[-1] - state.ets[-2]) / 2, # counter 2 + (23 * state.ets[-1] - 16 * state.ets[-2] + 5 * state.ets[-3]) / 12, # counter 3 + (1 / 24) + * (55 * state.ets[-1] - 59 * state.ets[-2] + 37 * state.ets[-3] - 9 * state.ets[-4]), # counter >= 4 + ), + ) + + sample = state.cur_sample + model_output = state.cur_model_output + prev_sample = self._get_prev_sample(state, sample, timestep, prev_timestep, model_output) + state = state.replace(counter=state.counter + 1) + + return (prev_sample, state) + + def _get_prev_sample(self, state: PNDMSchedulerState, sample, timestep, prev_timestep, model_output): + # See formula (9) of PNDM paper https://huggingface.co/papers/2202.09778 + # this function computes x_(t−δ) using the formula of (9) + # Note that x_t needs to be added to both sides of the equation + + # Notation ( -> + # alpha_prod_t -> α_t + # alpha_prod_t_prev -> α_(t−δ) + # beta_prod_t -> (1 - α_t) + # beta_prod_t_prev -> (1 - α_(t−δ)) + # sample -> x_t + # model_output -> e_θ(x_t, t) + # prev_sample -> x_(t−δ) + alpha_prod_t = state.common.alphas_cumprod[timestep] + alpha_prod_t_prev = jnp.where( + prev_timestep >= 0, state.common.alphas_cumprod[prev_timestep], state.final_alpha_cumprod + ) + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + if self.config.prediction_type == "v_prediction": + model_output = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + elif self.config.prediction_type != "epsilon": + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon` or `v_prediction`" + ) + + # corresponds to (α_(t−δ) - α_t) divided by + # denominator of x_t in formula (9) and plus 1 + # Note: (α_(t−δ) - α_t) / (sqrt(α_t) * (sqrt(α_(t−δ)) + sqr(α_t))) = + # sqrt(α_(t−δ)) / sqrt(α_t)) + sample_coeff = (alpha_prod_t_prev / alpha_prod_t) ** (0.5) + + # corresponds to denominator of e_θ(x_t, t) in formula (9) + model_output_denom_coeff = alpha_prod_t * beta_prod_t_prev ** (0.5) + ( + alpha_prod_t * beta_prod_t * alpha_prod_t_prev + ) ** (0.5) + + # full formula (9) + prev_sample = ( + sample_coeff * sample - (alpha_prod_t_prev - alpha_prod_t) * model_output / model_output_denom_coeff + ) + + return prev_sample + + def add_noise( + self, + state: PNDMSchedulerState, + original_samples: jnp.ndarray, + noise: jnp.ndarray, + timesteps: jnp.ndarray, + ) -> jnp.ndarray: + return add_noise_common(state.common, original_samples, noise, timesteps) + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_repaint.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_repaint.py new file mode 100644 index 0000000000000000000000000000000000000000..6530c5af9e5bc48657de83e1cf21620472534fd3 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_repaint.py @@ -0,0 +1,365 @@ +# Copyright 2025 ETH Zurich Computer Vision Lab and The HuggingFace Team. All rights reserved. +# +# 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 dataclasses import dataclass +from typing import Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +@dataclass +class RePaintSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's step function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample (x_{0}) based on the model output from + the current timestep. `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class RePaintScheduler(SchedulerMixin, ConfigMixin): + """ + `RePaintScheduler` is a scheduler for DDPM inpainting inside a given mask. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, `squaredcos_cap_v2`, or `sigmoid`. + eta (`float`): + The weight of noise for added noise in diffusion step. If its value is between 0.0 and 1.0 it corresponds + to the DDIM scheduler, and if its value is between -0.0 and 1.0 it corresponds to the DDPM scheduler. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample between -1 and 1 for numerical stability. + + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + eta: float = 0.0, + trained_betas: Optional[np.ndarray] = None, + clip_sample: bool = True, + ): + if trained_betas is not None: + self.betas = torch.from_numpy(trained_betas) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + elif beta_schedule == "sigmoid": + # GeoDiff sigmoid schedule + betas = torch.linspace(-6, 6, num_train_timesteps) + self.betas = torch.sigmoid(betas) * (beta_end - beta_start) + beta_start + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + self.one = torch.tensor(1.0) + + self.final_alpha_cumprod = torch.tensor(1.0) + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy()) + + self.eta = eta + + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps( + self, + num_inference_steps: int, + jump_length: int = 10, + jump_n_sample: int = 10, + device: Union[str, torch.device] = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, + `timesteps` must be `None`. + jump_length (`int`, defaults to 10): + The number of steps taken forward in time before going backward in time for a single jump (“j” in + RePaint paper). Take a look at Figure 9 and 10 in the paper. + jump_n_sample (`int`, defaults to 10): + The number of times to make a forward time jump for a given chosen time sample. Take a look at Figure 9 + and 10 in the paper. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + + """ + num_inference_steps = min(self.config.num_train_timesteps, num_inference_steps) + self.num_inference_steps = num_inference_steps + + timesteps = [] + + jumps = {} + for j in range(0, num_inference_steps - jump_length, jump_length): + jumps[j] = jump_n_sample - 1 + + t = num_inference_steps + while t >= 1: + t = t - 1 + timesteps.append(t) + + if jumps.get(t, 0) > 0: + jumps[t] = jumps[t] - 1 + for _ in range(jump_length): + t = t + 1 + timesteps.append(t) + + timesteps = np.array(timesteps) * (self.config.num_train_timesteps // self.num_inference_steps) + self.timesteps = torch.from_numpy(timesteps).to(device) + + def _get_variance(self, t): + prev_timestep = t - self.config.num_train_timesteps // self.num_inference_steps + + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + # For t > 0, compute predicted variance βt (see formula (6) and (7) from + # https://huggingface.co/papers/2006.11239) and sample from it to get + # previous sample x_{t-1} ~ N(pred_prev_sample, variance) == add + # variance to pred_sample + # Is equivalent to formula (16) in https://huggingface.co/papers/2010.02502 + # without eta. + # variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * self.betas[t] + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + original_image: torch.Tensor, + mask: torch.Tensor, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[RePaintSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + original_image (`torch.Tensor`): + The original image to inpaint on. + mask (`torch.Tensor`): + The mask where a value of 0.0 indicates which part of the original image to inpaint. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_repaint.RePaintSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_repaint.RePaintSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_repaint.RePaintSchedulerOutput`] is returned, + otherwise a tuple is returned where the first element is the sample tensor. + + """ + t = timestep + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + # 1. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + + # 2. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (15) from https://huggingface.co/papers/2006.11239 + pred_original_sample = (sample - beta_prod_t**0.5 * model_output) / alpha_prod_t**0.5 + + # 3. Clip "predicted x_0" + if self.config.clip_sample: + pred_original_sample = torch.clamp(pred_original_sample, -1, 1) + + # We choose to follow RePaint Algorithm 1 to get x_{t-1}, however we + # substitute formula (7) in the algorithm coming from DDPM paper + # (formula (4) Algorithm 2 - Sampling) with formula (12) from DDIM paper. + # DDIM schedule gives the same results as DDPM with eta = 1.0 + # Noise is being reused in 7. and 8., but no impact on quality has + # been observed. + + # 5. Add noise + device = model_output.device + noise = randn_tensor(model_output.shape, generator=generator, device=device, dtype=model_output.dtype) + std_dev_t = self.eta * self._get_variance(timestep) ** 0.5 + + variance = 0 + if t > 0 and self.eta > 0: + variance = std_dev_t * noise + + # 6. compute "direction pointing to x_t" of formula (12) + # from https://huggingface.co/papers/2010.02502 + pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** 0.5 * model_output + + # 7. compute x_{t-1} of formula (12) from https://huggingface.co/papers/2010.02502 + prev_unknown_part = alpha_prod_t_prev**0.5 * pred_original_sample + pred_sample_direction + variance + + # 8. Algorithm 1 Line 5 https://huggingface.co/papers/2201.09865 + # The computation reported in Algorithm 1 Line 5 is incorrect. Line 5 refers to formula (8a) of the same paper, + # which tells to sample from a Gaussian distribution with mean "(alpha_prod_t_prev**0.5) * original_image" + # and variance "(1 - alpha_prod_t_prev)". This means that the standard Gaussian distribution "noise" should be + # scaled by the square root of the variance (as it is done here), however Algorithm 1 Line 5 tells to scale by the variance. + prev_known_part = (alpha_prod_t_prev**0.5) * original_image + ((1 - alpha_prod_t_prev) ** 0.5) * noise + + # 9. Algorithm 1 Line 8 https://huggingface.co/papers/2201.09865 + pred_prev_sample = mask * prev_known_part + (1.0 - mask) * prev_unknown_part + + if not return_dict: + return ( + pred_prev_sample, + pred_original_sample, + ) + + return RePaintSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample) + + def undo_step(self, sample, timestep, generator=None): + n = self.config.num_train_timesteps // self.num_inference_steps + + for i in range(n): + beta = self.betas[timestep + i] + if sample.device.type == "mps": + # randn does not work reproducibly on mps + noise = randn_tensor(sample.shape, dtype=sample.dtype, generator=generator) + noise = noise.to(sample.device) + else: + noise = randn_tensor(sample.shape, generator=generator, device=sample.device, dtype=sample.dtype) + + # 10. Algorithm 1 Line 10 https://huggingface.co/papers/2201.09865 + sample = (1 - beta) ** 0.5 * sample + beta**0.5 * noise + + return sample + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + raise NotImplementedError("Use `DDPMScheduler.add_noise()` to train for sampling with RePaint.") + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_sasolver.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_sasolver.py new file mode 100644 index 0000000000000000000000000000000000000000..2df7d560ddfb68d786ad3e96227a56ead053ae4b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_sasolver.py @@ -0,0 +1,1221 @@ +# Copyright 2025 Shuchen Xue, etc. in University of Chinese Academy of Sciences Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: check https://huggingface.co/papers/2309.05019 +# The codebase is modified based on https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_dpmsolver_multistep.py + +import math +from typing import Callable, List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import deprecate, is_scipy_available +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +if is_scipy_available(): + import scipy.stats + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class SASolverScheduler(SchedulerMixin, ConfigMixin): + """ + `SASolverScheduler` is a fast dedicated high-order solver for diffusion SDEs. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + predictor_order (`int`, defaults to 2): + The predictor order which can be `1` or `2` or `3` or '4'. It is recommended to use `predictor_order=2` for + guided sampling, and `predictor_order=3` for unconditional sampling. + corrector_order (`int`, defaults to 2): + The corrector order which can be `1` or `2` or `3` or '4'. It is recommended to use `corrector_order=2` for + guided sampling, and `corrector_order=3` for unconditional sampling. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + tau_func (`Callable`, *optional*): + Stochasticity during the sampling. Default in init is `lambda t: 1 if t >= 200 and t <= 800 else 0`. + SA-Solver will sample from vanilla diffusion ODE if tau_func is set to `lambda t: 0`. SA-Solver will sample + from vanilla diffusion SDE if tau_func is set to `lambda t: 1`. For more details, please check + https://huggingface.co/papers/2309.05019 + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True` and + `algorithm_type="dpmsolver++"`. + algorithm_type (`str`, defaults to `data_prediction`): + Algorithm type for the solver; can be `data_prediction` or `noise_prediction`. It is recommended to use + `data_prediction` with `solver_order=2` for guided sampling like in Stable Diffusion. + lower_order_final (`bool`, defaults to `True`): + Whether to use lower-order solvers in the final steps. Default = True. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + lambda_min_clipped (`float`, defaults to `-inf`): + Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the + cosine (`squaredcos_cap_v2`) noise schedule. + variance_type (`str`, *optional*): + Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output + contains the predicted Gaussian variance. + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + predictor_order: int = 2, + corrector_order: int = 2, + prediction_type: str = "epsilon", + tau_func: Optional[Callable] = None, + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + algorithm_type: str = "data_prediction", + lower_order_final: bool = True, + use_karras_sigmas: Optional[bool] = False, + use_exponential_sigmas: Optional[bool] = False, + use_beta_sigmas: Optional[bool] = False, + use_flow_sigmas: Optional[bool] = False, + flow_shift: Optional[float] = 1.0, + lambda_min_clipped: float = -float("inf"), + variance_type: Optional[str] = None, + timestep_spacing: str = "linspace", + steps_offset: int = 0, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = ( + torch.linspace( + beta_start**0.5, + beta_end**0.5, + num_train_timesteps, + dtype=torch.float32, + ) + ** 2 + ) + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + # Currently we only support VP-type noise schedule + self.alpha_t = torch.sqrt(self.alphas_cumprod) + self.sigma_t = torch.sqrt(1 - self.alphas_cumprod) + self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t) + self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5 + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + if algorithm_type not in ["data_prediction", "noise_prediction"]: + raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}") + + # setable values + self.num_inference_steps = None + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps) + self.timestep_list = [None] * max(predictor_order, corrector_order - 1) + self.model_outputs = [None] * max(predictor_order, corrector_order - 1) + + if tau_func is None: + self.tau_func = lambda t: 1 if t >= 200 and t <= 800 else 0 + else: + self.tau_func = tau_func + self.predict_x0 = algorithm_type == "data_prediction" + self.lower_order_nums = 0 + self.last_sample = None + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_timesteps(self, num_inference_steps: int = None, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + # Clipping the minimum of all lambda(t) for numerical stability. + # This is critical for cosine (squaredcos_cap_v2) noise schedule. + clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped) + last_timestep = ((self.config.num_train_timesteps - clipped_idx).numpy()).item() + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, last_timestep - 1, num_inference_steps + 1).round()[::-1][:-1].copy().astype(np.int64) + ) + + elif self.config.timestep_spacing == "leading": + step_ratio = last_timestep // (num_inference_steps + 1) + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.arange(last_timestep, 0, -step_ratio).round().copy().astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + if self.config.use_karras_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round() + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_exponential_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_beta_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_flow_sigmas: + alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1) + sigmas = 1.0 - alphas + sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy() + timesteps = (sigmas * self.config.num_train_timesteps).copy() + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + else: + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5 + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + + self.sigmas = torch.from_numpy(sigmas) + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64) + + self.num_inference_steps = len(timesteps) + self.model_outputs = [ + None, + ] * max(self.config.predictor_order, self.config.corrector_order - 1) + self.lower_order_nums = 0 + self.last_sample = None + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t + def _sigma_to_alpha_sigma_t(self, sigma): + if self.config.use_flow_sigmas: + alpha_t = 1 - sigma + sigma_t = sigma + else: + alpha_t = 1 / ((sigma**2 + 1) ** 0.5) + sigma_t = sigma * alpha_t + + return alpha_t, sigma_t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """Constructs an exponential noise schedule.""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def convert_model_output( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + """ + Convert the model output to the corresponding type the data_prediction/noise_prediction algorithm needs. + Noise_prediction is designed to discretize an integral of the noise prediction model, and data_prediction is + designed to discretize an integral of the data prediction model. + + + + The algorithm and model type are decoupled. You can use either data_prediction or noise_prediction for both + noise prediction and data prediction models. + + + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + # SA-Solver_data_prediction needs to solve an integral of the data prediction model. + if self.config.algorithm_type in ["data_prediction"]: + if self.config.prediction_type == "epsilon": + # SA-Solver only needs the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + model_output = model_output[:, :3] + x0_pred = (sample - sigma_t * model_output) / alpha_t + elif self.config.prediction_type == "sample": + x0_pred = model_output + elif self.config.prediction_type == "v_prediction": + x0_pred = alpha_t * sample - sigma_t * model_output + elif self.config.prediction_type == "flow_prediction": + sigma_t = self.sigmas[self.step_index] + x0_pred = sample - sigma_t * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + "`v_prediction`, or `flow_prediction` for the SASolverScheduler." + ) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + return x0_pred + + # SA-Solver_noise_prediction needs to solve an integral of the noise prediction model. + elif self.config.algorithm_type in ["noise_prediction"]: + if self.config.prediction_type == "epsilon": + # SA-Solver only needs the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + epsilon = model_output[:, :3] + else: + epsilon = model_output + elif self.config.prediction_type == "sample": + epsilon = (sample - alpha_t * model_output) / sigma_t + elif self.config.prediction_type == "v_prediction": + epsilon = alpha_t * model_output + sigma_t * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction` for the SASolverScheduler." + ) + + if self.config.thresholding: + alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep] + x0_pred = (sample - sigma_t * epsilon) / alpha_t + x0_pred = self._threshold_sample(x0_pred) + epsilon = (sample - alpha_t * x0_pred) / sigma_t + + return epsilon + + def get_coefficients_exponential_negative(self, order, interval_start, interval_end): + """ + Calculate the integral of exp(-x) * x^order dx from interval_start to interval_end + """ + assert order in [0, 1, 2, 3], "order is only supported for 0, 1, 2 and 3" + + if order == 0: + return torch.exp(-interval_end) * (torch.exp(interval_end - interval_start) - 1) + elif order == 1: + return torch.exp(-interval_end) * ( + (interval_start + 1) * torch.exp(interval_end - interval_start) - (interval_end + 1) + ) + elif order == 2: + return torch.exp(-interval_end) * ( + (interval_start**2 + 2 * interval_start + 2) * torch.exp(interval_end - interval_start) + - (interval_end**2 + 2 * interval_end + 2) + ) + elif order == 3: + return torch.exp(-interval_end) * ( + (interval_start**3 + 3 * interval_start**2 + 6 * interval_start + 6) + * torch.exp(interval_end - interval_start) + - (interval_end**3 + 3 * interval_end**2 + 6 * interval_end + 6) + ) + + def get_coefficients_exponential_positive(self, order, interval_start, interval_end, tau): + """ + Calculate the integral of exp(x(1+tau^2)) * x^order dx from interval_start to interval_end + """ + assert order in [0, 1, 2, 3], "order is only supported for 0, 1, 2 and 3" + + # after change of variable(cov) + interval_end_cov = (1 + tau**2) * interval_end + interval_start_cov = (1 + tau**2) * interval_start + + if order == 0: + return ( + torch.exp(interval_end_cov) * (1 - torch.exp(-(interval_end_cov - interval_start_cov))) / (1 + tau**2) + ) + elif order == 1: + return ( + torch.exp(interval_end_cov) + * ( + (interval_end_cov - 1) + - (interval_start_cov - 1) * torch.exp(-(interval_end_cov - interval_start_cov)) + ) + / ((1 + tau**2) ** 2) + ) + elif order == 2: + return ( + torch.exp(interval_end_cov) + * ( + (interval_end_cov**2 - 2 * interval_end_cov + 2) + - (interval_start_cov**2 - 2 * interval_start_cov + 2) + * torch.exp(-(interval_end_cov - interval_start_cov)) + ) + / ((1 + tau**2) ** 3) + ) + elif order == 3: + return ( + torch.exp(interval_end_cov) + * ( + (interval_end_cov**3 - 3 * interval_end_cov**2 + 6 * interval_end_cov - 6) + - (interval_start_cov**3 - 3 * interval_start_cov**2 + 6 * interval_start_cov - 6) + * torch.exp(-(interval_end_cov - interval_start_cov)) + ) + / ((1 + tau**2) ** 4) + ) + + def lagrange_polynomial_coefficient(self, order, lambda_list): + """ + Calculate the coefficient of lagrange polynomial + """ + + assert order in [0, 1, 2, 3] + assert order == len(lambda_list) - 1 + if order == 0: + return [[1]] + elif order == 1: + return [ + [ + 1 / (lambda_list[0] - lambda_list[1]), + -lambda_list[1] / (lambda_list[0] - lambda_list[1]), + ], + [ + 1 / (lambda_list[1] - lambda_list[0]), + -lambda_list[0] / (lambda_list[1] - lambda_list[0]), + ], + ] + elif order == 2: + denominator1 = (lambda_list[0] - lambda_list[1]) * (lambda_list[0] - lambda_list[2]) + denominator2 = (lambda_list[1] - lambda_list[0]) * (lambda_list[1] - lambda_list[2]) + denominator3 = (lambda_list[2] - lambda_list[0]) * (lambda_list[2] - lambda_list[1]) + return [ + [ + 1 / denominator1, + (-lambda_list[1] - lambda_list[2]) / denominator1, + lambda_list[1] * lambda_list[2] / denominator1, + ], + [ + 1 / denominator2, + (-lambda_list[0] - lambda_list[2]) / denominator2, + lambda_list[0] * lambda_list[2] / denominator2, + ], + [ + 1 / denominator3, + (-lambda_list[0] - lambda_list[1]) / denominator3, + lambda_list[0] * lambda_list[1] / denominator3, + ], + ] + elif order == 3: + denominator1 = ( + (lambda_list[0] - lambda_list[1]) + * (lambda_list[0] - lambda_list[2]) + * (lambda_list[0] - lambda_list[3]) + ) + denominator2 = ( + (lambda_list[1] - lambda_list[0]) + * (lambda_list[1] - lambda_list[2]) + * (lambda_list[1] - lambda_list[3]) + ) + denominator3 = ( + (lambda_list[2] - lambda_list[0]) + * (lambda_list[2] - lambda_list[1]) + * (lambda_list[2] - lambda_list[3]) + ) + denominator4 = ( + (lambda_list[3] - lambda_list[0]) + * (lambda_list[3] - lambda_list[1]) + * (lambda_list[3] - lambda_list[2]) + ) + return [ + [ + 1 / denominator1, + (-lambda_list[1] - lambda_list[2] - lambda_list[3]) / denominator1, + ( + lambda_list[1] * lambda_list[2] + + lambda_list[1] * lambda_list[3] + + lambda_list[2] * lambda_list[3] + ) + / denominator1, + (-lambda_list[1] * lambda_list[2] * lambda_list[3]) / denominator1, + ], + [ + 1 / denominator2, + (-lambda_list[0] - lambda_list[2] - lambda_list[3]) / denominator2, + ( + lambda_list[0] * lambda_list[2] + + lambda_list[0] * lambda_list[3] + + lambda_list[2] * lambda_list[3] + ) + / denominator2, + (-lambda_list[0] * lambda_list[2] * lambda_list[3]) / denominator2, + ], + [ + 1 / denominator3, + (-lambda_list[0] - lambda_list[1] - lambda_list[3]) / denominator3, + ( + lambda_list[0] * lambda_list[1] + + lambda_list[0] * lambda_list[3] + + lambda_list[1] * lambda_list[3] + ) + / denominator3, + (-lambda_list[0] * lambda_list[1] * lambda_list[3]) / denominator3, + ], + [ + 1 / denominator4, + (-lambda_list[0] - lambda_list[1] - lambda_list[2]) / denominator4, + ( + lambda_list[0] * lambda_list[1] + + lambda_list[0] * lambda_list[2] + + lambda_list[1] * lambda_list[2] + ) + / denominator4, + (-lambda_list[0] * lambda_list[1] * lambda_list[2]) / denominator4, + ], + ] + + def get_coefficients_fn(self, order, interval_start, interval_end, lambda_list, tau): + assert order in [1, 2, 3, 4] + assert order == len(lambda_list), "the length of lambda list must be equal to the order" + coefficients = [] + lagrange_coefficient = self.lagrange_polynomial_coefficient(order - 1, lambda_list) + for i in range(order): + coefficient = 0 + for j in range(order): + if self.predict_x0: + coefficient += lagrange_coefficient[i][j] * self.get_coefficients_exponential_positive( + order - 1 - j, interval_start, interval_end, tau + ) + else: + coefficient += lagrange_coefficient[i][j] * self.get_coefficients_exponential_negative( + order - 1 - j, interval_start, interval_end + ) + coefficients.append(coefficient) + assert len(coefficients) == order, "the length of coefficients does not match the order" + return coefficients + + def stochastic_adams_bashforth_update( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor, + noise: torch.Tensor, + order: int, + tau: torch.Tensor, + **kwargs, + ) -> torch.Tensor: + """ + One step for the SA-Predictor. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model at the current timestep. + prev_timestep (`int`): + The previous discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + order (`int`): + The order of SA-Predictor at this timestep. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + prev_timestep = args[0] if len(args) > 0 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if noise is None: + if len(args) > 2: + noise = args[2] + else: + raise ValueError("missing `noise` as a required keyword argument") + if order is None: + if len(args) > 3: + order = args[3] + else: + raise ValueError("missing `order` as a required keyword argument") + if tau is None: + if len(args) > 4: + tau = args[4] + else: + raise ValueError("missing `tau` as a required keyword argument") + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + model_output_list = self.model_outputs + sigma_t, sigma_s0 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + ) + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + + gradient_part = torch.zeros_like(sample) + h = lambda_t - lambda_s0 + lambda_list = [] + + for i in range(order): + si = self.step_index - i + alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si]) + lambda_si = torch.log(alpha_si) - torch.log(sigma_si) + lambda_list.append(lambda_si) + + gradient_coefficients = self.get_coefficients_fn(order, lambda_s0, lambda_t, lambda_list, tau) + + x = sample + + if self.predict_x0: + if ( + order == 2 + ): ## if order = 2 we do a modification that does not influence the convergence order similar to unipc. Note: This is used only for few steps sampling. + # The added term is O(h^3). Empirically we find it will slightly improve the image quality. + # ODE case + # gradient_coefficients[0] += 1.0 * torch.exp(lambda_t) * (h ** 2 / 2 - (h - 1 + torch.exp(-h))) / (ns.marginal_lambda(t_prev_list[-1]) - ns.marginal_lambda(t_prev_list[-2])) + # gradient_coefficients[1] -= 1.0 * torch.exp(lambda_t) * (h ** 2 / 2 - (h - 1 + torch.exp(-h))) / (ns.marginal_lambda(t_prev_list[-1]) - ns.marginal_lambda(t_prev_list[-2])) + temp_sigma = self.sigmas[self.step_index - 1] + temp_alpha_s, temp_sigma_s = self._sigma_to_alpha_sigma_t(temp_sigma) + temp_lambda_s = torch.log(temp_alpha_s) - torch.log(temp_sigma_s) + gradient_coefficients[0] += ( + 1.0 + * torch.exp((1 + tau**2) * lambda_t) + * (h**2 / 2 - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h))) / ((1 + tau**2) ** 2)) + / (lambda_s0 - temp_lambda_s) + ) + gradient_coefficients[1] -= ( + 1.0 + * torch.exp((1 + tau**2) * lambda_t) + * (h**2 / 2 - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h))) / ((1 + tau**2) ** 2)) + / (lambda_s0 - temp_lambda_s) + ) + + for i in range(order): + if self.predict_x0: + gradient_part += ( + (1 + tau**2) + * sigma_t + * torch.exp(-(tau**2) * lambda_t) + * gradient_coefficients[i] + * model_output_list[-(i + 1)] + ) + else: + gradient_part += -(1 + tau**2) * alpha_t * gradient_coefficients[i] * model_output_list[-(i + 1)] + + if self.predict_x0: + noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise + else: + noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise + + if self.predict_x0: + x_t = torch.exp(-(tau**2) * h) * (sigma_t / sigma_s0) * x + gradient_part + noise_part + else: + x_t = (alpha_t / alpha_s0) * x + gradient_part + noise_part + + x_t = x_t.to(x.dtype) + return x_t + + def stochastic_adams_moulton_update( + self, + this_model_output: torch.Tensor, + *args, + last_sample: torch.Tensor, + last_noise: torch.Tensor, + this_sample: torch.Tensor, + order: int, + tau: torch.Tensor, + **kwargs, + ) -> torch.Tensor: + """ + One step for the SA-Corrector. + + Args: + this_model_output (`torch.Tensor`): + The model outputs at `x_t`. + this_timestep (`int`): + The current timestep `t`. + last_sample (`torch.Tensor`): + The generated sample before the last predictor `x_{t-1}`. + this_sample (`torch.Tensor`): + The generated sample after the last predictor `x_{t}`. + order (`int`): + The order of SA-Corrector at this step. + + Returns: + `torch.Tensor`: + The corrected sample tensor at the current timestep. + """ + + this_timestep = args[0] if len(args) > 0 else kwargs.pop("this_timestep", None) + if last_sample is None: + if len(args) > 1: + last_sample = args[1] + else: + raise ValueError("missing `last_sample` as a required keyword argument") + if last_noise is None: + if len(args) > 2: + last_noise = args[2] + else: + raise ValueError("missing `last_noise` as a required keyword argument") + if this_sample is None: + if len(args) > 3: + this_sample = args[3] + else: + raise ValueError("missing `this_sample` as a required keyword argument") + if order is None: + if len(args) > 4: + order = args[4] + else: + raise ValueError("missing `order` as a required keyword argument") + if tau is None: + if len(args) > 5: + tau = args[5] + else: + raise ValueError("missing `tau` as a required keyword argument") + if this_timestep is not None: + deprecate( + "this_timestep", + "1.0.0", + "Passing `this_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + model_output_list = self.model_outputs + sigma_t, sigma_s0 = ( + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + ) + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + gradient_part = torch.zeros_like(this_sample) + h = lambda_t - lambda_s0 + lambda_list = [] + for i in range(order): + si = self.step_index - i + alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si]) + lambda_si = torch.log(alpha_si) - torch.log(sigma_si) + lambda_list.append(lambda_si) + + model_prev_list = model_output_list + [this_model_output] + + gradient_coefficients = self.get_coefficients_fn(order, lambda_s0, lambda_t, lambda_list, tau) + + x = last_sample + + if self.predict_x0: + if ( + order == 2 + ): ## if order = 2 we do a modification that does not influence the convergence order similar to UniPC. Note: This is used only for few steps sampling. + # The added term is O(h^3). Empirically we find it will slightly improve the image quality. + # ODE case + # gradient_coefficients[0] += 1.0 * torch.exp(lambda_t) * (h / 2 - (h - 1 + torch.exp(-h)) / h) + # gradient_coefficients[1] -= 1.0 * torch.exp(lambda_t) * (h / 2 - (h - 1 + torch.exp(-h)) / h) + gradient_coefficients[0] += ( + 1.0 + * torch.exp((1 + tau**2) * lambda_t) + * (h / 2 - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h))) / ((1 + tau**2) ** 2 * h)) + ) + gradient_coefficients[1] -= ( + 1.0 + * torch.exp((1 + tau**2) * lambda_t) + * (h / 2 - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h))) / ((1 + tau**2) ** 2 * h)) + ) + + for i in range(order): + if self.predict_x0: + gradient_part += ( + (1 + tau**2) + * sigma_t + * torch.exp(-(tau**2) * lambda_t) + * gradient_coefficients[i] + * model_prev_list[-(i + 1)] + ) + else: + gradient_part += -(1 + tau**2) * alpha_t * gradient_coefficients[i] * model_prev_list[-(i + 1)] + + if self.predict_x0: + noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * last_noise + else: + noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * last_noise + + if self.predict_x0: + x_t = torch.exp(-(tau**2) * h) * (sigma_t / sigma_s0) * x + gradient_part + noise_part + else: + x_t = (alpha_t / alpha_s0) * x + gradient_part + noise_part + + x_t = x_t.to(x.dtype) + return x_t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + index_candidates = (schedule_timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + return step_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index + def _init_step_index(self, timestep): + """ + Initialize the step_index counter for the scheduler. + """ + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + generator=None, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the SA-Solver. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + use_corrector = self.step_index > 0 and self.last_sample is not None + + model_output_convert = self.convert_model_output(model_output, sample=sample) + + if use_corrector: + current_tau = self.tau_func(self.timestep_list[-1]) + sample = self.stochastic_adams_moulton_update( + this_model_output=model_output_convert, + last_sample=self.last_sample, + last_noise=self.last_noise, + this_sample=sample, + order=self.this_corrector_order, + tau=current_tau, + ) + + for i in range(max(self.config.predictor_order, self.config.corrector_order - 1) - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.timestep_list[i] = self.timestep_list[i + 1] + + self.model_outputs[-1] = model_output_convert + self.timestep_list[-1] = timestep + + noise = randn_tensor( + model_output.shape, + generator=generator, + device=model_output.device, + dtype=model_output.dtype, + ) + + if self.config.lower_order_final: + this_predictor_order = min(self.config.predictor_order, len(self.timesteps) - self.step_index) + this_corrector_order = min(self.config.corrector_order, len(self.timesteps) - self.step_index + 1) + else: + this_predictor_order = self.config.predictor_order + this_corrector_order = self.config.corrector_order + + self.this_predictor_order = min(this_predictor_order, self.lower_order_nums + 1) # warmup for multistep + self.this_corrector_order = min(this_corrector_order, self.lower_order_nums + 2) # warmup for multistep + assert self.this_predictor_order > 0 + assert self.this_corrector_order > 0 + + self.last_sample = sample + self.last_noise = noise + + current_tau = self.tau_func(self.timestep_list[-1]) + prev_sample = self.stochastic_adams_bashforth_update( + model_output=model_output_convert, + sample=sample, + noise=noise, + order=self.this_predictor_order, + tau=current_tau, + ) + + if self.lower_order_nums < max(self.config.predictor_order, self.config.corrector_order - 1): + self.lower_order_nums += 1 + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_scm.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_scm.py new file mode 100644 index 0000000000000000000000000000000000000000..acff268c9b9231e3f28a6431574c64521f757b0c --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_scm.py @@ -0,0 +1,265 @@ +# # Copyright 2025 Sana-Sprint Authors and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +from dataclasses import dataclass +from typing import Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..schedulers.scheduling_utils import SchedulerMixin +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->SCM +class SCMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +class SCMScheduler(SchedulerMixin, ConfigMixin): + """ + `SCMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with + non-Markovian guidance. This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass + documentation for the generic methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + prediction_type (`str`, defaults to `trigflow`): + Prediction type of the scheduler function. Currently only supports "trigflow". + sigma_data (`float`, defaults to 0.5): + The standard deviation of the noise added during multi-step inference. + """ + + # _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + prediction_type: str = "trigflow", + sigma_data: float = 0.5, + ): + """ + Initialize the SCM scheduler. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + prediction_type (`str`, defaults to `trigflow`): + Prediction type of the scheduler function. Currently only supports "trigflow". + sigma_data (`float`, defaults to 0.5): + The standard deviation of the noise added during multi-step inference. + """ + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + + self._step_index = None + self._begin_index = None + + @property + def step_index(self): + return self._step_index + + @property + def begin_index(self): + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_timesteps( + self, + num_inference_steps: int, + timesteps: torch.Tensor = None, + device: Union[str, torch.device] = None, + max_timesteps: float = 1.57080, + intermediate_timesteps: float = 1.3, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + timesteps (`torch.Tensor`, *optional*): + Custom timesteps to use for the denoising process. + max_timesteps (`float`, defaults to 1.57080): + The maximum timestep value used in the SCM scheduler. + intermediate_timesteps (`float`, *optional*, defaults to 1.3): + The intermediate timestep value used in SCM scheduler (only used when num_inference_steps=2). + """ + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + if timesteps is not None and len(timesteps) != num_inference_steps + 1: + raise ValueError("If providing custom timesteps, `timesteps` must be of length `num_inference_steps + 1`.") + + if timesteps is not None and max_timesteps is not None: + raise ValueError("If providing custom timesteps, `max_timesteps` should not be provided.") + + if timesteps is None and max_timesteps is None: + raise ValueError("Should provide either `timesteps` or `max_timesteps`.") + + if intermediate_timesteps is not None and num_inference_steps != 2: + raise ValueError("Intermediate timesteps for SCM is not supported when num_inference_steps != 2.") + + self.num_inference_steps = num_inference_steps + + if timesteps is not None: + if isinstance(timesteps, list): + self.timesteps = torch.tensor(timesteps, device=device).float() + elif isinstance(timesteps, torch.Tensor): + self.timesteps = timesteps.to(device).float() + else: + raise ValueError(f"Unsupported timesteps type: {type(timesteps)}") + elif intermediate_timesteps is not None: + self.timesteps = torch.tensor([max_timesteps, intermediate_timesteps, 0], device=device).float() + else: + # max_timesteps=arctan(80/0.5)=1.56454 is the default from sCM paper, we choose a different value here + self.timesteps = torch.linspace(max_timesteps, 0, num_inference_steps + 1, device=device).float() + print(f"Set timesteps: {self.timesteps}") + + self._step_index = None + self._begin_index = None + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + def step( + self, + model_output: torch.FloatTensor, + timestep: float, + sample: torch.FloatTensor, + generator: torch.Generator = None, + return_dict: bool = True, + ) -> Union[SCMSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_scm.SCMSchedulerOutput`] or `tuple`. + Returns: + [`~schedulers.scheduling_utils.SCMSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_scm.SCMSchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # 2. compute alphas, betas + t = self.timesteps[self.step_index + 1] + s = self.timesteps[self.step_index] + + # 4. Different Parameterization: + parameterization = self.config.prediction_type + + if parameterization == "trigflow": + pred_x0 = torch.cos(s) * sample - torch.sin(s) * model_output + else: + raise ValueError(f"Unsupported parameterization: {parameterization}") + + # 5. Sample z ~ N(0, I), For MultiStep Inference + # Noise is not used for one-step sampling. + if len(self.timesteps) > 1: + noise = ( + randn_tensor(model_output.shape, device=model_output.device, generator=generator) + * self.config.sigma_data + ) + prev_sample = torch.cos(t) * pred_x0 + torch.sin(t) * noise + else: + prev_sample = pred_x0 + + self._step_index += 1 + + if not return_dict: + return (prev_sample, pred_x0) + + return SCMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_x0) + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_sde_ve.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_sde_ve.py new file mode 100644 index 0000000000000000000000000000000000000000..1bfc08cce5e97c3efcacbb4f18744cac441ff6fa --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_sde_ve.py @@ -0,0 +1,301 @@ +# Copyright 2025 Google Brain and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/yang-song/score_sde_pytorch + +import math +from dataclasses import dataclass +from typing import Optional, Tuple, Union + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin, SchedulerOutput + + +@dataclass +class SdeVeOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + prev_sample_mean (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Mean averaged `prev_sample` over previous timesteps. + """ + + prev_sample: torch.Tensor + prev_sample_mean: torch.Tensor + + +class ScoreSdeVeScheduler(SchedulerMixin, ConfigMixin): + """ + `ScoreSdeVeScheduler` is a variance exploding stochastic differential equation (SDE) scheduler. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + snr (`float`, defaults to 0.15): + A coefficient weighting the step from the `model_output` sample (from the network) to the random noise. + sigma_min (`float`, defaults to 0.01): + The initial noise scale for the sigma sequence in the sampling procedure. The minimum sigma should mirror + the distribution of the data. + sigma_max (`float`, defaults to 1348.0): + The maximum value used for the range of continuous timesteps passed into the model. + sampling_eps (`float`, defaults to 1e-5): + The end value of sampling where timesteps decrease progressively from 1 to epsilon. + correct_steps (`int`, defaults to 1): + The number of correction steps performed on a produced sample. + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 2000, + snr: float = 0.15, + sigma_min: float = 0.01, + sigma_max: float = 1348.0, + sampling_eps: float = 1e-5, + correct_steps: int = 1, + ): + # standard deviation of the initial noise distribution + self.init_noise_sigma = sigma_max + + # setable values + self.timesteps = None + + self.set_sigmas(num_train_timesteps, sigma_min, sigma_max, sampling_eps) + + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps( + self, num_inference_steps: int, sampling_eps: float = None, device: Union[str, torch.device] = None + ): + """ + Sets the continuous timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + sampling_eps (`float`, *optional*): + The final timestep value (overrides value given during scheduler instantiation). + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + + """ + sampling_eps = sampling_eps if sampling_eps is not None else self.config.sampling_eps + + self.timesteps = torch.linspace(1, sampling_eps, num_inference_steps, device=device) + + def set_sigmas( + self, num_inference_steps: int, sigma_min: float = None, sigma_max: float = None, sampling_eps: float = None + ): + """ + Sets the noise scales used for the diffusion chain (to be run before inference). The sigmas control the weight + of the `drift` and `diffusion` components of the sample update. + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + sigma_min (`float`, optional): + The initial noise scale value (overrides value given during scheduler instantiation). + sigma_max (`float`, optional): + The final noise scale value (overrides value given during scheduler instantiation). + sampling_eps (`float`, optional): + The final timestep value (overrides value given during scheduler instantiation). + + """ + sigma_min = sigma_min if sigma_min is not None else self.config.sigma_min + sigma_max = sigma_max if sigma_max is not None else self.config.sigma_max + sampling_eps = sampling_eps if sampling_eps is not None else self.config.sampling_eps + if self.timesteps is None: + self.set_timesteps(num_inference_steps, sampling_eps) + + self.sigmas = sigma_min * (sigma_max / sigma_min) ** (self.timesteps / sampling_eps) + self.discrete_sigmas = torch.exp(torch.linspace(math.log(sigma_min), math.log(sigma_max), num_inference_steps)) + self.sigmas = torch.tensor([sigma_min * (sigma_max / sigma_min) ** t for t in self.timesteps]) + + def get_adjacent_sigma(self, timesteps, t): + return torch.where( + timesteps == 0, + torch.zeros_like(t.to(timesteps.device)), + self.discrete_sigmas[timesteps - 1].to(timesteps.device), + ) + + def step_pred( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[SdeVeOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_sde_ve.SdeVeOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_sde_ve.SdeVeOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_sde_ve.SdeVeOutput`] is returned, otherwise a tuple + is returned where the first element is the sample tensor. + + """ + if self.timesteps is None: + raise ValueError( + "`self.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler" + ) + + timestep = timestep * torch.ones( + sample.shape[0], device=sample.device + ) # torch.repeat_interleave(timestep, sample.shape[0]) + timesteps = (timestep * (len(self.timesteps) - 1)).long() + + # mps requires indices to be in the same device, so we use cpu as is the default with cuda + timesteps = timesteps.to(self.discrete_sigmas.device) + + sigma = self.discrete_sigmas[timesteps].to(sample.device) + adjacent_sigma = self.get_adjacent_sigma(timesteps, timestep).to(sample.device) + drift = torch.zeros_like(sample) + diffusion = (sigma**2 - adjacent_sigma**2) ** 0.5 + + # equation 6 in the paper: the model_output modeled by the network is grad_x log pt(x) + # also equation 47 shows the analog from SDE models to ancestral sampling methods + diffusion = diffusion.flatten() + while len(diffusion.shape) < len(sample.shape): + diffusion = diffusion.unsqueeze(-1) + drift = drift - diffusion**2 * model_output + + # equation 6: sample noise for the diffusion term of + noise = randn_tensor( + sample.shape, layout=sample.layout, generator=generator, device=sample.device, dtype=sample.dtype + ) + prev_sample_mean = sample - drift # subtract because `dt` is a small negative timestep + # TODO is the variable diffusion the correct scaling term for the noise? + prev_sample = prev_sample_mean + diffusion * noise # add impact of diffusion field g + + if not return_dict: + return (prev_sample, prev_sample_mean) + + return SdeVeOutput(prev_sample=prev_sample, prev_sample_mean=prev_sample_mean) + + def step_correct( + self, + model_output: torch.Tensor, + sample: torch.Tensor, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Correct the predicted sample based on the `model_output` of the network. This is often run repeatedly after + making the prediction for the previous timestep. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_sde_ve.SdeVeOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_sde_ve.SdeVeOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_sde_ve.SdeVeOutput`] is returned, otherwise a tuple + is returned where the first element is the sample tensor. + + """ + if self.timesteps is None: + raise ValueError( + "`self.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler" + ) + + # For small batch sizes, the paper "suggest replacing norm(z) with sqrt(d), where d is the dim. of z" + # sample noise for correction + noise = randn_tensor(sample.shape, layout=sample.layout, generator=generator).to(sample.device) + + # compute step size from the model_output, the noise, and the snr + grad_norm = torch.norm(model_output.reshape(model_output.shape[0], -1), dim=-1).mean() + noise_norm = torch.norm(noise.reshape(noise.shape[0], -1), dim=-1).mean() + step_size = (self.config.snr * noise_norm / grad_norm) ** 2 * 2 + step_size = step_size * torch.ones(sample.shape[0]).to(sample.device) + # self.repeat_scalar(step_size, sample.shape[0]) + + # compute corrected sample: model_output term and noise term + step_size = step_size.flatten() + while len(step_size.shape) < len(sample.shape): + step_size = step_size.unsqueeze(-1) + prev_sample_mean = sample + step_size * model_output + prev_sample = prev_sample_mean + ((step_size * 2) ** 0.5) * noise + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + timesteps = timesteps.to(original_samples.device) + sigmas = self.discrete_sigmas.to(original_samples.device)[timesteps] + noise = ( + noise * sigmas[:, None, None, None] + if noise is not None + else torch.randn_like(original_samples) * sigmas[:, None, None, None] + ) + noisy_samples = noise + original_samples + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_sde_ve_flax.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_sde_ve_flax.py new file mode 100644 index 0000000000000000000000000000000000000000..09cd081462b3fa5c5d005e3737735e9f34b4a5ac --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_sde_ve_flax.py @@ -0,0 +1,280 @@ +# Copyright 2025 Google Brain and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This file is strongly influenced by https://github.com/yang-song/score_sde_pytorch + +from dataclasses import dataclass +from typing import Optional, Tuple, Union + +import flax +import jax +import jax.numpy as jnp +from jax import random + +from ..configuration_utils import ConfigMixin, register_to_config +from .scheduling_utils_flax import FlaxSchedulerMixin, FlaxSchedulerOutput, broadcast_to_shape_from_left + + +@flax.struct.dataclass +class ScoreSdeVeSchedulerState: + # setable values + timesteps: Optional[jnp.ndarray] = None + discrete_sigmas: Optional[jnp.ndarray] = None + sigmas: Optional[jnp.ndarray] = None + + @classmethod + def create(cls): + return cls() + + +@dataclass +class FlaxSdeVeOutput(FlaxSchedulerOutput): + """ + Output class for the ScoreSdeVeScheduler's step function output. + + Args: + state (`ScoreSdeVeSchedulerState`): + prev_sample (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + prev_sample_mean (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)` for images): + Mean averaged `prev_sample`. Same as `prev_sample`, only mean-averaged over previous timesteps. + """ + + state: ScoreSdeVeSchedulerState + prev_sample: jnp.ndarray + prev_sample_mean: Optional[jnp.ndarray] = None + + +class FlaxScoreSdeVeScheduler(FlaxSchedulerMixin, ConfigMixin): + """ + The variance exploding stochastic differential equation (SDE) scheduler. + + For more information, see the original paper: https://huggingface.co/papers/2011.13456 + + [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` + function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. + [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and + [`~SchedulerMixin.from_pretrained`] functions. + + Args: + num_train_timesteps (`int`): number of diffusion steps used to train the model. + snr (`float`): + coefficient weighting the step from the model_output sample (from the network) to the random noise. + sigma_min (`float`): + initial noise scale for sigma sequence in sampling procedure. The minimum sigma should mirror the + distribution of the data. + sigma_max (`float`): maximum value used for the range of continuous timesteps passed into the model. + sampling_eps (`float`): the end value of sampling, where timesteps decrease progressively from 1 to + epsilon. + correct_steps (`int`): number of correction steps performed on a produced sample. + """ + + @property + def has_state(self): + return True + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 2000, + snr: float = 0.15, + sigma_min: float = 0.01, + sigma_max: float = 1348.0, + sampling_eps: float = 1e-5, + correct_steps: int = 1, + ): + pass + + def create_state(self): + state = ScoreSdeVeSchedulerState.create() + return self.set_sigmas( + state, + self.config.num_train_timesteps, + self.config.sigma_min, + self.config.sigma_max, + self.config.sampling_eps, + ) + + def set_timesteps( + self, state: ScoreSdeVeSchedulerState, num_inference_steps: int, shape: Tuple = (), sampling_eps: float = None + ) -> ScoreSdeVeSchedulerState: + """ + Sets the continuous timesteps used for the diffusion chain. Supporting function to be run before inference. + + Args: + state (`ScoreSdeVeSchedulerState`): the `FlaxScoreSdeVeScheduler` state data class instance. + num_inference_steps (`int`): + the number of diffusion steps used when generating samples with a pre-trained model. + sampling_eps (`float`, optional): + final timestep value (overrides value given at Scheduler instantiation). + + """ + sampling_eps = sampling_eps if sampling_eps is not None else self.config.sampling_eps + + timesteps = jnp.linspace(1, sampling_eps, num_inference_steps) + return state.replace(timesteps=timesteps) + + def set_sigmas( + self, + state: ScoreSdeVeSchedulerState, + num_inference_steps: int, + sigma_min: float = None, + sigma_max: float = None, + sampling_eps: float = None, + ) -> ScoreSdeVeSchedulerState: + """ + Sets the noise scales used for the diffusion chain. Supporting function to be run before inference. + + The sigmas control the weight of the `drift` and `diffusion` components of sample update. + + Args: + state (`ScoreSdeVeSchedulerState`): the `FlaxScoreSdeVeScheduler` state data class instance. + num_inference_steps (`int`): + the number of diffusion steps used when generating samples with a pre-trained model. + sigma_min (`float`, optional): + initial noise scale value (overrides value given at Scheduler instantiation). + sigma_max (`float`, optional): + final noise scale value (overrides value given at Scheduler instantiation). + sampling_eps (`float`, optional): + final timestep value (overrides value given at Scheduler instantiation). + """ + sigma_min = sigma_min if sigma_min is not None else self.config.sigma_min + sigma_max = sigma_max if sigma_max is not None else self.config.sigma_max + sampling_eps = sampling_eps if sampling_eps is not None else self.config.sampling_eps + if state.timesteps is None: + state = self.set_timesteps(state, num_inference_steps, sampling_eps) + + discrete_sigmas = jnp.exp(jnp.linspace(jnp.log(sigma_min), jnp.log(sigma_max), num_inference_steps)) + sigmas = jnp.array([sigma_min * (sigma_max / sigma_min) ** t for t in state.timesteps]) + + return state.replace(discrete_sigmas=discrete_sigmas, sigmas=sigmas) + + def get_adjacent_sigma(self, state, timesteps, t): + return jnp.where(timesteps == 0, jnp.zeros_like(t), state.discrete_sigmas[timesteps - 1]) + + def step_pred( + self, + state: ScoreSdeVeSchedulerState, + model_output: jnp.ndarray, + timestep: int, + sample: jnp.ndarray, + key: jax.Array, + return_dict: bool = True, + ) -> Union[FlaxSdeVeOutput, Tuple]: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + state (`ScoreSdeVeSchedulerState`): the `FlaxScoreSdeVeScheduler` state data class instance. + model_output (`jnp.ndarray`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + generator: random number generator. + return_dict (`bool`): option for returning tuple rather than FlaxSdeVeOutput class + + Returns: + [`FlaxSdeVeOutput`] or `tuple`: [`FlaxSdeVeOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is the sample tensor. + + """ + if state.timesteps is None: + raise ValueError( + "`state.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler" + ) + + timestep = timestep * jnp.ones( + sample.shape[0], + ) + timesteps = (timestep * (len(state.timesteps) - 1)).long() + + sigma = state.discrete_sigmas[timesteps] + adjacent_sigma = self.get_adjacent_sigma(state, timesteps, timestep) + drift = jnp.zeros_like(sample) + diffusion = (sigma**2 - adjacent_sigma**2) ** 0.5 + + # equation 6 in the paper: the model_output modeled by the network is grad_x log pt(x) + # also equation 47 shows the analog from SDE models to ancestral sampling methods + diffusion = diffusion.flatten() + diffusion = broadcast_to_shape_from_left(diffusion, sample.shape) + drift = drift - diffusion**2 * model_output + + # equation 6: sample noise for the diffusion term of + key = random.split(key, num=1) + noise = random.normal(key=key, shape=sample.shape) + prev_sample_mean = sample - drift # subtract because `dt` is a small negative timestep + # TODO is the variable diffusion the correct scaling term for the noise? + prev_sample = prev_sample_mean + diffusion * noise # add impact of diffusion field g + + if not return_dict: + return (prev_sample, prev_sample_mean, state) + + return FlaxSdeVeOutput(prev_sample=prev_sample, prev_sample_mean=prev_sample_mean, state=state) + + def step_correct( + self, + state: ScoreSdeVeSchedulerState, + model_output: jnp.ndarray, + sample: jnp.ndarray, + key: jax.Array, + return_dict: bool = True, + ) -> Union[FlaxSdeVeOutput, Tuple]: + """ + Correct the predicted sample based on the output model_output of the network. This is often run repeatedly + after making the prediction for the previous timestep. + + Args: + state (`ScoreSdeVeSchedulerState`): the `FlaxScoreSdeVeScheduler` state data class instance. + model_output (`jnp.ndarray`): direct output from learned diffusion model. + sample (`jnp.ndarray`): + current instance of sample being created by diffusion process. + generator: random number generator. + return_dict (`bool`): option for returning tuple rather than FlaxSdeVeOutput class + + Returns: + [`FlaxSdeVeOutput`] or `tuple`: [`FlaxSdeVeOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is the sample tensor. + + """ + if state.timesteps is None: + raise ValueError( + "`state.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler" + ) + + # For small batch sizes, the paper "suggest replacing norm(z) with sqrt(d), where d is the dim. of z" + # sample noise for correction + key = random.split(key, num=1) + noise = random.normal(key=key, shape=sample.shape) + + # compute step size from the model_output, the noise, and the snr + grad_norm = jnp.linalg.norm(model_output) + noise_norm = jnp.linalg.norm(noise) + step_size = (self.config.snr * noise_norm / grad_norm) ** 2 * 2 + step_size = step_size * jnp.ones(sample.shape[0]) + + # compute corrected sample: model_output term and noise term + step_size = step_size.flatten() + step_size = broadcast_to_shape_from_left(step_size, sample.shape) + prev_sample_mean = sample + step_size * model_output + prev_sample = prev_sample_mean + ((step_size * 2) ** 0.5) * noise + + if not return_dict: + return (prev_sample, state) + + return FlaxSdeVeOutput(prev_sample=prev_sample, state=state) + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_tcd.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_tcd.py new file mode 100644 index 0000000000000000000000000000000000000000..3fd5c341eca99d72c45e49ea72ac6ba2488dc9c2 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_tcd.py @@ -0,0 +1,690 @@ +# Copyright 2025 Stanford University Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +import math +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..schedulers.scheduling_utils import SchedulerMixin +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class TCDSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_noised_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted noised sample `(x_{s})` based on the model output from the current timestep. + """ + + prev_sample: torch.Tensor + pred_noised_sample: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas: torch.Tensor) -> torch.Tensor: + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + + Args: + betas (`torch.Tensor`): + the betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class TCDScheduler(SchedulerMixin, ConfigMixin): + """ + `TCDScheduler` incorporates the `Strategic Stochastic Sampling` introduced by the paper `Trajectory Consistency + Distillation`, extending the original Multistep Consistency Sampling to enable unrestricted trajectory traversal. + + This code is based on the official repo of TCD(https://github.com/jabir-zheng/TCD). + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. [`~ConfigMixin`] takes care of storing all config + attributes that are passed in the scheduler's `__init__` function, such as `num_train_timesteps`. They can be + accessed via `scheduler.config.num_train_timesteps`. [`SchedulerMixin`] provides general loading and saving + functionality via the [`SchedulerMixin.save_pretrained`] and [`~SchedulerMixin.from_pretrained`] functions. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + original_inference_steps (`int`, *optional*, defaults to 50): + The default number of inference steps used to generate a linearly-spaced timestep schedule, from which we + will ultimately take `num_inference_steps` evenly spaced timesteps to form the final timestep schedule. + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to 1.0): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, defaults to `True`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + timestep_scaling (`float`, defaults to 10.0): + The factor the timesteps will be multiplied by when calculating the consistency model boundary conditions + `c_skip` and `c_out`. Increasing this will decrease the approximation error (although the approximation + error at the default of `10.0` is already pretty small). + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, + beta_end: float = 0.012, + beta_schedule: str = "scaled_linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + original_inference_steps: int = 50, + clip_sample: bool = False, + clip_sample_range: float = 1.0, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: str = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + timestep_spacing: str = "leading", + timestep_scaling: float = 10.0, + rescale_betas_zero_snr: bool = False, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # At every step in ddim, we are looking into the previous alphas_cumprod + # For the final step, there is no previous alphas_cumprod because we are already at 0 + # `set_alpha_to_one` decides whether we set this parameter simply to one or + # whether we use the final alpha of the "non-previous" one. + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + self.custom_timesteps = False + + self._step_index = None + self._begin_index = None + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + @property + def step_index(self): + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler._get_variance + def _get_variance(self, timestep, prev_timestep): + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + def set_timesteps( + self, + num_inference_steps: Optional[int] = None, + device: Union[str, torch.device] = None, + original_inference_steps: Optional[int] = None, + timesteps: Optional[List[int]] = None, + strength: float = 1.0, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. If used, + `timesteps` must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + original_inference_steps (`int`, *optional*): + The original number of inference steps, which will be used to generate a linearly-spaced timestep + schedule (which is different from the standard `diffusers` implementation). We will then take + `num_inference_steps` timesteps from this schedule, evenly spaced in terms of indices, and use that as + our final timestep schedule. If not set, this will default to the `original_inference_steps` attribute. + timesteps (`List[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of equal spacing between timesteps on the training/distillation timestep + schedule is used. If `timesteps` is passed, `num_inference_steps` must be `None`. + strength (`float`, *optional*, defaults to 1.0): + Used to determine the number of timesteps used for inference when using img2img, inpaint, etc. + """ + # 0. Check inputs + if num_inference_steps is None and timesteps is None: + raise ValueError("Must pass exactly one of `num_inference_steps` or `custom_timesteps`.") + + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.") + + # 1. Calculate the TCD original training/distillation timestep schedule. + original_steps = ( + original_inference_steps if original_inference_steps is not None else self.config.original_inference_steps + ) + + if original_inference_steps is None: + # default option, timesteps align with discrete inference steps + if original_steps > self.config.num_train_timesteps: + raise ValueError( + f"`original_steps`: {original_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + # TCD Timesteps Setting + # The skipping step parameter k from the paper. + k = self.config.num_train_timesteps // original_steps + # TCD Training/Distillation Steps Schedule + tcd_origin_timesteps = np.asarray(list(range(1, int(original_steps * strength) + 1))) * k - 1 + else: + # customised option, sampled timesteps can be any arbitrary value + tcd_origin_timesteps = np.asarray(list(range(0, int(self.config.num_train_timesteps * strength)))) + + # 2. Calculate the TCD inference timestep schedule. + if timesteps is not None: + # 2.1 Handle custom timestep schedules. + train_timesteps = set(tcd_origin_timesteps) + non_train_timesteps = [] + for i in range(1, len(timesteps)): + if timesteps[i] >= timesteps[i - 1]: + raise ValueError("`custom_timesteps` must be in descending order.") + + if timesteps[i] not in train_timesteps: + non_train_timesteps.append(timesteps[i]) + + if timesteps[0] >= self.config.num_train_timesteps: + raise ValueError( + f"`timesteps` must start before `self.config.train_timesteps`: {self.config.num_train_timesteps}." + ) + + # Raise warning if timestep schedule does not start with self.config.num_train_timesteps - 1 + if strength == 1.0 and timesteps[0] != self.config.num_train_timesteps - 1: + logger.warning( + f"The first timestep on the custom timestep schedule is {timesteps[0]}, not" + f" `self.config.num_train_timesteps - 1`: {self.config.num_train_timesteps - 1}. You may get" + f" unexpected results when using this timestep schedule." + ) + + # Raise warning if custom timestep schedule contains timesteps not on original timestep schedule + if non_train_timesteps: + logger.warning( + f"The custom timestep schedule contains the following timesteps which are not on the original" + f" training/distillation timestep schedule: {non_train_timesteps}. You may get unexpected results" + f" when using this timestep schedule." + ) + + # Raise warning if custom timestep schedule is longer than original_steps + if original_steps is not None: + if len(timesteps) > original_steps: + logger.warning( + f"The number of timesteps in the custom timestep schedule is {len(timesteps)}, which exceeds the" + f" the length of the timestep schedule used for training: {original_steps}. You may get some" + f" unexpected results when using this timestep schedule." + ) + else: + if len(timesteps) > self.config.num_train_timesteps: + logger.warning( + f"The number of timesteps in the custom timestep schedule is {len(timesteps)}, which exceeds the" + f" the length of the timestep schedule used for training: {self.config.num_train_timesteps}. You may get some" + f" unexpected results when using this timestep schedule." + ) + + timesteps = np.array(timesteps, dtype=np.int64) + self.num_inference_steps = len(timesteps) + self.custom_timesteps = True + + # Apply strength (e.g. for img2img pipelines) (see StableDiffusionImg2ImgPipeline.get_timesteps) + init_timestep = min(int(self.num_inference_steps * strength), self.num_inference_steps) + t_start = max(self.num_inference_steps - init_timestep, 0) + timesteps = timesteps[t_start * self.order :] + # TODO: also reset self.num_inference_steps? + else: + # 2.2 Create the "standard" TCD inference timestep schedule. + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + if original_steps is not None: + skipping_step = len(tcd_origin_timesteps) // num_inference_steps + + if skipping_step < 1: + raise ValueError( + f"The combination of `original_steps x strength`: {original_steps} x {strength} is smaller than `num_inference_steps`: {num_inference_steps}. Make sure to either reduce `num_inference_steps` to a value smaller than {int(original_steps * strength)} or increase `strength` to a value higher than {float(num_inference_steps / original_steps)}." + ) + + self.num_inference_steps = num_inference_steps + + if original_steps is not None: + if num_inference_steps > original_steps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `original_inference_steps`:" + f" {original_steps} because the final timestep schedule will be a subset of the" + f" `original_inference_steps`-sized initial timestep schedule." + ) + else: + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `num_train_timesteps`:" + f" {self.config.num_train_timesteps} because the final timestep schedule will be a subset of the" + f" `num_train_timesteps`-sized initial timestep schedule." + ) + + # TCD Inference Steps Schedule + tcd_origin_timesteps = tcd_origin_timesteps[::-1].copy() + # Select (approximately) evenly spaced indices from tcd_origin_timesteps. + inference_indices = np.linspace(0, len(tcd_origin_timesteps), num=num_inference_steps, endpoint=False) + inference_indices = np.floor(inference_indices).astype(np.int64) + timesteps = tcd_origin_timesteps[inference_indices] + + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.long) + + self._step_index = None + self._begin_index = None + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + eta: float = 0.3, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[TCDSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + eta (`float`): + A stochastic parameter (referred to as `gamma` in the paper) used to control the stochasticity in every + step. When eta = 0, it represents deterministic sampling, whereas eta = 1 indicates full stochastic + sampling. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_tcd.TCDSchedulerOutput`] or `tuple`. + Returns: + [`~schedulers.scheduling_utils.TCDSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_tcd.TCDSchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + assert 0 <= eta <= 1.0, "gamma must be less than or equal to 1.0" + + # 1. get previous step value + prev_step_index = self.step_index + 1 + if prev_step_index < len(self.timesteps): + prev_timestep = self.timesteps[prev_step_index] + else: + prev_timestep = torch.tensor(0) + + timestep_s = torch.floor((1 - eta) * prev_timestep).to(dtype=torch.long) + + # 2. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[timestep] + beta_prod_t = 1 - alpha_prod_t + + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + + alpha_prod_s = self.alphas_cumprod[timestep_s] + beta_prod_s = 1 - alpha_prod_s + + # 3. Compute the predicted noised sample x_s based on the model parameterization + if self.config.prediction_type == "epsilon": # noise-prediction + pred_original_sample = (sample - beta_prod_t.sqrt() * model_output) / alpha_prod_t.sqrt() + pred_epsilon = model_output + pred_noised_sample = alpha_prod_s.sqrt() * pred_original_sample + beta_prod_s.sqrt() * pred_epsilon + elif self.config.prediction_type == "sample": # x-prediction + pred_original_sample = model_output + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + pred_noised_sample = alpha_prod_s.sqrt() * pred_original_sample + beta_prod_s.sqrt() * pred_epsilon + elif self.config.prediction_type == "v_prediction": # v-prediction + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + pred_noised_sample = alpha_prod_s.sqrt() * pred_original_sample + beta_prod_s.sqrt() * pred_epsilon + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or" + " `v_prediction` for `TCDScheduler`." + ) + + # 4. Sample and inject noise z ~ N(0, I) for MultiStep Inference + # Noise is not used on the final timestep of the timestep schedule. + # This also means that noise is not used for one-step sampling. + # Eta (referred to as "gamma" in the paper) was introduced to control the stochasticity in every step. + # When eta = 0, it represents deterministic sampling, whereas eta = 1 indicates full stochastic sampling. + if eta > 0: + if self.step_index != self.num_inference_steps - 1: + noise = randn_tensor( + model_output.shape, generator=generator, device=model_output.device, dtype=pred_noised_sample.dtype + ) + prev_sample = (alpha_prod_t_prev / alpha_prod_s).sqrt() * pred_noised_sample + ( + 1 - alpha_prod_t_prev / alpha_prod_s + ).sqrt() * noise + else: + prev_sample = pred_noised_sample + else: + prev_sample = pred_noised_sample + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample, pred_noised_sample) + + return TCDSchedulerOutput(prev_sample=prev_sample, pred_noised_sample=pred_noised_sample) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self): + return self.config.num_train_timesteps + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.previous_timestep + def previous_timestep(self, timestep): + if self.custom_timesteps or self.num_inference_steps: + index = (self.timesteps == timestep).nonzero(as_tuple=True)[0][0] + if index == self.timesteps.shape[0] - 1: + prev_t = torch.tensor(-1) + else: + prev_t = self.timesteps[index + 1] + else: + prev_t = timestep - 1 + return prev_t diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_unclip.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_unclip.py new file mode 100644 index 0000000000000000000000000000000000000000..d78efabfbc57ba215b1b4d3d0348e3db85543e08 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_unclip.py @@ -0,0 +1,355 @@ +# Copyright 2025 Kakao Brain and The HuggingFace Team. All rights reserved. +# +# 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 dataclasses import dataclass +from typing import Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->UnCLIP +class UnCLIPSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: Optional[torch.Tensor] = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class UnCLIPScheduler(SchedulerMixin, ConfigMixin): + """ + NOTE: do not use this scheduler. The DDPM scheduler has been updated to support the changes made here. This + scheduler will be removed and replaced with DDPM. + + This is a modified DDPM Scheduler specifically for the karlo unCLIP model. + + This scheduler has some minor variations in how it calculates the learned range variance and dynamically + re-calculates betas based off the timesteps it is skipping. + + The scheduler also uses a slightly different step ratio when computing timesteps to use for inference. + + See [`~DDPMScheduler`] for more information on DDPM scheduling + + Args: + num_train_timesteps (`int`): number of diffusion steps used to train the model. + variance_type (`str`): + options to clip the variance used when adding noise to the denoised sample. Choose from `fixed_small_log` + or `learned_range`. + clip_sample (`bool`, default `True`): + option to clip predicted sample between `-clip_sample_range` and `clip_sample_range` for numerical + stability. + clip_sample_range (`float`, default `1.0`): + The range to clip the sample between. See `clip_sample`. + prediction_type (`str`, default `epsilon`, optional): + prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion process) + or `sample` (directly predicting the noisy sample`) + """ + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + variance_type: str = "fixed_small_log", + clip_sample: bool = True, + clip_sample_range: Optional[float] = 1.0, + prediction_type: str = "epsilon", + beta_schedule: str = "squaredcos_cap_v2", + ): + if beta_schedule != "squaredcos_cap_v2": + raise ValueError("UnCLIPScheduler only supports `beta_schedule`: 'squaredcos_cap_v2'") + + self.betas = betas_for_alpha_bar(num_train_timesteps) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + self.one = torch.tensor(1.0) + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy()) + + self.variance_type = variance_type + + def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): input sample + timestep (`int`, optional): current timestep + + Returns: + `torch.Tensor`: scaled input sample + """ + return sample + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference. + + Note that this scheduler uses a slightly different step ratio than the other diffusers schedulers. The + different step ratio is to mimic the original karlo implementation and does not affect the quality or accuracy + of the results. + + Args: + num_inference_steps (`int`): + the number of diffusion steps used when generating samples with a pre-trained model. + """ + self.num_inference_steps = num_inference_steps + step_ratio = (self.config.num_train_timesteps - 1) / (self.num_inference_steps - 1) + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + self.timesteps = torch.from_numpy(timesteps).to(device) + + def _get_variance(self, t, prev_timestep=None, predicted_variance=None, variance_type=None): + if prev_timestep is None: + prev_timestep = t - 1 + + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.one + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + if prev_timestep == t - 1: + beta = self.betas[t] + else: + beta = 1 - alpha_prod_t / alpha_prod_t_prev + + # For t > 0, compute predicted variance βt (see formula (6) and (7) from https://huggingface.co/papers/2006.11239) + # and sample from it to get previous sample + # x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample + variance = beta_prod_t_prev / beta_prod_t * beta + + if variance_type is None: + variance_type = self.config.variance_type + + # hacks - were probably added for training stability + if variance_type == "fixed_small_log": + variance = torch.log(torch.clamp(variance, min=1e-20)) + variance = torch.exp(0.5 * variance) + elif variance_type == "learned_range": + # NOTE difference with DDPM scheduler + min_log = variance.log() + max_log = beta.log() + + frac = (predicted_variance + 1) / 2 + variance = frac * max_log + (1 - frac) * min_log + + return variance + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + prev_timestep: Optional[int] = None, + generator=None, + return_dict: bool = True, + ) -> Union[UnCLIPSchedulerOutput, Tuple]: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + current instance of sample being created by diffusion process. + prev_timestep (`int`, *optional*): The previous timestep to predict the previous sample at. + Used to dynamically compute beta. If not given, `t-1` is used and the pre-computed beta is used. + generator: random number generator. + return_dict (`bool`): option for returning tuple rather than UnCLIPSchedulerOutput class + + Returns: + [`~schedulers.scheduling_utils.UnCLIPSchedulerOutput`] or `tuple`: + [`~schedulers.scheduling_utils.UnCLIPSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is the sample tensor. + + """ + t = timestep + + if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type == "learned_range": + model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1) + else: + predicted_variance = None + + # 1. compute alphas, betas + if prev_timestep is None: + prev_timestep = t - 1 + + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.one + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + if prev_timestep == t - 1: + beta = self.betas[t] + alpha = self.alphas[t] + else: + beta = 1 - alpha_prod_t / alpha_prod_t_prev + alpha = 1 - beta + + # 2. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (15) from https://huggingface.co/papers/2006.11239 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon` or `sample`" + " for the UnCLIPScheduler." + ) + + # 3. Clip "predicted x_0" + if self.config.clip_sample: + pred_original_sample = torch.clamp( + pred_original_sample, -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * beta) / beta_prod_t + current_sample_coeff = alpha ** (0.5) * beta_prod_t_prev / beta_prod_t + + # 5. Compute predicted previous sample µ_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample + + # 6. Add noise + variance = 0 + if t > 0: + variance_noise = randn_tensor( + model_output.shape, dtype=model_output.dtype, generator=generator, device=model_output.device + ) + + variance = self._get_variance( + t, + predicted_variance=predicted_variance, + prev_timestep=prev_timestep, + ) + + if self.variance_type == "fixed_small_log": + variance = variance + elif self.variance_type == "learned_range": + variance = (0.5 * variance).exp() + else: + raise ValueError( + f"variance_type given as {self.variance_type} must be one of `fixed_small_log` or `learned_range`" + " for the UnCLIPScheduler." + ) + + variance = variance * variance_noise + + pred_prev_sample = pred_prev_sample + variance + + if not return_dict: + return ( + pred_prev_sample, + pred_original_sample, + ) + + return UnCLIPSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_unipc_multistep.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_unipc_multistep.py new file mode 100644 index 0000000000000000000000000000000000000000..8b1f699b101a00c50787f1f0ac04d5f062a3b922 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_unipc_multistep.py @@ -0,0 +1,1076 @@ +# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved. +# +# 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. + +# DISCLAIMER: check https://huggingface.co/papers/2302.04867 and https://github.com/wl-zhao/UniPC for more info +# The codebase is modified based on https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_dpmsolver_multistep.py + +import math +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import deprecate, is_scipy_available +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +if is_scipy_available(): + import scipy.stats + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas): + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + + Args: + betas (`torch.Tensor`): + the betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin): + """ + `UniPCMultistepScheduler` is a training-free framework designed for the fast sampling of diffusion models. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + solver_order (`int`, default `2`): + The UniPC order which can be any positive integer. The effective order of accuracy is `solver_order + 1` + due to the UniC. It is recommended to use `solver_order=2` for guided sampling, and `solver_order=3` for + unconditional sampling. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://imagen.research.google/video/paper.pdf) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True` and `predict_x0=True`. + predict_x0 (`bool`, defaults to `True`): + Whether to use the updating algorithm on the predicted x0. + solver_type (`str`, default `bh2`): + Solver type for UniPC. It is recommended to use `bh1` for unconditional sampling when steps < 10, and `bh2` + otherwise. + lower_order_final (`bool`, default `True`): + Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can + stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10. + disable_corrector (`list`, default `[]`): + Decides which step to disable the corrector to mitigate the misalignment between `epsilon_theta(x_t, c)` + and `epsilon_theta(x_t^c, c)` which can influence convergence for a large guidance scale. Corrector is + usually disabled during the first few steps. + solver_p (`SchedulerMixin`, default `None`): + Any other scheduler that if specified, the algorithm becomes `solver_p + UniC`. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + final_sigmas_type (`str`, defaults to `"zero"`): + The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final + sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + solver_order: int = 2, + prediction_type: str = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + predict_x0: bool = True, + solver_type: str = "bh2", + lower_order_final: bool = True, + disable_corrector: List[int] = [], + solver_p: SchedulerMixin = None, + use_karras_sigmas: Optional[bool] = False, + use_exponential_sigmas: Optional[bool] = False, + use_beta_sigmas: Optional[bool] = False, + use_flow_sigmas: Optional[bool] = False, + flow_shift: Optional[float] = 1.0, + timestep_spacing: str = "linspace", + steps_offset: int = 0, + final_sigmas_type: Optional[str] = "zero", # "zero", "sigma_min" + rescale_betas_zero_snr: bool = False, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + if rescale_betas_zero_snr: + # Close to 0 without being 0 so first sigma is not inf + # FP16 smallest positive subnormal works well here + self.alphas_cumprod[-1] = 2**-24 + + # Currently we only support VP-type noise schedule + self.alpha_t = torch.sqrt(self.alphas_cumprod) + self.sigma_t = torch.sqrt(1 - self.alphas_cumprod) + self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t) + self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5 + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + if solver_type not in ["bh1", "bh2"]: + if solver_type in ["midpoint", "heun", "logrho"]: + self.register_to_config(solver_type="bh2") + else: + raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}") + + self.predict_x0 = predict_x0 + # setable values + self.num_inference_steps = None + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps) + self.model_outputs = [None] * solver_order + self.timestep_list = [None] * solver_order + self.lower_order_nums = 0 + self.disable_corrector = disable_corrector + self.solver_p = solver_p + self.last_sample = None + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps + 1) + .round()[::-1][:-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // (num_inference_steps + 1) + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.arange(self.config.num_train_timesteps, 0, -step_ratio).round().copy().astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + if self.config.use_karras_sigmas: + log_sigmas = np.log(sigmas) + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round() + if self.config.final_sigmas_type == "sigma_min": + sigma_last = sigmas[-1] + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + elif self.config.use_exponential_sigmas: + log_sigmas = np.log(sigmas) + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + if self.config.final_sigmas_type == "sigma_min": + sigma_last = sigmas[-1] + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + elif self.config.use_beta_sigmas: + log_sigmas = np.log(sigmas) + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + if self.config.final_sigmas_type == "sigma_min": + sigma_last = sigmas[-1] + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + elif self.config.use_flow_sigmas: + alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1) + sigmas = 1.0 - alphas + sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy() + timesteps = (sigmas * self.config.num_train_timesteps).copy() + if self.config.final_sigmas_type == "sigma_min": + sigma_last = sigmas[-1] + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + else: + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + if self.config.final_sigmas_type == "sigma_min": + sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5 + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + + self.sigmas = torch.from_numpy(sigmas) + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64) + + self.num_inference_steps = len(timesteps) + + self.model_outputs = [ + None, + ] * self.config.solver_order + self.lower_order_nums = 0 + self.last_sample = None + if self.solver_p: + self.solver_p.set_timesteps(self.num_inference_steps, device=device) + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t + def _sigma_to_alpha_sigma_t(self, sigma): + if self.config.use_flow_sigmas: + alpha_t = 1 - sigma + sigma_t = sigma + else: + alpha_t = 1 / ((sigma**2 + 1) ** 0.5) + sigma_t = sigma * alpha_t + + return alpha_t, sigma_t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor: + """Constructs the noise schedule of Karras et al. (2022).""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """Constructs an exponential noise schedule.""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)""" + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def convert_model_output( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + r""" + Convert the model output to the corresponding type the UniPC algorithm needs. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + + if self.predict_x0: + if self.config.prediction_type == "epsilon": + x0_pred = (sample - sigma_t * model_output) / alpha_t + elif self.config.prediction_type == "sample": + x0_pred = model_output + elif self.config.prediction_type == "v_prediction": + x0_pred = alpha_t * sample - sigma_t * model_output + elif self.config.prediction_type == "flow_prediction": + sigma_t = self.sigmas[self.step_index] + x0_pred = sample - sigma_t * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + "`v_prediction`, or `flow_prediction` for the UniPCMultistepScheduler." + ) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + return x0_pred + else: + if self.config.prediction_type == "epsilon": + return model_output + elif self.config.prediction_type == "sample": + epsilon = (sample - alpha_t * model_output) / sigma_t + return epsilon + elif self.config.prediction_type == "v_prediction": + epsilon = alpha_t * model_output + sigma_t * sample + return epsilon + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction` for the UniPCMultistepScheduler." + ) + + def multistep_uni_p_bh_update( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + order: int = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the UniP (B(h) version). Alternatively, `self.solver_p` is used if is specified. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model at the current timestep. + prev_timestep (`int`): + The previous discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + order (`int`): + The order of UniP at this timestep (corresponds to the *p* in UniPC-p). + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + prev_timestep = args[0] if len(args) > 0 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if order is None: + if len(args) > 2: + order = args[2] + else: + raise ValueError("missing `order` as a required keyword argument") + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + model_output_list = self.model_outputs + + s0 = self.timestep_list[-1] + m0 = model_output_list[-1] + x = sample + + if self.solver_p: + x_t = self.solver_p.step(model_output, s0, x).prev_sample + return x_t + + sigma_t, sigma_s0 = self.sigmas[self.step_index + 1], self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + + h = lambda_t - lambda_s0 + device = sample.device + + rks = [] + D1s = [] + for i in range(1, order): + si = self.step_index - i + mi = model_output_list[-(i + 1)] + alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si]) + lambda_si = torch.log(alpha_si) - torch.log(sigma_si) + rk = (lambda_si - lambda_s0) / h + rks.append(rk) + D1s.append((mi - m0) / rk) + + rks.append(1.0) + rks = torch.tensor(rks, device=device) + + R = [] + b = [] + + hh = -h if self.predict_x0 else h + h_phi_1 = torch.expm1(hh) # h\phi_1(h) = e^h - 1 + h_phi_k = h_phi_1 / hh - 1 + + factorial_i = 1 + + if self.config.solver_type == "bh1": + B_h = hh + elif self.config.solver_type == "bh2": + B_h = torch.expm1(hh) + else: + raise NotImplementedError() + + for i in range(1, order + 1): + R.append(torch.pow(rks, i - 1)) + b.append(h_phi_k * factorial_i / B_h) + factorial_i *= i + 1 + h_phi_k = h_phi_k / hh - 1 / factorial_i + + R = torch.stack(R) + b = torch.tensor(b, device=device) + + if len(D1s) > 0: + D1s = torch.stack(D1s, dim=1) # (B, K) + # for order 2, we use a simplified version + if order == 2: + rhos_p = torch.tensor([0.5], dtype=x.dtype, device=device) + else: + rhos_p = torch.linalg.solve(R[:-1, :-1], b[:-1]).to(device).to(x.dtype) + else: + D1s = None + + if self.predict_x0: + x_t_ = sigma_t / sigma_s0 * x - alpha_t * h_phi_1 * m0 + if D1s is not None: + pred_res = torch.einsum("k,bkc...->bc...", rhos_p, D1s) + else: + pred_res = 0 + x_t = x_t_ - alpha_t * B_h * pred_res + else: + x_t_ = alpha_t / alpha_s0 * x - sigma_t * h_phi_1 * m0 + if D1s is not None: + pred_res = torch.einsum("k,bkc...->bc...", rhos_p, D1s) + else: + pred_res = 0 + x_t = x_t_ - sigma_t * B_h * pred_res + + x_t = x_t.to(x.dtype) + return x_t + + def multistep_uni_c_bh_update( + self, + this_model_output: torch.Tensor, + *args, + last_sample: torch.Tensor = None, + this_sample: torch.Tensor = None, + order: int = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the UniC (B(h) version). + + Args: + this_model_output (`torch.Tensor`): + The model outputs at `x_t`. + this_timestep (`int`): + The current timestep `t`. + last_sample (`torch.Tensor`): + The generated sample before the last predictor `x_{t-1}`. + this_sample (`torch.Tensor`): + The generated sample after the last predictor `x_{t}`. + order (`int`): + The `p` of UniC-p at this step. The effective order of accuracy should be `order + 1`. + + Returns: + `torch.Tensor`: + The corrected sample tensor at the current timestep. + """ + this_timestep = args[0] if len(args) > 0 else kwargs.pop("this_timestep", None) + if last_sample is None: + if len(args) > 1: + last_sample = args[1] + else: + raise ValueError("missing `last_sample` as a required keyword argument") + if this_sample is None: + if len(args) > 2: + this_sample = args[2] + else: + raise ValueError("missing `this_sample` as a required keyword argument") + if order is None: + if len(args) > 3: + order = args[3] + else: + raise ValueError("missing `order` as a required keyword argument") + if this_timestep is not None: + deprecate( + "this_timestep", + "1.0.0", + "Passing `this_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + model_output_list = self.model_outputs + + m0 = model_output_list[-1] + x = last_sample + x_t = this_sample + model_t = this_model_output + + sigma_t, sigma_s0 = self.sigmas[self.step_index], self.sigmas[self.step_index - 1] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + + h = lambda_t - lambda_s0 + device = this_sample.device + + rks = [] + D1s = [] + for i in range(1, order): + si = self.step_index - (i + 1) + mi = model_output_list[-(i + 1)] + alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si]) + lambda_si = torch.log(alpha_si) - torch.log(sigma_si) + rk = (lambda_si - lambda_s0) / h + rks.append(rk) + D1s.append((mi - m0) / rk) + + rks.append(1.0) + rks = torch.tensor(rks, device=device) + + R = [] + b = [] + + hh = -h if self.predict_x0 else h + h_phi_1 = torch.expm1(hh) # h\phi_1(h) = e^h - 1 + h_phi_k = h_phi_1 / hh - 1 + + factorial_i = 1 + + if self.config.solver_type == "bh1": + B_h = hh + elif self.config.solver_type == "bh2": + B_h = torch.expm1(hh) + else: + raise NotImplementedError() + + for i in range(1, order + 1): + R.append(torch.pow(rks, i - 1)) + b.append(h_phi_k * factorial_i / B_h) + factorial_i *= i + 1 + h_phi_k = h_phi_k / hh - 1 / factorial_i + + R = torch.stack(R) + b = torch.tensor(b, device=device) + + if len(D1s) > 0: + D1s = torch.stack(D1s, dim=1) + else: + D1s = None + + # for order 1, we use a simplified version + if order == 1: + rhos_c = torch.tensor([0.5], dtype=x.dtype, device=device) + else: + rhos_c = torch.linalg.solve(R, b).to(device).to(x.dtype) + + if self.predict_x0: + x_t_ = sigma_t / sigma_s0 * x - alpha_t * h_phi_1 * m0 + if D1s is not None: + corr_res = torch.einsum("k,bkc...->bc...", rhos_c[:-1], D1s) + else: + corr_res = 0 + D1_t = model_t - m0 + x_t = x_t_ - alpha_t * B_h * (corr_res + rhos_c[-1] * D1_t) + else: + x_t_ = alpha_t / alpha_s0 * x - sigma_t * h_phi_1 * m0 + if D1s is not None: + corr_res = torch.einsum("k,bkc...->bc...", rhos_c[:-1], D1s) + else: + corr_res = 0 + D1_t = model_t - m0 + x_t = x_t_ - sigma_t * B_h * (corr_res + rhos_c[-1] * D1_t) + x_t = x_t.to(x.dtype) + return x_t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + index_candidates = (schedule_timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + return step_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index + def _init_step_index(self, timestep): + """ + Initialize the step_index counter for the scheduler. + """ + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: Union[int, torch.Tensor], + sample: torch.Tensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the multistep UniPC. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + use_corrector = ( + self.step_index > 0 and self.step_index - 1 not in self.disable_corrector and self.last_sample is not None + ) + + model_output_convert = self.convert_model_output(model_output, sample=sample) + if use_corrector: + sample = self.multistep_uni_c_bh_update( + this_model_output=model_output_convert, + last_sample=self.last_sample, + this_sample=sample, + order=self.this_order, + ) + + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.timestep_list[i] = self.timestep_list[i + 1] + + self.model_outputs[-1] = model_output_convert + self.timestep_list[-1] = timestep + + if self.config.lower_order_final: + this_order = min(self.config.solver_order, len(self.timesteps) - self.step_index) + else: + this_order = self.config.solver_order + + self.this_order = min(this_order, self.lower_order_nums + 1) # warmup for multistep + assert self.this_order > 0 + + self.last_sample = sample + prev_sample = self.multistep_uni_p_bh_update( + model_output=model_output, # pass the original non-converted model output, in case solver-p is used + sample=sample, + order=self.this_order, + ) + + if self.lower_order_nums < self.config.solver_order: + self.lower_order_nums += 1 + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # begin_index is None when the scheduler is used for training or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + noisy_samples = alpha_t * original_samples + sigma_t * noise + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_utils.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..61d3e5a22fb07f41faae3afad47269d409aa0fc2 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_utils.py @@ -0,0 +1,194 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 importlib +import os +from dataclasses import dataclass +from enum import Enum +from typing import Optional, Union + +import torch +from huggingface_hub.utils import validate_hf_hub_args +from typing_extensions import Self + +from ..utils import BaseOutput, PushToHubMixin + + +SCHEDULER_CONFIG_NAME = "scheduler_config.json" + + +# NOTE: We make this type an enum because it simplifies usage in docs and prevents +# circular imports when used for `_compatibles` within the schedulers module. +# When it's used as a type in pipelines, it really is a Union because the actual +# scheduler instance is passed in. +class KarrasDiffusionSchedulers(Enum): + DDIMScheduler = 1 + DDPMScheduler = 2 + PNDMScheduler = 3 + LMSDiscreteScheduler = 4 + EulerDiscreteScheduler = 5 + HeunDiscreteScheduler = 6 + EulerAncestralDiscreteScheduler = 7 + DPMSolverMultistepScheduler = 8 + DPMSolverSinglestepScheduler = 9 + KDPM2DiscreteScheduler = 10 + KDPM2AncestralDiscreteScheduler = 11 + DEISMultistepScheduler = 12 + UniPCMultistepScheduler = 13 + DPMSolverSDEScheduler = 14 + EDMEulerScheduler = 15 + + +AysSchedules = { + "StableDiffusionTimesteps": [999, 850, 736, 645, 545, 455, 343, 233, 124, 24], + "StableDiffusionSigmas": [14.615, 6.475, 3.861, 2.697, 1.886, 1.396, 0.963, 0.652, 0.399, 0.152, 0.0], + "StableDiffusionXLTimesteps": [999, 845, 730, 587, 443, 310, 193, 116, 53, 13], + "StableDiffusionXLSigmas": [14.615, 6.315, 3.771, 2.181, 1.342, 0.862, 0.555, 0.380, 0.234, 0.113, 0.0], + "StableDiffusionVideoSigmas": [700.00, 54.5, 15.886, 7.977, 4.248, 1.789, 0.981, 0.403, 0.173, 0.034, 0.0], +} + + +@dataclass +class SchedulerOutput(BaseOutput): + """ + Base class for the output of a scheduler's `step` function. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.Tensor + + +class SchedulerMixin(PushToHubMixin): + """ + Base class for all schedulers. + + [`SchedulerMixin`] contains common functions shared by all schedulers such as general loading and saving + functionalities. + + [`ConfigMixin`] takes care of storing the configuration attributes (like `num_train_timesteps`) that are passed to + the scheduler's `__init__` function, and the attributes can be accessed by `scheduler.config.num_train_timesteps`. + + Class attributes: + - **_compatibles** (`List[str]`) -- A list of scheduler classes that are compatible with the parent scheduler + class. Use [`~ConfigMixin.from_config`] to load a different compatible scheduler class (should be overridden + by parent class). + """ + + config_name = SCHEDULER_CONFIG_NAME + _compatibles = [] + has_compatibles = True + + @classmethod + @validate_hf_hub_args + def from_pretrained( + cls, + pretrained_model_name_or_path: Optional[Union[str, os.PathLike]] = None, + subfolder: Optional[str] = None, + return_unused_kwargs=False, + **kwargs, + ) -> Self: + r""" + Instantiate a scheduler from a pre-defined JSON configuration file in a local directory or Hub repository. + + Parameters: + pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*): + Can be either: + + - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on + the Hub. + - A path to a *directory* (for example `./my_model_directory`) containing the scheduler + configuration saved with [`~SchedulerMixin.save_pretrained`]. + subfolder (`str`, *optional*): + The subfolder location of a model file within a larger model repository on the Hub or locally. + return_unused_kwargs (`bool`, *optional*, defaults to `False`): + Whether kwargs that are not consumed by the Python class should be returned or not. + cache_dir (`Union[str, os.PathLike]`, *optional*): + Path to a directory where a downloaded pretrained model configuration is cached if the standard cache + is not used. + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. + output_loading_info(`bool`, *optional*, defaults to `False`): + Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages. + local_files_only(`bool`, *optional*, defaults to `False`): + Whether to only load local model weights and configuration files or not. If set to `True`, the model + won't be downloaded from the Hub. + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from + `diffusers-cli login` (stored in `~/.huggingface`) is used. + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier + allowed by Git. + + + + To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in with + `huggingface-cli login`. You can also activate the special + ["offline-mode"](https://huggingface.co/diffusers/installation.html#offline-mode) to use this method in a + firewalled environment. + + + + """ + config, kwargs, commit_hash = cls.load_config( + pretrained_model_name_or_path=pretrained_model_name_or_path, + subfolder=subfolder, + return_unused_kwargs=True, + return_commit_hash=True, + **kwargs, + ) + return cls.from_config(config, return_unused_kwargs=return_unused_kwargs, **kwargs) + + def save_pretrained(self, save_directory: Union[str, os.PathLike], push_to_hub: bool = False, **kwargs): + """ + Save a scheduler configuration object to a directory so that it can be reloaded using the + [`~SchedulerMixin.from_pretrained`] class method. + + Args: + save_directory (`str` or `os.PathLike`): + Directory where the configuration JSON file will be saved (will be created if it does not exist). + push_to_hub (`bool`, *optional*, defaults to `False`): + Whether or not to push your model to the Hugging Face Hub after saving it. You can specify the + repository you want to push to with `repo_id` (will default to the name of `save_directory` in your + namespace). + kwargs (`Dict[str, Any]`, *optional*): + Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method. + """ + self.save_config(save_directory=save_directory, push_to_hub=push_to_hub, **kwargs) + + @property + def compatibles(self): + """ + Returns all schedulers that are compatible with this scheduler + + Returns: + `List[SchedulerMixin]`: List of compatible schedulers + """ + return self._get_compatibles() + + @classmethod + def _get_compatibles(cls): + compatible_classes_str = list(set([cls.__name__] + cls._compatibles)) + diffusers_library = importlib.import_module(__name__.split(".")[0]) + compatible_classes = [ + getattr(diffusers_library, c) for c in compatible_classes_str if hasattr(diffusers_library, c) + ] + return compatible_classes diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_utils_flax.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_utils_flax.py new file mode 100644 index 0000000000000000000000000000000000000000..abcde6c386654a764538ac476e30af2d0e472799 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_utils_flax.py @@ -0,0 +1,291 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 importlib +import math +import os +from dataclasses import dataclass +from enum import Enum +from typing import Optional, Tuple, Union + +import flax +import jax.numpy as jnp +from huggingface_hub.utils import validate_hf_hub_args + +from ..utils import BaseOutput, PushToHubMixin + + +SCHEDULER_CONFIG_NAME = "scheduler_config.json" + + +# NOTE: We make this type an enum because it simplifies usage in docs and prevents +# circular imports when used for `_compatibles` within the schedulers module. +# When it's used as a type in pipelines, it really is a Union because the actual +# scheduler instance is passed in. +class FlaxKarrasDiffusionSchedulers(Enum): + FlaxDDIMScheduler = 1 + FlaxDDPMScheduler = 2 + FlaxPNDMScheduler = 3 + FlaxLMSDiscreteScheduler = 4 + FlaxDPMSolverMultistepScheduler = 5 + FlaxEulerDiscreteScheduler = 6 + + +@dataclass +class FlaxSchedulerOutput(BaseOutput): + """ + Base class for the scheduler's step function output. + + Args: + prev_sample (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: jnp.ndarray + + +class FlaxSchedulerMixin(PushToHubMixin): + """ + Mixin containing common functions for the schedulers. + + Class attributes: + - **_compatibles** (`List[str]`) -- A list of classes that are compatible with the parent class, so that + `from_config` can be used from a class different than the one used to save the config (should be overridden + by parent class). + """ + + config_name = SCHEDULER_CONFIG_NAME + ignore_for_config = ["dtype"] + _compatibles = [] + has_compatibles = True + + @classmethod + @validate_hf_hub_args + def from_pretrained( + cls, + pretrained_model_name_or_path: Optional[Union[str, os.PathLike]] = None, + subfolder: Optional[str] = None, + return_unused_kwargs=False, + **kwargs, + ): + r""" + Instantiate a Scheduler class from a pre-defined JSON-file. + + Parameters: + pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*): + Can be either: + + - A string, the *model id* of a model repo on huggingface.co. Valid model ids should have an + organization name, like `google/ddpm-celebahq-256`. + - A path to a *directory* containing model weights saved using [`~SchedulerMixin.save_pretrained`], + e.g., `./my_model_directory/`. + subfolder (`str`, *optional*): + In case the relevant files are located inside a subfolder of the model repo (either remote in + huggingface.co or downloaded locally), you can specify the folder name here. + return_unused_kwargs (`bool`, *optional*, defaults to `False`): + Whether kwargs that are not consumed by the Python class should be returned or not. + + cache_dir (`Union[str, os.PathLike]`, *optional*): + Path to a directory in which a downloaded pretrained model configuration should be cached if the + standard cache should not be used. + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. + output_loading_info(`bool`, *optional*, defaults to `False`): + Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages. + local_files_only(`bool`, *optional*, defaults to `False`): + Whether or not to only look at local files (i.e., do not try to download the model). + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated + when running `transformers-cli login` (stored in `~/.huggingface`). + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a + git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any + identifier allowed by git. + + + + It is required to be logged in (`huggingface-cli login`) when you want to use private or [gated + models](https://huggingface.co/docs/hub/models-gated#gated-models). + + + + + + Activate the special ["offline-mode"](https://huggingface.co/transformers/installation.html#offline-mode) to + use this method in a firewalled environment. + + + + """ + config, kwargs = cls.load_config( + pretrained_model_name_or_path=pretrained_model_name_or_path, + subfolder=subfolder, + return_unused_kwargs=True, + **kwargs, + ) + scheduler, unused_kwargs = cls.from_config(config, return_unused_kwargs=True, **kwargs) + + if hasattr(scheduler, "create_state") and getattr(scheduler, "has_state", False): + state = scheduler.create_state() + + if return_unused_kwargs: + return scheduler, state, unused_kwargs + + return scheduler, state + + def save_pretrained(self, save_directory: Union[str, os.PathLike], push_to_hub: bool = False, **kwargs): + """ + Save a scheduler configuration object to the directory `save_directory`, so that it can be re-loaded using the + [`~FlaxSchedulerMixin.from_pretrained`] class method. + + Args: + save_directory (`str` or `os.PathLike`): + Directory where the configuration JSON file will be saved (will be created if it does not exist). + push_to_hub (`bool`, *optional*, defaults to `False`): + Whether or not to push your model to the Hugging Face Hub after saving it. You can specify the + repository you want to push to with `repo_id` (will default to the name of `save_directory` in your + namespace). + kwargs (`Dict[str, Any]`, *optional*): + Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method. + """ + self.save_config(save_directory=save_directory, push_to_hub=push_to_hub, **kwargs) + + @property + def compatibles(self): + """ + Returns all schedulers that are compatible with this scheduler + + Returns: + `List[SchedulerMixin]`: List of compatible schedulers + """ + return self._get_compatibles() + + @classmethod + def _get_compatibles(cls): + compatible_classes_str = list(set([cls.__name__] + cls._compatibles)) + diffusers_library = importlib.import_module(__name__.split(".")[0]) + compatible_classes = [ + getattr(diffusers_library, c) for c in compatible_classes_str if hasattr(diffusers_library, c) + ] + return compatible_classes + + +def broadcast_to_shape_from_left(x: jnp.ndarray, shape: Tuple[int]) -> jnp.ndarray: + assert len(shape) >= x.ndim + return jnp.broadcast_to(x.reshape(x.shape + (1,) * (len(shape) - x.ndim)), shape) + + +def betas_for_alpha_bar(num_diffusion_timesteps: int, max_beta=0.999, dtype=jnp.float32) -> jnp.ndarray: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + + Returns: + betas (`jnp.ndarray`): the betas used by the scheduler to step the model outputs + """ + + def alpha_bar(time_step): + return math.cos((time_step + 0.008) / 1.008 * math.pi / 2) ** 2 + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta)) + return jnp.array(betas, dtype=dtype) + + +@flax.struct.dataclass +class CommonSchedulerState: + alphas: jnp.ndarray + betas: jnp.ndarray + alphas_cumprod: jnp.ndarray + + @classmethod + def create(cls, scheduler): + config = scheduler.config + + if config.trained_betas is not None: + betas = jnp.asarray(config.trained_betas, dtype=scheduler.dtype) + elif config.beta_schedule == "linear": + betas = jnp.linspace(config.beta_start, config.beta_end, config.num_train_timesteps, dtype=scheduler.dtype) + elif config.beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + betas = ( + jnp.linspace( + config.beta_start**0.5, config.beta_end**0.5, config.num_train_timesteps, dtype=scheduler.dtype + ) + ** 2 + ) + elif config.beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + betas = betas_for_alpha_bar(config.num_train_timesteps, dtype=scheduler.dtype) + else: + raise NotImplementedError( + f"beta_schedule {config.beta_schedule} is not implemented for scheduler {scheduler.__class__.__name__}" + ) + + alphas = 1.0 - betas + + alphas_cumprod = jnp.cumprod(alphas, axis=0) + + return cls( + alphas=alphas, + betas=betas, + alphas_cumprod=alphas_cumprod, + ) + + +def get_sqrt_alpha_prod( + state: CommonSchedulerState, original_samples: jnp.ndarray, noise: jnp.ndarray, timesteps: jnp.ndarray +): + alphas_cumprod = state.alphas_cumprod + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + sqrt_alpha_prod = broadcast_to_shape_from_left(sqrt_alpha_prod, original_samples.shape) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + sqrt_one_minus_alpha_prod = broadcast_to_shape_from_left(sqrt_one_minus_alpha_prod, original_samples.shape) + + return sqrt_alpha_prod, sqrt_one_minus_alpha_prod + + +def add_noise_common( + state: CommonSchedulerState, original_samples: jnp.ndarray, noise: jnp.ndarray, timesteps: jnp.ndarray +): + sqrt_alpha_prod, sqrt_one_minus_alpha_prod = get_sqrt_alpha_prod(state, original_samples, noise, timesteps) + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + +def get_velocity_common(state: CommonSchedulerState, sample: jnp.ndarray, noise: jnp.ndarray, timesteps: jnp.ndarray): + sqrt_alpha_prod, sqrt_one_minus_alpha_prod = get_sqrt_alpha_prod(state, sample, noise, timesteps) + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity diff --git a/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_vq_diffusion.py b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_vq_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..57306301d023defde8ace84c62bc0f0662ee91f7 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_vq_diffusion.py @@ -0,0 +1,467 @@ +# Copyright 2025 Microsoft and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass +from typing import Optional, Tuple, Union + +import numpy as np +import torch +import torch.nn.functional as F + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from .scheduling_utils import SchedulerMixin + + +@dataclass +class VQDiffusionSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's step function output. + + Args: + prev_sample (`torch.LongTensor` of shape `(batch size, num latent pixels)`): + Computed sample x_{t-1} of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.LongTensor + + +def index_to_log_onehot(x: torch.LongTensor, num_classes: int) -> torch.Tensor: + """ + Convert batch of vector of class indices into batch of log onehot vectors + + Args: + x (`torch.LongTensor` of shape `(batch size, vector length)`): + Batch of class indices + + num_classes (`int`): + number of classes to be used for the onehot vectors + + Returns: + `torch.Tensor` of shape `(batch size, num classes, vector length)`: + Log onehot vectors + """ + x_onehot = F.one_hot(x, num_classes) + x_onehot = x_onehot.permute(0, 2, 1) + log_x = torch.log(x_onehot.float().clamp(min=1e-30)) + return log_x + + +def gumbel_noised(logits: torch.Tensor, generator: Optional[torch.Generator]) -> torch.Tensor: + """ + Apply gumbel noise to `logits` + """ + uniform = torch.rand(logits.shape, device=logits.device, generator=generator) + gumbel_noise = -torch.log(-torch.log(uniform + 1e-30) + 1e-30) + noised = gumbel_noise + logits + return noised + + +def alpha_schedules(num_diffusion_timesteps: int, alpha_cum_start=0.99999, alpha_cum_end=0.000009): + """ + Cumulative and non-cumulative alpha schedules. + + See section 4.1. + """ + att = ( + np.arange(0, num_diffusion_timesteps) / (num_diffusion_timesteps - 1) * (alpha_cum_end - alpha_cum_start) + + alpha_cum_start + ) + att = np.concatenate(([1], att)) + at = att[1:] / att[:-1] + att = np.concatenate((att[1:], [1])) + return at, att + + +def gamma_schedules(num_diffusion_timesteps: int, gamma_cum_start=0.000009, gamma_cum_end=0.99999): + """ + Cumulative and non-cumulative gamma schedules. + + See section 4.1. + """ + ctt = ( + np.arange(0, num_diffusion_timesteps) / (num_diffusion_timesteps - 1) * (gamma_cum_end - gamma_cum_start) + + gamma_cum_start + ) + ctt = np.concatenate(([0], ctt)) + one_minus_ctt = 1 - ctt + one_minus_ct = one_minus_ctt[1:] / one_minus_ctt[:-1] + ct = 1 - one_minus_ct + ctt = np.concatenate((ctt[1:], [0])) + return ct, ctt + + +class VQDiffusionScheduler(SchedulerMixin, ConfigMixin): + """ + A scheduler for vector quantized diffusion. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_vec_classes (`int`): + The number of classes of the vector embeddings of the latent pixels. Includes the class for the masked + latent pixel. + num_train_timesteps (`int`, defaults to 100): + The number of diffusion steps to train the model. + alpha_cum_start (`float`, defaults to 0.99999): + The starting cumulative alpha value. + alpha_cum_end (`float`, defaults to 0.00009): + The ending cumulative alpha value. + gamma_cum_start (`float`, defaults to 0.00009): + The starting cumulative gamma value. + gamma_cum_end (`float`, defaults to 0.99999): + The ending cumulative gamma value. + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_vec_classes: int, + num_train_timesteps: int = 100, + alpha_cum_start: float = 0.99999, + alpha_cum_end: float = 0.000009, + gamma_cum_start: float = 0.000009, + gamma_cum_end: float = 0.99999, + ): + self.num_embed = num_vec_classes + + # By convention, the index for the mask class is the last class index + self.mask_class = self.num_embed - 1 + + at, att = alpha_schedules(num_train_timesteps, alpha_cum_start=alpha_cum_start, alpha_cum_end=alpha_cum_end) + ct, ctt = gamma_schedules(num_train_timesteps, gamma_cum_start=gamma_cum_start, gamma_cum_end=gamma_cum_end) + + num_non_mask_classes = self.num_embed - 1 + bt = (1 - at - ct) / num_non_mask_classes + btt = (1 - att - ctt) / num_non_mask_classes + + at = torch.tensor(at.astype("float64")) + bt = torch.tensor(bt.astype("float64")) + ct = torch.tensor(ct.astype("float64")) + log_at = torch.log(at) + log_bt = torch.log(bt) + log_ct = torch.log(ct) + + att = torch.tensor(att.astype("float64")) + btt = torch.tensor(btt.astype("float64")) + ctt = torch.tensor(ctt.astype("float64")) + log_cumprod_at = torch.log(att) + log_cumprod_bt = torch.log(btt) + log_cumprod_ct = torch.log(ctt) + + self.log_at = log_at.float() + self.log_bt = log_bt.float() + self.log_ct = log_ct.float() + self.log_cumprod_at = log_cumprod_at.float() + self.log_cumprod_bt = log_cumprod_bt.float() + self.log_cumprod_ct = log_cumprod_ct.float() + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy()) + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps and diffusion process parameters (alpha, beta, gamma) should be moved + to. + """ + self.num_inference_steps = num_inference_steps + timesteps = np.arange(0, self.num_inference_steps)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps).to(device) + + self.log_at = self.log_at.to(device) + self.log_bt = self.log_bt.to(device) + self.log_ct = self.log_ct.to(device) + self.log_cumprod_at = self.log_cumprod_at.to(device) + self.log_cumprod_bt = self.log_cumprod_bt.to(device) + self.log_cumprod_ct = self.log_cumprod_ct.to(device) + + def step( + self, + model_output: torch.Tensor, + timestep: torch.long, + sample: torch.LongTensor, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[VQDiffusionSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by the reverse transition distribution. See + [`~VQDiffusionScheduler.q_posterior`] for more details about how the distribution is computer. + + Args: + log_p_x_0: (`torch.Tensor` of shape `(batch size, num classes - 1, num latent pixels)`): + The log probabilities for the predicted classes of the initial latent pixels. Does not include a + prediction for the masked class as the initial unnoised image cannot be masked. + t (`torch.long`): + The timestep that determines which transition matrices are used. + x_t (`torch.LongTensor` of shape `(batch size, num latent pixels)`): + The classes of each latent pixel at time `t`. + generator (`torch.Generator`, or `None`): + A random number generator for the noise applied to `p(x_{t-1} | x_t)` before it is sampled from. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_vq_diffusion.VQDiffusionSchedulerOutput`] or + `tuple`. + + Returns: + [`~schedulers.scheduling_vq_diffusion.VQDiffusionSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_vq_diffusion.VQDiffusionSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + """ + if timestep == 0: + log_p_x_t_min_1 = model_output + else: + log_p_x_t_min_1 = self.q_posterior(model_output, sample, timestep) + + log_p_x_t_min_1 = gumbel_noised(log_p_x_t_min_1, generator) + + x_t_min_1 = log_p_x_t_min_1.argmax(dim=1) + + if not return_dict: + return (x_t_min_1,) + + return VQDiffusionSchedulerOutput(prev_sample=x_t_min_1) + + def q_posterior(self, log_p_x_0, x_t, t): + """ + Calculates the log probabilities for the predicted classes of the image at timestep `t-1`: + + ``` + p(x_{t-1} | x_t) = sum( q(x_t | x_{t-1}) * q(x_{t-1} | x_0) * p(x_0) / q(x_t | x_0) ) + ``` + + Args: + log_p_x_0 (`torch.Tensor` of shape `(batch size, num classes - 1, num latent pixels)`): + The log probabilities for the predicted classes of the initial latent pixels. Does not include a + prediction for the masked class as the initial unnoised image cannot be masked. + x_t (`torch.LongTensor` of shape `(batch size, num latent pixels)`): + The classes of each latent pixel at time `t`. + t (`torch.Long`): + The timestep that determines which transition matrix is used. + + Returns: + `torch.Tensor` of shape `(batch size, num classes, num latent pixels)`: + The log probabilities for the predicted classes of the image at timestep `t-1`. + """ + log_onehot_x_t = index_to_log_onehot(x_t, self.num_embed) + + log_q_x_t_given_x_0 = self.log_Q_t_transitioning_to_known_class( + t=t, x_t=x_t, log_onehot_x_t=log_onehot_x_t, cumulative=True + ) + + log_q_t_given_x_t_min_1 = self.log_Q_t_transitioning_to_known_class( + t=t, x_t=x_t, log_onehot_x_t=log_onehot_x_t, cumulative=False + ) + + # p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) ... p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0) + # . . . + # . . . + # . . . + # p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) ... p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) + q = log_p_x_0 - log_q_x_t_given_x_0 + + # sum_0 = p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) + ... + p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}), ... , + # sum_n = p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0) + ... + p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) + q_log_sum_exp = torch.logsumexp(q, dim=1, keepdim=True) + + # p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_0 ... p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_n + # . . . + # . . . + # . . . + # p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_0 ... p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_n + q = q - q_log_sum_exp + + # (p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_0) * a_cumulative_{t-1} + b_cumulative_{t-1} ... (p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_n) * a_cumulative_{t-1} + b_cumulative_{t-1} + # . . . + # . . . + # . . . + # (p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_0) * a_cumulative_{t-1} + b_cumulative_{t-1} ... (p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_n) * a_cumulative_{t-1} + b_cumulative_{t-1} + # c_cumulative_{t-1} ... c_cumulative_{t-1} + q = self.apply_cumulative_transitions(q, t - 1) + + # ((p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_0) * a_cumulative_{t-1} + b_cumulative_{t-1}) * q(x_t | x_{t-1}=C_0) * sum_0 ... ((p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_n) * a_cumulative_{t-1} + b_cumulative_{t-1}) * q(x_t | x_{t-1}=C_0) * sum_n + # . . . + # . . . + # . . . + # ((p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_0) * a_cumulative_{t-1} + b_cumulative_{t-1}) * q(x_t | x_{t-1}=C_{k-1}) * sum_0 ... ((p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_n) * a_cumulative_{t-1} + b_cumulative_{t-1}) * q(x_t | x_{t-1}=C_{k-1}) * sum_n + # c_cumulative_{t-1} * q(x_t | x_{t-1}=C_k) * sum_0 ... c_cumulative_{t-1} * q(x_t | x_{t-1}=C_k) * sum_0 + log_p_x_t_min_1 = q + log_q_t_given_x_t_min_1 + q_log_sum_exp + + # For each column, there are two possible cases. + # + # Where: + # - sum(p_n(x_0))) is summing over all classes for x_0 + # - C_i is the class transitioning from (not to be confused with c_t and c_cumulative_t being used for gamma's) + # - C_j is the class transitioning to + # + # 1. x_t is masked i.e. x_t = c_k + # + # Simplifying the expression, the column vector is: + # . + # . + # . + # (c_t / c_cumulative_t) * (a_cumulative_{t-1} * p_n(x_0 = C_i | x_t) + b_cumulative_{t-1} * sum(p_n(x_0))) + # . + # . + # . + # (c_cumulative_{t-1} / c_cumulative_t) * sum(p_n(x_0)) + # + # From equation (11) stated in terms of forward probabilities, the last row is trivially verified. + # + # For the other rows, we can state the equation as ... + # + # (c_t / c_cumulative_t) * [b_cumulative_{t-1} * p(x_0=c_0) + ... + (a_cumulative_{t-1} + b_cumulative_{t-1}) * p(x_0=C_i) + ... + b_cumulative_{k-1} * p(x_0=c_{k-1})] + # + # This verifies the other rows. + # + # 2. x_t is not masked + # + # Simplifying the expression, there are two cases for the rows of the column vector, where C_j = C_i and where C_j != C_i: + # . + # . + # . + # C_j != C_i: b_t * ((b_cumulative_{t-1} / b_cumulative_t) * p_n(x_0 = c_0) + ... + ((a_cumulative_{t-1} + b_cumulative_{t-1}) / b_cumulative_t) * p_n(x_0 = C_i) + ... + (b_cumulative_{t-1} / (a_cumulative_t + b_cumulative_t)) * p_n(c_0=C_j) + ... + (b_cumulative_{t-1} / b_cumulative_t) * p_n(x_0 = c_{k-1})) + # . + # . + # . + # C_j = C_i: (a_t + b_t) * ((b_cumulative_{t-1} / b_cumulative_t) * p_n(x_0 = c_0) + ... + ((a_cumulative_{t-1} + b_cumulative_{t-1}) / (a_cumulative_t + b_cumulative_t)) * p_n(x_0 = C_i = C_j) + ... + (b_cumulative_{t-1} / b_cumulative_t) * p_n(x_0 = c_{k-1})) + # . + # . + # . + # 0 + # + # The last row is trivially verified. The other rows can be verified by directly expanding equation (11) stated in terms of forward probabilities. + return log_p_x_t_min_1 + + def log_Q_t_transitioning_to_known_class( + self, *, t: torch.int, x_t: torch.LongTensor, log_onehot_x_t: torch.Tensor, cumulative: bool + ): + """ + Calculates the log probabilities of the rows from the (cumulative or non-cumulative) transition matrix for each + latent pixel in `x_t`. + + Args: + t (`torch.Long`): + The timestep that determines which transition matrix is used. + x_t (`torch.LongTensor` of shape `(batch size, num latent pixels)`): + The classes of each latent pixel at time `t`. + log_onehot_x_t (`torch.Tensor` of shape `(batch size, num classes, num latent pixels)`): + The log one-hot vectors of `x_t`. + cumulative (`bool`): + If cumulative is `False`, the single step transition matrix `t-1`->`t` is used. If cumulative is + `True`, the cumulative transition matrix `0`->`t` is used. + + Returns: + `torch.Tensor` of shape `(batch size, num classes - 1, num latent pixels)`: + Each _column_ of the returned matrix is a _row_ of log probabilities of the complete probability + transition matrix. + + When non cumulative, returns `self.num_classes - 1` rows because the initial latent pixel cannot be + masked. + + Where: + - `q_n` is the probability distribution for the forward process of the `n`th latent pixel. + - C_0 is a class of a latent pixel embedding + - C_k is the class of the masked latent pixel + + non-cumulative result (omitting logarithms): + ``` + q_0(x_t | x_{t-1} = C_0) ... q_n(x_t | x_{t-1} = C_0) + . . . + . . . + . . . + q_0(x_t | x_{t-1} = C_k) ... q_n(x_t | x_{t-1} = C_k) + ``` + + cumulative result (omitting logarithms): + ``` + q_0_cumulative(x_t | x_0 = C_0) ... q_n_cumulative(x_t | x_0 = C_0) + . . . + . . . + . . . + q_0_cumulative(x_t | x_0 = C_{k-1}) ... q_n_cumulative(x_t | x_0 = C_{k-1}) + ``` + """ + if cumulative: + a = self.log_cumprod_at[t] + b = self.log_cumprod_bt[t] + c = self.log_cumprod_ct[t] + else: + a = self.log_at[t] + b = self.log_bt[t] + c = self.log_ct[t] + + if not cumulative: + # The values in the onehot vector can also be used as the logprobs for transitioning + # from masked latent pixels. If we are not calculating the cumulative transitions, + # we need to save these vectors to be re-appended to the final matrix so the values + # aren't overwritten. + # + # `P(x_t!=mask|x_{t-1=mask}) = 0` and 0 will be the value of the last row of the onehot vector + # if x_t is not masked + # + # `P(x_t=mask|x_{t-1=mask}) = 1` and 1 will be the value of the last row of the onehot vector + # if x_t is masked + log_onehot_x_t_transitioning_from_masked = log_onehot_x_t[:, -1, :].unsqueeze(1) + + # `index_to_log_onehot` will add onehot vectors for masked pixels, + # so the default one hot matrix has one too many rows. See the doc string + # for an explanation of the dimensionality of the returned matrix. + log_onehot_x_t = log_onehot_x_t[:, :-1, :] + + # this is a cheeky trick to produce the transition probabilities using log one-hot vectors. + # + # Don't worry about what values this sets in the columns that mark transitions + # to masked latent pixels. They are overwrote later with the `mask_class_mask`. + # + # Looking at the below logspace formula in non-logspace, each value will evaluate to either + # `1 * a + b = a + b` where `log_Q_t` has the one hot value in the column + # or + # `0 * a + b = b` where `log_Q_t` has the 0 values in the column. + # + # See equation 7 for more details. + log_Q_t = (log_onehot_x_t + a).logaddexp(b) + + # The whole column of each masked pixel is `c` + mask_class_mask = x_t == self.mask_class + mask_class_mask = mask_class_mask.unsqueeze(1).expand(-1, self.num_embed - 1, -1) + log_Q_t[mask_class_mask] = c + + if not cumulative: + log_Q_t = torch.cat((log_Q_t, log_onehot_x_t_transitioning_from_masked), dim=1) + + return log_Q_t + + def apply_cumulative_transitions(self, q, t): + bsz = q.shape[0] + a = self.log_cumprod_at[t] + b = self.log_cumprod_bt[t] + c = self.log_cumprod_ct[t] + + num_latent_pixels = q.shape[2] + c = c.expand(bsz, 1, num_latent_pixels) + + q = (q + a).logaddexp(b) + q = torch.cat((q, c), dim=1) + + return q diff --git a/venv/lib/python3.10/site-packages/diffusers/training_utils.py b/venv/lib/python3.10/site-packages/diffusers/training_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..bc30411d87267901a24795d30d86bed2d53da9a0 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/training_utils.py @@ -0,0 +1,652 @@ +import contextlib +import copy +import gc +import math +import random +from typing import Any, Dict, Iterable, List, Optional, Tuple, Union + +import numpy as np +import torch + +from .models import UNet2DConditionModel +from .schedulers import SchedulerMixin +from .utils import ( + convert_state_dict_to_diffusers, + convert_state_dict_to_peft, + deprecate, + is_peft_available, + is_torch_npu_available, + is_torchvision_available, + is_transformers_available, +) + + +if is_transformers_available(): + import transformers + + if transformers.integrations.deepspeed.is_deepspeed_zero3_enabled(): + import deepspeed + +if is_peft_available(): + from peft import set_peft_model_state_dict + +if is_torchvision_available(): + from torchvision import transforms + +if is_torch_npu_available(): + import torch_npu # noqa: F401 + + +def set_seed(seed: int): + """ + Helper function for reproducible behavior to set the seed in `random`, `numpy`, `torch`. + + Args: + seed (`int`): The seed to set. + + Returns: + `None` + """ + random.seed(seed) + np.random.seed(seed) + torch.manual_seed(seed) + if is_torch_npu_available(): + torch.npu.manual_seed_all(seed) + else: + torch.cuda.manual_seed_all(seed) + # ^^ safe to call this function even if cuda is not available + + +def compute_snr(noise_scheduler, timesteps): + """ + Computes SNR as per + https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849 + for the given timesteps using the provided noise scheduler. + + Args: + noise_scheduler (`NoiseScheduler`): + An object containing the noise schedule parameters, specifically `alphas_cumprod`, which is used to compute + the SNR values. + timesteps (`torch.Tensor`): + A tensor of timesteps for which the SNR is computed. + + Returns: + `torch.Tensor`: A tensor containing the computed SNR values for each timestep. + """ + alphas_cumprod = noise_scheduler.alphas_cumprod + sqrt_alphas_cumprod = alphas_cumprod**0.5 + sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5 + + # Expand the tensors. + # Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026 + sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float() + while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape): + sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None] + alpha = sqrt_alphas_cumprod.expand(timesteps.shape) + + sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float() + while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape): + sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None] + sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape) + + # Compute SNR. + snr = (alpha / sigma) ** 2 + return snr + + +def resolve_interpolation_mode(interpolation_type: str): + """ + Maps a string describing an interpolation function to the corresponding torchvision `InterpolationMode` enum. The + full list of supported enums is documented at + https://pytorch.org/vision/0.9/transforms.html#torchvision.transforms.functional.InterpolationMode. + + Args: + interpolation_type (`str`): + A string describing an interpolation method. Currently, `bilinear`, `bicubic`, `box`, `nearest`, + `nearest_exact`, `hamming`, and `lanczos` are supported, corresponding to the supported interpolation modes + in torchvision. + + Returns: + `torchvision.transforms.InterpolationMode`: an `InterpolationMode` enum used by torchvision's `resize` + transform. + """ + if not is_torchvision_available(): + raise ImportError( + "Please make sure to install `torchvision` to be able to use the `resolve_interpolation_mode()` function." + ) + + if interpolation_type == "bilinear": + interpolation_mode = transforms.InterpolationMode.BILINEAR + elif interpolation_type == "bicubic": + interpolation_mode = transforms.InterpolationMode.BICUBIC + elif interpolation_type == "box": + interpolation_mode = transforms.InterpolationMode.BOX + elif interpolation_type == "nearest": + interpolation_mode = transforms.InterpolationMode.NEAREST + elif interpolation_type == "nearest_exact": + interpolation_mode = transforms.InterpolationMode.NEAREST_EXACT + elif interpolation_type == "hamming": + interpolation_mode = transforms.InterpolationMode.HAMMING + elif interpolation_type == "lanczos": + interpolation_mode = transforms.InterpolationMode.LANCZOS + else: + raise ValueError( + f"The given interpolation mode {interpolation_type} is not supported. Currently supported interpolation" + f" modes are `bilinear`, `bicubic`, `box`, `nearest`, `nearest_exact`, `hamming`, and `lanczos`." + ) + + return interpolation_mode + + +def compute_dream_and_update_latents( + unet: UNet2DConditionModel, + noise_scheduler: SchedulerMixin, + timesteps: torch.Tensor, + noise: torch.Tensor, + noisy_latents: torch.Tensor, + target: torch.Tensor, + encoder_hidden_states: torch.Tensor, + dream_detail_preservation: float = 1.0, +) -> Tuple[Optional[torch.Tensor], Optional[torch.Tensor]]: + """ + Implements "DREAM (Diffusion Rectification and Estimation-Adaptive Models)" from + https://huggingface.co/papers/2312.00210. DREAM helps align training with sampling to help training be more + efficient and accurate at the cost of an extra forward step without gradients. + + Args: + `unet`: The state unet to use to make a prediction. + `noise_scheduler`: The noise scheduler used to add noise for the given timestep. + `timesteps`: The timesteps for the noise_scheduler to user. + `noise`: A tensor of noise in the shape of noisy_latents. + `noisy_latents`: Previously noise latents from the training loop. + `target`: The ground-truth tensor to predict after eps is removed. + `encoder_hidden_states`: Text embeddings from the text model. + `dream_detail_preservation`: A float value that indicates detail preservation level. + See reference. + + Returns: + `tuple[torch.Tensor, torch.Tensor]`: Adjusted noisy_latents and target. + """ + alphas_cumprod = noise_scheduler.alphas_cumprod.to(timesteps.device)[timesteps, None, None, None] + sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5 + + # The paper uses lambda = sqrt(1 - alpha) ** p, with p = 1 in their experiments. + dream_lambda = sqrt_one_minus_alphas_cumprod**dream_detail_preservation + + pred = None + with torch.no_grad(): + pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample + + _noisy_latents, _target = (None, None) + if noise_scheduler.config.prediction_type == "epsilon": + predicted_noise = pred + delta_noise = (noise - predicted_noise).detach() + delta_noise.mul_(dream_lambda) + _noisy_latents = noisy_latents.add(sqrt_one_minus_alphas_cumprod * delta_noise) + _target = target.add(delta_noise) + elif noise_scheduler.config.prediction_type == "v_prediction": + raise NotImplementedError("DREAM has not been implemented for v-prediction") + else: + raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}") + + return _noisy_latents, _target + + +def unet_lora_state_dict(unet: UNet2DConditionModel) -> Dict[str, torch.Tensor]: + r""" + Returns: + A state dict containing just the LoRA parameters. + """ + lora_state_dict = {} + + for name, module in unet.named_modules(): + if hasattr(module, "set_lora_layer"): + lora_layer = getattr(module, "lora_layer") + if lora_layer is not None: + current_lora_layer_sd = lora_layer.state_dict() + for lora_layer_matrix_name, lora_param in current_lora_layer_sd.items(): + # The matrix name can either be "down" or "up". + lora_state_dict[f"{name}.lora.{lora_layer_matrix_name}"] = lora_param + + return lora_state_dict + + +def cast_training_params(model: Union[torch.nn.Module, List[torch.nn.Module]], dtype=torch.float32): + """ + Casts the training parameters of the model to the specified data type. + + Args: + model: The PyTorch model whose parameters will be cast. + dtype: The data type to which the model parameters will be cast. + """ + if not isinstance(model, list): + model = [model] + for m in model: + for param in m.parameters(): + # only upcast trainable parameters into fp32 + if param.requires_grad: + param.data = param.to(dtype) + + +def _set_state_dict_into_text_encoder( + lora_state_dict: Dict[str, torch.Tensor], prefix: str, text_encoder: torch.nn.Module +): + """ + Sets the `lora_state_dict` into `text_encoder` coming from `transformers`. + + Args: + lora_state_dict: The state dictionary to be set. + prefix: String identifier to retrieve the portion of the state dict that belongs to `text_encoder`. + text_encoder: Where the `lora_state_dict` is to be set. + """ + + text_encoder_state_dict = { + f"{k.replace(prefix, '')}": v for k, v in lora_state_dict.items() if k.startswith(prefix) + } + text_encoder_state_dict = convert_state_dict_to_peft(convert_state_dict_to_diffusers(text_encoder_state_dict)) + set_peft_model_state_dict(text_encoder, text_encoder_state_dict, adapter_name="default") + + +def _collate_lora_metadata(modules_to_save: Dict[str, torch.nn.Module]) -> Dict[str, Any]: + metadatas = {} + for module_name, module in modules_to_save.items(): + if module is not None: + metadatas[f"{module_name}_lora_adapter_metadata"] = module.peft_config["default"].to_dict() + return metadatas + + +def compute_density_for_timestep_sampling( + weighting_scheme: str, + batch_size: int, + logit_mean: float = None, + logit_std: float = None, + mode_scale: float = None, + device: Union[torch.device, str] = "cpu", + generator: Optional[torch.Generator] = None, +): + """ + Compute the density for sampling the timesteps when doing SD3 training. + + Courtesy: This was contributed by Rafie Walker in https://github.com/huggingface/diffusers/pull/8528. + + SD3 paper reference: https://huggingface.co/papers/2403.03206v1. + """ + if weighting_scheme == "logit_normal": + u = torch.normal(mean=logit_mean, std=logit_std, size=(batch_size,), device=device, generator=generator) + u = torch.nn.functional.sigmoid(u) + elif weighting_scheme == "mode": + u = torch.rand(size=(batch_size,), device=device, generator=generator) + u = 1 - u - mode_scale * (torch.cos(math.pi * u / 2) ** 2 - 1 + u) + else: + u = torch.rand(size=(batch_size,), device=device, generator=generator) + return u + + +def compute_loss_weighting_for_sd3(weighting_scheme: str, sigmas=None): + """ + Computes loss weighting scheme for SD3 training. + + Courtesy: This was contributed by Rafie Walker in https://github.com/huggingface/diffusers/pull/8528. + + SD3 paper reference: https://huggingface.co/papers/2403.03206v1. + """ + if weighting_scheme == "sigma_sqrt": + weighting = (sigmas**-2.0).float() + elif weighting_scheme == "cosmap": + bot = 1 - 2 * sigmas + 2 * sigmas**2 + weighting = 2 / (math.pi * bot) + else: + weighting = torch.ones_like(sigmas) + return weighting + + +def free_memory(): + """ + Runs garbage collection. Then clears the cache of the available accelerator. + """ + gc.collect() + + if torch.cuda.is_available(): + torch.cuda.empty_cache() + elif torch.backends.mps.is_available(): + torch.mps.empty_cache() + elif is_torch_npu_available(): + torch_npu.npu.empty_cache() + elif hasattr(torch, "xpu") and torch.xpu.is_available(): + torch.xpu.empty_cache() + + +# Adapted from torch-ema https://github.com/fadel/pytorch_ema/blob/master/torch_ema/ema.py#L14 +class EMAModel: + """ + Exponential Moving Average of models weights + """ + + def __init__( + self, + parameters: Iterable[torch.nn.Parameter], + decay: float = 0.9999, + min_decay: float = 0.0, + update_after_step: int = 0, + use_ema_warmup: bool = False, + inv_gamma: Union[float, int] = 1.0, + power: Union[float, int] = 2 / 3, + foreach: bool = False, + model_cls: Optional[Any] = None, + model_config: Dict[str, Any] = None, + **kwargs, + ): + """ + Args: + parameters (Iterable[torch.nn.Parameter]): The parameters to track. + decay (float): The decay factor for the exponential moving average. + min_decay (float): The minimum decay factor for the exponential moving average. + update_after_step (int): The number of steps to wait before starting to update the EMA weights. + use_ema_warmup (bool): Whether to use EMA warmup. + inv_gamma (float): + Inverse multiplicative factor of EMA warmup. Default: 1. Only used if `use_ema_warmup` is True. + power (float): Exponential factor of EMA warmup. Default: 2/3. Only used if `use_ema_warmup` is True. + foreach (bool): Use torch._foreach functions for updating shadow parameters. Should be faster. + device (Optional[Union[str, torch.device]]): The device to store the EMA weights on. If None, the EMA + weights will be stored on CPU. + + @crowsonkb's notes on EMA Warmup: + If gamma=1 and power=1, implements a simple average. gamma=1, power=2/3 are good values for models you plan + to train for a million or more steps (reaches decay factor 0.999 at 31.6K steps, 0.9999 at 1M steps), + gamma=1, power=3/4 for models you plan to train for less (reaches decay factor 0.999 at 10K steps, 0.9999 + at 215.4k steps). + """ + + if isinstance(parameters, torch.nn.Module): + deprecation_message = ( + "Passing a `torch.nn.Module` to `ExponentialMovingAverage` is deprecated. " + "Please pass the parameters of the module instead." + ) + deprecate( + "passing a `torch.nn.Module` to `ExponentialMovingAverage`", + "1.0.0", + deprecation_message, + standard_warn=False, + ) + parameters = parameters.parameters() + + # set use_ema_warmup to True if a torch.nn.Module is passed for backwards compatibility + use_ema_warmup = True + + if kwargs.get("max_value", None) is not None: + deprecation_message = "The `max_value` argument is deprecated. Please use `decay` instead." + deprecate("max_value", "1.0.0", deprecation_message, standard_warn=False) + decay = kwargs["max_value"] + + if kwargs.get("min_value", None) is not None: + deprecation_message = "The `min_value` argument is deprecated. Please use `min_decay` instead." + deprecate("min_value", "1.0.0", deprecation_message, standard_warn=False) + min_decay = kwargs["min_value"] + + parameters = list(parameters) + self.shadow_params = [p.clone().detach() for p in parameters] + + if kwargs.get("device", None) is not None: + deprecation_message = "The `device` argument is deprecated. Please use `to` instead." + deprecate("device", "1.0.0", deprecation_message, standard_warn=False) + self.to(device=kwargs["device"]) + + self.temp_stored_params = None + + self.decay = decay + self.min_decay = min_decay + self.update_after_step = update_after_step + self.use_ema_warmup = use_ema_warmup + self.inv_gamma = inv_gamma + self.power = power + self.optimization_step = 0 + self.cur_decay_value = None # set in `step()` + self.foreach = foreach + + self.model_cls = model_cls + self.model_config = model_config + + @classmethod + def from_pretrained(cls, path, model_cls, foreach=False) -> "EMAModel": + _, ema_kwargs = model_cls.from_config(path, return_unused_kwargs=True) + model = model_cls.from_pretrained(path) + + ema_model = cls(model.parameters(), model_cls=model_cls, model_config=model.config, foreach=foreach) + + ema_model.load_state_dict(ema_kwargs) + return ema_model + + def save_pretrained(self, path): + if self.model_cls is None: + raise ValueError("`save_pretrained` can only be used if `model_cls` was defined at __init__.") + + if self.model_config is None: + raise ValueError("`save_pretrained` can only be used if `model_config` was defined at __init__.") + + model = self.model_cls.from_config(self.model_config) + state_dict = self.state_dict() + state_dict.pop("shadow_params", None) + + model.register_to_config(**state_dict) + self.copy_to(model.parameters()) + model.save_pretrained(path) + + def get_decay(self, optimization_step: int) -> float: + """ + Compute the decay factor for the exponential moving average. + """ + step = max(0, optimization_step - self.update_after_step - 1) + + if step <= 0: + return 0.0 + + if self.use_ema_warmup: + cur_decay_value = 1 - (1 + step / self.inv_gamma) ** -self.power + else: + cur_decay_value = (1 + step) / (10 + step) + + cur_decay_value = min(cur_decay_value, self.decay) + # make sure decay is not smaller than min_decay + cur_decay_value = max(cur_decay_value, self.min_decay) + return cur_decay_value + + @torch.no_grad() + def step(self, parameters: Iterable[torch.nn.Parameter]): + if isinstance(parameters, torch.nn.Module): + deprecation_message = ( + "Passing a `torch.nn.Module` to `ExponentialMovingAverage.step` is deprecated. " + "Please pass the parameters of the module instead." + ) + deprecate( + "passing a `torch.nn.Module` to `ExponentialMovingAverage.step`", + "1.0.0", + deprecation_message, + standard_warn=False, + ) + parameters = parameters.parameters() + + parameters = list(parameters) + + self.optimization_step += 1 + + # Compute the decay factor for the exponential moving average. + decay = self.get_decay(self.optimization_step) + self.cur_decay_value = decay + one_minus_decay = 1 - decay + + context_manager = contextlib.nullcontext() + + if self.foreach: + if is_transformers_available() and transformers.integrations.deepspeed.is_deepspeed_zero3_enabled(): + context_manager = deepspeed.zero.GatheredParameters(parameters, modifier_rank=None) + + with context_manager: + params_grad = [param for param in parameters if param.requires_grad] + s_params_grad = [ + s_param for s_param, param in zip(self.shadow_params, parameters) if param.requires_grad + ] + + if len(params_grad) < len(parameters): + torch._foreach_copy_( + [s_param for s_param, param in zip(self.shadow_params, parameters) if not param.requires_grad], + [param for param in parameters if not param.requires_grad], + non_blocking=True, + ) + + torch._foreach_sub_( + s_params_grad, torch._foreach_sub(s_params_grad, params_grad), alpha=one_minus_decay + ) + + else: + for s_param, param in zip(self.shadow_params, parameters): + if is_transformers_available() and transformers.integrations.deepspeed.is_deepspeed_zero3_enabled(): + context_manager = deepspeed.zero.GatheredParameters(param, modifier_rank=None) + + with context_manager: + if param.requires_grad: + s_param.sub_(one_minus_decay * (s_param - param)) + else: + s_param.copy_(param) + + def copy_to(self, parameters: Iterable[torch.nn.Parameter]) -> None: + """ + Copy current averaged parameters into given collection of parameters. + + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + updated with the stored moving averages. If `None`, the parameters with which this + `ExponentialMovingAverage` was initialized will be used. + """ + parameters = list(parameters) + if self.foreach: + torch._foreach_copy_( + [param.data for param in parameters], + [s_param.to(param.device).data for s_param, param in zip(self.shadow_params, parameters)], + ) + else: + for s_param, param in zip(self.shadow_params, parameters): + param.data.copy_(s_param.to(param.device).data) + + def pin_memory(self) -> None: + r""" + Move internal buffers of the ExponentialMovingAverage to pinned memory. Useful for non-blocking transfers for + offloading EMA params to the host. + """ + + self.shadow_params = [p.pin_memory() for p in self.shadow_params] + + def to(self, device=None, dtype=None, non_blocking=False) -> None: + r""" + Move internal buffers of the ExponentialMovingAverage to `device`. + + Args: + device: like `device` argument to `torch.Tensor.to` + """ + # .to() on the tensors handles None correctly + self.shadow_params = [ + p.to(device=device, dtype=dtype, non_blocking=non_blocking) + if p.is_floating_point() + else p.to(device=device, non_blocking=non_blocking) + for p in self.shadow_params + ] + + def state_dict(self) -> dict: + r""" + Returns the state of the ExponentialMovingAverage as a dict. This method is used by accelerate during + checkpointing to save the ema state dict. + """ + # Following PyTorch conventions, references to tensors are returned: + # "returns a reference to the state and not its copy!" - + # https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict + return { + "decay": self.decay, + "min_decay": self.min_decay, + "optimization_step": self.optimization_step, + "update_after_step": self.update_after_step, + "use_ema_warmup": self.use_ema_warmup, + "inv_gamma": self.inv_gamma, + "power": self.power, + "shadow_params": self.shadow_params, + } + + def store(self, parameters: Iterable[torch.nn.Parameter]) -> None: + r""" + Saves the current parameters for restoring later. + + Args: + parameters: Iterable of `torch.nn.Parameter`. The parameters to be temporarily stored. + """ + self.temp_stored_params = [param.detach().cpu().clone() for param in parameters] + + def restore(self, parameters: Iterable[torch.nn.Parameter]) -> None: + r""" + Restore the parameters stored with the `store` method. Useful to validate the model with EMA parameters + without: affecting the original optimization process. Store the parameters before the `copy_to()` method. After + validation (or model saving), use this to restore the former parameters. + + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + updated with the stored parameters. If `None`, the parameters with which this + `ExponentialMovingAverage` was initialized will be used. + """ + + if self.temp_stored_params is None: + raise RuntimeError("This ExponentialMovingAverage has no `store()`ed weights to `restore()`") + if self.foreach: + torch._foreach_copy_( + [param.data for param in parameters], [c_param.data for c_param in self.temp_stored_params] + ) + else: + for c_param, param in zip(self.temp_stored_params, parameters): + param.data.copy_(c_param.data) + + # Better memory-wise. + self.temp_stored_params = None + + def load_state_dict(self, state_dict: dict) -> None: + r""" + Loads the ExponentialMovingAverage state. This method is used by accelerate during checkpointing to save the + ema state dict. + + Args: + state_dict (dict): EMA state. Should be an object returned + from a call to :meth:`state_dict`. + """ + # deepcopy, to be consistent with module API + state_dict = copy.deepcopy(state_dict) + + self.decay = state_dict.get("decay", self.decay) + if self.decay < 0.0 or self.decay > 1.0: + raise ValueError("Decay must be between 0 and 1") + + self.min_decay = state_dict.get("min_decay", self.min_decay) + if not isinstance(self.min_decay, float): + raise ValueError("Invalid min_decay") + + self.optimization_step = state_dict.get("optimization_step", self.optimization_step) + if not isinstance(self.optimization_step, int): + raise ValueError("Invalid optimization_step") + + self.update_after_step = state_dict.get("update_after_step", self.update_after_step) + if not isinstance(self.update_after_step, int): + raise ValueError("Invalid update_after_step") + + self.use_ema_warmup = state_dict.get("use_ema_warmup", self.use_ema_warmup) + if not isinstance(self.use_ema_warmup, bool): + raise ValueError("Invalid use_ema_warmup") + + self.inv_gamma = state_dict.get("inv_gamma", self.inv_gamma) + if not isinstance(self.inv_gamma, (float, int)): + raise ValueError("Invalid inv_gamma") + + self.power = state_dict.get("power", self.power) + if not isinstance(self.power, (float, int)): + raise ValueError("Invalid power") + + shadow_params = state_dict.get("shadow_params", None) + if shadow_params is not None: + self.shadow_params = shadow_params + if not isinstance(self.shadow_params, list): + raise ValueError("shadow_params must be a list") + if not all(isinstance(p, torch.Tensor) for p in self.shadow_params): + raise ValueError("shadow_params must all be Tensors") diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__init__.py b/venv/lib/python3.10/site-packages/diffusers/utils/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..2df05cb8eb36fecaa22d12435008dd9903fecc3f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/__init__.py @@ -0,0 +1,153 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# 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 os + +from packaging import version + +from .. import __version__ +from .constants import ( + CONFIG_NAME, + DEPRECATED_REVISION_ARGS, + DIFFUSERS_DYNAMIC_MODULE_NAME, + FLAX_WEIGHTS_NAME, + GGUF_FILE_EXTENSION, + HF_MODULES_CACHE, + HUGGINGFACE_CO_RESOLVE_ENDPOINT, + MIN_PEFT_VERSION, + ONNX_EXTERNAL_WEIGHTS_NAME, + ONNX_WEIGHTS_NAME, + SAFE_WEIGHTS_INDEX_NAME, + SAFETENSORS_FILE_EXTENSION, + SAFETENSORS_WEIGHTS_NAME, + USE_PEFT_BACKEND, + WEIGHTS_INDEX_NAME, + WEIGHTS_NAME, +) +from .deprecation_utils import deprecate +from .doc_utils import replace_example_docstring +from .dynamic_modules_utils import get_class_from_dynamic_module +from .export_utils import export_to_gif, export_to_obj, export_to_ply, export_to_video +from .hub_utils import ( + PushToHubMixin, + _add_variant, + _get_checkpoint_shard_files, + _get_model_file, + extract_commit_hash, + http_user_agent, +) +from .import_utils import ( + BACKENDS_MAPPING, + DIFFUSERS_SLOW_IMPORT, + ENV_VARS_TRUE_AND_AUTO_VALUES, + ENV_VARS_TRUE_VALUES, + USE_JAX, + USE_TF, + USE_TORCH, + DummyObject, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_accelerate_available, + is_accelerate_version, + is_better_profanity_available, + is_bitsandbytes_available, + is_bitsandbytes_version, + is_bs4_available, + is_cosmos_guardrail_available, + is_flax_available, + is_ftfy_available, + is_gguf_available, + is_gguf_version, + is_google_colab, + is_hf_hub_version, + is_hpu_available, + is_inflect_available, + is_invisible_watermark_available, + is_k_diffusion_available, + is_k_diffusion_version, + is_librosa_available, + is_matplotlib_available, + is_nltk_available, + is_note_seq_available, + is_onnx_available, + is_opencv_available, + is_optimum_quanto_available, + is_optimum_quanto_version, + is_peft_available, + is_peft_version, + is_pytorch_retinaface_available, + is_safetensors_available, + is_scipy_available, + is_sentencepiece_available, + is_tensorboard_available, + is_timm_available, + is_torch_available, + is_torch_npu_available, + is_torch_version, + is_torch_xla_available, + is_torch_xla_version, + is_torchao_available, + is_torchao_version, + is_torchsde_available, + is_torchvision_available, + is_transformers_available, + is_transformers_version, + is_unidecode_available, + is_wandb_available, + is_xformers_available, + requires_backends, +) +from .loading_utils import get_module_from_name, get_submodule_by_name, load_image, load_video +from .logging import get_logger +from .outputs import BaseOutput +from .peft_utils import ( + check_peft_version, + delete_adapter_layers, + get_adapter_name, + get_peft_kwargs, + recurse_remove_peft_layers, + scale_lora_layers, + set_adapter_layers, + set_weights_and_activate_adapters, + unscale_lora_layers, +) +from .pil_utils import PIL_INTERPOLATION, make_image_grid, numpy_to_pil, pt_to_pil +from .remote_utils import remote_decode +from .state_dict_utils import ( + convert_all_state_dict_to_peft, + convert_state_dict_to_diffusers, + convert_state_dict_to_kohya, + convert_state_dict_to_peft, + convert_unet_state_dict_to_peft, + state_dict_all_zero, +) +from .typing_utils import _get_detailed_type, _is_valid_type + + +logger = get_logger(__name__) + + +def check_min_version(min_version): + if version.parse(__version__) < version.parse(min_version): + if "dev" in min_version: + error_message = ( + "This example requires a source install from HuggingFace diffusers (see " + "`https://huggingface.co/docs/diffusers/installation#install-from-source`)," + ) + else: + error_message = f"This example requires a minimum version of {min_version}," + error_message += f" but the version found is {__version__}.\n" + raise ImportError(error_message) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3c947a1b8d10a6b52b0f92bc15247ef3a40feba2 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/accelerate_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/accelerate_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0c228a9560595acdf8330bab0785a45cecfacef4 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/accelerate_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/constants.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/constants.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..709f27d7002b42c88d57a01a20322ac3d49565a6 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/constants.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/deprecation_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/deprecation_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6e5b0e03f9a7bc6780e1ffd86907a4fc4ad62655 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/deprecation_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/doc_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/doc_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d63b4bf2eda49aec987fde25b9ac481770b66b50 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/doc_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_bitsandbytes_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_bitsandbytes_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d62ed68f13a8a0ba0a114c9d51aed0cf4d46ac65 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_bitsandbytes_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_flax_and_transformers_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_flax_and_transformers_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..254e3338b33373e40f177ad8b543b782c0916913 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_flax_and_transformers_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_flax_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_flax_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1e1f2ac479b3131d3b170982a3a9f5e7909057ea Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_flax_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_gguf_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_gguf_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e4ccdbb6a8bc1bbd918559b09e3372e33bcfcf95 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_gguf_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_note_seq_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_note_seq_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..28d7fd3c23b4f5b7ca952541c1525159302880ae Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_note_seq_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_onnx_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_onnx_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0d033e5deac32116d23b3a6f72bea544b31a5e5b Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_onnx_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_optimum_quanto_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_optimum_quanto_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..29fce5d87e4f629c9ab399de7dd343ff9c4f6a95 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_optimum_quanto_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_pt_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_pt_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d87295777cb804ab5a42c4c680ff8f13434a0c17 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_pt_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_librosa_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_librosa_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9f349ed14ae0eefbc22c8be3218baf2c8bb8bb41 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_librosa_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_scipy_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_scipy_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..26e80b1cb05816f141528b3f249ec8ea1abc332b Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_scipy_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_torchsde_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_torchsde_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6c10fb31a29ac4a678df6e16fbefc6caf0b682e3 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_torchsde_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_and_k_diffusion_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_and_k_diffusion_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..30bc1ea543362470e87000ec067caaa2f783a41d Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_and_k_diffusion_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_and_onnx_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_and_onnx_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..14f3dd4e0ceee422e94cd2854b7e705bde7299e1 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_and_onnx_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_and_opencv_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_and_opencv_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c7ecf189f1827e8e66c6ef8ef654ab2db9a37455 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_and_opencv_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_and_sentencepiece_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_and_sentencepiece_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a217d9202f9ac718103e3d10dabd6cabec71dc1d Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_and_sentencepiece_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b18ef521d041c7cbe8b51d9355b63ffa5506bc36 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torch_and_transformers_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torchao_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torchao_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..20a832c305bb7e923bd4bec124154c3cc6826de1 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_torchao_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_transformers_and_torch_and_note_seq_objects.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_transformers_and_torch_and_note_seq_objects.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..35dc4e735a3eff7d34dd40b404a58697cd2d9d05 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dummy_transformers_and_torch_and_note_seq_objects.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dynamic_modules_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dynamic_modules_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cfffcc2ca5f78f05ab67ebd586143e19a6212ecf Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/dynamic_modules_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/export_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/export_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..818f1df202da3d13e90270a6c7e7b901ddd786cb Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/export_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/hub_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/hub_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7fc19531ab4caf46fc8fad53e0a7a4641adeaf44 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/hub_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/import_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/import_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..628a69fa74c4f2fd7378743a43f11cde9f75386f Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/import_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/loading_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/loading_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..66b6d3c8c912cd82386c3c50551759c0529b7695 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/loading_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/logging.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/logging.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..674198c5a2c9d7a8f0b6fa216da3c94806245f9c Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/logging.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/outputs.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/outputs.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bb0afc8b05cca2d6605a87462119955a1fe1eec5 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/outputs.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/peft_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/peft_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1d0fda61f1edaacf9d5ad0b4b1befeaf3089ffe9 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/peft_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/pil_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/pil_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a1d737a4ec97a99b6183dde272f02a89237b1d44 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/pil_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/remote_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/remote_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..73af0c80556031ba1cb408028bc9c324268db2bf Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/remote_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/source_code_parsing_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/source_code_parsing_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fbb1dfea715ed3130ac9027630ff2173ee86bd0f Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/source_code_parsing_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/state_dict_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/state_dict_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..45b94483fad9553cb94830eff71d216de1432845 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/state_dict_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/testing_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/testing_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..78583ddf8184585b07b0b1621603180772651231 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/testing_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/torch_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/torch_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..94a5d1d830ca4dd8db952a9edf40707672fd259b Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/torch_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/typing_utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/typing_utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c4028f05881a68eded00b3612615529bb29bc91d Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/typing_utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/versions.cpython-310.pyc b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/versions.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bfe94a8cb595a20fc21f45fef5a68c964f9620f0 Binary files /dev/null and b/venv/lib/python3.10/site-packages/diffusers/utils/__pycache__/versions.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/accelerate_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/accelerate_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..af3b712b5a9ea71ace657c5189fb6ad9f10d8148 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/accelerate_utils.py @@ -0,0 +1,48 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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. +""" +Accelerate utilities: Utilities related to accelerate +""" + +from packaging import version + +from .import_utils import is_accelerate_available + + +if is_accelerate_available(): + import accelerate + + +def apply_forward_hook(method): + """ + Decorator that applies a registered CpuOffload hook to an arbitrary function rather than `forward`. This is useful + for cases where a PyTorch module provides functions other than `forward` that should trigger a move to the + appropriate acceleration device. This is the case for `encode` and `decode` in [`AutoencoderKL`]. + + This decorator looks inside the internal `_hf_hook` property to find a registered offload hook. + + :param method: The method to decorate. This method should be a method of a PyTorch module. + """ + if not is_accelerate_available(): + return method + accelerate_version = version.parse(accelerate.__version__).base_version + if version.parse(accelerate_version) < version.parse("0.17.0"): + return method + + def wrapper(self, *args, **kwargs): + if hasattr(self, "_hf_hook") and hasattr(self._hf_hook, "pre_forward"): + self._hf_hook.pre_forward(self) + return method(self, *args, **kwargs) + + return wrapper diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/constants.py b/venv/lib/python3.10/site-packages/diffusers/utils/constants.py new file mode 100644 index 0000000000000000000000000000000000000000..7c04287d33ed9e4b2030d8c76aed5ff06cbbb285 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/constants.py @@ -0,0 +1,70 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# 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 importlib +import os + +from huggingface_hub.constants import HF_HOME +from packaging import version + +from ..dependency_versions_check import dep_version_check +from .import_utils import ENV_VARS_TRUE_VALUES, is_peft_available, is_transformers_available + + +MIN_PEFT_VERSION = "0.6.0" +MIN_TRANSFORMERS_VERSION = "4.34.0" +_CHECK_PEFT = os.environ.get("_CHECK_PEFT", "1") in ENV_VARS_TRUE_VALUES + + +CONFIG_NAME = "config.json" +WEIGHTS_NAME = "diffusion_pytorch_model.bin" +WEIGHTS_INDEX_NAME = "diffusion_pytorch_model.bin.index.json" +FLAX_WEIGHTS_NAME = "diffusion_flax_model.msgpack" +ONNX_WEIGHTS_NAME = "model.onnx" +SAFETENSORS_WEIGHTS_NAME = "diffusion_pytorch_model.safetensors" +SAFE_WEIGHTS_INDEX_NAME = "diffusion_pytorch_model.safetensors.index.json" +SAFETENSORS_FILE_EXTENSION = "safetensors" +GGUF_FILE_EXTENSION = "gguf" +ONNX_EXTERNAL_WEIGHTS_NAME = "weights.pb" +HUGGINGFACE_CO_RESOLVE_ENDPOINT = os.environ.get("HF_ENDPOINT", "https://huggingface.co") +DIFFUSERS_DYNAMIC_MODULE_NAME = "diffusers_modules" +HF_MODULES_CACHE = os.getenv("HF_MODULES_CACHE", os.path.join(HF_HOME, "modules")) +DEPRECATED_REVISION_ARGS = ["fp16", "non-ema"] +DIFFUSERS_REQUEST_TIMEOUT = 60 + +# Below should be `True` if the current version of `peft` and `transformers` are compatible with +# PEFT backend. Will automatically fall back to PEFT backend if the correct versions of the libraries are +# available. +# For PEFT it is has to be greater than or equal to 0.6.0 and for transformers it has to be greater than or equal to 4.34.0. +_required_peft_version = is_peft_available() and version.parse( + version.parse(importlib.metadata.version("peft")).base_version +) >= version.parse(MIN_PEFT_VERSION) +_required_transformers_version = is_transformers_available() and version.parse( + version.parse(importlib.metadata.version("transformers")).base_version +) >= version.parse(MIN_TRANSFORMERS_VERSION) + +USE_PEFT_BACKEND = _required_peft_version and _required_transformers_version + +if USE_PEFT_BACKEND and _CHECK_PEFT: + dep_version_check("peft") + + +DECODE_ENDPOINT_SD_V1 = "https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud/" +DECODE_ENDPOINT_SD_XL = "https://x2dmsqunjd6k9prw.us-east-1.aws.endpoints.huggingface.cloud/" +DECODE_ENDPOINT_FLUX = "https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud/" +DECODE_ENDPOINT_HUNYUAN_VIDEO = "https://o7ywnmrahorts457.us-east-1.aws.endpoints.huggingface.cloud/" + + +ENCODE_ENDPOINT_SD_V1 = "https://qc6479g0aac6qwy9.us-east-1.aws.endpoints.huggingface.cloud/" +ENCODE_ENDPOINT_SD_XL = "https://xjqqhmyn62rog84g.us-east-1.aws.endpoints.huggingface.cloud/" +ENCODE_ENDPOINT_FLUX = "https://ptccx55jz97f9zgo.us-east-1.aws.endpoints.huggingface.cloud/" diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/deprecation_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/deprecation_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..4f001b3047d60c7572d770ad93595af3e62ae45e --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/deprecation_utils.py @@ -0,0 +1,49 @@ +import inspect +import warnings +from typing import Any, Dict, Optional, Union + +from packaging import version + + +def deprecate(*args, take_from: Optional[Union[Dict, Any]] = None, standard_warn=True, stacklevel=2): + from .. import __version__ + + deprecated_kwargs = take_from + values = () + if not isinstance(args[0], tuple): + args = (args,) + + for attribute, version_name, message in args: + if version.parse(version.parse(__version__).base_version) >= version.parse(version_name): + raise ValueError( + f"The deprecation tuple {(attribute, version_name, message)} should be removed since diffusers'" + f" version {__version__} is >= {version_name}" + ) + + warning = None + if isinstance(deprecated_kwargs, dict) and attribute in deprecated_kwargs: + values += (deprecated_kwargs.pop(attribute),) + warning = f"The `{attribute}` argument is deprecated and will be removed in version {version_name}." + elif hasattr(deprecated_kwargs, attribute): + values += (getattr(deprecated_kwargs, attribute),) + warning = f"The `{attribute}` attribute is deprecated and will be removed in version {version_name}." + elif deprecated_kwargs is None: + warning = f"`{attribute}` is deprecated and will be removed in version {version_name}." + + if warning is not None: + warning = warning + " " if standard_warn else "" + warnings.warn(warning + message, FutureWarning, stacklevel=stacklevel) + + if isinstance(deprecated_kwargs, dict) and len(deprecated_kwargs) > 0: + call_frame = inspect.getouterframes(inspect.currentframe())[1] + filename = call_frame.filename + line_number = call_frame.lineno + function = call_frame.function + key, value = next(iter(deprecated_kwargs.items())) + raise TypeError(f"{function} in {filename} line {line_number - 1} got an unexpected keyword argument `{key}`") + + if len(values) == 0: + return + elif len(values) == 1: + return values[0] + return values diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/doc_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/doc_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..815083e1425801a5715da8ecd99af017427ff740 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/doc_utils.py @@ -0,0 +1,39 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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. +""" +Doc utilities: Utilities related to documentation +""" + +import re + + +def replace_example_docstring(example_docstring): + def docstring_decorator(fn): + func_doc = fn.__doc__ + lines = func_doc.split("\n") + i = 0 + while i < len(lines) and re.search(r"^\s*Examples?:\s*$", lines[i]) is None: + i += 1 + if i < len(lines): + lines[i] = example_docstring + func_doc = "\n".join(lines) + else: + raise ValueError( + f"The function {fn} should have an empty 'Examples:' in its docstring as placeholder, " + f"current docstring is:\n{func_doc}" + ) + fn.__doc__ = func_doc + return fn + + return docstring_decorator diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_bitsandbytes_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_bitsandbytes_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..2dc589428de960f9c46f5a22874e2ef6475c418c --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_bitsandbytes_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class BitsAndBytesConfig(metaclass=DummyObject): + _backends = ["bitsandbytes"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["bitsandbytes"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["bitsandbytes"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["bitsandbytes"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_flax_and_transformers_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_flax_and_transformers_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..5e65e5349bb0a6a0bac62cddf0ce0fad64237c68 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_flax_and_transformers_objects.py @@ -0,0 +1,77 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class FlaxStableDiffusionControlNetPipeline(metaclass=DummyObject): + _backends = ["flax", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax", "transformers"]) + + +class FlaxStableDiffusionImg2ImgPipeline(metaclass=DummyObject): + _backends = ["flax", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax", "transformers"]) + + +class FlaxStableDiffusionInpaintPipeline(metaclass=DummyObject): + _backends = ["flax", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax", "transformers"]) + + +class FlaxStableDiffusionPipeline(metaclass=DummyObject): + _backends = ["flax", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax", "transformers"]) + + +class FlaxStableDiffusionXLPipeline(metaclass=DummyObject): + _backends = ["flax", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax", "transformers"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_flax_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_flax_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..5fa8dbc819316e96f7483addba43f90b9d8f397b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_flax_objects.py @@ -0,0 +1,212 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class FlaxControlNetModel(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + +class FlaxModelMixin(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + +class FlaxUNet2DConditionModel(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + +class FlaxAutoencoderKL(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + +class FlaxDiffusionPipeline(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + +class FlaxDDIMScheduler(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + +class FlaxDDPMScheduler(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + +class FlaxDPMSolverMultistepScheduler(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + +class FlaxEulerDiscreteScheduler(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + +class FlaxKarrasVeScheduler(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + +class FlaxLMSDiscreteScheduler(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + +class FlaxPNDMScheduler(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + +class FlaxSchedulerMixin(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + +class FlaxScoreSdeVeScheduler(metaclass=DummyObject): + _backends = ["flax"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["flax"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["flax"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_gguf_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_gguf_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..4a6d9a060a1397c5ef0e402e67170960bb3de516 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_gguf_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class GGUFQuantizationConfig(metaclass=DummyObject): + _backends = ["gguf"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["gguf"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["gguf"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["gguf"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_note_seq_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_note_seq_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..c02d0b015aedc37c01fb3b843bc79547aae5da68 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_note_seq_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class MidiProcessor(metaclass=DummyObject): + _backends = ["note_seq"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["note_seq"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["note_seq"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["note_seq"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_onnx_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_onnx_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..bde5f6ad0793e2d81bc638600b46ff81748d09ee --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_onnx_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class OnnxRuntimeModel(metaclass=DummyObject): + _backends = ["onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["onnx"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_optimum_quanto_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_optimum_quanto_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..44f8eaffc2460360d17720578b8224496f91cf94 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_optimum_quanto_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class QuantoConfig(metaclass=DummyObject): + _backends = ["optimum_quanto"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["optimum_quanto"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["optimum_quanto"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["optimum_quanto"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_pt_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_pt_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..2981f3a420d6b1f3c3cf057018f12d52e727f3c8 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_pt_objects.py @@ -0,0 +1,2108 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class FasterCacheConfig(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HookRegistry(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class PyramidAttentionBroadcastConfig(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +def apply_faster_cache(*args, **kwargs): + requires_backends(apply_faster_cache, ["torch"]) + + +def apply_pyramid_attention_broadcast(*args, **kwargs): + requires_backends(apply_pyramid_attention_broadcast, ["torch"]) + + +class AllegroTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AsymmetricAutoencoderKL(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AuraFlowTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderDC(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKL(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLAllegro(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLCogVideoX(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLCosmos(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLHunyuanVideo(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLLTXVideo(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLMagvit(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLMochi(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLTemporalDecoder(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLWan(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderOobleck(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderTiny(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CacheMixin(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ChromaTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CogVideoXTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CogView3PlusTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CogView4Transformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ConsisIDTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ConsistencyDecoderVAE(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ControlNetUnionModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ControlNetXSAdapter(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CosmosTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DiTTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class EasyAnimateTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FluxControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FluxMultiControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FluxTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HiDreamImageTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HunyuanDiT2DControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HunyuanDiT2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HunyuanDiT2DMultiControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HunyuanVideoFramepackTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HunyuanVideoTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class I2VGenXLUNet(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class Kandinsky3UNet(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LatteTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LTXVideoTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class Lumina2Transformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LuminaNextDiT2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class MochiTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ModelMixin(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class MotionAdapter(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class MultiAdapter(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class MultiControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class OmniGenTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class PixArtTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class PriorTransformer(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SanaControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SanaTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SD3ControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SD3MultiControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SD3Transformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SparseControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class StableAudioDiTModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class T2IAdapter(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class T5FilmDecoder(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class Transformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class TransformerTemporalModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UNet1DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UNet2DConditionModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UNet2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UNet3DConditionModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UNetControlNetXSModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UNetMotionModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UNetSpatioTemporalConditionModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UVit2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class VQModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class WanTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class WanVACETransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +def get_constant_schedule(*args, **kwargs): + requires_backends(get_constant_schedule, ["torch"]) + + +def get_constant_schedule_with_warmup(*args, **kwargs): + requires_backends(get_constant_schedule_with_warmup, ["torch"]) + + +def get_cosine_schedule_with_warmup(*args, **kwargs): + requires_backends(get_cosine_schedule_with_warmup, ["torch"]) + + +def get_cosine_with_hard_restarts_schedule_with_warmup(*args, **kwargs): + requires_backends(get_cosine_with_hard_restarts_schedule_with_warmup, ["torch"]) + + +def get_linear_schedule_with_warmup(*args, **kwargs): + requires_backends(get_linear_schedule_with_warmup, ["torch"]) + + +def get_polynomial_decay_schedule_with_warmup(*args, **kwargs): + requires_backends(get_polynomial_decay_schedule_with_warmup, ["torch"]) + + +def get_scheduler(*args, **kwargs): + requires_backends(get_scheduler, ["torch"]) + + +class AudioPipelineOutput(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoPipelineForImage2Image(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoPipelineForInpainting(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoPipelineForText2Image(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class BlipDiffusionControlNetPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class BlipDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CLIPImageProjection(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ConsistencyModelPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DanceDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDIMPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDPMPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DiffusionPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DiTPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ImagePipelineOutput(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class KarrasVePipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LDMPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LDMSuperResolutionPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class PNDMPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class RePaintPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ScoreSdeVePipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class StableDiffusionMixin(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DiffusersQuantizer(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AmusedScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CMStochasticIterativeScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CogVideoXDDIMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CogVideoXDPMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDIMInverseScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDIMParallelScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDIMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDPMParallelScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDPMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDPMWuerstchenScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DEISMultistepScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DPMSolverMultistepInverseScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DPMSolverMultistepScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DPMSolverSinglestepScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class EDMDPMSolverMultistepScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class EDMEulerScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class EulerAncestralDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class EulerDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FlowMatchEulerDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FlowMatchHeunDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FlowMatchLCMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HeunDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class IPNDMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class KarrasVeScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class KDPM2AncestralDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class KDPM2DiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LCMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class PNDMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class RePaintScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SASolverScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SchedulerMixin(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SCMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ScoreSdeVeScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class TCDScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UnCLIPScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UniPCMultistepScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class VQDiffusionScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class EMAModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_librosa_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_librosa_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..2088bc4a744198284f22fe54e6f1055cf3568566 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_librosa_objects.py @@ -0,0 +1,32 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class AudioDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "librosa"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "librosa"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "librosa"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "librosa"]) + + +class Mel(metaclass=DummyObject): + _backends = ["torch", "librosa"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "librosa"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "librosa"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "librosa"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_scipy_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_scipy_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..a1ff25863822b04971d2c6dfdc17f5b28774cf05 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_scipy_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class LMSDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch", "scipy"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "scipy"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "scipy"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "scipy"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_torchsde_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_torchsde_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..6ff14231b9cc8c88effc4d11f7b439fa16bb7f20 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_torchsde_objects.py @@ -0,0 +1,32 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class CosineDPMSolverMultistepScheduler(metaclass=DummyObject): + _backends = ["torch", "torchsde"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "torchsde"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "torchsde"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "torchsde"]) + + +class DPMSolverSDEScheduler(metaclass=DummyObject): + _backends = ["torch", "torchsde"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "torchsde"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "torchsde"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "torchsde"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_and_k_diffusion_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_and_k_diffusion_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..2ab00c54ce2d7ca78f674085f0fe294d1f05c1e0 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_and_k_diffusion_objects.py @@ -0,0 +1,32 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class StableDiffusionKDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "k_diffusion"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "k_diffusion"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "k_diffusion"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "k_diffusion"]) + + +class StableDiffusionXLKDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "k_diffusion"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "k_diffusion"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "k_diffusion"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "k_diffusion"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_and_onnx_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_and_onnx_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..b7afad8226b87292100270e3e7daad6885be0e7f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_and_onnx_objects.py @@ -0,0 +1,92 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class OnnxStableDiffusionImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + +class OnnxStableDiffusionInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + +class OnnxStableDiffusionInpaintPipelineLegacy(metaclass=DummyObject): + _backends = ["torch", "transformers", "onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + +class OnnxStableDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + +class OnnxStableDiffusionUpscalePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + +class StableDiffusionOnnxPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_and_opencv_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_and_opencv_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..c0c4d9df5eeeca4cca0a5147058eba70f89bcbc2 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_and_opencv_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class ConsisIDPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "opencv"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "opencv"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "opencv"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "opencv"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_and_sentencepiece_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_and_sentencepiece_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..349db10d1fa5b8505173e1578d3c83ea17e95fbc --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_and_sentencepiece_objects.py @@ -0,0 +1,47 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class KolorsImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "sentencepiece"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "sentencepiece"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "sentencepiece"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "sentencepiece"]) + + +class KolorsPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "sentencepiece"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "sentencepiece"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "sentencepiece"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "sentencepiece"]) + + +class KolorsPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "sentencepiece"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "sentencepiece"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "sentencepiece"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "sentencepiece"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..656a8ac6c6f6581693eaa00c5a02de10a42ed513 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torch_and_transformers_objects.py @@ -0,0 +1,3032 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class AllegroPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AltDiffusionImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AltDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AmusedImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AmusedInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AmusedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimateDiffControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimateDiffPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimateDiffPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimateDiffSDXLPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimateDiffSparseControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimateDiffVideoToVideoControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimateDiffVideoToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AudioLDM2Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AudioLDM2ProjectionModel(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AudioLDM2UNet2DConditionModel(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AudioLDMPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AuraFlowPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ChromaImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ChromaPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CLIPImageProjection(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CogVideoXFunControlPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CogVideoXImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CogVideoXPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CogVideoXVideoToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CogView3PlusPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CogView4ControlPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CogView4Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ConsisIDPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Cosmos2TextToImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Cosmos2VideoToWorldPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CosmosTextToWorldPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CosmosVideoToWorldPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CycleDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class EasyAnimateControlPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class EasyAnimateInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class EasyAnimatePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxControlImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxControlInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxControlNetImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxControlNetInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxControlPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxFillPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxPriorReduxPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HiDreamImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanDiTControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanDiTPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanDiTPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanSkyreelsImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanVideoFramepackPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanVideoImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class I2VGenXLPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class IFImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class IFImg2ImgSuperResolutionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class IFInpaintingPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class IFInpaintingSuperResolutionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class IFPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class IFSuperResolutionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ImageTextPipelineOutput(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Kandinsky3Img2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Kandinsky3Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyCombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyImg2ImgCombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyInpaintCombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyPriorPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22CombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22ControlnetImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22ControlnetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22Img2ImgCombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22Img2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22InpaintCombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22InpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22PriorEmb2EmbPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22PriorPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LatentConsistencyModelImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LatentConsistencyModelPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LattePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LDMTextToImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LEditsPPPipelineStableDiffusion(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LEditsPPPipelineStableDiffusionXL(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTXConditionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTXImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTXLatentUpsamplePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTXPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Lumina2Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Lumina2Text2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LuminaPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LuminaText2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MarigoldDepthPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MarigoldIntrinsicsPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MarigoldNormalsPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MochiPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MusicLDMPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class OmniGenPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class PaintByExamplePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class PIAPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class PixArtAlphaPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class PixArtSigmaPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class PixArtSigmaPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ReduxImageEncoder(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SanaControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SanaPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SanaPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SanaSprintImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SanaSprintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SemanticStableDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ShapEImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ShapEPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableAudioPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableAudioProjectionModel(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableCascadeCombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableCascadeDecoderPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableCascadePriorPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3ControlNetInpaintingPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3ControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3Img2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3InpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3PAGImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3PAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionAdapterPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionAttendAndExcitePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionControlNetImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionControlNetInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionControlNetPAGInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionControlNetPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionControlNetXSPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionDepth2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionDiffEditPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionGLIGENPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionGLIGENTextImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionImageVariationPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionInpaintPipelineLegacy(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionInstructPix2PixPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionLatentUpscalePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionLDM3DPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionModelEditingPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionPAGImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionPAGInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionPanoramaPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionParadigmsPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionPipelineSafe(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionPix2PixZeroPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionSAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionUpscalePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLAdapterPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetPAGImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetUnionImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetUnionInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetUnionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetXSPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLInstructPix2PixPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLPAGImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLPAGInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableUnCLIPImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableUnCLIPPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableVideoDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class TextToVideoSDPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class TextToVideoZeroPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class TextToVideoZeroSDXLPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class UnCLIPImageVariationPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class UnCLIPPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class UniDiffuserModel(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class UniDiffuserPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class UniDiffuserTextDecoder(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VersatileDiffusionDualGuidedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VersatileDiffusionImageVariationPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VersatileDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VersatileDiffusionTextToImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VideoToVideoSDPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VisualClozeGenerationPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VisualClozePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VQDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WanImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WanPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WanVACEPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WanVideoToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WuerstchenCombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WuerstchenDecoderPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WuerstchenPriorPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torchao_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torchao_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..16f0f6a55f64099c73ce4d81af8c87053b833e44 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_torchao_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class TorchAoConfig(metaclass=DummyObject): + _backends = ["torchao"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torchao"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torchao"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torchao"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dummy_transformers_and_torch_and_note_seq_objects.py b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_transformers_and_torch_and_note_seq_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..fbde04e33f0abd86d12f3dee048a4f0585c9f19d --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dummy_transformers_and_torch_and_note_seq_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class SpectrogramDiffusionPipeline(metaclass=DummyObject): + _backends = ["transformers", "torch", "note_seq"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["transformers", "torch", "note_seq"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["transformers", "torch", "note_seq"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["transformers", "torch", "note_seq"]) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/dynamic_modules_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/dynamic_modules_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..4878937ab2026c1140378f342215a38077aba2d0 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/dynamic_modules_utils.py @@ -0,0 +1,475 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# +# 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. +"""Utilities to dynamically load objects from the Hub.""" + +import importlib +import inspect +import json +import os +import re +import shutil +import sys +from pathlib import Path +from typing import Dict, Optional, Union +from urllib import request + +from huggingface_hub import hf_hub_download, model_info +from huggingface_hub.utils import RevisionNotFoundError, validate_hf_hub_args +from packaging import version + +from .. import __version__ +from . import DIFFUSERS_DYNAMIC_MODULE_NAME, HF_MODULES_CACHE, logging + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +# See https://huggingface.co/datasets/diffusers/community-pipelines-mirror +COMMUNITY_PIPELINES_MIRROR_ID = "diffusers/community-pipelines-mirror" + + +def get_diffusers_versions(): + url = "https://pypi.org/pypi/diffusers/json" + releases = json.loads(request.urlopen(url).read())["releases"].keys() + return sorted(releases, key=lambda x: version.Version(x)) + + +def init_hf_modules(): + """ + Creates the cache directory for modules with an init, and adds it to the Python path. + """ + # This function has already been executed if HF_MODULES_CACHE already is in the Python path. + if HF_MODULES_CACHE in sys.path: + return + + sys.path.append(HF_MODULES_CACHE) + os.makedirs(HF_MODULES_CACHE, exist_ok=True) + init_path = Path(HF_MODULES_CACHE) / "__init__.py" + if not init_path.exists(): + init_path.touch() + + +def create_dynamic_module(name: Union[str, os.PathLike]): + """ + Creates a dynamic module in the cache directory for modules. + """ + init_hf_modules() + dynamic_module_path = Path(HF_MODULES_CACHE) / name + # If the parent module does not exist yet, recursively create it. + if not dynamic_module_path.parent.exists(): + create_dynamic_module(dynamic_module_path.parent) + os.makedirs(dynamic_module_path, exist_ok=True) + init_path = dynamic_module_path / "__init__.py" + if not init_path.exists(): + init_path.touch() + + +def get_relative_imports(module_file): + """ + Get the list of modules that are relatively imported in a module file. + + Args: + module_file (`str` or `os.PathLike`): The module file to inspect. + """ + with open(module_file, "r", encoding="utf-8") as f: + content = f.read() + + # Imports of the form `import .xxx` + relative_imports = re.findall(r"^\s*import\s+\.(\S+)\s*$", content, flags=re.MULTILINE) + # Imports of the form `from .xxx import yyy` + relative_imports += re.findall(r"^\s*from\s+\.(\S+)\s+import", content, flags=re.MULTILINE) + # Unique-ify + return list(set(relative_imports)) + + +def get_relative_import_files(module_file): + """ + Get the list of all files that are needed for a given module. Note that this function recurses through the relative + imports (if a imports b and b imports c, it will return module files for b and c). + + Args: + module_file (`str` or `os.PathLike`): The module file to inspect. + """ + no_change = False + files_to_check = [module_file] + all_relative_imports = [] + + # Let's recurse through all relative imports + while not no_change: + new_imports = [] + for f in files_to_check: + new_imports.extend(get_relative_imports(f)) + + module_path = Path(module_file).parent + new_import_files = [str(module_path / m) for m in new_imports] + new_import_files = [f for f in new_import_files if f not in all_relative_imports] + files_to_check = [f"{f}.py" for f in new_import_files] + + no_change = len(new_import_files) == 0 + all_relative_imports.extend(files_to_check) + + return all_relative_imports + + +def check_imports(filename): + """ + Check if the current Python environment contains all the libraries that are imported in a file. + """ + with open(filename, "r", encoding="utf-8") as f: + content = f.read() + + # Imports of the form `import xxx` + imports = re.findall(r"^\s*import\s+(\S+)\s*$", content, flags=re.MULTILINE) + # Imports of the form `from xxx import yyy` + imports += re.findall(r"^\s*from\s+(\S+)\s+import", content, flags=re.MULTILINE) + # Only keep the top-level module + imports = [imp.split(".")[0] for imp in imports if not imp.startswith(".")] + + # Unique-ify and test we got them all + imports = list(set(imports)) + missing_packages = [] + for imp in imports: + try: + importlib.import_module(imp) + except ImportError: + missing_packages.append(imp) + + if len(missing_packages) > 0: + raise ImportError( + "This modeling file requires the following packages that were not found in your environment: " + f"{', '.join(missing_packages)}. Run `pip install {' '.join(missing_packages)}`" + ) + + return get_relative_imports(filename) + + +def get_class_in_module(class_name, module_path, pretrained_model_name_or_path=None): + """ + Import a module on the cache directory for modules and extract a class from it. + """ + module_path = module_path.replace(os.path.sep, ".") + try: + module = importlib.import_module(module_path) + except ModuleNotFoundError as e: + # This can happen when the repo id contains ".", which Python's import machinery interprets as a directory + # separator. We do a bit of monkey patching to detect and fix this case. + if not ( + pretrained_model_name_or_path is not None + and "." in pretrained_model_name_or_path + and module_path.startswith("diffusers_modules") + and pretrained_model_name_or_path.replace("/", "--") in module_path + ): + raise e # We can't figure this one out, just reraise the original error + + corrected_path = os.path.join(HF_MODULES_CACHE, module_path.replace(".", "/")) + ".py" + corrected_path = corrected_path.replace( + pretrained_model_name_or_path.replace("/", "--").replace(".", "/"), + pretrained_model_name_or_path.replace("/", "--"), + ) + module = importlib.machinery.SourceFileLoader(module_path, corrected_path).load_module() + + if class_name is None: + return find_pipeline_class(module) + return getattr(module, class_name) + + +def find_pipeline_class(loaded_module): + """ + Retrieve pipeline class that inherits from `DiffusionPipeline`. Note that there has to be exactly one class + inheriting from `DiffusionPipeline`. + """ + from ..pipelines import DiffusionPipeline + + cls_members = dict(inspect.getmembers(loaded_module, inspect.isclass)) + + pipeline_class = None + for cls_name, cls in cls_members.items(): + if ( + cls_name != DiffusionPipeline.__name__ + and issubclass(cls, DiffusionPipeline) + and cls.__module__.split(".")[0] != "diffusers" + ): + if pipeline_class is not None: + raise ValueError( + f"Multiple classes that inherit from {DiffusionPipeline.__name__} have been found:" + f" {pipeline_class.__name__}, and {cls_name}. Please make sure to define only one in" + f" {loaded_module}." + ) + pipeline_class = cls + + return pipeline_class + + +@validate_hf_hub_args +def get_cached_module_file( + pretrained_model_name_or_path: Union[str, os.PathLike], + module_file: str, + cache_dir: Optional[Union[str, os.PathLike]] = None, + force_download: bool = False, + proxies: Optional[Dict[str, str]] = None, + token: Optional[Union[bool, str]] = None, + revision: Optional[str] = None, + local_files_only: bool = False, +): + """ + Prepares Downloads a module from a local folder or a distant repo and returns its path inside the cached + Transformers module. + + Args: + pretrained_model_name_or_path (`str` or `os.PathLike`): + This can be either: + + - a string, the *model id* of a pretrained model configuration hosted inside a model repo on + huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced + under a user or organization name, like `dbmdz/bert-base-german-cased`. + - a path to a *directory* containing a configuration file saved using the + [`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`. + + module_file (`str`): + The name of the module file containing the class to look for. + cache_dir (`str` or `os.PathLike`, *optional*): + Path to a directory in which a downloaded pretrained model configuration should be cached if the standard + cache should not be used. + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force to (re-)download the configuration files and override the cached versions if they + exist. + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request. + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated + when running `transformers-cli login` (stored in `~/.huggingface`). + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a + git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any + identifier allowed by git. + local_files_only (`bool`, *optional*, defaults to `False`): + If `True`, will only try to load the tokenizer configuration from local files. + + + + You may pass a token in `token` if you are not logged in (`huggingface-cli login`) and want to use private or + [gated models](https://huggingface.co/docs/hub/models-gated#gated-models). + + + + Returns: + `str`: The path to the module inside the cache. + """ + # Download and cache module_file from the repo `pretrained_model_name_or_path` of grab it if it's a local file. + pretrained_model_name_or_path = str(pretrained_model_name_or_path) + + module_file_or_url = os.path.join(pretrained_model_name_or_path, module_file) + + if os.path.isfile(module_file_or_url): + resolved_module_file = module_file_or_url + submodule = "local" + elif pretrained_model_name_or_path.count("/") == 0: + available_versions = get_diffusers_versions() + # cut ".dev0" + latest_version = "v" + ".".join(__version__.split(".")[:3]) + + # retrieve github version that matches + if revision is None: + revision = latest_version if latest_version[1:] in available_versions else "main" + logger.info(f"Defaulting to latest_version: {revision}.") + elif revision in available_versions: + revision = f"v{revision}" + elif revision == "main": + revision = revision + else: + raise ValueError( + f"`custom_revision`: {revision} does not exist. Please make sure to choose one of" + f" {', '.join(available_versions + ['main'])}." + ) + + try: + resolved_module_file = hf_hub_download( + repo_id=COMMUNITY_PIPELINES_MIRROR_ID, + repo_type="dataset", + filename=f"{revision}/{pretrained_model_name_or_path}.py", + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + ) + submodule = "git" + module_file = pretrained_model_name_or_path + ".py" + except RevisionNotFoundError as e: + raise EnvironmentError( + f"Revision '{revision}' not found in the community pipelines mirror. Check available revisions on" + " https://huggingface.co/datasets/diffusers/community-pipelines-mirror/tree/main." + " If you don't find the revision you are looking for, please open an issue on https://github.com/huggingface/diffusers/issues." + ) from e + except EnvironmentError: + logger.error(f"Could not locate the {module_file} inside {pretrained_model_name_or_path}.") + raise + else: + try: + # Load from URL or cache if already cached + resolved_module_file = hf_hub_download( + pretrained_model_name_or_path, + module_file, + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + ) + submodule = os.path.join("local", "--".join(pretrained_model_name_or_path.split("/"))) + except EnvironmentError: + logger.error(f"Could not locate the {module_file} inside {pretrained_model_name_or_path}.") + raise + + # Check we have all the requirements in our environment + modules_needed = check_imports(resolved_module_file) + + # Now we move the module inside our cached dynamic modules. + full_submodule = DIFFUSERS_DYNAMIC_MODULE_NAME + os.path.sep + submodule + create_dynamic_module(full_submodule) + submodule_path = Path(HF_MODULES_CACHE) / full_submodule + if submodule == "local" or submodule == "git": + # We always copy local files (we could hash the file to see if there was a change, and give them the name of + # that hash, to only copy when there is a modification but it seems overkill for now). + # The only reason we do the copy is to avoid putting too many folders in sys.path. + shutil.copyfile(resolved_module_file, submodule_path / module_file) + for module_needed in modules_needed: + if len(module_needed.split(".")) == 2: + module_needed = "/".join(module_needed.split(".")) + module_folder = module_needed.split("/")[0] + if not os.path.exists(submodule_path / module_folder): + os.makedirs(submodule_path / module_folder) + module_needed = f"{module_needed}.py" + shutil.copyfile(os.path.join(pretrained_model_name_or_path, module_needed), submodule_path / module_needed) + else: + # Get the commit hash + # TODO: we will get this info in the etag soon, so retrieve it from there and not here. + commit_hash = model_info(pretrained_model_name_or_path, revision=revision, token=token).sha + + # The module file will end up being placed in a subfolder with the git hash of the repo. This way we get the + # benefit of versioning. + submodule_path = submodule_path / commit_hash + full_submodule = full_submodule + os.path.sep + commit_hash + create_dynamic_module(full_submodule) + + if not (submodule_path / module_file).exists(): + if len(module_file.split("/")) == 2: + module_folder = module_file.split("/")[0] + if not os.path.exists(submodule_path / module_folder): + os.makedirs(submodule_path / module_folder) + shutil.copyfile(resolved_module_file, submodule_path / module_file) + + # Make sure we also have every file with relative + for module_needed in modules_needed: + if len(module_needed.split(".")) == 2: + module_needed = "/".join(module_needed.split(".")) + if not (submodule_path / module_needed).exists(): + get_cached_module_file( + pretrained_model_name_or_path, + f"{module_needed}.py", + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + token=token, + revision=revision, + local_files_only=local_files_only, + ) + return os.path.join(full_submodule, module_file) + + +@validate_hf_hub_args +def get_class_from_dynamic_module( + pretrained_model_name_or_path: Union[str, os.PathLike], + module_file: str, + class_name: Optional[str] = None, + cache_dir: Optional[Union[str, os.PathLike]] = None, + force_download: bool = False, + proxies: Optional[Dict[str, str]] = None, + token: Optional[Union[bool, str]] = None, + revision: Optional[str] = None, + local_files_only: bool = False, + **kwargs, +): + """ + Extracts a class from a module file, present in the local folder or repository of a model. + + + + Calling this function will execute the code in the module file found locally or downloaded from the Hub. It should + therefore only be called on trusted repos. + + + + Args: + pretrained_model_name_or_path (`str` or `os.PathLike`): + This can be either: + + - a string, the *model id* of a pretrained model configuration hosted inside a model repo on + huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced + under a user or organization name, like `dbmdz/bert-base-german-cased`. + - a path to a *directory* containing a configuration file saved using the + [`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`. + + module_file (`str`): + The name of the module file containing the class to look for. + class_name (`str`): + The name of the class to import in the module. + cache_dir (`str` or `os.PathLike`, *optional*): + Path to a directory in which a downloaded pretrained model configuration should be cached if the standard + cache should not be used. + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force to (re-)download the configuration files and override the cached versions if they + exist. + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request. + token (`str` or `bool`, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated + when running `transformers-cli login` (stored in `~/.huggingface`). + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a + git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any + identifier allowed by git. + local_files_only (`bool`, *optional*, defaults to `False`): + If `True`, will only try to load the tokenizer configuration from local files. + + + + You may pass a token in `token` if you are not logged in (`huggingface-cli login`) and want to use private or + [gated models](https://huggingface.co/docs/hub/models-gated#gated-models). + + + + Returns: + `type`: The class, dynamically imported from the module. + + Examples: + + ```python + # Download module `modeling.py` from huggingface.co and cache then extract the class `MyBertModel` from this + # module. + cls = get_class_from_dynamic_module("sgugger/my-bert-model", "modeling.py", "MyBertModel") + ```""" + # And lastly we get the class inside our newly created module + final_module = get_cached_module_file( + pretrained_model_name_or_path, + module_file, + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + token=token, + revision=revision, + local_files_only=local_files_only, + ) + return get_class_in_module(class_name, final_module.replace(".py", ""), pretrained_model_name_or_path) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/export_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/export_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..07cf46928a44df0b27337e5d007c6a9bdb0e7c28 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/export_utils.py @@ -0,0 +1,209 @@ +import io +import random +import struct +import tempfile +from contextlib import contextmanager +from typing import List, Optional, Union + +import numpy as np +import PIL.Image +import PIL.ImageOps + +from .import_utils import BACKENDS_MAPPING, is_imageio_available, is_opencv_available +from .logging import get_logger + + +global_rng = random.Random() + +logger = get_logger(__name__) + + +@contextmanager +def buffered_writer(raw_f): + f = io.BufferedWriter(raw_f) + yield f + f.flush() + + +def export_to_gif(image: List[PIL.Image.Image], output_gif_path: str = None, fps: int = 10) -> str: + if output_gif_path is None: + output_gif_path = tempfile.NamedTemporaryFile(suffix=".gif").name + + image[0].save( + output_gif_path, + save_all=True, + append_images=image[1:], + optimize=False, + duration=1000 // fps, + loop=0, + ) + return output_gif_path + + +def export_to_ply(mesh, output_ply_path: str = None): + """ + Write a PLY file for a mesh. + """ + if output_ply_path is None: + output_ply_path = tempfile.NamedTemporaryFile(suffix=".ply").name + + coords = mesh.verts.detach().cpu().numpy() + faces = mesh.faces.cpu().numpy() + rgb = np.stack([mesh.vertex_channels[x].detach().cpu().numpy() for x in "RGB"], axis=1) + + with buffered_writer(open(output_ply_path, "wb")) as f: + f.write(b"ply\n") + f.write(b"format binary_little_endian 1.0\n") + f.write(bytes(f"element vertex {len(coords)}\n", "ascii")) + f.write(b"property float x\n") + f.write(b"property float y\n") + f.write(b"property float z\n") + if rgb is not None: + f.write(b"property uchar red\n") + f.write(b"property uchar green\n") + f.write(b"property uchar blue\n") + if faces is not None: + f.write(bytes(f"element face {len(faces)}\n", "ascii")) + f.write(b"property list uchar int vertex_index\n") + f.write(b"end_header\n") + + if rgb is not None: + rgb = (rgb * 255.499).round().astype(int) + vertices = [ + (*coord, *rgb) + for coord, rgb in zip( + coords.tolist(), + rgb.tolist(), + ) + ] + format = struct.Struct("<3f3B") + for item in vertices: + f.write(format.pack(*item)) + else: + format = struct.Struct("<3f") + for vertex in coords.tolist(): + f.write(format.pack(*vertex)) + + if faces is not None: + format = struct.Struct(" str: + """ + quality: + Video output quality. Default is 5. Uses variable bit rate. Highest quality is 10, lowest is 0. Set to None to + prevent variable bitrate flags to FFMPEG so you can manually specify them using output_params instead. + Specifying a fixed bitrate using `bitrate` disables this parameter. + + bitrate: + Set a constant bitrate for the video encoding. Default is None causing `quality` parameter to be used instead. + Better quality videos with smaller file sizes will result from using the `quality` variable bitrate parameter + rather than specifying a fixed bitrate with this parameter. + + macro_block_size: + Size constraint for video. Width and height, must be divisible by this number. If not divisible by this number + imageio will tell ffmpeg to scale the image up to the next closest size divisible by this number. Most codecs + are compatible with a macroblock size of 16 (default), some can go smaller (4, 8). To disable this automatic + feature set it to None or 1, however be warned many players can't decode videos that are odd in size and some + codecs will produce poor results or fail. See https://en.wikipedia.org/wiki/Macroblock. + """ + # TODO: Dhruv. Remove by Diffusers release 0.33.0 + # Added to prevent breaking existing code + if not is_imageio_available(): + logger.warning( + ( + "It is recommended to use `export_to_video` with `imageio` and `imageio-ffmpeg` as a backend. \n" + "These libraries are not present in your environment. Attempting to use legacy OpenCV backend to export video. \n" + "Support for the OpenCV backend will be deprecated in a future Diffusers version" + ) + ) + return _legacy_export_to_video(video_frames, output_video_path, fps) + + if is_imageio_available(): + import imageio + else: + raise ImportError(BACKENDS_MAPPING["imageio"][1].format("export_to_video")) + + try: + imageio.plugins.ffmpeg.get_exe() + except AttributeError: + raise AttributeError( + ( + "Found an existing imageio backend in your environment. Attempting to export video with imageio. \n" + "Unable to find a compatible ffmpeg installation in your environment to use with imageio. Please install via `pip install imageio-ffmpeg" + ) + ) + + if output_video_path is None: + output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4").name + + if isinstance(video_frames[0], np.ndarray): + video_frames = [(frame * 255).astype(np.uint8) for frame in video_frames] + + elif isinstance(video_frames[0], PIL.Image.Image): + video_frames = [np.array(frame) for frame in video_frames] + + with imageio.get_writer( + output_video_path, fps=fps, quality=quality, bitrate=bitrate, macro_block_size=macro_block_size + ) as writer: + for frame in video_frames: + writer.append_data(frame) + + return output_video_path diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/hub_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/hub_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..f80f96a3425d880c54bcf95d65dc6f3d2ba0f312 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/hub_utils.py @@ -0,0 +1,557 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# +# 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 json +import os +import re +import sys +import tempfile +import warnings +from pathlib import Path +from typing import Dict, List, Optional, Union +from uuid import uuid4 + +from huggingface_hub import ( + DDUFEntry, + ModelCard, + ModelCardData, + create_repo, + hf_hub_download, + model_info, + snapshot_download, + upload_folder, +) +from huggingface_hub.constants import HF_HUB_DISABLE_TELEMETRY, HF_HUB_OFFLINE +from huggingface_hub.file_download import REGEX_COMMIT_HASH +from huggingface_hub.utils import ( + EntryNotFoundError, + RepositoryNotFoundError, + RevisionNotFoundError, + is_jinja_available, + validate_hf_hub_args, +) +from packaging import version +from requests import HTTPError + +from .. import __version__ +from .constants import ( + DEPRECATED_REVISION_ARGS, + HUGGINGFACE_CO_RESOLVE_ENDPOINT, + SAFETENSORS_WEIGHTS_NAME, + WEIGHTS_NAME, +) +from .import_utils import ( + ENV_VARS_TRUE_VALUES, + _flax_version, + _jax_version, + _onnxruntime_version, + _torch_version, + is_flax_available, + is_onnx_available, + is_torch_available, +) +from .logging import get_logger + + +logger = get_logger(__name__) + +MODEL_CARD_TEMPLATE_PATH = Path(__file__).parent / "model_card_template.md" +SESSION_ID = uuid4().hex + + +def http_user_agent(user_agent: Union[Dict, str, None] = None) -> str: + """ + Formats a user-agent string with basic info about a request. + """ + ua = f"diffusers/{__version__}; python/{sys.version.split()[0]}; session_id/{SESSION_ID}" + if HF_HUB_DISABLE_TELEMETRY or HF_HUB_OFFLINE: + return ua + "; telemetry/off" + if is_torch_available(): + ua += f"; torch/{_torch_version}" + if is_flax_available(): + ua += f"; jax/{_jax_version}" + ua += f"; flax/{_flax_version}" + if is_onnx_available(): + ua += f"; onnxruntime/{_onnxruntime_version}" + # CI will set this value to True + if os.environ.get("DIFFUSERS_IS_CI", "").upper() in ENV_VARS_TRUE_VALUES: + ua += "; is_ci/true" + if isinstance(user_agent, dict): + ua += "; " + "; ".join(f"{k}/{v}" for k, v in user_agent.items()) + elif isinstance(user_agent, str): + ua += "; " + user_agent + return ua + + +def load_or_create_model_card( + repo_id_or_path: str = None, + token: Optional[str] = None, + is_pipeline: bool = False, + from_training: bool = False, + model_description: Optional[str] = None, + base_model: str = None, + prompt: Optional[str] = None, + license: Optional[str] = None, + widget: Optional[List[dict]] = None, + inference: Optional[bool] = None, +) -> ModelCard: + """ + Loads or creates a model card. + + Args: + repo_id_or_path (`str`): + The repo id (e.g., "runwayml/stable-diffusion-v1-5") or local path where to look for the model card. + token (`str`, *optional*): + Authentication token. Will default to the stored token. See https://huggingface.co/settings/token for more + details. + is_pipeline (`bool`): + Boolean to indicate if we're adding tag to a [`DiffusionPipeline`]. + from_training: (`bool`): Boolean flag to denote if the model card is being created from a training script. + model_description (`str`, *optional*): Model description to add to the model card. Helpful when using + `load_or_create_model_card` from a training script. + base_model (`str`): Base model identifier (e.g., "stabilityai/stable-diffusion-xl-base-1.0"). Useful + for DreamBooth-like training. + prompt (`str`, *optional*): Prompt used for training. Useful for DreamBooth-like training. + license: (`str`, *optional*): License of the output artifact. Helpful when using + `load_or_create_model_card` from a training script. + widget (`List[dict]`, *optional*): Widget to accompany a gallery template. + inference: (`bool`, optional): Whether to turn on inference widget. Helpful when using + `load_or_create_model_card` from a training script. + """ + if not is_jinja_available(): + raise ValueError( + "Modelcard rendering is based on Jinja templates." + " Please make sure to have `jinja` installed before using `load_or_create_model_card`." + " To install it, please run `pip install Jinja2`." + ) + + try: + # Check if the model card is present on the remote repo + model_card = ModelCard.load(repo_id_or_path, token=token) + except (EntryNotFoundError, RepositoryNotFoundError): + # Otherwise create a model card from template + if from_training: + model_card = ModelCard.from_template( + card_data=ModelCardData( # Card metadata object that will be converted to YAML block + license=license, + library_name="diffusers", + inference=inference, + base_model=base_model, + instance_prompt=prompt, + widget=widget, + ), + template_path=MODEL_CARD_TEMPLATE_PATH, + model_description=model_description, + ) + else: + card_data = ModelCardData() + component = "pipeline" if is_pipeline else "model" + if model_description is None: + model_description = f"This is the model card of a 🧨 diffusers {component} that has been pushed on the Hub. This model card has been automatically generated." + model_card = ModelCard.from_template(card_data, model_description=model_description) + + return model_card + + +def populate_model_card(model_card: ModelCard, tags: Union[str, List[str]] = None) -> ModelCard: + """Populates the `model_card` with library name and optional tags.""" + if model_card.data.library_name is None: + model_card.data.library_name = "diffusers" + + if tags is not None: + if isinstance(tags, str): + tags = [tags] + if model_card.data.tags is None: + model_card.data.tags = [] + for tag in tags: + model_card.data.tags.append(tag) + + return model_card + + +def extract_commit_hash(resolved_file: Optional[str], commit_hash: Optional[str] = None): + """ + Extracts the commit hash from a resolved filename toward a cache file. + """ + if resolved_file is None or commit_hash is not None: + return commit_hash + resolved_file = str(Path(resolved_file).as_posix()) + search = re.search(r"snapshots/([^/]+)/", resolved_file) + if search is None: + return None + commit_hash = search.groups()[0] + return commit_hash if REGEX_COMMIT_HASH.match(commit_hash) else None + + +def _add_variant(weights_name: str, variant: Optional[str] = None) -> str: + if variant is not None: + splits = weights_name.split(".") + splits = splits[:-1] + [variant] + splits[-1:] + weights_name = ".".join(splits) + + return weights_name + + +@validate_hf_hub_args +def _get_model_file( + pretrained_model_name_or_path: Union[str, Path], + *, + weights_name: str, + subfolder: Optional[str] = None, + cache_dir: Optional[str] = None, + force_download: bool = False, + proxies: Optional[Dict] = None, + local_files_only: bool = False, + token: Optional[str] = None, + user_agent: Optional[Union[Dict, str]] = None, + revision: Optional[str] = None, + commit_hash: Optional[str] = None, + dduf_entries: Optional[Dict[str, DDUFEntry]] = None, +): + pretrained_model_name_or_path = str(pretrained_model_name_or_path) + + if dduf_entries: + if subfolder is not None: + raise ValueError( + "DDUF file only allow for 1 level of directory (e.g transformer/model1/model.safetentors is not allowed). " + "Please check the DDUF structure" + ) + model_file = ( + weights_name + if pretrained_model_name_or_path == "" + else "/".join([pretrained_model_name_or_path, weights_name]) + ) + if model_file in dduf_entries: + return model_file + else: + raise EnvironmentError(f"Error no file named {weights_name} found in archive {dduf_entries.keys()}.") + elif os.path.isfile(pretrained_model_name_or_path): + return pretrained_model_name_or_path + elif os.path.isdir(pretrained_model_name_or_path): + if os.path.isfile(os.path.join(pretrained_model_name_or_path, weights_name)): + # Load from a PyTorch checkpoint + model_file = os.path.join(pretrained_model_name_or_path, weights_name) + return model_file + elif subfolder is not None and os.path.isfile( + os.path.join(pretrained_model_name_or_path, subfolder, weights_name) + ): + model_file = os.path.join(pretrained_model_name_or_path, subfolder, weights_name) + return model_file + else: + raise EnvironmentError( + f"Error no file named {weights_name} found in directory {pretrained_model_name_or_path}." + ) + else: + # 1. First check if deprecated way of loading from branches is used + if ( + revision in DEPRECATED_REVISION_ARGS + and (weights_name == WEIGHTS_NAME or weights_name == SAFETENSORS_WEIGHTS_NAME) + and version.parse(version.parse(__version__).base_version) >= version.parse("0.22.0") + ): + try: + model_file = hf_hub_download( + pretrained_model_name_or_path, + filename=_add_variant(weights_name, revision), + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + user_agent=user_agent, + subfolder=subfolder, + revision=revision or commit_hash, + ) + warnings.warn( + f"Loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'` is deprecated. Loading instead from `revision='main'` with `variant={revision}`. Loading model variants via `revision='{revision}'` will be removed in diffusers v1. Please use `variant='{revision}'` instead.", + FutureWarning, + ) + return model_file + except: # noqa: E722 + warnings.warn( + f"You are loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'`. This behavior is deprecated and will be removed in diffusers v1. One should use `variant='{revision}'` instead. However, it appears that {pretrained_model_name_or_path} currently does not have a {_add_variant(weights_name, revision)} file in the 'main' branch of {pretrained_model_name_or_path}. \n The Diffusers team and community would be very grateful if you could open an issue: https://github.com/huggingface/diffusers/issues/new with the title '{pretrained_model_name_or_path} is missing {_add_variant(weights_name, revision)}' so that the correct variant file can be added.", + FutureWarning, + ) + try: + # 2. Load model file as usual + model_file = hf_hub_download( + pretrained_model_name_or_path, + filename=weights_name, + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + user_agent=user_agent, + subfolder=subfolder, + revision=revision or commit_hash, + ) + return model_file + + except RepositoryNotFoundError as e: + raise EnvironmentError( + f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier " + "listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a " + "token having permission to this repo with `token` or log in with `huggingface-cli " + "login`." + ) from e + except RevisionNotFoundError as e: + raise EnvironmentError( + f"{revision} is not a valid git identifier (branch name, tag name or commit id) that exists for " + "this model name. Check the model page at " + f"'https://huggingface.co/{pretrained_model_name_or_path}' for available revisions." + ) from e + except EntryNotFoundError as e: + raise EnvironmentError( + f"{pretrained_model_name_or_path} does not appear to have a file named {weights_name}." + ) from e + except HTTPError as e: + raise EnvironmentError( + f"There was a specific connection error when trying to load {pretrained_model_name_or_path}:\n{e}" + ) from e + except ValueError as e: + raise EnvironmentError( + f"We couldn't connect to '{HUGGINGFACE_CO_RESOLVE_ENDPOINT}' to load this model, couldn't find it" + f" in the cached files and it looks like {pretrained_model_name_or_path} is not the path to a" + f" directory containing a file named {weights_name} or" + " \nCheckout your internet connection or see how to run the library in" + " offline mode at 'https://huggingface.co/docs/diffusers/installation#offline-mode'." + ) from e + except EnvironmentError as e: + raise EnvironmentError( + f"Can't load the model for '{pretrained_model_name_or_path}'. If you were trying to load it from " + "'https://huggingface.co/models', make sure you don't have a local directory with the same name. " + f"Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a directory " + f"containing a file named {weights_name}" + ) from e + + +def _get_checkpoint_shard_files( + pretrained_model_name_or_path, + index_filename, + cache_dir=None, + proxies=None, + local_files_only=False, + token=None, + user_agent=None, + revision=None, + subfolder="", + dduf_entries: Optional[Dict[str, DDUFEntry]] = None, +): + """ + For a given model: + + - download and cache all the shards of a sharded checkpoint if `pretrained_model_name_or_path` is a model ID on the + Hub + - returns the list of paths to all the shards, as well as some metadata. + + For the description of each arg, see [`PreTrainedModel.from_pretrained`]. `index_filename` is the full path to the + index (downloaded and cached if `pretrained_model_name_or_path` is a model ID on the Hub). + """ + if dduf_entries: + if index_filename not in dduf_entries: + raise ValueError(f"Can't find a checkpoint index ({index_filename}) in {pretrained_model_name_or_path}.") + else: + if not os.path.isfile(index_filename): + raise ValueError(f"Can't find a checkpoint index ({index_filename}) in {pretrained_model_name_or_path}.") + + if dduf_entries: + index = json.loads(dduf_entries[index_filename].read_text()) + else: + with open(index_filename, "r") as f: + index = json.loads(f.read()) + + original_shard_filenames = sorted(set(index["weight_map"].values())) + sharded_metadata = index["metadata"] + sharded_metadata["all_checkpoint_keys"] = list(index["weight_map"].keys()) + sharded_metadata["weight_map"] = index["weight_map"].copy() + shards_path = os.path.join(pretrained_model_name_or_path, subfolder) + + # First, let's deal with local folder. + if os.path.isdir(pretrained_model_name_or_path) or dduf_entries: + shard_filenames = [os.path.join(shards_path, f) for f in original_shard_filenames] + for shard_file in shard_filenames: + if dduf_entries: + if shard_file not in dduf_entries: + raise FileNotFoundError( + f"{shards_path} does not appear to have a file named {shard_file} which is " + "required according to the checkpoint index." + ) + else: + if not os.path.exists(shard_file): + raise FileNotFoundError( + f"{shards_path} does not appear to have a file named {shard_file} which is " + "required according to the checkpoint index." + ) + return shard_filenames, sharded_metadata + + # At this stage pretrained_model_name_or_path is a model identifier on the Hub + allow_patterns = original_shard_filenames + if subfolder is not None: + allow_patterns = [os.path.join(subfolder, p) for p in allow_patterns] + + ignore_patterns = ["*.json", "*.md"] + # `model_info` call must guarded with the above condition. + model_files_info = model_info(pretrained_model_name_or_path, revision=revision, token=token) + for shard_file in original_shard_filenames: + shard_file_present = any(shard_file in k.rfilename for k in model_files_info.siblings) + if not shard_file_present: + raise EnvironmentError( + f"{shards_path} does not appear to have a file named {shard_file} which is " + "required according to the checkpoint index." + ) + + try: + # Load from URL + cached_folder = snapshot_download( + pretrained_model_name_or_path, + cache_dir=cache_dir, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + allow_patterns=allow_patterns, + ignore_patterns=ignore_patterns, + user_agent=user_agent, + ) + if subfolder is not None: + cached_folder = os.path.join(cached_folder, subfolder) + + # We have already dealt with RepositoryNotFoundError and RevisionNotFoundError when getting the index, so + # we don't have to catch them here. We have also dealt with EntryNotFoundError. + except HTTPError as e: + raise EnvironmentError( + f"We couldn't connect to '{HUGGINGFACE_CO_RESOLVE_ENDPOINT}' to load {pretrained_model_name_or_path}. You should try" + " again after checking your internet connection." + ) from e + + cached_filenames = [os.path.join(cached_folder, f) for f in original_shard_filenames] + + return cached_filenames, sharded_metadata + + +def _check_legacy_sharding_variant_format(folder: str = None, filenames: List[str] = None, variant: str = None): + if filenames and folder: + raise ValueError("Both `filenames` and `folder` cannot be provided.") + if not filenames: + filenames = [] + for _, _, files in os.walk(folder): + for file in files: + filenames.append(os.path.basename(file)) + transformers_index_format = r"\d{5}-of-\d{5}" + variant_file_re = re.compile(rf".*-{transformers_index_format}\.{variant}\.[a-z]+$") + return any(variant_file_re.match(f) is not None for f in filenames) + + +class PushToHubMixin: + """ + A Mixin to push a model, scheduler, or pipeline to the Hugging Face Hub. + """ + + def _upload_folder( + self, + working_dir: Union[str, os.PathLike], + repo_id: str, + token: Optional[str] = None, + commit_message: Optional[str] = None, + create_pr: bool = False, + ): + """ + Uploads all files in `working_dir` to `repo_id`. + """ + if commit_message is None: + if "Model" in self.__class__.__name__: + commit_message = "Upload model" + elif "Scheduler" in self.__class__.__name__: + commit_message = "Upload scheduler" + else: + commit_message = f"Upload {self.__class__.__name__}" + + logger.info(f"Uploading the files of {working_dir} to {repo_id}.") + return upload_folder( + repo_id=repo_id, folder_path=working_dir, token=token, commit_message=commit_message, create_pr=create_pr + ) + + def push_to_hub( + self, + repo_id: str, + commit_message: Optional[str] = None, + private: Optional[bool] = None, + token: Optional[str] = None, + create_pr: bool = False, + safe_serialization: bool = True, + variant: Optional[str] = None, + ) -> str: + """ + Upload model, scheduler, or pipeline files to the 🤗 Hugging Face Hub. + + Parameters: + repo_id (`str`): + The name of the repository you want to push your model, scheduler, or pipeline files to. It should + contain your organization name when pushing to an organization. `repo_id` can also be a path to a local + directory. + commit_message (`str`, *optional*): + Message to commit while pushing. Default to `"Upload {object}"`. + private (`bool`, *optional*): + Whether to make the repo private. If `None` (default), the repo will be public unless the + organization's default is private. This value is ignored if the repo already exists. + token (`str`, *optional*): + The token to use as HTTP bearer authorization for remote files. The token generated when running + `huggingface-cli login` (stored in `~/.huggingface`). + create_pr (`bool`, *optional*, defaults to `False`): + Whether or not to create a PR with the uploaded files or directly commit. + safe_serialization (`bool`, *optional*, defaults to `True`): + Whether or not to convert the model weights to the `safetensors` format. + variant (`str`, *optional*): + If specified, weights are saved in the format `pytorch_model..bin`. + + Examples: + + ```python + from diffusers import UNet2DConditionModel + + unet = UNet2DConditionModel.from_pretrained("stabilityai/stable-diffusion-2", subfolder="unet") + + # Push the `unet` to your namespace with the name "my-finetuned-unet". + unet.push_to_hub("my-finetuned-unet") + + # Push the `unet` to an organization with the name "my-finetuned-unet". + unet.push_to_hub("your-org/my-finetuned-unet") + ``` + """ + repo_id = create_repo(repo_id, private=private, token=token, exist_ok=True).repo_id + + # Create a new empty model card and eventually tag it + model_card = load_or_create_model_card(repo_id, token=token) + model_card = populate_model_card(model_card) + + # Save all files. + save_kwargs = {"safe_serialization": safe_serialization} + if "Scheduler" not in self.__class__.__name__: + save_kwargs.update({"variant": variant}) + + with tempfile.TemporaryDirectory() as tmpdir: + self.save_pretrained(tmpdir, **save_kwargs) + + # Update model card if needed: + model_card.save(os.path.join(tmpdir, "README.md")) + + return self._upload_folder( + tmpdir, + repo_id, + token=token, + commit_message=commit_message, + create_pr=create_pr, + ) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/import_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/import_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..f12e9de33172a27ea7e809b1ff8c30effa14c9bc --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/import_utils.py @@ -0,0 +1,891 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 utilities: Utilities related to imports and our lazy inits. +""" + +import importlib.util +import inspect +import operator as op +import os +import sys +from collections import OrderedDict, defaultdict +from itertools import chain +from types import ModuleType +from typing import Any, Tuple, Union + +from huggingface_hub.utils import is_jinja_available # noqa: F401 +from packaging.version import Version, parse + +from . import logging + + +# The package importlib_metadata is in a different place, depending on the python version. +if sys.version_info < (3, 8): + import importlib_metadata +else: + import importlib.metadata as importlib_metadata +try: + _package_map = importlib_metadata.packages_distributions() # load-once to avoid expensive calls +except Exception: + _package_map = None + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +ENV_VARS_TRUE_VALUES = {"1", "ON", "YES", "TRUE"} +ENV_VARS_TRUE_AND_AUTO_VALUES = ENV_VARS_TRUE_VALUES.union({"AUTO"}) + +USE_TF = os.environ.get("USE_TF", "AUTO").upper() +USE_TORCH = os.environ.get("USE_TORCH", "AUTO").upper() +USE_JAX = os.environ.get("USE_FLAX", "AUTO").upper() +USE_SAFETENSORS = os.environ.get("USE_SAFETENSORS", "AUTO").upper() +DIFFUSERS_SLOW_IMPORT = os.environ.get("DIFFUSERS_SLOW_IMPORT", "FALSE").upper() +DIFFUSERS_SLOW_IMPORT = DIFFUSERS_SLOW_IMPORT in ENV_VARS_TRUE_VALUES + +STR_OPERATION_TO_FUNC = {">": op.gt, ">=": op.ge, "==": op.eq, "!=": op.ne, "<=": op.le, "<": op.lt} + +_is_google_colab = "google.colab" in sys.modules or any(k.startswith("COLAB_") for k in os.environ) + + +def _is_package_available(pkg_name: str, get_dist_name: bool = False) -> Tuple[bool, str]: + global _package_map + pkg_exists = importlib.util.find_spec(pkg_name) is not None + pkg_version = "N/A" + + if pkg_exists: + if _package_map is None: + _package_map = defaultdict(list) + try: + # Fallback for Python < 3.10 + for dist in importlib_metadata.distributions(): + _top_level_declared = (dist.read_text("top_level.txt") or "").split() + _infered_opt_names = { + f.parts[0] if len(f.parts) > 1 else inspect.getmodulename(f) for f in (dist.files or []) + } - {None} + _top_level_inferred = filter(lambda name: "." not in name, _infered_opt_names) + for pkg in _top_level_declared or _top_level_inferred: + _package_map[pkg].append(dist.metadata["Name"]) + except Exception as _: + pass + try: + if get_dist_name and pkg_name in _package_map and _package_map[pkg_name]: + if len(_package_map[pkg_name]) > 1: + logger.warning( + f"Multiple distributions found for package {pkg_name}. Picked distribution: {_package_map[pkg_name][0]}" + ) + pkg_name = _package_map[pkg_name][0] + pkg_version = importlib_metadata.version(pkg_name) + logger.debug(f"Successfully imported {pkg_name} version {pkg_version}") + except (ImportError, importlib_metadata.PackageNotFoundError): + pkg_exists = False + + return pkg_exists, pkg_version + + +if USE_TORCH in ENV_VARS_TRUE_AND_AUTO_VALUES and USE_TF not in ENV_VARS_TRUE_VALUES: + _torch_available, _torch_version = _is_package_available("torch") + +else: + logger.info("Disabling PyTorch because USE_TORCH is set") + _torch_available = False + _torch_version = "N/A" + +_jax_version = "N/A" +_flax_version = "N/A" +if USE_JAX in ENV_VARS_TRUE_AND_AUTO_VALUES: + _flax_available = importlib.util.find_spec("jax") is not None and importlib.util.find_spec("flax") is not None + if _flax_available: + try: + _jax_version = importlib_metadata.version("jax") + _flax_version = importlib_metadata.version("flax") + logger.info(f"JAX version {_jax_version}, Flax version {_flax_version} available.") + except importlib_metadata.PackageNotFoundError: + _flax_available = False +else: + _flax_available = False + +if USE_SAFETENSORS in ENV_VARS_TRUE_AND_AUTO_VALUES: + _safetensors_available, _safetensors_version = _is_package_available("safetensors") + +else: + logger.info("Disabling Safetensors because USE_TF is set") + _safetensors_available = False + +_onnxruntime_version = "N/A" +_onnx_available = importlib.util.find_spec("onnxruntime") is not None +if _onnx_available: + candidates = ( + "onnxruntime", + "onnxruntime-cann", + "onnxruntime-directml", + "ort_nightly_directml", + "onnxruntime-gpu", + "ort_nightly_gpu", + "onnxruntime-migraphx", + "onnxruntime-openvino", + "onnxruntime-qnn", + "onnxruntime-rocm", + "onnxruntime-training", + "onnxruntime-vitisai", + ) + _onnxruntime_version = None + # For the metadata, we have to look for both onnxruntime and onnxruntime-x + for pkg in candidates: + try: + _onnxruntime_version = importlib_metadata.version(pkg) + break + except importlib_metadata.PackageNotFoundError: + pass + _onnx_available = _onnxruntime_version is not None + if _onnx_available: + logger.debug(f"Successfully imported onnxruntime version {_onnxruntime_version}") + +# (sayakpaul): importlib.util.find_spec("opencv-python") returns None even when it's installed. +# _opencv_available = importlib.util.find_spec("opencv-python") is not None +try: + candidates = ( + "opencv-python", + "opencv-contrib-python", + "opencv-python-headless", + "opencv-contrib-python-headless", + ) + _opencv_version = None + for pkg in candidates: + try: + _opencv_version = importlib_metadata.version(pkg) + break + except importlib_metadata.PackageNotFoundError: + pass + _opencv_available = _opencv_version is not None + if _opencv_available: + logger.debug(f"Successfully imported cv2 version {_opencv_version}") +except importlib_metadata.PackageNotFoundError: + _opencv_available = False + +_bs4_available = importlib.util.find_spec("bs4") is not None +try: + # importlib metadata under different name + _bs4_version = importlib_metadata.version("beautifulsoup4") + logger.debug(f"Successfully imported ftfy version {_bs4_version}") +except importlib_metadata.PackageNotFoundError: + _bs4_available = False + +_invisible_watermark_available = importlib.util.find_spec("imwatermark") is not None +try: + _invisible_watermark_version = importlib_metadata.version("invisible-watermark") + logger.debug(f"Successfully imported invisible-watermark version {_invisible_watermark_version}") +except importlib_metadata.PackageNotFoundError: + _invisible_watermark_available = False + +_torch_xla_available, _torch_xla_version = _is_package_available("torch_xla") +_torch_npu_available, _torch_npu_version = _is_package_available("torch_npu") +_transformers_available, _transformers_version = _is_package_available("transformers") +_hf_hub_available, _hf_hub_version = _is_package_available("huggingface_hub") +_inflect_available, _inflect_version = _is_package_available("inflect") +_unidecode_available, _unidecode_version = _is_package_available("unidecode") +_k_diffusion_available, _k_diffusion_version = _is_package_available("k_diffusion") +_note_seq_available, _note_seq_version = _is_package_available("note_seq") +_wandb_available, _wandb_version = _is_package_available("wandb") +_tensorboard_available, _tensorboard_version = _is_package_available("tensorboard") +_compel_available, _compel_version = _is_package_available("compel") +_sentencepiece_available, _sentencepiece_version = _is_package_available("sentencepiece") +_torchsde_available, _torchsde_version = _is_package_available("torchsde") +_peft_available, _peft_version = _is_package_available("peft") +_torchvision_available, _torchvision_version = _is_package_available("torchvision") +_matplotlib_available, _matplotlib_version = _is_package_available("matplotlib") +_timm_available, _timm_version = _is_package_available("timm") +_bitsandbytes_available, _bitsandbytes_version = _is_package_available("bitsandbytes") +_imageio_available, _imageio_version = _is_package_available("imageio") +_ftfy_available, _ftfy_version = _is_package_available("ftfy") +_scipy_available, _scipy_version = _is_package_available("scipy") +_librosa_available, _librosa_version = _is_package_available("librosa") +_accelerate_available, _accelerate_version = _is_package_available("accelerate") +_xformers_available, _xformers_version = _is_package_available("xformers") +_gguf_available, _gguf_version = _is_package_available("gguf") +_torchao_available, _torchao_version = _is_package_available("torchao") +_bitsandbytes_available, _bitsandbytes_version = _is_package_available("bitsandbytes") +_optimum_quanto_available, _optimum_quanto_version = _is_package_available("optimum", get_dist_name=True) +_pytorch_retinaface_available, _pytorch_retinaface_version = _is_package_available("pytorch_retinaface") +_better_profanity_available, _better_profanity_version = _is_package_available("better_profanity") +_nltk_available, _nltk_version = _is_package_available("nltk") +_cosmos_guardrail_available, _cosmos_guardrail_version = _is_package_available("cosmos_guardrail") + + +def is_torch_available(): + return _torch_available + + +def is_torch_xla_available(): + return _torch_xla_available + + +def is_torch_npu_available(): + return _torch_npu_available + + +def is_flax_available(): + return _flax_available + + +def is_transformers_available(): + return _transformers_available + + +def is_inflect_available(): + return _inflect_available + + +def is_unidecode_available(): + return _unidecode_available + + +def is_onnx_available(): + return _onnx_available + + +def is_opencv_available(): + return _opencv_available + + +def is_scipy_available(): + return _scipy_available + + +def is_librosa_available(): + return _librosa_available + + +def is_xformers_available(): + return _xformers_available + + +def is_accelerate_available(): + return _accelerate_available + + +def is_k_diffusion_available(): + return _k_diffusion_available + + +def is_note_seq_available(): + return _note_seq_available + + +def is_wandb_available(): + return _wandb_available + + +def is_tensorboard_available(): + return _tensorboard_available + + +def is_compel_available(): + return _compel_available + + +def is_ftfy_available(): + return _ftfy_available + + +def is_bs4_available(): + return _bs4_available + + +def is_torchsde_available(): + return _torchsde_available + + +def is_invisible_watermark_available(): + return _invisible_watermark_available + + +def is_peft_available(): + return _peft_available + + +def is_torchvision_available(): + return _torchvision_available + + +def is_matplotlib_available(): + return _matplotlib_available + + +def is_safetensors_available(): + return _safetensors_available + + +def is_bitsandbytes_available(): + return _bitsandbytes_available + + +def is_google_colab(): + return _is_google_colab + + +def is_sentencepiece_available(): + return _sentencepiece_available + + +def is_imageio_available(): + return _imageio_available + + +def is_gguf_available(): + return _gguf_available + + +def is_torchao_available(): + return _torchao_available + + +def is_optimum_quanto_available(): + return _optimum_quanto_available + + +def is_timm_available(): + return _timm_available + + +def is_pytorch_retinaface_available(): + return _pytorch_retinaface_available + + +def is_better_profanity_available(): + return _better_profanity_available + + +def is_nltk_available(): + return _nltk_available + + +def is_cosmos_guardrail_available(): + return _cosmos_guardrail_available + + +def is_hpu_available(): + return all(importlib.util.find_spec(lib) for lib in ("habana_frameworks", "habana_frameworks.torch")) + + +# docstyle-ignore +FLAX_IMPORT_ERROR = """ +{0} requires the FLAX library but it was not found in your environment. Checkout the instructions on the +installation page: https://github.com/google/flax and follow the ones that match your environment. +""" + +# docstyle-ignore +INFLECT_IMPORT_ERROR = """ +{0} requires the inflect library but it was not found in your environment. You can install it with pip: `pip install +inflect` +""" + +# docstyle-ignore +PYTORCH_IMPORT_ERROR = """ +{0} requires the PyTorch library but it was not found in your environment. Checkout the instructions on the +installation page: https://pytorch.org/get-started/locally/ and follow the ones that match your environment. +""" + +# docstyle-ignore +ONNX_IMPORT_ERROR = """ +{0} requires the onnxruntime library but it was not found in your environment. You can install it with pip: `pip +install onnxruntime` +""" + +# docstyle-ignore +OPENCV_IMPORT_ERROR = """ +{0} requires the OpenCV library but it was not found in your environment. You can install it with pip: `pip +install opencv-python` +""" + +# docstyle-ignore +SCIPY_IMPORT_ERROR = """ +{0} requires the scipy library but it was not found in your environment. You can install it with pip: `pip install +scipy` +""" + +# docstyle-ignore +LIBROSA_IMPORT_ERROR = """ +{0} requires the librosa library but it was not found in your environment. Checkout the instructions on the +installation page: https://librosa.org/doc/latest/install.html and follow the ones that match your environment. +""" + +# docstyle-ignore +TRANSFORMERS_IMPORT_ERROR = """ +{0} requires the transformers library but it was not found in your environment. You can install it with pip: `pip +install transformers` +""" + +# docstyle-ignore +UNIDECODE_IMPORT_ERROR = """ +{0} requires the unidecode library but it was not found in your environment. You can install it with pip: `pip install +Unidecode` +""" + +# docstyle-ignore +K_DIFFUSION_IMPORT_ERROR = """ +{0} requires the k-diffusion library but it was not found in your environment. You can install it with pip: `pip +install k-diffusion` +""" + +# docstyle-ignore +NOTE_SEQ_IMPORT_ERROR = """ +{0} requires the note-seq library but it was not found in your environment. You can install it with pip: `pip +install note-seq` +""" + +# docstyle-ignore +WANDB_IMPORT_ERROR = """ +{0} requires the wandb library but it was not found in your environment. You can install it with pip: `pip +install wandb` +""" + +# docstyle-ignore +TENSORBOARD_IMPORT_ERROR = """ +{0} requires the tensorboard library but it was not found in your environment. You can install it with pip: `pip +install tensorboard` +""" + + +# docstyle-ignore +COMPEL_IMPORT_ERROR = """ +{0} requires the compel library but it was not found in your environment. You can install it with pip: `pip install compel` +""" + +# docstyle-ignore +BS4_IMPORT_ERROR = """ +{0} requires the Beautiful Soup library but it was not found in your environment. You can install it with pip: +`pip install beautifulsoup4`. Please note that you may need to restart your runtime after installation. +""" + +# docstyle-ignore +FTFY_IMPORT_ERROR = """ +{0} requires the ftfy library but it was not found in your environment. Checkout the instructions on the +installation section: https://github.com/rspeer/python-ftfy/tree/master#installing and follow the ones +that match your environment. Please note that you may need to restart your runtime after installation. +""" + +# docstyle-ignore +TORCHSDE_IMPORT_ERROR = """ +{0} requires the torchsde library but it was not found in your environment. You can install it with pip: `pip install torchsde` +""" + +# docstyle-ignore +INVISIBLE_WATERMARK_IMPORT_ERROR = """ +{0} requires the invisible-watermark library but it was not found in your environment. You can install it with pip: `pip install invisible-watermark>=0.2.0` +""" + +# docstyle-ignore +PEFT_IMPORT_ERROR = """ +{0} requires the peft library but it was not found in your environment. You can install it with pip: `pip install peft` +""" + +# docstyle-ignore +SAFETENSORS_IMPORT_ERROR = """ +{0} requires the safetensors library but it was not found in your environment. You can install it with pip: `pip install safetensors` +""" + +# docstyle-ignore +SENTENCEPIECE_IMPORT_ERROR = """ +{0} requires the sentencepiece library but it was not found in your environment. You can install it with pip: `pip install sentencepiece` +""" + + +# docstyle-ignore +BITSANDBYTES_IMPORT_ERROR = """ +{0} requires the bitsandbytes library but it was not found in your environment. You can install it with pip: `pip install bitsandbytes` +""" + +# docstyle-ignore +IMAGEIO_IMPORT_ERROR = """ +{0} requires the imageio library and ffmpeg but it was not found in your environment. You can install it with pip: `pip install imageio imageio-ffmpeg` +""" + +# docstyle-ignore +GGUF_IMPORT_ERROR = """ +{0} requires the gguf library but it was not found in your environment. You can install it with pip: `pip install gguf` +""" + +TORCHAO_IMPORT_ERROR = """ +{0} requires the torchao library but it was not found in your environment. You can install it with pip: `pip install +torchao` +""" + +QUANTO_IMPORT_ERROR = """ +{0} requires the optimum-quanto library but it was not found in your environment. You can install it with pip: `pip +install optimum-quanto` +""" + +# docstyle-ignore +PYTORCH_RETINAFACE_IMPORT_ERROR = """ +{0} requires the pytorch_retinaface library but it was not found in your environment. You can install it with pip: `pip install pytorch_retinaface` +""" + +# docstyle-ignore +BETTER_PROFANITY_IMPORT_ERROR = """ +{0} requires the better_profanity library but it was not found in your environment. You can install it with pip: `pip install better_profanity` +""" + +# docstyle-ignore +NLTK_IMPORT_ERROR = """ +{0} requires the nltk library but it was not found in your environment. You can install it with pip: `pip install nltk` +""" + + +BACKENDS_MAPPING = OrderedDict( + [ + ("bs4", (is_bs4_available, BS4_IMPORT_ERROR)), + ("flax", (is_flax_available, FLAX_IMPORT_ERROR)), + ("inflect", (is_inflect_available, INFLECT_IMPORT_ERROR)), + ("onnx", (is_onnx_available, ONNX_IMPORT_ERROR)), + ("opencv", (is_opencv_available, OPENCV_IMPORT_ERROR)), + ("scipy", (is_scipy_available, SCIPY_IMPORT_ERROR)), + ("torch", (is_torch_available, PYTORCH_IMPORT_ERROR)), + ("transformers", (is_transformers_available, TRANSFORMERS_IMPORT_ERROR)), + ("unidecode", (is_unidecode_available, UNIDECODE_IMPORT_ERROR)), + ("librosa", (is_librosa_available, LIBROSA_IMPORT_ERROR)), + ("k_diffusion", (is_k_diffusion_available, K_DIFFUSION_IMPORT_ERROR)), + ("note_seq", (is_note_seq_available, NOTE_SEQ_IMPORT_ERROR)), + ("wandb", (is_wandb_available, WANDB_IMPORT_ERROR)), + ("tensorboard", (is_tensorboard_available, TENSORBOARD_IMPORT_ERROR)), + ("compel", (is_compel_available, COMPEL_IMPORT_ERROR)), + ("ftfy", (is_ftfy_available, FTFY_IMPORT_ERROR)), + ("torchsde", (is_torchsde_available, TORCHSDE_IMPORT_ERROR)), + ("invisible_watermark", (is_invisible_watermark_available, INVISIBLE_WATERMARK_IMPORT_ERROR)), + ("peft", (is_peft_available, PEFT_IMPORT_ERROR)), + ("safetensors", (is_safetensors_available, SAFETENSORS_IMPORT_ERROR)), + ("bitsandbytes", (is_bitsandbytes_available, BITSANDBYTES_IMPORT_ERROR)), + ("sentencepiece", (is_sentencepiece_available, SENTENCEPIECE_IMPORT_ERROR)), + ("imageio", (is_imageio_available, IMAGEIO_IMPORT_ERROR)), + ("gguf", (is_gguf_available, GGUF_IMPORT_ERROR)), + ("torchao", (is_torchao_available, TORCHAO_IMPORT_ERROR)), + ("quanto", (is_optimum_quanto_available, QUANTO_IMPORT_ERROR)), + ("pytorch_retinaface", (is_pytorch_retinaface_available, PYTORCH_RETINAFACE_IMPORT_ERROR)), + ("better_profanity", (is_better_profanity_available, BETTER_PROFANITY_IMPORT_ERROR)), + ("nltk", (is_nltk_available, NLTK_IMPORT_ERROR)), + ] +) + + +def requires_backends(obj, backends): + if not isinstance(backends, (list, tuple)): + backends = [backends] + + name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__ + checks = (BACKENDS_MAPPING[backend] for backend in backends) + failed = [msg.format(name) for available, msg in checks if not available()] + if failed: + raise ImportError("".join(failed)) + + if name in [ + "VersatileDiffusionTextToImagePipeline", + "VersatileDiffusionPipeline", + "VersatileDiffusionDualGuidedPipeline", + "StableDiffusionImageVariationPipeline", + "UnCLIPPipeline", + ] and is_transformers_version("<", "4.25.0"): + raise ImportError( + f"You need to install `transformers>=4.25` in order to use {name}: \n```\n pip install" + " --upgrade transformers \n```" + ) + + if name in ["StableDiffusionDepth2ImgPipeline", "StableDiffusionPix2PixZeroPipeline"] and is_transformers_version( + "<", "4.26.0" + ): + raise ImportError( + f"You need to install `transformers>=4.26` in order to use {name}: \n```\n pip install" + " --upgrade transformers \n```" + ) + + +class DummyObject(type): + """ + Metaclass for the dummy objects. Any class inheriting from it will return the ImportError generated by + `requires_backend` each time a user tries to access any method of that class. + """ + + def __getattr__(cls, key): + if key.startswith("_") and key not in ["_load_connected_pipes", "_is_onnx"]: + return super().__getattr__(cls, key) + requires_backends(cls, cls._backends) + + +# This function was copied from: https://github.com/huggingface/accelerate/blob/874c4967d94badd24f893064cc3bef45f57cadf7/src/accelerate/utils/versions.py#L319 +def compare_versions(library_or_version: Union[str, Version], operation: str, requirement_version: str): + """ + Compares a library version to some requirement using a given operation. + + Args: + library_or_version (`str` or `packaging.version.Version`): + A library name or a version to check. + operation (`str`): + A string representation of an operator, such as `">"` or `"<="`. + requirement_version (`str`): + The version to compare the library version against + """ + if operation not in STR_OPERATION_TO_FUNC.keys(): + raise ValueError(f"`operation` must be one of {list(STR_OPERATION_TO_FUNC.keys())}, received {operation}") + operation = STR_OPERATION_TO_FUNC[operation] + if isinstance(library_or_version, str): + library_or_version = parse(importlib_metadata.version(library_or_version)) + return operation(library_or_version, parse(requirement_version)) + + +# This function was copied from: https://github.com/huggingface/accelerate/blob/874c4967d94badd24f893064cc3bef45f57cadf7/src/accelerate/utils/versions.py#L338 +def is_torch_version(operation: str, version: str): + """ + Compares the current PyTorch version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A string version of PyTorch + """ + return compare_versions(parse(_torch_version), operation, version) + + +def is_torch_xla_version(operation: str, version: str): + """ + Compares the current torch_xla version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A string version of torch_xla + """ + if not is_torch_xla_available: + return False + return compare_versions(parse(_torch_xla_version), operation, version) + + +def is_transformers_version(operation: str, version: str): + """ + Compares the current Transformers version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _transformers_available: + return False + return compare_versions(parse(_transformers_version), operation, version) + + +def is_hf_hub_version(operation: str, version: str): + """ + Compares the current Hugging Face Hub version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _hf_hub_available: + return False + return compare_versions(parse(_hf_hub_version), operation, version) + + +def is_accelerate_version(operation: str, version: str): + """ + Compares the current Accelerate version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _accelerate_available: + return False + return compare_versions(parse(_accelerate_version), operation, version) + + +def is_peft_version(operation: str, version: str): + """ + Compares the current PEFT version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _peft_available: + return False + return compare_versions(parse(_peft_version), operation, version) + + +def is_bitsandbytes_version(operation: str, version: str): + """ + Args: + Compares the current bitsandbytes version to a given reference with an operation. + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _bitsandbytes_available: + return False + return compare_versions(parse(_bitsandbytes_version), operation, version) + + +def is_gguf_version(operation: str, version: str): + """ + Compares the current Accelerate version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _gguf_available: + return False + return compare_versions(parse(_gguf_version), operation, version) + + +def is_torchao_version(operation: str, version: str): + """ + Compares the current torchao version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _torchao_available: + return False + return compare_versions(parse(_torchao_version), operation, version) + + +def is_k_diffusion_version(operation: str, version: str): + """ + Compares the current k-diffusion version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _k_diffusion_available: + return False + return compare_versions(parse(_k_diffusion_version), operation, version) + + +def is_optimum_quanto_version(operation: str, version: str): + """ + Compares the current Accelerate version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _optimum_quanto_available: + return False + return compare_versions(parse(_optimum_quanto_version), operation, version) + + +def get_objects_from_module(module): + """ + Returns a dict of object names and values in a module, while skipping private/internal objects + + Args: + module (ModuleType): + Module to extract the objects from. + + Returns: + dict: Dictionary of object names and corresponding values + """ + + objects = {} + for name in dir(module): + if name.startswith("_"): + continue + objects[name] = getattr(module, name) + + return objects + + +class OptionalDependencyNotAvailable(BaseException): + """ + An error indicating that an optional dependency of Diffusers was not found in the environment. + """ + + +class _LazyModule(ModuleType): + """ + Module class that surfaces all objects but only performs associated imports when the objects are requested. + """ + + # Very heavily inspired by optuna.integration._IntegrationModule + # https://github.com/optuna/optuna/blob/master/optuna/integration/__init__.py + def __init__(self, name, module_file, import_structure, module_spec=None, extra_objects=None): + super().__init__(name) + self._modules = set(import_structure.keys()) + self._class_to_module = {} + for key, values in import_structure.items(): + for value in values: + self._class_to_module[value] = key + # Needed for autocompletion in an IDE + self.__all__ = list(import_structure.keys()) + list(chain(*import_structure.values())) + self.__file__ = module_file + self.__spec__ = module_spec + self.__path__ = [os.path.dirname(module_file)] + self._objects = {} if extra_objects is None else extra_objects + self._name = name + self._import_structure = import_structure + + # Needed for autocompletion in an IDE + def __dir__(self): + result = super().__dir__() + # The elements of self.__all__ that are submodules may or may not be in the dir already, depending on whether + # they have been accessed or not. So we only add the elements of self.__all__ that are not already in the dir. + for attr in self.__all__: + if attr not in result: + result.append(attr) + return result + + def __getattr__(self, name: str) -> Any: + if name in self._objects: + return self._objects[name] + if name in self._modules: + value = self._get_module(name) + elif name in self._class_to_module.keys(): + module = self._get_module(self._class_to_module[name]) + value = getattr(module, name) + else: + raise AttributeError(f"module {self.__name__} has no attribute {name}") + + setattr(self, name, value) + return value + + def _get_module(self, module_name: str): + try: + return importlib.import_module("." + module_name, self.__name__) + except Exception as e: + raise RuntimeError( + f"Failed to import {self.__name__}.{module_name} because of the following error (look up to see its" + f" traceback):\n{e}" + ) from e + + def __reduce__(self): + return (self.__class__, (self._name, self.__file__, self._import_structure)) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/loading_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/loading_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..dd23ae73c86175a5736af5f98d27bcd3c7a63859 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/loading_utils.py @@ -0,0 +1,163 @@ +import os +import tempfile +from typing import Any, Callable, List, Optional, Tuple, Union +from urllib.parse import unquote, urlparse + +import PIL.Image +import PIL.ImageOps +import requests + +from .constants import DIFFUSERS_REQUEST_TIMEOUT +from .import_utils import BACKENDS_MAPPING, is_imageio_available + + +def load_image( + image: Union[str, PIL.Image.Image], convert_method: Optional[Callable[[PIL.Image.Image], PIL.Image.Image]] = None +) -> PIL.Image.Image: + """ + Loads `image` to a PIL Image. + + Args: + image (`str` or `PIL.Image.Image`): + The image to convert to the PIL Image format. + convert_method (Callable[[PIL.Image.Image], PIL.Image.Image], *optional*): + A conversion method to apply to the image after loading it. When set to `None` the image will be converted + "RGB". + + Returns: + `PIL.Image.Image`: + A PIL Image. + """ + if isinstance(image, str): + if image.startswith("http://") or image.startswith("https://"): + image = PIL.Image.open(requests.get(image, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw) + elif os.path.isfile(image): + image = PIL.Image.open(image) + else: + raise ValueError( + f"Incorrect path or URL. URLs must start with `http://` or `https://`, and {image} is not a valid path." + ) + elif isinstance(image, PIL.Image.Image): + image = image + else: + raise ValueError( + "Incorrect format used for the image. Should be a URL linking to an image, a local path, or a PIL image." + ) + + image = PIL.ImageOps.exif_transpose(image) + + if convert_method is not None: + image = convert_method(image) + else: + image = image.convert("RGB") + + return image + + +def load_video( + video: str, + convert_method: Optional[Callable[[List[PIL.Image.Image]], List[PIL.Image.Image]]] = None, +) -> List[PIL.Image.Image]: + """ + Loads `video` to a list of PIL Image. + + Args: + video (`str`): + A URL or Path to a video to convert to a list of PIL Image format. + convert_method (Callable[[List[PIL.Image.Image]], List[PIL.Image.Image]], *optional*): + A conversion method to apply to the video after loading it. When set to `None` the images will be converted + to "RGB". + + Returns: + `List[PIL.Image.Image]`: + The video as a list of PIL images. + """ + is_url = video.startswith("http://") or video.startswith("https://") + is_file = os.path.isfile(video) + was_tempfile_created = False + + if not (is_url or is_file): + raise ValueError( + f"Incorrect path or URL. URLs must start with `http://` or `https://`, and {video} is not a valid path." + ) + + if is_url: + response = requests.get(video, stream=True) + if response.status_code != 200: + raise ValueError(f"Failed to download video. Status code: {response.status_code}") + + parsed_url = urlparse(video) + file_name = os.path.basename(unquote(parsed_url.path)) + + suffix = os.path.splitext(file_name)[1] or ".mp4" + video_path = tempfile.NamedTemporaryFile(suffix=suffix, delete=False).name + + was_tempfile_created = True + + video_data = response.iter_content(chunk_size=8192) + with open(video_path, "wb") as f: + for chunk in video_data: + f.write(chunk) + + video = video_path + + pil_images = [] + if video.endswith(".gif"): + gif = PIL.Image.open(video) + try: + while True: + pil_images.append(gif.copy()) + gif.seek(gif.tell() + 1) + except EOFError: + pass + + else: + if is_imageio_available(): + import imageio + else: + raise ImportError(BACKENDS_MAPPING["imageio"][1].format("load_video")) + + try: + imageio.plugins.ffmpeg.get_exe() + except AttributeError: + raise AttributeError( + "`Unable to find an ffmpeg installation on your machine. Please install via `pip install imageio-ffmpeg" + ) + + with imageio.get_reader(video) as reader: + # Read all frames + for frame in reader: + pil_images.append(PIL.Image.fromarray(frame)) + + if was_tempfile_created: + os.remove(video_path) + + if convert_method is not None: + pil_images = convert_method(pil_images) + + return pil_images + + +# Taken from `transformers`. +def get_module_from_name(module, tensor_name: str) -> Tuple[Any, str]: + if "." in tensor_name: + splits = tensor_name.split(".") + for split in splits[:-1]: + new_module = getattr(module, split) + if new_module is None: + raise ValueError(f"{module} has no attribute {split}.") + module = new_module + tensor_name = splits[-1] + return module, tensor_name + + +def get_submodule_by_name(root_module, module_path: str): + current = root_module + parts = module_path.split(".") + for part in parts: + if part.isdigit(): + idx = int(part) + current = current[idx] # e.g., for nn.ModuleList or nn.Sequential + else: + current = getattr(current, part) + return current diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/logging.py b/venv/lib/python3.10/site-packages/diffusers/utils/logging.py new file mode 100644 index 0000000000000000000000000000000000000000..2ad6d3a476076dbc7f4d3e9ab1ab9f48de24d76f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/logging.py @@ -0,0 +1,340 @@ +# coding=utf-8 +# Copyright 2025 Optuna, Hugging Face +# +# 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 utilities.""" + +import logging +import os +import sys +import threading +from logging import ( + CRITICAL, # NOQA + DEBUG, # NOQA + ERROR, # NOQA + FATAL, # NOQA + INFO, # NOQA + NOTSET, # NOQA + WARN, # NOQA + WARNING, # NOQA +) +from typing import Dict, Optional + +from tqdm import auto as tqdm_lib + + +_lock = threading.Lock() +_default_handler: Optional[logging.Handler] = None + +log_levels = { + "debug": logging.DEBUG, + "info": logging.INFO, + "warning": logging.WARNING, + "error": logging.ERROR, + "critical": logging.CRITICAL, +} + +_default_log_level = logging.WARNING + +_tqdm_active = True + + +def _get_default_logging_level() -> int: + """ + If DIFFUSERS_VERBOSITY env var is set to one of the valid choices return that as the new default level. If it is + not - fall back to `_default_log_level` + """ + env_level_str = os.getenv("DIFFUSERS_VERBOSITY", None) + if env_level_str: + if env_level_str in log_levels: + return log_levels[env_level_str] + else: + logging.getLogger().warning( + f"Unknown option DIFFUSERS_VERBOSITY={env_level_str}, has to be one of: {', '.join(log_levels.keys())}" + ) + return _default_log_level + + +def _get_library_name() -> str: + return __name__.split(".")[0] + + +def _get_library_root_logger() -> logging.Logger: + return logging.getLogger(_get_library_name()) + + +def _configure_library_root_logger() -> None: + global _default_handler + + with _lock: + if _default_handler: + # This library has already configured the library root logger. + return + _default_handler = logging.StreamHandler() # Set sys.stderr as stream. + + if sys.stderr: # only if sys.stderr exists, e.g. when not using pythonw in windows + _default_handler.flush = sys.stderr.flush + + # Apply our default configuration to the library root logger. + library_root_logger = _get_library_root_logger() + library_root_logger.addHandler(_default_handler) + library_root_logger.setLevel(_get_default_logging_level()) + library_root_logger.propagate = False + + +def _reset_library_root_logger() -> None: + global _default_handler + + with _lock: + if not _default_handler: + return + + library_root_logger = _get_library_root_logger() + library_root_logger.removeHandler(_default_handler) + library_root_logger.setLevel(logging.NOTSET) + _default_handler = None + + +def get_log_levels_dict() -> Dict[str, int]: + return log_levels + + +def get_logger(name: Optional[str] = None) -> logging.Logger: + """ + Return a logger with the specified name. + + This function is not supposed to be directly accessed unless you are writing a custom diffusers module. + """ + + if name is None: + name = _get_library_name() + + _configure_library_root_logger() + return logging.getLogger(name) + + +def get_verbosity() -> int: + """ + Return the current level for the 🤗 Diffusers' root logger as an `int`. + + Returns: + `int`: + Logging level integers which can be one of: + + - `50`: `diffusers.logging.CRITICAL` or `diffusers.logging.FATAL` + - `40`: `diffusers.logging.ERROR` + - `30`: `diffusers.logging.WARNING` or `diffusers.logging.WARN` + - `20`: `diffusers.logging.INFO` + - `10`: `diffusers.logging.DEBUG` + + """ + + _configure_library_root_logger() + return _get_library_root_logger().getEffectiveLevel() + + +def set_verbosity(verbosity: int) -> None: + """ + Set the verbosity level for the 🤗 Diffusers' root logger. + + Args: + verbosity (`int`): + Logging level which can be one of: + + - `diffusers.logging.CRITICAL` or `diffusers.logging.FATAL` + - `diffusers.logging.ERROR` + - `diffusers.logging.WARNING` or `diffusers.logging.WARN` + - `diffusers.logging.INFO` + - `diffusers.logging.DEBUG` + """ + + _configure_library_root_logger() + _get_library_root_logger().setLevel(verbosity) + + +def set_verbosity_info() -> None: + """Set the verbosity to the `INFO` level.""" + return set_verbosity(INFO) + + +def set_verbosity_warning() -> None: + """Set the verbosity to the `WARNING` level.""" + return set_verbosity(WARNING) + + +def set_verbosity_debug() -> None: + """Set the verbosity to the `DEBUG` level.""" + return set_verbosity(DEBUG) + + +def set_verbosity_error() -> None: + """Set the verbosity to the `ERROR` level.""" + return set_verbosity(ERROR) + + +def disable_default_handler() -> None: + """Disable the default handler of the 🤗 Diffusers' root logger.""" + + _configure_library_root_logger() + + assert _default_handler is not None + _get_library_root_logger().removeHandler(_default_handler) + + +def enable_default_handler() -> None: + """Enable the default handler of the 🤗 Diffusers' root logger.""" + + _configure_library_root_logger() + + assert _default_handler is not None + _get_library_root_logger().addHandler(_default_handler) + + +def add_handler(handler: logging.Handler) -> None: + """adds a handler to the HuggingFace Diffusers' root logger.""" + + _configure_library_root_logger() + + assert handler is not None + _get_library_root_logger().addHandler(handler) + + +def remove_handler(handler: logging.Handler) -> None: + """removes given handler from the HuggingFace Diffusers' root logger.""" + + _configure_library_root_logger() + + assert handler is not None and handler in _get_library_root_logger().handlers + _get_library_root_logger().removeHandler(handler) + + +def disable_propagation() -> None: + """ + Disable propagation of the library log outputs. Note that log propagation is disabled by default. + """ + + _configure_library_root_logger() + _get_library_root_logger().propagate = False + + +def enable_propagation() -> None: + """ + Enable propagation of the library log outputs. Please disable the HuggingFace Diffusers' default handler to prevent + double logging if the root logger has been configured. + """ + + _configure_library_root_logger() + _get_library_root_logger().propagate = True + + +def enable_explicit_format() -> None: + """ + Enable explicit formatting for every 🤗 Diffusers' logger. The explicit formatter is as follows: + ``` + [LEVELNAME|FILENAME|LINE NUMBER] TIME >> MESSAGE + ``` + All handlers currently bound to the root logger are affected by this method. + """ + handlers = _get_library_root_logger().handlers + + for handler in handlers: + formatter = logging.Formatter("[%(levelname)s|%(filename)s:%(lineno)s] %(asctime)s >> %(message)s") + handler.setFormatter(formatter) + + +def reset_format() -> None: + """ + Resets the formatting for 🤗 Diffusers' loggers. + + All handlers currently bound to the root logger are affected by this method. + """ + handlers = _get_library_root_logger().handlers + + for handler in handlers: + handler.setFormatter(None) + + +def warning_advice(self, *args, **kwargs) -> None: + """ + This method is identical to `logger.warning()`, but if env var DIFFUSERS_NO_ADVISORY_WARNINGS=1 is set, this + warning will not be printed + """ + no_advisory_warnings = os.getenv("DIFFUSERS_NO_ADVISORY_WARNINGS", False) + if no_advisory_warnings: + return + self.warning(*args, **kwargs) + + +logging.Logger.warning_advice = warning_advice + + +class EmptyTqdm: + """Dummy tqdm which doesn't do anything.""" + + def __init__(self, *args, **kwargs): # pylint: disable=unused-argument + self._iterator = args[0] if args else None + + def __iter__(self): + return iter(self._iterator) + + def __getattr__(self, _): + """Return empty function.""" + + def empty_fn(*args, **kwargs): # pylint: disable=unused-argument + return + + return empty_fn + + def __enter__(self): + return self + + def __exit__(self, type_, value, traceback): + return + + +class _tqdm_cls: + def __call__(self, *args, **kwargs): + if _tqdm_active: + return tqdm_lib.tqdm(*args, **kwargs) + else: + return EmptyTqdm(*args, **kwargs) + + def set_lock(self, *args, **kwargs): + self._lock = None + if _tqdm_active: + return tqdm_lib.tqdm.set_lock(*args, **kwargs) + + def get_lock(self): + if _tqdm_active: + return tqdm_lib.tqdm.get_lock() + + +tqdm = _tqdm_cls() + + +def is_progress_bar_enabled() -> bool: + """Return a boolean indicating whether tqdm progress bars are enabled.""" + global _tqdm_active + return bool(_tqdm_active) + + +def enable_progress_bar() -> None: + """Enable tqdm progress bar.""" + global _tqdm_active + _tqdm_active = True + + +def disable_progress_bar() -> None: + """Disable tqdm progress bar.""" + global _tqdm_active + _tqdm_active = False diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/model_card_template.md b/venv/lib/python3.10/site-packages/diffusers/utils/model_card_template.md new file mode 100644 index 0000000000000000000000000000000000000000..f41b71e24e2081425d3049ae51b50036d5d28b6a --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/model_card_template.md @@ -0,0 +1,24 @@ +--- +{{ card_data }} +--- + + + +{{ model_description }} + +## Intended uses & limitations + +#### How to use + +```python +# TODO: add an example code snippet for running this diffusion pipeline +``` + +#### Limitations and bias + +[TODO: provide examples of latent issues and potential remediations] + +## Training details + +[TODO: describe the data used to train the model] diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/outputs.py b/venv/lib/python3.10/site-packages/diffusers/utils/outputs.py new file mode 100644 index 0000000000000000000000000000000000000000..35691496a182cacc209f300deb4fbf1b1a889cb6 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/outputs.py @@ -0,0 +1,138 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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. +""" +Generic utilities +""" + +from collections import OrderedDict +from dataclasses import fields, is_dataclass +from typing import Any, Tuple + +import numpy as np + +from .import_utils import is_torch_available, is_torch_version + + +def is_tensor(x) -> bool: + """ + Tests if `x` is a `torch.Tensor` or `np.ndarray`. + """ + if is_torch_available(): + import torch + + if isinstance(x, torch.Tensor): + return True + + return isinstance(x, np.ndarray) + + +class BaseOutput(OrderedDict): + """ + Base class for all model outputs as dataclass. Has a `__getitem__` that allows indexing by integer or slice (like a + tuple) or strings (like a dictionary) that will ignore the `None` attributes. Otherwise behaves like a regular + Python dictionary. + + + + You can't unpack a [`BaseOutput`] directly. Use the [`~utils.BaseOutput.to_tuple`] method to convert it to a tuple + first. + + + """ + + def __init_subclass__(cls) -> None: + """Register subclasses as pytree nodes. + + This is necessary to synchronize gradients when using `torch.nn.parallel.DistributedDataParallel` with + `static_graph=True` with modules that output `ModelOutput` subclasses. + """ + if is_torch_available(): + import torch.utils._pytree + + if is_torch_version("<", "2.2"): + torch.utils._pytree._register_pytree_node( + cls, + torch.utils._pytree._dict_flatten, + lambda values, context: cls(**torch.utils._pytree._dict_unflatten(values, context)), + ) + else: + torch.utils._pytree.register_pytree_node( + cls, + torch.utils._pytree._dict_flatten, + lambda values, context: cls(**torch.utils._pytree._dict_unflatten(values, context)), + serialized_type_name=f"{cls.__module__}.{cls.__name__}", + ) + + def __post_init__(self) -> None: + class_fields = fields(self) + + # Safety and consistency checks + if not len(class_fields): + raise ValueError(f"{self.__class__.__name__} has no fields.") + + first_field = getattr(self, class_fields[0].name) + other_fields_are_none = all(getattr(self, field.name) is None for field in class_fields[1:]) + + if other_fields_are_none and isinstance(first_field, dict): + for key, value in first_field.items(): + self[key] = value + else: + for field in class_fields: + v = getattr(self, field.name) + if v is not None: + self[field.name] = v + + def __delitem__(self, *args, **kwargs): + raise Exception(f"You cannot use ``__delitem__`` on a {self.__class__.__name__} instance.") + + def setdefault(self, *args, **kwargs): + raise Exception(f"You cannot use ``setdefault`` on a {self.__class__.__name__} instance.") + + def pop(self, *args, **kwargs): + raise Exception(f"You cannot use ``pop`` on a {self.__class__.__name__} instance.") + + def update(self, *args, **kwargs): + raise Exception(f"You cannot use ``update`` on a {self.__class__.__name__} instance.") + + def __getitem__(self, k: Any) -> Any: + if isinstance(k, str): + inner_dict = dict(self.items()) + return inner_dict[k] + else: + return self.to_tuple()[k] + + def __setattr__(self, name: Any, value: Any) -> None: + if name in self.keys() and value is not None: + # Don't call self.__setitem__ to avoid recursion errors + super().__setitem__(name, value) + super().__setattr__(name, value) + + def __setitem__(self, key, value): + # Will raise a KeyException if needed + super().__setitem__(key, value) + # Don't call self.__setattr__ to avoid recursion errors + super().__setattr__(key, value) + + def __reduce__(self): + if not is_dataclass(self): + return super().__reduce__() + callable, _args, *remaining = super().__reduce__() + args = tuple(getattr(self, field.name) for field in fields(self)) + return callable, args, *remaining + + def to_tuple(self) -> Tuple[Any, ...]: + """ + Convert self to a tuple containing all the attributes/keys that are not `None`. + """ + return tuple(self[k] for k in self.keys()) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/peft_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/peft_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..3907bdd5b333e1f554a2ac70cad2827a41c6cf21 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/peft_utils.py @@ -0,0 +1,373 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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. +""" +PEFT utilities: Utilities related to peft library +""" + +import collections +import importlib +from typing import Optional + +from packaging import version + +from . import logging +from .import_utils import is_peft_available, is_peft_version, is_torch_available +from .torch_utils import empty_device_cache + + +logger = logging.get_logger(__name__) + +if is_torch_available(): + import torch + + +def recurse_remove_peft_layers(model): + r""" + Recursively replace all instances of `LoraLayer` with corresponding new layers in `model`. + """ + from peft.tuners.tuners_utils import BaseTunerLayer + + has_base_layer_pattern = False + for module in model.modules(): + if isinstance(module, BaseTunerLayer): + has_base_layer_pattern = hasattr(module, "base_layer") + break + + if has_base_layer_pattern: + from peft.utils import _get_submodules + + key_list = [key for key, _ in model.named_modules() if "lora" not in key] + for key in key_list: + try: + parent, target, target_name = _get_submodules(model, key) + except AttributeError: + continue + if hasattr(target, "base_layer"): + setattr(parent, target_name, target.get_base_layer()) + else: + # This is for backwards compatibility with PEFT <= 0.6.2. + # TODO can be removed once that PEFT version is no longer supported. + from peft.tuners.lora import LoraLayer + + for name, module in model.named_children(): + if len(list(module.children())) > 0: + ## compound module, go inside it + recurse_remove_peft_layers(module) + + module_replaced = False + + if isinstance(module, LoraLayer) and isinstance(module, torch.nn.Linear): + new_module = torch.nn.Linear( + module.in_features, + module.out_features, + bias=module.bias is not None, + ).to(module.weight.device) + new_module.weight = module.weight + if module.bias is not None: + new_module.bias = module.bias + + module_replaced = True + elif isinstance(module, LoraLayer) and isinstance(module, torch.nn.Conv2d): + new_module = torch.nn.Conv2d( + module.in_channels, + module.out_channels, + module.kernel_size, + module.stride, + module.padding, + module.dilation, + module.groups, + ).to(module.weight.device) + + new_module.weight = module.weight + if module.bias is not None: + new_module.bias = module.bias + + module_replaced = True + + if module_replaced: + setattr(model, name, new_module) + del module + + empty_device_cache() + return model + + +def scale_lora_layers(model, weight): + """ + Adjust the weightage given to the LoRA layers of the model. + + Args: + model (`torch.nn.Module`): + The model to scale. + weight (`float`): + The weight to be given to the LoRA layers. + """ + from peft.tuners.tuners_utils import BaseTunerLayer + + if weight == 1.0: + return + + for module in model.modules(): + if isinstance(module, BaseTunerLayer): + module.scale_layer(weight) + + +def unscale_lora_layers(model, weight: Optional[float] = None): + """ + Removes the previously passed weight given to the LoRA layers of the model. + + Args: + model (`torch.nn.Module`): + The model to scale. + weight (`float`, *optional*): + The weight to be given to the LoRA layers. If no scale is passed the scale of the lora layer will be + re-initialized to the correct value. If 0.0 is passed, we will re-initialize the scale with the correct + value. + """ + from peft.tuners.tuners_utils import BaseTunerLayer + + if weight is None or weight == 1.0: + return + + for module in model.modules(): + if isinstance(module, BaseTunerLayer): + if weight != 0: + module.unscale_layer(weight) + else: + for adapter_name in module.active_adapters: + # if weight == 0 unscale should re-set the scale to the original value. + module.set_scale(adapter_name, 1.0) + + +def get_peft_kwargs(rank_dict, network_alpha_dict, peft_state_dict, is_unet=True): + rank_pattern = {} + alpha_pattern = {} + r = lora_alpha = list(rank_dict.values())[0] + + if len(set(rank_dict.values())) > 1: + # get the rank occurring the most number of times + r = collections.Counter(rank_dict.values()).most_common()[0][0] + + # for modules with rank different from the most occurring rank, add it to the `rank_pattern` + rank_pattern = dict(filter(lambda x: x[1] != r, rank_dict.items())) + rank_pattern = {k.split(".lora_B.")[0]: v for k, v in rank_pattern.items()} + + if network_alpha_dict is not None and len(network_alpha_dict) > 0: + if len(set(network_alpha_dict.values())) > 1: + # get the alpha occurring the most number of times + lora_alpha = collections.Counter(network_alpha_dict.values()).most_common()[0][0] + + # for modules with alpha different from the most occurring alpha, add it to the `alpha_pattern` + alpha_pattern = dict(filter(lambda x: x[1] != lora_alpha, network_alpha_dict.items())) + if is_unet: + alpha_pattern = { + ".".join(k.split(".lora_A.")[0].split(".")).replace(".alpha", ""): v + for k, v in alpha_pattern.items() + } + else: + alpha_pattern = {".".join(k.split(".down.")[0].split(".")[:-1]): v for k, v in alpha_pattern.items()} + else: + lora_alpha = set(network_alpha_dict.values()).pop() + + # layer names without the Diffusers specific + target_modules = list({name.split(".lora")[0] for name in peft_state_dict.keys()}) + use_dora = any("lora_magnitude_vector" in k for k in peft_state_dict) + # for now we know that the "bias" keys are only associated with `lora_B`. + lora_bias = any("lora_B" in k and k.endswith(".bias") for k in peft_state_dict) + + lora_config_kwargs = { + "r": r, + "lora_alpha": lora_alpha, + "rank_pattern": rank_pattern, + "alpha_pattern": alpha_pattern, + "target_modules": target_modules, + "use_dora": use_dora, + "lora_bias": lora_bias, + } + return lora_config_kwargs + + +def get_adapter_name(model): + from peft.tuners.tuners_utils import BaseTunerLayer + + for module in model.modules(): + if isinstance(module, BaseTunerLayer): + return f"default_{len(module.r)}" + return "default_0" + + +def set_adapter_layers(model, enabled=True): + from peft.tuners.tuners_utils import BaseTunerLayer + + for module in model.modules(): + if isinstance(module, BaseTunerLayer): + # The recent version of PEFT needs to call `enable_adapters` instead + if hasattr(module, "enable_adapters"): + module.enable_adapters(enabled=enabled) + else: + module.disable_adapters = not enabled + + +def delete_adapter_layers(model, adapter_name): + from peft.tuners.tuners_utils import BaseTunerLayer + + for module in model.modules(): + if isinstance(module, BaseTunerLayer): + if hasattr(module, "delete_adapter"): + module.delete_adapter(adapter_name) + else: + raise ValueError( + "The version of PEFT you are using is not compatible, please use a version that is greater than 0.6.1" + ) + + # For transformers integration - we need to pop the adapter from the config + if getattr(model, "_hf_peft_config_loaded", False) and hasattr(model, "peft_config"): + model.peft_config.pop(adapter_name, None) + # In case all adapters are deleted, we need to delete the config + # and make sure to set the flag to False + if len(model.peft_config) == 0: + del model.peft_config + model._hf_peft_config_loaded = None + + +def set_weights_and_activate_adapters(model, adapter_names, weights): + from peft.tuners.tuners_utils import BaseTunerLayer + + def get_module_weight(weight_for_adapter, module_name): + if not isinstance(weight_for_adapter, dict): + # If weight_for_adapter is a single number, always return it. + return weight_for_adapter + + for layer_name, weight_ in weight_for_adapter.items(): + if layer_name in module_name: + return weight_ + + parts = module_name.split(".") + # e.g. key = "down_blocks.1.attentions.0" + key = f"{parts[0]}.{parts[1]}.attentions.{parts[3]}" + block_weight = weight_for_adapter.get(key, 1.0) + + return block_weight + + for module_name, module in model.named_modules(): + if isinstance(module, BaseTunerLayer): + # For backward compatibility with previous PEFT versions, set multiple active adapters + if hasattr(module, "set_adapter"): + module.set_adapter(adapter_names) + else: + module.active_adapter = adapter_names + + # Set the scaling weight for each adapter for this module + for adapter_name, weight in zip(adapter_names, weights): + module.set_scale(adapter_name, get_module_weight(weight, module_name)) + + +def check_peft_version(min_version: str) -> None: + r""" + Checks if the version of PEFT is compatible. + + Args: + version (`str`): + The version of PEFT to check against. + """ + if not is_peft_available(): + raise ValueError("PEFT is not installed. Please install it with `pip install peft`") + + is_peft_version_compatible = version.parse(importlib.metadata.version("peft")) > version.parse(min_version) + + if not is_peft_version_compatible: + raise ValueError( + f"The version of PEFT you are using is not compatible, please use a version that is greater" + f" than {min_version}" + ) + + +def _create_lora_config( + state_dict, + network_alphas, + metadata, + rank_pattern_dict, + is_unet: bool = True, +): + from peft import LoraConfig + + if metadata is not None: + lora_config_kwargs = metadata + else: + lora_config_kwargs = get_peft_kwargs( + rank_pattern_dict, network_alpha_dict=network_alphas, peft_state_dict=state_dict, is_unet=is_unet + ) + + _maybe_raise_error_for_ambiguous_keys(lora_config_kwargs) + + # Version checks for DoRA and lora_bias + if "use_dora" in lora_config_kwargs and lora_config_kwargs["use_dora"]: + if is_peft_version("<", "0.9.0"): + raise ValueError("DoRA requires PEFT >= 0.9.0. Please upgrade.") + + if "lora_bias" in lora_config_kwargs and lora_config_kwargs["lora_bias"]: + if is_peft_version("<=", "0.13.2"): + raise ValueError("lora_bias requires PEFT >= 0.14.0. Please upgrade.") + + try: + return LoraConfig(**lora_config_kwargs) + except TypeError as e: + raise TypeError("`LoraConfig` class could not be instantiated.") from e + + +def _maybe_raise_error_for_ambiguous_keys(config): + rank_pattern = config["rank_pattern"].copy() + target_modules = config["target_modules"] + + for key in list(rank_pattern.keys()): + # try to detect ambiguity + # `target_modules` can also be a str, in which case this loop would loop + # over the chars of the str. The technically correct way to match LoRA keys + # in PEFT is to use LoraModel._check_target_module_exists (lora_config, key). + # But this cuts it for now. + exact_matches = [mod for mod in target_modules if mod == key] + substring_matches = [mod for mod in target_modules if key in mod and mod != key] + + if exact_matches and substring_matches: + if is_peft_version("<", "0.14.1"): + raise ValueError( + "There are ambiguous keys present in this LoRA. To load it, please update your `peft` installation - `pip install -U peft`." + ) + + +def _maybe_warn_for_unhandled_keys(incompatible_keys, adapter_name): + warn_msg = "" + if incompatible_keys is not None: + # Check only for unexpected keys. + unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None) + if unexpected_keys: + lora_unexpected_keys = [k for k in unexpected_keys if "lora_" in k and adapter_name in k] + if lora_unexpected_keys: + warn_msg = ( + f"Loading adapter weights from state_dict led to unexpected keys found in the model:" + f" {', '.join(lora_unexpected_keys)}. " + ) + + # Filter missing keys specific to the current adapter. + missing_keys = getattr(incompatible_keys, "missing_keys", None) + if missing_keys: + lora_missing_keys = [k for k in missing_keys if "lora_" in k and adapter_name in k] + if lora_missing_keys: + warn_msg += ( + f"Loading adapter weights from state_dict led to missing keys in the model:" + f" {', '.join(lora_missing_keys)}." + ) + + if warn_msg: + logger.warning(warn_msg) diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/pil_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/pil_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..76678070b697c7d87fc3691d9bc5bb3bea83c5b1 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/pil_utils.py @@ -0,0 +1,67 @@ +from typing import List + +import PIL.Image +import PIL.ImageOps +from packaging import version +from PIL import Image + + +if version.parse(version.parse(PIL.__version__).base_version) >= version.parse("9.1.0"): + PIL_INTERPOLATION = { + "linear": PIL.Image.Resampling.BILINEAR, + "bilinear": PIL.Image.Resampling.BILINEAR, + "bicubic": PIL.Image.Resampling.BICUBIC, + "lanczos": PIL.Image.Resampling.LANCZOS, + "nearest": PIL.Image.Resampling.NEAREST, + } +else: + PIL_INTERPOLATION = { + "linear": PIL.Image.LINEAR, + "bilinear": PIL.Image.BILINEAR, + "bicubic": PIL.Image.BICUBIC, + "lanczos": PIL.Image.LANCZOS, + "nearest": PIL.Image.NEAREST, + } + + +def pt_to_pil(images): + """ + Convert a torch image to a PIL image. + """ + images = (images / 2 + 0.5).clamp(0, 1) + images = images.cpu().permute(0, 2, 3, 1).float().numpy() + images = numpy_to_pil(images) + return images + + +def numpy_to_pil(images): + """ + Convert a numpy image or a batch of images to a PIL image. + """ + if images.ndim == 3: + images = images[None, ...] + images = (images * 255).round().astype("uint8") + if images.shape[-1] == 1: + # special case for grayscale (single channel) images + pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images] + else: + pil_images = [Image.fromarray(image) for image in images] + + return pil_images + + +def make_image_grid(images: List[PIL.Image.Image], rows: int, cols: int, resize: int = None) -> PIL.Image.Image: + """ + Prepares a single grid of images. Useful for visualization purposes. + """ + assert len(images) == rows * cols + + if resize is not None: + images = [img.resize((resize, resize)) for img in images] + + w, h = images[0].size + grid = Image.new("RGB", size=(cols * w, rows * h)) + + for i, img in enumerate(images): + grid.paste(img, box=(i % cols * w, i // cols * h)) + return grid diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/remote_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/remote_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..6494dc14171a935cc7efc247f8b4870e76b575d8 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/remote_utils.py @@ -0,0 +1,425 @@ +# coding=utf-8 +# Copyright 2025 HuggingFace Inc. +# +# 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 io +import json +from typing import List, Literal, Optional, Union, cast + +import requests + +from .deprecation_utils import deprecate +from .import_utils import is_safetensors_available, is_torch_available + + +if is_torch_available(): + import torch + + from ..image_processor import VaeImageProcessor + from ..video_processor import VideoProcessor + + if is_safetensors_available(): + import safetensors.torch + + DTYPE_MAP = { + "float16": torch.float16, + "float32": torch.float32, + "bfloat16": torch.bfloat16, + "uint8": torch.uint8, + } + + +from PIL import Image + + +def detect_image_type(data: bytes) -> str: + if data.startswith(b"\xff\xd8"): + return "jpeg" + elif data.startswith(b"\x89PNG\r\n\x1a\n"): + return "png" + elif data.startswith(b"GIF87a") or data.startswith(b"GIF89a"): + return "gif" + elif data.startswith(b"BM"): + return "bmp" + return "unknown" + + +def check_inputs_decode( + endpoint: str, + tensor: "torch.Tensor", + processor: Optional[Union["VaeImageProcessor", "VideoProcessor"]] = None, + do_scaling: bool = True, + scaling_factor: Optional[float] = None, + shift_factor: Optional[float] = None, + output_type: Literal["mp4", "pil", "pt"] = "pil", + return_type: Literal["mp4", "pil", "pt"] = "pil", + image_format: Literal["png", "jpg"] = "jpg", + partial_postprocess: bool = False, + input_tensor_type: Literal["binary"] = "binary", + output_tensor_type: Literal["binary"] = "binary", + height: Optional[int] = None, + width: Optional[int] = None, +): + if tensor.ndim == 3 and height is None and width is None: + raise ValueError("`height` and `width` required for packed latents.") + if ( + output_type == "pt" + and return_type == "pil" + and not partial_postprocess + and not isinstance(processor, (VaeImageProcessor, VideoProcessor)) + ): + raise ValueError("`processor` is required.") + if do_scaling and scaling_factor is None: + deprecate( + "do_scaling", + "1.0.0", + "`do_scaling` is deprecated, pass `scaling_factor` and `shift_factor` if required.", + standard_warn=False, + ) + + +def postprocess_decode( + response: requests.Response, + processor: Optional[Union["VaeImageProcessor", "VideoProcessor"]] = None, + output_type: Literal["mp4", "pil", "pt"] = "pil", + return_type: Literal["mp4", "pil", "pt"] = "pil", + partial_postprocess: bool = False, +): + if output_type == "pt" or (output_type == "pil" and processor is not None): + output_tensor = response.content + parameters = response.headers + shape = json.loads(parameters["shape"]) + dtype = parameters["dtype"] + torch_dtype = DTYPE_MAP[dtype] + output_tensor = torch.frombuffer(bytearray(output_tensor), dtype=torch_dtype).reshape(shape) + if output_type == "pt": + if partial_postprocess: + if return_type == "pil": + output = [Image.fromarray(image.numpy()) for image in output_tensor] + if len(output) == 1: + output = output[0] + elif return_type == "pt": + output = output_tensor + else: + if processor is None or return_type == "pt": + output = output_tensor + else: + if isinstance(processor, VideoProcessor): + output = cast( + List[Image.Image], + processor.postprocess_video(output_tensor, output_type="pil")[0], + ) + else: + output = cast( + Image.Image, + processor.postprocess(output_tensor, output_type="pil")[0], + ) + elif output_type == "pil" and return_type == "pil" and processor is None: + output = Image.open(io.BytesIO(response.content)).convert("RGB") + detected_format = detect_image_type(response.content) + output.format = detected_format + elif output_type == "pil" and processor is not None: + if return_type == "pil": + output = [ + Image.fromarray(image) + for image in (output_tensor.permute(0, 2, 3, 1).float().numpy() * 255).round().astype("uint8") + ] + elif return_type == "pt": + output = output_tensor + elif output_type == "mp4" and return_type == "mp4": + output = response.content + return output + + +def prepare_decode( + tensor: "torch.Tensor", + processor: Optional[Union["VaeImageProcessor", "VideoProcessor"]] = None, + do_scaling: bool = True, + scaling_factor: Optional[float] = None, + shift_factor: Optional[float] = None, + output_type: Literal["mp4", "pil", "pt"] = "pil", + image_format: Literal["png", "jpg"] = "jpg", + partial_postprocess: bool = False, + height: Optional[int] = None, + width: Optional[int] = None, +): + headers = {} + parameters = { + "image_format": image_format, + "output_type": output_type, + "partial_postprocess": partial_postprocess, + "shape": list(tensor.shape), + "dtype": str(tensor.dtype).split(".")[-1], + } + if do_scaling and scaling_factor is not None: + parameters["scaling_factor"] = scaling_factor + if do_scaling and shift_factor is not None: + parameters["shift_factor"] = shift_factor + if do_scaling and scaling_factor is None: + parameters["do_scaling"] = do_scaling + elif do_scaling and scaling_factor is None and shift_factor is None: + parameters["do_scaling"] = do_scaling + if height is not None and width is not None: + parameters["height"] = height + parameters["width"] = width + headers["Content-Type"] = "tensor/binary" + headers["Accept"] = "tensor/binary" + if output_type == "pil" and image_format == "jpg" and processor is None: + headers["Accept"] = "image/jpeg" + elif output_type == "pil" and image_format == "png" and processor is None: + headers["Accept"] = "image/png" + elif output_type == "mp4": + headers["Accept"] = "text/plain" + tensor_data = safetensors.torch._tobytes(tensor, "tensor") + return {"data": tensor_data, "params": parameters, "headers": headers} + + +def remote_decode( + endpoint: str, + tensor: "torch.Tensor", + processor: Optional[Union["VaeImageProcessor", "VideoProcessor"]] = None, + do_scaling: bool = True, + scaling_factor: Optional[float] = None, + shift_factor: Optional[float] = None, + output_type: Literal["mp4", "pil", "pt"] = "pil", + return_type: Literal["mp4", "pil", "pt"] = "pil", + image_format: Literal["png", "jpg"] = "jpg", + partial_postprocess: bool = False, + input_tensor_type: Literal["binary"] = "binary", + output_tensor_type: Literal["binary"] = "binary", + height: Optional[int] = None, + width: Optional[int] = None, +) -> Union[Image.Image, List[Image.Image], bytes, "torch.Tensor"]: + """ + Hugging Face Hybrid Inference that allow running VAE decode remotely. + + Args: + endpoint (`str`): + Endpoint for Remote Decode. + tensor (`torch.Tensor`): + Tensor to be decoded. + processor (`VaeImageProcessor` or `VideoProcessor`, *optional*): + Used with `return_type="pt"`, and `return_type="pil"` for Video models. + do_scaling (`bool`, default `True`, *optional*): + **DEPRECATED**. **pass `scaling_factor`/`shift_factor` instead.** **still set + do_scaling=None/do_scaling=False for no scaling until option is removed** When `True` scaling e.g. `latents + / self.vae.config.scaling_factor` is applied remotely. If `False`, input must be passed with scaling + applied. + scaling_factor (`float`, *optional*): + Scaling is applied when passed e.g. [`latents / + self.vae.config.scaling_factor`](https://github.com/huggingface/diffusers/blob/7007febae5cff000d4df9059d9cf35133e8b2ca9/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py#L1083C37-L1083C77). + - SD v1: 0.18215 + - SD XL: 0.13025 + - Flux: 0.3611 + If `None`, input must be passed with scaling applied. + shift_factor (`float`, *optional*): + Shift is applied when passed e.g. `latents + self.vae.config.shift_factor`. + - Flux: 0.1159 + If `None`, input must be passed with scaling applied. + output_type (`"mp4"` or `"pil"` or `"pt", default `"pil"): + **Endpoint** output type. Subject to change. Report feedback on preferred type. + + `"mp4": Supported by video models. Endpoint returns `bytes` of video. `"pil"`: Supported by image and video + models. + Image models: Endpoint returns `bytes` of an image in `image_format`. Video models: Endpoint returns + `torch.Tensor` with partial `postprocessing` applied. + Requires `processor` as a flag (any `None` value will work). + `"pt"`: Support by image and video models. Endpoint returns `torch.Tensor`. + With `partial_postprocess=True` the tensor is postprocessed `uint8` image tensor. + + Recommendations: + `"pt"` with `partial_postprocess=True` is the smallest transfer for full quality. `"pt"` with + `partial_postprocess=False` is the most compatible with third party code. `"pil"` with + `image_format="jpg"` is the smallest transfer overall. + + return_type (`"mp4"` or `"pil"` or `"pt", default `"pil"): + **Function** return type. + + `"mp4": Function returns `bytes` of video. `"pil"`: Function returns `PIL.Image.Image`. + With `output_type="pil" no further processing is applied. With `output_type="pt" a `PIL.Image.Image` is + created. + `partial_postprocess=False` `processor` is required. `partial_postprocess=True` `processor` is + **not** required. + `"pt"`: Function returns `torch.Tensor`. + `processor` is **not** required. `partial_postprocess=False` tensor is `float16` or `bfloat16`, without + denormalization. `partial_postprocess=True` tensor is `uint8`, denormalized. + + image_format (`"png"` or `"jpg"`, default `jpg`): + Used with `output_type="pil"`. Endpoint returns `jpg` or `png`. + + partial_postprocess (`bool`, default `False`): + Used with `output_type="pt"`. `partial_postprocess=False` tensor is `float16` or `bfloat16`, without + denormalization. `partial_postprocess=True` tensor is `uint8`, denormalized. + + input_tensor_type (`"binary"`, default `"binary"`): + Tensor transfer type. + + output_tensor_type (`"binary"`, default `"binary"`): + Tensor transfer type. + + height (`int`, **optional**): + Required for `"packed"` latents. + + width (`int`, **optional**): + Required for `"packed"` latents. + + Returns: + output (`Image.Image` or `List[Image.Image]` or `bytes` or `torch.Tensor`). + """ + if input_tensor_type == "base64": + deprecate( + "input_tensor_type='base64'", + "1.0.0", + "input_tensor_type='base64' is deprecated. Using `binary`.", + standard_warn=False, + ) + input_tensor_type = "binary" + if output_tensor_type == "base64": + deprecate( + "output_tensor_type='base64'", + "1.0.0", + "output_tensor_type='base64' is deprecated. Using `binary`.", + standard_warn=False, + ) + output_tensor_type = "binary" + check_inputs_decode( + endpoint, + tensor, + processor, + do_scaling, + scaling_factor, + shift_factor, + output_type, + return_type, + image_format, + partial_postprocess, + input_tensor_type, + output_tensor_type, + height, + width, + ) + kwargs = prepare_decode( + tensor=tensor, + processor=processor, + do_scaling=do_scaling, + scaling_factor=scaling_factor, + shift_factor=shift_factor, + output_type=output_type, + image_format=image_format, + partial_postprocess=partial_postprocess, + height=height, + width=width, + ) + response = requests.post(endpoint, **kwargs) + if not response.ok: + raise RuntimeError(response.json()) + output = postprocess_decode( + response=response, + processor=processor, + output_type=output_type, + return_type=return_type, + partial_postprocess=partial_postprocess, + ) + return output + + +def check_inputs_encode( + endpoint: str, + image: Union["torch.Tensor", Image.Image], + scaling_factor: Optional[float] = None, + shift_factor: Optional[float] = None, +): + pass + + +def postprocess_encode( + response: requests.Response, +): + output_tensor = response.content + parameters = response.headers + shape = json.loads(parameters["shape"]) + dtype = parameters["dtype"] + torch_dtype = DTYPE_MAP[dtype] + output_tensor = torch.frombuffer(bytearray(output_tensor), dtype=torch_dtype).reshape(shape) + return output_tensor + + +def prepare_encode( + image: Union["torch.Tensor", Image.Image], + scaling_factor: Optional[float] = None, + shift_factor: Optional[float] = None, +): + headers = {} + parameters = {} + if scaling_factor is not None: + parameters["scaling_factor"] = scaling_factor + if shift_factor is not None: + parameters["shift_factor"] = shift_factor + if isinstance(image, torch.Tensor): + data = safetensors.torch._tobytes(image.contiguous(), "tensor") + parameters["shape"] = list(image.shape) + parameters["dtype"] = str(image.dtype).split(".")[-1] + else: + buffer = io.BytesIO() + image.save(buffer, format="PNG") + data = buffer.getvalue() + return {"data": data, "params": parameters, "headers": headers} + + +def remote_encode( + endpoint: str, + image: Union["torch.Tensor", Image.Image], + scaling_factor: Optional[float] = None, + shift_factor: Optional[float] = None, +) -> "torch.Tensor": + """ + Hugging Face Hybrid Inference that allow running VAE encode remotely. + + Args: + endpoint (`str`): + Endpoint for Remote Decode. + image (`torch.Tensor` or `PIL.Image.Image`): + Image to be encoded. + scaling_factor (`float`, *optional*): + Scaling is applied when passed e.g. [`latents * self.vae.config.scaling_factor`]. + - SD v1: 0.18215 + - SD XL: 0.13025 + - Flux: 0.3611 + If `None`, input must be passed with scaling applied. + shift_factor (`float`, *optional*): + Shift is applied when passed e.g. `latents - self.vae.config.shift_factor`. + - Flux: 0.1159 + If `None`, input must be passed with scaling applied. + + Returns: + output (`torch.Tensor`). + """ + check_inputs_encode( + endpoint, + image, + scaling_factor, + shift_factor, + ) + kwargs = prepare_encode( + image=image, + scaling_factor=scaling_factor, + shift_factor=shift_factor, + ) + response = requests.post(endpoint, **kwargs) + if not response.ok: + raise RuntimeError(response.json()) + output = postprocess_encode( + response=response, + ) + return output diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/source_code_parsing_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/source_code_parsing_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..5f94711c21d825f4baa0b32bbc37fd94c89aba03 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/source_code_parsing_utils.py @@ -0,0 +1,52 @@ +import ast +import importlib +import inspect +import textwrap + + +class ReturnNameVisitor(ast.NodeVisitor): + """Thanks to ChatGPT for pairing.""" + + def __init__(self): + self.return_names = [] + + def visit_Return(self, node): + # Check if the return value is a tuple. + if isinstance(node.value, ast.Tuple): + for elt in node.value.elts: + if isinstance(elt, ast.Name): + self.return_names.append(elt.id) + else: + try: + self.return_names.append(ast.unparse(elt)) + except Exception: + self.return_names.append(str(elt)) + else: + if isinstance(node.value, ast.Name): + self.return_names.append(node.value.id) + else: + try: + self.return_names.append(ast.unparse(node.value)) + except Exception: + self.return_names.append(str(node.value)) + self.generic_visit(node) + + def _determine_parent_module(self, cls): + from diffusers import DiffusionPipeline + from diffusers.models.modeling_utils import ModelMixin + + if issubclass(cls, DiffusionPipeline): + return "pipelines" + elif issubclass(cls, ModelMixin): + return "models" + else: + raise NotImplementedError + + def get_ast_tree(self, cls, attribute_name="encode_prompt"): + parent_module_name = self._determine_parent_module(cls) + main_module = importlib.import_module(f"diffusers.{parent_module_name}") + current_cls_module = getattr(main_module, cls.__name__) + source_code = inspect.getsource(getattr(current_cls_module, attribute_name)) + source_code = textwrap.dedent(source_code) + tree = ast.parse(source_code) + return tree diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/state_dict_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/state_dict_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..50bfce8b15eb09fc2289f7eefa1d497d97bdbd5b --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/state_dict_utils.py @@ -0,0 +1,366 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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. +""" +State dict utilities: utility methods for converting state dicts easily +""" + +import enum +import json + +from .import_utils import is_torch_available +from .logging import get_logger + + +if is_torch_available(): + import torch + + +logger = get_logger(__name__) + + +class StateDictType(enum.Enum): + """ + The mode to use when converting state dicts. + """ + + DIFFUSERS_OLD = "diffusers_old" + KOHYA_SS = "kohya_ss" + PEFT = "peft" + DIFFUSERS = "diffusers" + + +# We need to define a proper mapping for Unet since it uses different output keys than text encoder +# e.g. to_q_lora -> q_proj / to_q +UNET_TO_DIFFUSERS = { + ".to_out_lora.up": ".to_out.0.lora_B", + ".to_out_lora.down": ".to_out.0.lora_A", + ".to_q_lora.down": ".to_q.lora_A", + ".to_q_lora.up": ".to_q.lora_B", + ".to_k_lora.down": ".to_k.lora_A", + ".to_k_lora.up": ".to_k.lora_B", + ".to_v_lora.down": ".to_v.lora_A", + ".to_v_lora.up": ".to_v.lora_B", + ".lora.up": ".lora_B", + ".lora.down": ".lora_A", + ".to_out.lora_magnitude_vector": ".to_out.0.lora_magnitude_vector", +} + + +DIFFUSERS_TO_PEFT = { + ".q_proj.lora_linear_layer.up": ".q_proj.lora_B", + ".q_proj.lora_linear_layer.down": ".q_proj.lora_A", + ".k_proj.lora_linear_layer.up": ".k_proj.lora_B", + ".k_proj.lora_linear_layer.down": ".k_proj.lora_A", + ".v_proj.lora_linear_layer.up": ".v_proj.lora_B", + ".v_proj.lora_linear_layer.down": ".v_proj.lora_A", + ".out_proj.lora_linear_layer.up": ".out_proj.lora_B", + ".out_proj.lora_linear_layer.down": ".out_proj.lora_A", + ".lora_linear_layer.up": ".lora_B", + ".lora_linear_layer.down": ".lora_A", + "text_projection.lora.down.weight": "text_projection.lora_A.weight", + "text_projection.lora.up.weight": "text_projection.lora_B.weight", +} + +DIFFUSERS_OLD_TO_PEFT = { + ".to_q_lora.up": ".q_proj.lora_B", + ".to_q_lora.down": ".q_proj.lora_A", + ".to_k_lora.up": ".k_proj.lora_B", + ".to_k_lora.down": ".k_proj.lora_A", + ".to_v_lora.up": ".v_proj.lora_B", + ".to_v_lora.down": ".v_proj.lora_A", + ".to_out_lora.up": ".out_proj.lora_B", + ".to_out_lora.down": ".out_proj.lora_A", + ".lora_linear_layer.up": ".lora_B", + ".lora_linear_layer.down": ".lora_A", +} + +PEFT_TO_DIFFUSERS = { + ".q_proj.lora_B": ".q_proj.lora_linear_layer.up", + ".q_proj.lora_A": ".q_proj.lora_linear_layer.down", + ".k_proj.lora_B": ".k_proj.lora_linear_layer.up", + ".k_proj.lora_A": ".k_proj.lora_linear_layer.down", + ".v_proj.lora_B": ".v_proj.lora_linear_layer.up", + ".v_proj.lora_A": ".v_proj.lora_linear_layer.down", + ".out_proj.lora_B": ".out_proj.lora_linear_layer.up", + ".out_proj.lora_A": ".out_proj.lora_linear_layer.down", + "to_k.lora_A": "to_k.lora.down", + "to_k.lora_B": "to_k.lora.up", + "to_q.lora_A": "to_q.lora.down", + "to_q.lora_B": "to_q.lora.up", + "to_v.lora_A": "to_v.lora.down", + "to_v.lora_B": "to_v.lora.up", + "to_out.0.lora_A": "to_out.0.lora.down", + "to_out.0.lora_B": "to_out.0.lora.up", +} + +DIFFUSERS_OLD_TO_DIFFUSERS = { + ".to_q_lora.up": ".q_proj.lora_linear_layer.up", + ".to_q_lora.down": ".q_proj.lora_linear_layer.down", + ".to_k_lora.up": ".k_proj.lora_linear_layer.up", + ".to_k_lora.down": ".k_proj.lora_linear_layer.down", + ".to_v_lora.up": ".v_proj.lora_linear_layer.up", + ".to_v_lora.down": ".v_proj.lora_linear_layer.down", + ".to_out_lora.up": ".out_proj.lora_linear_layer.up", + ".to_out_lora.down": ".out_proj.lora_linear_layer.down", + ".to_k.lora_magnitude_vector": ".k_proj.lora_magnitude_vector", + ".to_v.lora_magnitude_vector": ".v_proj.lora_magnitude_vector", + ".to_q.lora_magnitude_vector": ".q_proj.lora_magnitude_vector", + ".to_out.lora_magnitude_vector": ".out_proj.lora_magnitude_vector", +} + +PEFT_TO_KOHYA_SS = { + "lora_A": "lora_down", + "lora_B": "lora_up", + # This is not a comprehensive dict as kohya format requires replacing `.` with `_` in keys, + # adding prefixes and adding alpha values + # Check `convert_state_dict_to_kohya` for more +} + +PEFT_STATE_DICT_MAPPINGS = { + StateDictType.DIFFUSERS_OLD: DIFFUSERS_OLD_TO_PEFT, + StateDictType.DIFFUSERS: DIFFUSERS_TO_PEFT, +} + +DIFFUSERS_STATE_DICT_MAPPINGS = { + StateDictType.DIFFUSERS_OLD: DIFFUSERS_OLD_TO_DIFFUSERS, + StateDictType.PEFT: PEFT_TO_DIFFUSERS, +} + +KOHYA_STATE_DICT_MAPPINGS = {StateDictType.PEFT: PEFT_TO_KOHYA_SS} + +KEYS_TO_ALWAYS_REPLACE = { + ".processor.": ".", +} + + +def convert_state_dict(state_dict, mapping): + r""" + Simply iterates over the state dict and replaces the patterns in `mapping` with the corresponding values. + + Args: + state_dict (`dict[str, torch.Tensor]`): + The state dict to convert. + mapping (`dict[str, str]`): + The mapping to use for conversion, the mapping should be a dictionary with the following structure: + - key: the pattern to replace + - value: the pattern to replace with + + Returns: + converted_state_dict (`dict`) + The converted state dict. + """ + converted_state_dict = {} + for k, v in state_dict.items(): + # First, filter out the keys that we always want to replace + for pattern in KEYS_TO_ALWAYS_REPLACE.keys(): + if pattern in k: + new_pattern = KEYS_TO_ALWAYS_REPLACE[pattern] + k = k.replace(pattern, new_pattern) + + for pattern in mapping.keys(): + if pattern in k: + new_pattern = mapping[pattern] + k = k.replace(pattern, new_pattern) + break + converted_state_dict[k] = v + return converted_state_dict + + +def convert_state_dict_to_peft(state_dict, original_type=None, **kwargs): + r""" + Converts a state dict to the PEFT format The state dict can be from previous diffusers format (`OLD_DIFFUSERS`), or + new diffusers format (`DIFFUSERS`). The method only supports the conversion from diffusers old/new to PEFT for now. + + Args: + state_dict (`dict[str, torch.Tensor]`): + The state dict to convert. + original_type (`StateDictType`, *optional*): + The original type of the state dict, if not provided, the method will try to infer it automatically. + """ + if original_type is None: + # Old diffusers to PEFT + if any("to_out_lora" in k for k in state_dict.keys()): + original_type = StateDictType.DIFFUSERS_OLD + elif any("lora_linear_layer" in k for k in state_dict.keys()): + original_type = StateDictType.DIFFUSERS + else: + raise ValueError("Could not automatically infer state dict type") + + if original_type not in PEFT_STATE_DICT_MAPPINGS.keys(): + raise ValueError(f"Original type {original_type} is not supported") + + mapping = PEFT_STATE_DICT_MAPPINGS[original_type] + return convert_state_dict(state_dict, mapping) + + +def convert_state_dict_to_diffusers(state_dict, original_type=None, **kwargs): + r""" + Converts a state dict to new diffusers format. The state dict can be from previous diffusers format + (`OLD_DIFFUSERS`), or PEFT format (`PEFT`) or new diffusers format (`DIFFUSERS`). In the last case the method will + return the state dict as is. + + The method only supports the conversion from diffusers old, PEFT to diffusers new for now. + + Args: + state_dict (`dict[str, torch.Tensor]`): + The state dict to convert. + original_type (`StateDictType`, *optional*): + The original type of the state dict, if not provided, the method will try to infer it automatically. + kwargs (`dict`, *args*): + Additional arguments to pass to the method. + + - **adapter_name**: For example, in case of PEFT, some keys will be prepended + with the adapter name, therefore needs a special handling. By default PEFT also takes care of that in + `get_peft_model_state_dict` method: + https://github.com/huggingface/peft/blob/ba0477f2985b1ba311b83459d29895c809404e99/src/peft/utils/save_and_load.py#L92 + but we add it here in case we don't want to rely on that method. + """ + peft_adapter_name = kwargs.pop("adapter_name", None) + if peft_adapter_name is not None: + peft_adapter_name = "." + peft_adapter_name + else: + peft_adapter_name = "" + + if original_type is None: + # Old diffusers to PEFT + if any("to_out_lora" in k for k in state_dict.keys()): + original_type = StateDictType.DIFFUSERS_OLD + elif any(f".lora_A{peft_adapter_name}.weight" in k for k in state_dict.keys()): + original_type = StateDictType.PEFT + elif any("lora_linear_layer" in k for k in state_dict.keys()): + # nothing to do + return state_dict + else: + raise ValueError("Could not automatically infer state dict type") + + if original_type not in DIFFUSERS_STATE_DICT_MAPPINGS.keys(): + raise ValueError(f"Original type {original_type} is not supported") + + mapping = DIFFUSERS_STATE_DICT_MAPPINGS[original_type] + return convert_state_dict(state_dict, mapping) + + +def convert_unet_state_dict_to_peft(state_dict): + r""" + Converts a state dict from UNet format to diffusers format - i.e. by removing some keys + """ + mapping = UNET_TO_DIFFUSERS + return convert_state_dict(state_dict, mapping) + + +def convert_all_state_dict_to_peft(state_dict): + r""" + Attempts to first `convert_state_dict_to_peft`, and if it doesn't detect `lora_linear_layer` for a valid + `DIFFUSERS` LoRA for example, attempts to exclusively convert the Unet `convert_unet_state_dict_to_peft` + """ + try: + peft_dict = convert_state_dict_to_peft(state_dict) + except Exception as e: + if str(e) == "Could not automatically infer state dict type": + peft_dict = convert_unet_state_dict_to_peft(state_dict) + else: + raise + + if not any("lora_A" in key or "lora_B" in key for key in peft_dict.keys()): + raise ValueError("Your LoRA was not converted to PEFT") + + return peft_dict + + +def convert_state_dict_to_kohya(state_dict, original_type=None, **kwargs): + r""" + Converts a `PEFT` state dict to `Kohya` format that can be used in AUTOMATIC1111, ComfyUI, SD.Next, InvokeAI, etc. + The method only supports the conversion from PEFT to Kohya for now. + + Args: + state_dict (`dict[str, torch.Tensor]`): + The state dict to convert. + original_type (`StateDictType`, *optional*): + The original type of the state dict, if not provided, the method will try to infer it automatically. + kwargs (`dict`, *args*): + Additional arguments to pass to the method. + + - **adapter_name**: For example, in case of PEFT, some keys will be prepended + with the adapter name, therefore needs a special handling. By default PEFT also takes care of that in + `get_peft_model_state_dict` method: + https://github.com/huggingface/peft/blob/ba0477f2985b1ba311b83459d29895c809404e99/src/peft/utils/save_and_load.py#L92 + but we add it here in case we don't want to rely on that method. + """ + try: + import torch + except ImportError: + logger.error("Converting PEFT state dicts to Kohya requires torch to be installed.") + raise + + peft_adapter_name = kwargs.pop("adapter_name", None) + if peft_adapter_name is not None: + peft_adapter_name = "." + peft_adapter_name + else: + peft_adapter_name = "" + + if original_type is None: + if any(f".lora_A{peft_adapter_name}.weight" in k for k in state_dict.keys()): + original_type = StateDictType.PEFT + + if original_type not in KOHYA_STATE_DICT_MAPPINGS.keys(): + raise ValueError(f"Original type {original_type} is not supported") + + # Use the convert_state_dict function with the appropriate mapping + kohya_ss_partial_state_dict = convert_state_dict(state_dict, KOHYA_STATE_DICT_MAPPINGS[StateDictType.PEFT]) + kohya_ss_state_dict = {} + + # Additional logic for replacing header, alpha parameters `.` with `_` in all keys + for kohya_key, weight in kohya_ss_partial_state_dict.items(): + if "text_encoder_2." in kohya_key: + kohya_key = kohya_key.replace("text_encoder_2.", "lora_te2.") + elif "text_encoder." in kohya_key: + kohya_key = kohya_key.replace("text_encoder.", "lora_te1.") + elif "unet" in kohya_key: + kohya_key = kohya_key.replace("unet", "lora_unet") + elif "lora_magnitude_vector" in kohya_key: + kohya_key = kohya_key.replace("lora_magnitude_vector", "dora_scale") + + kohya_key = kohya_key.replace(".", "_", kohya_key.count(".") - 2) + kohya_key = kohya_key.replace(peft_adapter_name, "") # Kohya doesn't take names + kohya_ss_state_dict[kohya_key] = weight + if "lora_down" in kohya_key: + alpha_key = f"{kohya_key.split('.')[0]}.alpha" + kohya_ss_state_dict[alpha_key] = torch.tensor(len(weight)) + + return kohya_ss_state_dict + + +def state_dict_all_zero(state_dict, filter_str=None): + if filter_str is not None: + if isinstance(filter_str, str): + filter_str = [filter_str] + state_dict = {k: v for k, v in state_dict.items() if any(f in k for f in filter_str)} + + return all(torch.all(param == 0).item() for param in state_dict.values()) + + +def _load_sft_state_dict_metadata(model_file: str): + import safetensors.torch + + from ..loaders.lora_base import LORA_ADAPTER_METADATA_KEY + + with safetensors.torch.safe_open(model_file, framework="pt", device="cpu") as f: + metadata = f.metadata() or {} + + metadata.pop("format", None) + if metadata: + raw = metadata.get(LORA_ADAPTER_METADATA_KEY) + return json.loads(raw) if raw else None + else: + return None diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/testing_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/testing_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..e5da39c1d8654aaedc7b81c78369bd21b2053c31 --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/testing_utils.py @@ -0,0 +1,1431 @@ +import functools +import importlib +import importlib.metadata +import inspect +import io +import logging +import multiprocessing +import os +import random +import re +import struct +import sys +import tempfile +import time +import unittest +import urllib.parse +from collections import UserDict +from contextlib import contextmanager +from io import BytesIO, StringIO +from pathlib import Path +from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Tuple, Union + +import numpy as np +import PIL.Image +import PIL.ImageOps +import requests +from numpy.linalg import norm +from packaging import version + +from .constants import DIFFUSERS_REQUEST_TIMEOUT +from .import_utils import ( + BACKENDS_MAPPING, + is_accelerate_available, + is_bitsandbytes_available, + is_compel_available, + is_flax_available, + is_gguf_available, + is_note_seq_available, + is_onnx_available, + is_opencv_available, + is_optimum_quanto_available, + is_peft_available, + is_timm_available, + is_torch_available, + is_torch_version, + is_torchao_available, + is_torchsde_available, + is_transformers_available, +) +from .logging import get_logger + + +if is_torch_available(): + import torch + + IS_ROCM_SYSTEM = torch.version.hip is not None + IS_CUDA_SYSTEM = torch.version.cuda is not None + IS_XPU_SYSTEM = getattr(torch.version, "xpu", None) is not None +else: + IS_ROCM_SYSTEM = False + IS_CUDA_SYSTEM = False + IS_XPU_SYSTEM = False + +global_rng = random.Random() + +logger = get_logger(__name__) + +_required_peft_version = is_peft_available() and version.parse( + version.parse(importlib.metadata.version("peft")).base_version +) > version.parse("0.5") +_required_transformers_version = is_transformers_available() and version.parse( + version.parse(importlib.metadata.version("transformers")).base_version +) > version.parse("4.33") + +USE_PEFT_BACKEND = _required_peft_version and _required_transformers_version +BIG_GPU_MEMORY = int(os.getenv("BIG_GPU_MEMORY", 40)) + +if is_torch_available(): + import torch + + # Set a backend environment variable for any extra module import required for a custom accelerator + if "DIFFUSERS_TEST_BACKEND" in os.environ: + backend = os.environ["DIFFUSERS_TEST_BACKEND"] + try: + _ = importlib.import_module(backend) + except ModuleNotFoundError as e: + raise ModuleNotFoundError( + f"Failed to import `DIFFUSERS_TEST_BACKEND` '{backend}'! This should be the name of an installed module \ + to enable a specified backend.):\n{e}" + ) from e + + if "DIFFUSERS_TEST_DEVICE" in os.environ: + torch_device = os.environ["DIFFUSERS_TEST_DEVICE"] + try: + # try creating device to see if provided device is valid + _ = torch.device(torch_device) + except RuntimeError as e: + raise RuntimeError( + f"Unknown testing device specified by environment variable `DIFFUSERS_TEST_DEVICE`: {torch_device}" + ) from e + logger.info(f"torch_device overrode to {torch_device}") + else: + if torch.cuda.is_available(): + torch_device = "cuda" + elif torch.xpu.is_available(): + torch_device = "xpu" + else: + torch_device = "cpu" + is_torch_higher_equal_than_1_12 = version.parse( + version.parse(torch.__version__).base_version + ) >= version.parse("1.12") + + if is_torch_higher_equal_than_1_12: + # Some builds of torch 1.12 don't have the mps backend registered. See #892 for more details + mps_backend_registered = hasattr(torch.backends, "mps") + torch_device = "mps" if (mps_backend_registered and torch.backends.mps.is_available()) else torch_device + + from .torch_utils import get_torch_cuda_device_capability + + +def torch_all_close(a, b, *args, **kwargs): + if not is_torch_available(): + raise ValueError("PyTorch needs to be installed to use this function.") + if not torch.allclose(a, b, *args, **kwargs): + assert False, f"Max diff is absolute {(a - b).abs().max()}. Diff tensor is {(a - b).abs()}." + return True + + +def numpy_cosine_similarity_distance(a, b): + similarity = np.dot(a, b) / (norm(a) * norm(b)) + distance = 1.0 - similarity.mean() + + return distance + + +def check_if_dicts_are_equal(dict1, dict2): + dict1, dict2 = dict1.copy(), dict2.copy() + + for key, value in dict1.items(): + if isinstance(value, set): + dict1[key] = sorted(value) + for key, value in dict2.items(): + if isinstance(value, set): + dict2[key] = sorted(value) + + for key in dict1: + if key not in dict2: + return False + if dict1[key] != dict2[key]: + return False + + for key in dict2: + if key not in dict1: + return False + + return True + + +def print_tensor_test( + tensor, + limit_to_slices=None, + max_torch_print=None, + filename="test_corrections.txt", + expected_tensor_name="expected_slice", +): + if max_torch_print: + torch.set_printoptions(threshold=10_000) + + test_name = os.environ.get("PYTEST_CURRENT_TEST") + if not torch.is_tensor(tensor): + tensor = torch.from_numpy(tensor) + if limit_to_slices: + tensor = tensor[0, -3:, -3:, -1] + + tensor_str = str(tensor.detach().cpu().flatten().to(torch.float32)).replace("\n", "") + # format is usually: + # expected_slice = np.array([-0.5713, -0.3018, -0.9814, 0.04663, -0.879, 0.76, -1.734, 0.1044, 1.161]) + output_str = tensor_str.replace("tensor", f"{expected_tensor_name} = np.array") + test_file, test_class, test_fn = test_name.split("::") + test_fn = test_fn.split()[0] + with open(filename, "a") as f: + print("::".join([test_file, test_class, test_fn, output_str]), file=f) + + +def get_tests_dir(append_path=None): + """ + Args: + append_path: optional path to append to the tests dir path + Return: + The full path to the `tests` dir, so that the tests can be invoked from anywhere. Optionally `append_path` is + joined after the `tests` dir the former is provided. + """ + # this function caller's __file__ + caller__file__ = inspect.stack()[1][1] + tests_dir = os.path.abspath(os.path.dirname(caller__file__)) + + while not tests_dir.endswith("tests"): + tests_dir = os.path.dirname(tests_dir) + + if append_path: + return Path(tests_dir, append_path).as_posix() + else: + return tests_dir + + +# Taken from the following PR: +# https://github.com/huggingface/accelerate/pull/1964 +def str_to_bool(value) -> int: + """ + Converts a string representation of truth to `True` (1) or `False` (0). True values are `y`, `yes`, `t`, `true`, + `on`, and `1`; False value are `n`, `no`, `f`, `false`, `off`, and `0`; + """ + value = value.lower() + if value in ("y", "yes", "t", "true", "on", "1"): + return 1 + elif value in ("n", "no", "f", "false", "off", "0"): + return 0 + else: + raise ValueError(f"invalid truth value {value}") + + +def parse_flag_from_env(key, default=False): + try: + value = os.environ[key] + except KeyError: + # KEY isn't set, default to `default`. + _value = default + else: + # KEY is set, convert it to True or False. + try: + _value = str_to_bool(value) + except ValueError: + # More values are supported, but let's keep the message simple. + raise ValueError(f"If set, {key} must be yes or no.") + return _value + + +_run_slow_tests = parse_flag_from_env("RUN_SLOW", default=False) +_run_nightly_tests = parse_flag_from_env("RUN_NIGHTLY", default=False) +_run_compile_tests = parse_flag_from_env("RUN_COMPILE", default=False) + + +def floats_tensor(shape, scale=1.0, rng=None, name=None): + """Creates a random float32 tensor""" + if rng is None: + rng = global_rng + + total_dims = 1 + for dim in shape: + total_dims *= dim + + values = [] + for _ in range(total_dims): + values.append(rng.random() * scale) + + return torch.tensor(data=values, dtype=torch.float).view(shape).contiguous() + + +def slow(test_case): + """ + Decorator marking a test as slow. + + Slow tests are skipped by default. Set the RUN_SLOW environment variable to a truthy value to run them. + + """ + return unittest.skipUnless(_run_slow_tests, "test is slow")(test_case) + + +def nightly(test_case): + """ + Decorator marking a test that runs nightly in the diffusers CI. + + Slow tests are skipped by default. Set the RUN_NIGHTLY environment variable to a truthy value to run them. + + """ + return unittest.skipUnless(_run_nightly_tests, "test is nightly")(test_case) + + +def is_torch_compile(test_case): + """ + Decorator marking a test that runs compile tests in the diffusers CI. + + Compile tests are skipped by default. Set the RUN_COMPILE environment variable to a truthy value to run them. + + """ + return unittest.skipUnless(_run_compile_tests, "test is torch compile")(test_case) + + +def require_torch(test_case): + """ + Decorator marking a test that requires PyTorch. These tests are skipped when PyTorch isn't installed. + """ + return unittest.skipUnless(is_torch_available(), "test requires PyTorch")(test_case) + + +def require_torch_2(test_case): + """ + Decorator marking a test that requires PyTorch 2. These tests are skipped when it isn't installed. + """ + return unittest.skipUnless(is_torch_available() and is_torch_version(">=", "2.0.0"), "test requires PyTorch 2")( + test_case + ) + + +def require_torch_version_greater_equal(torch_version): + """Decorator marking a test that requires torch with a specific version or greater.""" + + def decorator(test_case): + correct_torch_version = is_torch_available() and is_torch_version(">=", torch_version) + return unittest.skipUnless( + correct_torch_version, f"test requires torch with the version greater than or equal to {torch_version}" + )(test_case) + + return decorator + + +def require_torch_version_greater(torch_version): + """Decorator marking a test that requires torch with a specific version greater.""" + + def decorator(test_case): + correct_torch_version = is_torch_available() and is_torch_version(">", torch_version) + return unittest.skipUnless( + correct_torch_version, f"test requires torch with the version greater than {torch_version}" + )(test_case) + + return decorator + + +def require_torch_gpu(test_case): + """Decorator marking a test that requires CUDA and PyTorch.""" + return unittest.skipUnless(is_torch_available() and torch_device == "cuda", "test requires PyTorch+CUDA")( + test_case + ) + + +def require_torch_cuda_compatibility(expected_compute_capability): + def decorator(test_case): + if torch.cuda.is_available(): + current_compute_capability = get_torch_cuda_device_capability() + return unittest.skipUnless( + float(current_compute_capability) == float(expected_compute_capability), + "Test not supported for this compute capability.", + ) + + return decorator + + +# These decorators are for accelerator-specific behaviours that are not GPU-specific +def require_torch_accelerator(test_case): + """Decorator marking a test that requires an accelerator backend and PyTorch.""" + return unittest.skipUnless(is_torch_available() and torch_device != "cpu", "test requires accelerator+PyTorch")( + test_case + ) + + +def require_torch_multi_gpu(test_case): + """ + Decorator marking a test that requires a multi-GPU setup (in PyTorch). These tests are skipped on a machine without + multiple GPUs. To run *only* the multi_gpu tests, assuming all test names contain multi_gpu: $ pytest -sv ./tests + -k "multi_gpu" + """ + if not is_torch_available(): + return unittest.skip("test requires PyTorch")(test_case) + + import torch + + return unittest.skipUnless(torch.cuda.device_count() > 1, "test requires multiple GPUs")(test_case) + + +def require_torch_multi_accelerator(test_case): + """ + Decorator marking a test that requires a multi-accelerator setup (in PyTorch). These tests are skipped on a machine + without multiple hardware accelerators. + """ + if not is_torch_available(): + return unittest.skip("test requires PyTorch")(test_case) + + import torch + + return unittest.skipUnless( + torch.cuda.device_count() > 1 or torch.xpu.device_count() > 1, "test requires multiple hardware accelerators" + )(test_case) + + +def require_torch_accelerator_with_fp16(test_case): + """Decorator marking a test that requires an accelerator with support for the FP16 data type.""" + return unittest.skipUnless(_is_torch_fp16_available(torch_device), "test requires accelerator with fp16 support")( + test_case + ) + + +def require_torch_accelerator_with_fp64(test_case): + """Decorator marking a test that requires an accelerator with support for the FP64 data type.""" + return unittest.skipUnless(_is_torch_fp64_available(torch_device), "test requires accelerator with fp64 support")( + test_case + ) + + +def require_big_gpu_with_torch_cuda(test_case): + """ + Decorator marking a test that requires a bigger GPU (24GB) for execution. Some example pipelines: Flux, SD3, Cog, + etc. + """ + if not is_torch_available(): + return unittest.skip("test requires PyTorch")(test_case) + + import torch + + if not torch.cuda.is_available(): + return unittest.skip("test requires PyTorch CUDA")(test_case) + + device_properties = torch.cuda.get_device_properties(0) + total_memory = device_properties.total_memory / (1024**3) + return unittest.skipUnless( + total_memory >= BIG_GPU_MEMORY, f"test requires a GPU with at least {BIG_GPU_MEMORY} GB memory" + )(test_case) + + +def require_big_accelerator(test_case): + """ + Decorator marking a test that requires a bigger hardware accelerator (24GB) for execution. Some example pipelines: + Flux, SD3, Cog, etc. + """ + if not is_torch_available(): + return unittest.skip("test requires PyTorch")(test_case) + + import torch + + if not (torch.cuda.is_available() or torch.xpu.is_available()): + return unittest.skip("test requires PyTorch CUDA")(test_case) + + if torch.xpu.is_available(): + device_properties = torch.xpu.get_device_properties(0) + else: + device_properties = torch.cuda.get_device_properties(0) + + total_memory = device_properties.total_memory / (1024**3) + return unittest.skipUnless( + total_memory >= BIG_GPU_MEMORY, + f"test requires a hardware accelerator with at least {BIG_GPU_MEMORY} GB memory", + )(test_case) + + +def require_torch_accelerator_with_training(test_case): + """Decorator marking a test that requires an accelerator with support for training.""" + return unittest.skipUnless( + is_torch_available() and backend_supports_training(torch_device), + "test requires accelerator with training support", + )(test_case) + + +def skip_mps(test_case): + """Decorator marking a test to skip if torch_device is 'mps'""" + return unittest.skipUnless(torch_device != "mps", "test requires non 'mps' device")(test_case) + + +def require_flax(test_case): + """ + Decorator marking a test that requires JAX & Flax. These tests are skipped when one / both are not installed + """ + return unittest.skipUnless(is_flax_available(), "test requires JAX & Flax")(test_case) + + +def require_compel(test_case): + """ + Decorator marking a test that requires compel: https://github.com/damian0815/compel. These tests are skipped when + the library is not installed. + """ + return unittest.skipUnless(is_compel_available(), "test requires compel")(test_case) + + +def require_onnxruntime(test_case): + """ + Decorator marking a test that requires onnxruntime. These tests are skipped when onnxruntime isn't installed. + """ + return unittest.skipUnless(is_onnx_available(), "test requires onnxruntime")(test_case) + + +def require_note_seq(test_case): + """ + Decorator marking a test that requires note_seq. These tests are skipped when note_seq isn't installed. + """ + return unittest.skipUnless(is_note_seq_available(), "test requires note_seq")(test_case) + + +def require_accelerator(test_case): + """ + Decorator marking a test that requires a hardware accelerator backend. These tests are skipped when there are no + hardware accelerator available. + """ + return unittest.skipUnless(torch_device != "cpu", "test requires a hardware accelerator")(test_case) + + +def require_torchsde(test_case): + """ + Decorator marking a test that requires torchsde. These tests are skipped when torchsde isn't installed. + """ + return unittest.skipUnless(is_torchsde_available(), "test requires torchsde")(test_case) + + +def require_peft_backend(test_case): + """ + Decorator marking a test that requires PEFT backend, this would require some specific versions of PEFT and + transformers. + """ + return unittest.skipUnless(USE_PEFT_BACKEND, "test requires PEFT backend")(test_case) + + +def require_timm(test_case): + """ + Decorator marking a test that requires timm. These tests are skipped when timm isn't installed. + """ + return unittest.skipUnless(is_timm_available(), "test requires timm")(test_case) + + +def require_bitsandbytes(test_case): + """ + Decorator marking a test that requires bitsandbytes. These tests are skipped when bitsandbytes isn't installed. + """ + return unittest.skipUnless(is_bitsandbytes_available(), "test requires bitsandbytes")(test_case) + + +def require_quanto(test_case): + """ + Decorator marking a test that requires quanto. These tests are skipped when quanto isn't installed. + """ + return unittest.skipUnless(is_optimum_quanto_available(), "test requires quanto")(test_case) + + +def require_accelerate(test_case): + """ + Decorator marking a test that requires accelerate. These tests are skipped when accelerate isn't installed. + """ + return unittest.skipUnless(is_accelerate_available(), "test requires accelerate")(test_case) + + +def require_peft_version_greater(peft_version): + """ + Decorator marking a test that requires PEFT backend with a specific version, this would require some specific + versions of PEFT and transformers. + """ + + def decorator(test_case): + correct_peft_version = is_peft_available() and version.parse( + version.parse(importlib.metadata.version("peft")).base_version + ) > version.parse(peft_version) + return unittest.skipUnless( + correct_peft_version, f"test requires PEFT backend with the version greater than {peft_version}" + )(test_case) + + return decorator + + +def require_transformers_version_greater(transformers_version): + """ + Decorator marking a test that requires transformers with a specific version, this would require some specific + versions of PEFT and transformers. + """ + + def decorator(test_case): + correct_transformers_version = is_transformers_available() and version.parse( + version.parse(importlib.metadata.version("transformers")).base_version + ) > version.parse(transformers_version) + return unittest.skipUnless( + correct_transformers_version, + f"test requires transformers with the version greater than {transformers_version}", + )(test_case) + + return decorator + + +def require_accelerate_version_greater(accelerate_version): + def decorator(test_case): + correct_accelerate_version = is_accelerate_available() and version.parse( + version.parse(importlib.metadata.version("accelerate")).base_version + ) > version.parse(accelerate_version) + return unittest.skipUnless( + correct_accelerate_version, f"Test requires accelerate with the version greater than {accelerate_version}." + )(test_case) + + return decorator + + +def require_bitsandbytes_version_greater(bnb_version): + def decorator(test_case): + correct_bnb_version = is_bitsandbytes_available() and version.parse( + version.parse(importlib.metadata.version("bitsandbytes")).base_version + ) > version.parse(bnb_version) + return unittest.skipUnless( + correct_bnb_version, f"Test requires bitsandbytes with the version greater than {bnb_version}." + )(test_case) + + return decorator + + +def require_hf_hub_version_greater(hf_hub_version): + def decorator(test_case): + correct_hf_hub_version = version.parse( + version.parse(importlib.metadata.version("huggingface_hub")).base_version + ) > version.parse(hf_hub_version) + return unittest.skipUnless( + correct_hf_hub_version, f"Test requires huggingface_hub with the version greater than {hf_hub_version}." + )(test_case) + + return decorator + + +def require_gguf_version_greater_or_equal(gguf_version): + def decorator(test_case): + correct_gguf_version = is_gguf_available() and version.parse( + version.parse(importlib.metadata.version("gguf")).base_version + ) >= version.parse(gguf_version) + return unittest.skipUnless( + correct_gguf_version, f"Test requires gguf with the version greater than {gguf_version}." + )(test_case) + + return decorator + + +def require_torchao_version_greater_or_equal(torchao_version): + def decorator(test_case): + correct_torchao_version = is_torchao_available() and version.parse( + version.parse(importlib.metadata.version("torchao")).base_version + ) >= version.parse(torchao_version) + return unittest.skipUnless( + correct_torchao_version, f"Test requires torchao with version greater than {torchao_version}." + )(test_case) + + return decorator + + +def deprecate_after_peft_backend(test_case): + """ + Decorator marking a test that will be skipped after PEFT backend + """ + return unittest.skipUnless(not USE_PEFT_BACKEND, "test skipped in favor of PEFT backend")(test_case) + + +def get_python_version(): + sys_info = sys.version_info + major, minor = sys_info.major, sys_info.minor + return major, minor + + +def load_numpy(arry: Union[str, np.ndarray], local_path: Optional[str] = None) -> np.ndarray: + if isinstance(arry, str): + if local_path is not None: + # local_path can be passed to correct images of tests + return Path(local_path, arry.split("/")[-5], arry.split("/")[-2], arry.split("/")[-1]).as_posix() + elif arry.startswith("http://") or arry.startswith("https://"): + response = requests.get(arry, timeout=DIFFUSERS_REQUEST_TIMEOUT) + response.raise_for_status() + arry = np.load(BytesIO(response.content)) + elif os.path.isfile(arry): + arry = np.load(arry) + else: + raise ValueError( + f"Incorrect path or url, URLs must start with `http://` or `https://`, and {arry} is not a valid path" + ) + elif isinstance(arry, np.ndarray): + pass + else: + raise ValueError( + "Incorrect format used for numpy ndarray. Should be an url linking to an image, a local path, or a" + " ndarray." + ) + + return arry + + +def load_pt(url: str, map_location: Optional[str] = None, weights_only: Optional[bool] = True): + response = requests.get(url, timeout=DIFFUSERS_REQUEST_TIMEOUT) + response.raise_for_status() + arry = torch.load(BytesIO(response.content), map_location=map_location, weights_only=weights_only) + return arry + + +def load_image(image: Union[str, PIL.Image.Image]) -> PIL.Image.Image: + """ + Loads `image` to a PIL Image. + + Args: + image (`str` or `PIL.Image.Image`): + The image to convert to the PIL Image format. + Returns: + `PIL.Image.Image`: + A PIL Image. + """ + if isinstance(image, str): + if image.startswith("http://") or image.startswith("https://"): + image = PIL.Image.open(requests.get(image, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw) + elif os.path.isfile(image): + image = PIL.Image.open(image) + else: + raise ValueError( + f"Incorrect path or url, URLs must start with `http://` or `https://`, and {image} is not a valid path" + ) + elif isinstance(image, PIL.Image.Image): + image = image + else: + raise ValueError( + "Incorrect format used for image. Should be an url linking to an image, a local path, or a PIL image." + ) + image = PIL.ImageOps.exif_transpose(image) + image = image.convert("RGB") + return image + + +def preprocess_image(image: PIL.Image, batch_size: int): + w, h = image.size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + image = image.resize((w, h), resample=PIL.Image.LANCZOS) + image = np.array(image).astype(np.float32) / 255.0 + image = np.vstack([image[None].transpose(0, 3, 1, 2)] * batch_size) + image = torch.from_numpy(image) + return 2.0 * image - 1.0 + + +def export_to_gif(image: List[PIL.Image.Image], output_gif_path: str = None) -> str: + if output_gif_path is None: + output_gif_path = tempfile.NamedTemporaryFile(suffix=".gif").name + + image[0].save( + output_gif_path, + save_all=True, + append_images=image[1:], + optimize=False, + duration=100, + loop=0, + ) + return output_gif_path + + +@contextmanager +def buffered_writer(raw_f): + f = io.BufferedWriter(raw_f) + yield f + f.flush() + + +def export_to_ply(mesh, output_ply_path: str = None): + """ + Write a PLY file for a mesh. + """ + if output_ply_path is None: + output_ply_path = tempfile.NamedTemporaryFile(suffix=".ply").name + + coords = mesh.verts.detach().cpu().numpy() + faces = mesh.faces.cpu().numpy() + rgb = np.stack([mesh.vertex_channels[x].detach().cpu().numpy() for x in "RGB"], axis=1) + + with buffered_writer(open(output_ply_path, "wb")) as f: + f.write(b"ply\n") + f.write(b"format binary_little_endian 1.0\n") + f.write(bytes(f"element vertex {len(coords)}\n", "ascii")) + f.write(b"property float x\n") + f.write(b"property float y\n") + f.write(b"property float z\n") + if rgb is not None: + f.write(b"property uchar red\n") + f.write(b"property uchar green\n") + f.write(b"property uchar blue\n") + if faces is not None: + f.write(bytes(f"element face {len(faces)}\n", "ascii")) + f.write(b"property list uchar int vertex_index\n") + f.write(b"end_header\n") + + if rgb is not None: + rgb = (rgb * 255.499).round().astype(int) + vertices = [ + (*coord, *rgb) + for coord, rgb in zip( + coords.tolist(), + rgb.tolist(), + ) + ] + format = struct.Struct("<3f3B") + for item in vertices: + f.write(format.pack(*item)) + else: + format = struct.Struct("<3f") + for vertex in coords.tolist(): + f.write(format.pack(*vertex)) + + if faces is not None: + format = struct.Struct(" str: + if is_opencv_available(): + import cv2 + else: + raise ImportError(BACKENDS_MAPPING["opencv"][1].format("export_to_video")) + if output_video_path is None: + output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4").name + + fourcc = cv2.VideoWriter_fourcc(*"mp4v") + h, w, c = video_frames[0].shape + video_writer = cv2.VideoWriter(output_video_path, fourcc, fps=8, frameSize=(w, h)) + for i in range(len(video_frames)): + img = cv2.cvtColor(video_frames[i], cv2.COLOR_RGB2BGR) + video_writer.write(img) + return output_video_path + + +def load_hf_numpy(path) -> np.ndarray: + base_url = "https://huggingface.co/datasets/fusing/diffusers-testing/resolve/main" + + if not path.startswith("http://") and not path.startswith("https://"): + path = os.path.join(base_url, urllib.parse.quote(path)) + + return load_numpy(path) + + +# --- pytest conf functions --- # + +# to avoid multiple invocation from tests/conftest.py and examples/conftest.py - make sure it's called only once +pytest_opt_registered = {} + + +def pytest_addoption_shared(parser): + """ + This function is to be called from `conftest.py` via `pytest_addoption` wrapper that has to be defined there. + + It allows loading both `conftest.py` files at once without causing a failure due to adding the same `pytest` + option. + + """ + option = "--make-reports" + if option not in pytest_opt_registered: + parser.addoption( + option, + action="store", + default=False, + help="generate report files. The value of this option is used as a prefix to report names", + ) + pytest_opt_registered[option] = 1 + + +def pytest_terminal_summary_main(tr, id): + """ + Generate multiple reports at the end of test suite run - each report goes into a dedicated file in the current + directory. The report files are prefixed with the test suite name. + + This function emulates --duration and -rA pytest arguments. + + This function is to be called from `conftest.py` via `pytest_terminal_summary` wrapper that has to be defined + there. + + Args: + - tr: `terminalreporter` passed from `conftest.py` + - id: unique id like `tests` or `examples` that will be incorporated into the final reports filenames - this is + needed as some jobs have multiple runs of pytest, so we can't have them overwrite each other. + + NB: this functions taps into a private _pytest API and while unlikely, it could break should + pytest do internal changes - also it calls default internal methods of terminalreporter which + can be hijacked by various `pytest-` plugins and interfere. + + """ + from _pytest.config import create_terminal_writer + + if not len(id): + id = "tests" + + config = tr.config + orig_writer = config.get_terminal_writer() + orig_tbstyle = config.option.tbstyle + orig_reportchars = tr.reportchars + + dir = "reports" + Path(dir).mkdir(parents=True, exist_ok=True) + report_files = { + k: f"{dir}/{id}_{k}.txt" + for k in [ + "durations", + "errors", + "failures_long", + "failures_short", + "failures_line", + "passes", + "stats", + "summary_short", + "warnings", + ] + } + + # custom durations report + # note: there is no need to call pytest --durations=XX to get this separate report + # adapted from https://github.com/pytest-dev/pytest/blob/897f151e/src/_pytest/runner.py#L66 + dlist = [] + for replist in tr.stats.values(): + for rep in replist: + if hasattr(rep, "duration"): + dlist.append(rep) + if dlist: + dlist.sort(key=lambda x: x.duration, reverse=True) + with open(report_files["durations"], "w") as f: + durations_min = 0.05 # sec + f.write("slowest durations\n") + for i, rep in enumerate(dlist): + if rep.duration < durations_min: + f.write(f"{len(dlist) - i} durations < {durations_min} secs were omitted") + break + f.write(f"{rep.duration:02.2f}s {rep.when:<8} {rep.nodeid}\n") + + def summary_failures_short(tr): + # expecting that the reports were --tb=long (default) so we chop them off here to the last frame + reports = tr.getreports("failed") + if not reports: + return + tr.write_sep("=", "FAILURES SHORT STACK") + for rep in reports: + msg = tr._getfailureheadline(rep) + tr.write_sep("_", msg, red=True, bold=True) + # chop off the optional leading extra frames, leaving only the last one + longrepr = re.sub(r".*_ _ _ (_ ){10,}_ _ ", "", rep.longreprtext, 0, re.M | re.S) + tr._tw.line(longrepr) + # note: not printing out any rep.sections to keep the report short + + # use ready-made report funcs, we are just hijacking the filehandle to log to a dedicated file each + # adapted from https://github.com/pytest-dev/pytest/blob/897f151e/src/_pytest/terminal.py#L814 + # note: some pytest plugins may interfere by hijacking the default `terminalreporter` (e.g. + # pytest-instafail does that) + + # report failures with line/short/long styles + config.option.tbstyle = "auto" # full tb + with open(report_files["failures_long"], "w") as f: + tr._tw = create_terminal_writer(config, f) + tr.summary_failures() + + # config.option.tbstyle = "short" # short tb + with open(report_files["failures_short"], "w") as f: + tr._tw = create_terminal_writer(config, f) + summary_failures_short(tr) + + config.option.tbstyle = "line" # one line per error + with open(report_files["failures_line"], "w") as f: + tr._tw = create_terminal_writer(config, f) + tr.summary_failures() + + with open(report_files["errors"], "w") as f: + tr._tw = create_terminal_writer(config, f) + tr.summary_errors() + + with open(report_files["warnings"], "w") as f: + tr._tw = create_terminal_writer(config, f) + tr.summary_warnings() # normal warnings + tr.summary_warnings() # final warnings + + tr.reportchars = "wPpsxXEf" # emulate -rA (used in summary_passes() and short_test_summary()) + with open(report_files["passes"], "w") as f: + tr._tw = create_terminal_writer(config, f) + tr.summary_passes() + + with open(report_files["summary_short"], "w") as f: + tr._tw = create_terminal_writer(config, f) + tr.short_test_summary() + + with open(report_files["stats"], "w") as f: + tr._tw = create_terminal_writer(config, f) + tr.summary_stats() + + # restore: + tr._tw = orig_writer + tr.reportchars = orig_reportchars + config.option.tbstyle = orig_tbstyle + + +# Copied from https://github.com/huggingface/transformers/blob/000e52aec8850d3fe2f360adc6fd256e5b47fe4c/src/transformers/testing_utils.py#L1905 +def is_flaky(max_attempts: int = 5, wait_before_retry: Optional[float] = None, description: Optional[str] = None): + """ + To decorate flaky tests. They will be retried on failures. + + Args: + max_attempts (`int`, *optional*, defaults to 5): + The maximum number of attempts to retry the flaky test. + wait_before_retry (`float`, *optional*): + If provided, will wait that number of seconds before retrying the test. + description (`str`, *optional*): + A string to describe the situation (what / where / why is flaky, link to GH issue/PR comments, errors, + etc.) + """ + + def decorator(test_func_ref): + @functools.wraps(test_func_ref) + def wrapper(*args, **kwargs): + retry_count = 1 + + while retry_count < max_attempts: + try: + return test_func_ref(*args, **kwargs) + + except Exception as err: + print(f"Test failed with {err} at try {retry_count}/{max_attempts}.", file=sys.stderr) + if wait_before_retry is not None: + time.sleep(wait_before_retry) + retry_count += 1 + + return test_func_ref(*args, **kwargs) + + return wrapper + + return decorator + + +# Taken from: https://github.com/huggingface/transformers/blob/3658488ff77ff8d45101293e749263acf437f4d5/src/transformers/testing_utils.py#L1787 +def run_test_in_subprocess(test_case, target_func, inputs=None, timeout=None): + """ + To run a test in a subprocess. In particular, this can avoid (GPU) memory issue. + + Args: + test_case (`unittest.TestCase`): + The test that will run `target_func`. + target_func (`Callable`): + The function implementing the actual testing logic. + inputs (`dict`, *optional*, defaults to `None`): + The inputs that will be passed to `target_func` through an (input) queue. + timeout (`int`, *optional*, defaults to `None`): + The timeout (in seconds) that will be passed to the input and output queues. If not specified, the env. + variable `PYTEST_TIMEOUT` will be checked. If still `None`, its value will be set to `600`. + """ + if timeout is None: + timeout = int(os.environ.get("PYTEST_TIMEOUT", 600)) + + start_methohd = "spawn" + ctx = multiprocessing.get_context(start_methohd) + + input_queue = ctx.Queue(1) + output_queue = ctx.JoinableQueue(1) + + # We can't send `unittest.TestCase` to the child, otherwise we get issues regarding pickle. + input_queue.put(inputs, timeout=timeout) + + process = ctx.Process(target=target_func, args=(input_queue, output_queue, timeout)) + process.start() + # Kill the child process if we can't get outputs from it in time: otherwise, the hanging subprocess prevents + # the test to exit properly. + try: + results = output_queue.get(timeout=timeout) + output_queue.task_done() + except Exception as e: + process.terminate() + test_case.fail(e) + process.join(timeout=timeout) + + if results["error"] is not None: + test_case.fail(f"{results['error']}") + + +class CaptureLogger: + """ + Args: + Context manager to capture `logging` streams + logger: 'logging` logger object + Returns: + The captured output is available via `self.out` + Example: + ```python + >>> from diffusers import logging + >>> from diffusers.testing_utils import CaptureLogger + + >>> msg = "Testing 1, 2, 3" + >>> logging.set_verbosity_info() + >>> logger = logging.get_logger("diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.py") + >>> with CaptureLogger(logger) as cl: + ... logger.info(msg) + >>> assert cl.out, msg + "\n" + ``` + """ + + def __init__(self, logger): + self.logger = logger + self.io = StringIO() + self.sh = logging.StreamHandler(self.io) + self.out = "" + + def __enter__(self): + self.logger.addHandler(self.sh) + return self + + def __exit__(self, *exc): + self.logger.removeHandler(self.sh) + self.out = self.io.getvalue() + + def __repr__(self): + return f"captured: {self.out}\n" + + +def enable_full_determinism(): + """ + Helper function for reproducible behavior during distributed training. See + - https://pytorch.org/docs/stable/notes/randomness.html for pytorch + """ + # Enable PyTorch deterministic mode. This potentially requires either the environment + # variable 'CUDA_LAUNCH_BLOCKING' or 'CUBLAS_WORKSPACE_CONFIG' to be set, + # depending on the CUDA version, so we set them both here + os.environ["CUDA_LAUNCH_BLOCKING"] = "1" + os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":16:8" + torch.use_deterministic_algorithms(True) + + # Enable CUDNN deterministic mode + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + torch.backends.cuda.matmul.allow_tf32 = False + + +def disable_full_determinism(): + os.environ["CUDA_LAUNCH_BLOCKING"] = "0" + os.environ["CUBLAS_WORKSPACE_CONFIG"] = "" + torch.use_deterministic_algorithms(False) + + +# Utils for custom and alternative accelerator devices +def _is_torch_fp16_available(device): + if not is_torch_available(): + return False + + import torch + + device = torch.device(device) + + try: + x = torch.zeros((2, 2), dtype=torch.float16).to(device) + _ = torch.mul(x, x) + return True + + except Exception as e: + if device.type == "cuda": + raise ValueError( + f"You have passed a device of type 'cuda' which should work with 'fp16', but 'cuda' does not seem to be correctly installed on your machine: {e}" + ) + + return False + + +def _is_torch_fp64_available(device): + if not is_torch_available(): + return False + + import torch + + device = torch.device(device) + + try: + x = torch.zeros((2, 2), dtype=torch.float64).to(device) + _ = torch.mul(x, x) + return True + + except Exception as e: + if device.type == "cuda": + raise ValueError( + f"You have passed a device of type 'cuda' which should work with 'fp64', but 'cuda' does not seem to be correctly installed on your machine: {e}" + ) + + return False + + +# Guard these lookups for when Torch is not used - alternative accelerator support is for PyTorch +if is_torch_available(): + # Behaviour flags + BACKEND_SUPPORTS_TRAINING = {"cuda": True, "xpu": True, "cpu": True, "mps": False, "default": True} + + # Function definitions + BACKEND_EMPTY_CACHE = { + "cuda": torch.cuda.empty_cache, + "xpu": torch.xpu.empty_cache, + "cpu": None, + "mps": torch.mps.empty_cache, + "default": None, + } + BACKEND_DEVICE_COUNT = { + "cuda": torch.cuda.device_count, + "xpu": torch.xpu.device_count, + "cpu": lambda: 0, + "mps": lambda: 0, + "default": 0, + } + BACKEND_MANUAL_SEED = { + "cuda": torch.cuda.manual_seed, + "xpu": torch.xpu.manual_seed, + "cpu": torch.manual_seed, + "mps": torch.mps.manual_seed, + "default": torch.manual_seed, + } + BACKEND_RESET_PEAK_MEMORY_STATS = { + "cuda": torch.cuda.reset_peak_memory_stats, + "xpu": getattr(torch.xpu, "reset_peak_memory_stats", None), + "cpu": None, + "mps": None, + "default": None, + } + BACKEND_RESET_MAX_MEMORY_ALLOCATED = { + "cuda": torch.cuda.reset_max_memory_allocated, + "xpu": getattr(torch.xpu, "reset_peak_memory_stats", None), + "cpu": None, + "mps": None, + "default": None, + } + BACKEND_MAX_MEMORY_ALLOCATED = { + "cuda": torch.cuda.max_memory_allocated, + "xpu": getattr(torch.xpu, "max_memory_allocated", None), + "cpu": 0, + "mps": 0, + "default": 0, + } + BACKEND_SYNCHRONIZE = { + "cuda": torch.cuda.synchronize, + "xpu": getattr(torch.xpu, "synchronize", None), + "cpu": None, + "mps": None, + "default": None, + } + + +# This dispatches a defined function according to the accelerator from the function definitions. +def _device_agnostic_dispatch(device: str, dispatch_table: Dict[str, Callable], *args, **kwargs): + if device not in dispatch_table: + return dispatch_table["default"](*args, **kwargs) + + fn = dispatch_table[device] + + # Some device agnostic functions return values. Need to guard against 'None' instead at + # user level + if not callable(fn): + return fn + + return fn(*args, **kwargs) + + +# These are callables which automatically dispatch the function specific to the accelerator +def backend_manual_seed(device: str, seed: int): + return _device_agnostic_dispatch(device, BACKEND_MANUAL_SEED, seed) + + +def backend_synchronize(device: str): + return _device_agnostic_dispatch(device, BACKEND_SYNCHRONIZE) + + +def backend_empty_cache(device: str): + return _device_agnostic_dispatch(device, BACKEND_EMPTY_CACHE) + + +def backend_device_count(device: str): + return _device_agnostic_dispatch(device, BACKEND_DEVICE_COUNT) + + +def backend_reset_peak_memory_stats(device: str): + return _device_agnostic_dispatch(device, BACKEND_RESET_PEAK_MEMORY_STATS) + + +def backend_reset_max_memory_allocated(device: str): + return _device_agnostic_dispatch(device, BACKEND_RESET_MAX_MEMORY_ALLOCATED) + + +def backend_max_memory_allocated(device: str): + return _device_agnostic_dispatch(device, BACKEND_MAX_MEMORY_ALLOCATED) + + +# These are callables which return boolean behaviour flags and can be used to specify some +# device agnostic alternative where the feature is unsupported. +def backend_supports_training(device: str): + if not is_torch_available(): + return False + + if device not in BACKEND_SUPPORTS_TRAINING: + device = "default" + + return BACKEND_SUPPORTS_TRAINING[device] + + +# Guard for when Torch is not available +if is_torch_available(): + # Update device function dict mapping + def update_mapping_from_spec(device_fn_dict: Dict[str, Callable], attribute_name: str): + try: + # Try to import the function directly + spec_fn = getattr(device_spec_module, attribute_name) + device_fn_dict[torch_device] = spec_fn + except AttributeError as e: + # If the function doesn't exist, and there is no default, throw an error + if "default" not in device_fn_dict: + raise AttributeError( + f"`{attribute_name}` not found in '{device_spec_path}' and no default fallback function found." + ) from e + + if "DIFFUSERS_TEST_DEVICE_SPEC" in os.environ: + device_spec_path = os.environ["DIFFUSERS_TEST_DEVICE_SPEC"] + if not Path(device_spec_path).is_file(): + raise ValueError(f"Specified path to device specification file is not found. Received {device_spec_path}") + + try: + import_name = device_spec_path[: device_spec_path.index(".py")] + except ValueError as e: + raise ValueError(f"Provided device spec file is not a Python file! Received {device_spec_path}") from e + + device_spec_module = importlib.import_module(import_name) + + try: + device_name = device_spec_module.DEVICE_NAME + except AttributeError: + raise AttributeError("Device spec file did not contain `DEVICE_NAME`") + + if "DIFFUSERS_TEST_DEVICE" in os.environ and torch_device != device_name: + msg = f"Mismatch between environment variable `DIFFUSERS_TEST_DEVICE` '{torch_device}' and device found in spec '{device_name}'\n" + msg += "Either unset `DIFFUSERS_TEST_DEVICE` or ensure it matches device spec name." + raise ValueError(msg) + + torch_device = device_name + + # Add one entry here for each `BACKEND_*` dictionary. + update_mapping_from_spec(BACKEND_MANUAL_SEED, "MANUAL_SEED_FN") + update_mapping_from_spec(BACKEND_EMPTY_CACHE, "EMPTY_CACHE_FN") + update_mapping_from_spec(BACKEND_DEVICE_COUNT, "DEVICE_COUNT_FN") + update_mapping_from_spec(BACKEND_SUPPORTS_TRAINING, "SUPPORTS_TRAINING") + update_mapping_from_spec(BACKEND_RESET_PEAK_MEMORY_STATS, "RESET_PEAK_MEMORY_STATS_FN") + update_mapping_from_spec(BACKEND_RESET_MAX_MEMORY_ALLOCATED, "RESET_MAX_MEMORY_ALLOCATED_FN") + update_mapping_from_spec(BACKEND_MAX_MEMORY_ALLOCATED, "MAX_MEMORY_ALLOCATED_FN") + + +# Modified from https://github.com/huggingface/transformers/blob/cdfb018d0300fef3b07d9220f3efe9c2a9974662/src/transformers/testing_utils.py#L3090 + +# Type definition of key used in `Expectations` class. +DeviceProperties = Tuple[Union[str, None], Union[int, None]] + + +@functools.lru_cache +def get_device_properties() -> DeviceProperties: + """ + Get environment device properties. + """ + if IS_CUDA_SYSTEM or IS_ROCM_SYSTEM: + import torch + + major, _ = torch.cuda.get_device_capability() + if IS_ROCM_SYSTEM: + return ("rocm", major) + else: + return ("cuda", major) + elif IS_XPU_SYSTEM: + import torch + + # To get more info of the architecture meaning and bit allocation, refer to https://github.com/intel/llvm/blob/sycl/sycl/include/sycl/ext/oneapi/experimental/device_architecture.def + arch = torch.xpu.get_device_capability()["architecture"] + gen_mask = 0x000000FF00000000 + gen = (arch & gen_mask) >> 32 + return ("xpu", gen) + else: + return (torch_device, None) + + +if TYPE_CHECKING: + DevicePropertiesUserDict = UserDict[DeviceProperties, Any] +else: + DevicePropertiesUserDict = UserDict + + +class Expectations(DevicePropertiesUserDict): + def get_expectation(self) -> Any: + """ + Find best matching expectation based on environment device properties. + """ + return self.find_expectation(get_device_properties()) + + @staticmethod + def is_default(key: DeviceProperties) -> bool: + return all(p is None for p in key) + + @staticmethod + def score(key: DeviceProperties, other: DeviceProperties) -> int: + """ + Returns score indicating how similar two instances of the `Properties` tuple are. Points are calculated using + bits, but documented as int. Rules are as follows: + * Matching `type` gives 8 points. + * Semi-matching `type`, for example cuda and rocm, gives 4 points. + * Matching `major` (compute capability major version) gives 2 points. + * Default expectation (if present) gives 1 points. + """ + (device_type, major) = key + (other_device_type, other_major) = other + + score = 0b0 + if device_type == other_device_type: + score |= 0b1000 + elif device_type in ["cuda", "rocm"] and other_device_type in ["cuda", "rocm"]: + score |= 0b100 + + if major == other_major and other_major is not None: + score |= 0b10 + + if Expectations.is_default(other): + score |= 0b1 + + return int(score) + + def find_expectation(self, key: DeviceProperties = (None, None)) -> Any: + """ + Find best matching expectation based on provided device properties. + """ + (result_key, result) = max(self.data.items(), key=lambda x: Expectations.score(key, x[0])) + + if Expectations.score(key, result_key) == 0: + raise ValueError(f"No matching expectation found for {key}") + + return result + + def __repr__(self): + return f"{self.data}" diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/torch_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/torch_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..ffc1119727dfaf9b83beb4041427d6d4343649fd --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/torch_utils.py @@ -0,0 +1,181 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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. +""" +PyTorch utilities: Utilities related to PyTorch +""" + +from typing import List, Optional, Tuple, Union + +from . import logging +from .import_utils import is_torch_available, is_torch_npu_available, is_torch_version + + +if is_torch_available(): + import torch + from torch.fft import fftn, fftshift, ifftn, ifftshift + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +try: + from torch._dynamo import allow_in_graph as maybe_allow_in_graph +except (ImportError, ModuleNotFoundError): + + def maybe_allow_in_graph(cls): + return cls + + +def randn_tensor( + shape: Union[Tuple, List], + generator: Optional[Union[List["torch.Generator"], "torch.Generator"]] = None, + device: Optional[Union[str, "torch.device"]] = None, + dtype: Optional["torch.dtype"] = None, + layout: Optional["torch.layout"] = None, +): + """A helper function to create random tensors on the desired `device` with the desired `dtype`. When + passing a list of generators, you can seed each batch size individually. If CPU generators are passed, the tensor + is always created on the CPU. + """ + # device on which tensor is created defaults to device + if isinstance(device, str): + device = torch.device(device) + rand_device = device + batch_size = shape[0] + + layout = layout or torch.strided + device = device or torch.device("cpu") + + if generator is not None: + gen_device_type = generator.device.type if not isinstance(generator, list) else generator[0].device.type + if gen_device_type != device.type and gen_device_type == "cpu": + rand_device = "cpu" + if device != "mps": + logger.info( + f"The passed generator was created on 'cpu' even though a tensor on {device} was expected." + f" Tensors will be created on 'cpu' and then moved to {device}. Note that one can probably" + f" slightly speed up this function by passing a generator that was created on the {device} device." + ) + elif gen_device_type != device.type and gen_device_type == "cuda": + raise ValueError(f"Cannot generate a {device} tensor from a generator of type {gen_device_type}.") + + # make sure generator list of length 1 is treated like a non-list + if isinstance(generator, list) and len(generator) == 1: + generator = generator[0] + + if isinstance(generator, list): + shape = (1,) + shape[1:] + latents = [ + torch.randn(shape, generator=generator[i], device=rand_device, dtype=dtype, layout=layout) + for i in range(batch_size) + ] + latents = torch.cat(latents, dim=0).to(device) + else: + latents = torch.randn(shape, generator=generator, device=rand_device, dtype=dtype, layout=layout).to(device) + + return latents + + +def is_compiled_module(module) -> bool: + """Check whether the module was compiled with torch.compile()""" + if is_torch_version("<", "2.0.0") or not hasattr(torch, "_dynamo"): + return False + return isinstance(module, torch._dynamo.eval_frame.OptimizedModule) + + +def fourier_filter(x_in: "torch.Tensor", threshold: int, scale: int) -> "torch.Tensor": + """Fourier filter as introduced in FreeU (https://huggingface.co/papers/2309.11497). + + This version of the method comes from here: + https://github.com/huggingface/diffusers/pull/5164#issuecomment-1732638706 + """ + x = x_in + B, C, H, W = x.shape + + # Non-power of 2 images must be float32 + if (W & (W - 1)) != 0 or (H & (H - 1)) != 0: + x = x.to(dtype=torch.float32) + # fftn does not support bfloat16 + elif x.dtype == torch.bfloat16: + x = x.to(dtype=torch.float32) + + # FFT + x_freq = fftn(x, dim=(-2, -1)) + x_freq = fftshift(x_freq, dim=(-2, -1)) + + B, C, H, W = x_freq.shape + mask = torch.ones((B, C, H, W), device=x.device) + + crow, ccol = H // 2, W // 2 + mask[..., crow - threshold : crow + threshold, ccol - threshold : ccol + threshold] = scale + x_freq = x_freq * mask + + # IFFT + x_freq = ifftshift(x_freq, dim=(-2, -1)) + x_filtered = ifftn(x_freq, dim=(-2, -1)).real + + return x_filtered.to(dtype=x_in.dtype) + + +def apply_freeu( + resolution_idx: int, hidden_states: "torch.Tensor", res_hidden_states: "torch.Tensor", **freeu_kwargs +) -> Tuple["torch.Tensor", "torch.Tensor"]: + """Applies the FreeU mechanism as introduced in https: + //arxiv.org/abs/2309.11497. Adapted from the official code repository: https://github.com/ChenyangSi/FreeU. + + Args: + resolution_idx (`int`): Integer denoting the UNet block where FreeU is being applied. + hidden_states (`torch.Tensor`): Inputs to the underlying block. + res_hidden_states (`torch.Tensor`): Features from the skip block corresponding to the underlying block. + s1 (`float`): Scaling factor for stage 1 to attenuate the contributions of the skip features. + s2 (`float`): Scaling factor for stage 2 to attenuate the contributions of the skip features. + b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. + b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. + """ + if resolution_idx == 0: + num_half_channels = hidden_states.shape[1] // 2 + hidden_states[:, :num_half_channels] = hidden_states[:, :num_half_channels] * freeu_kwargs["b1"] + res_hidden_states = fourier_filter(res_hidden_states, threshold=1, scale=freeu_kwargs["s1"]) + if resolution_idx == 1: + num_half_channels = hidden_states.shape[1] // 2 + hidden_states[:, :num_half_channels] = hidden_states[:, :num_half_channels] * freeu_kwargs["b2"] + res_hidden_states = fourier_filter(res_hidden_states, threshold=1, scale=freeu_kwargs["s2"]) + + return hidden_states, res_hidden_states + + +def get_torch_cuda_device_capability(): + if torch.cuda.is_available(): + device = torch.device("cuda") + compute_capability = torch.cuda.get_device_capability(device) + compute_capability = f"{compute_capability[0]}.{compute_capability[1]}" + return float(compute_capability) + else: + return None + + +def get_device(): + if torch.cuda.is_available(): + return "cuda" + elif is_torch_npu_available(): + return "npu" + elif hasattr(torch, "xpu") and torch.xpu.is_available(): + return "xpu" + else: + return "cpu" + + +def empty_device_cache(device_type: Optional[str] = None): + if device_type is None: + device_type = get_device() + device_mod = getattr(torch, device_type, torch.cuda) + device_mod.empty_cache() diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/typing_utils.py b/venv/lib/python3.10/site-packages/diffusers/utils/typing_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..2b5b1a4f5ab5510856bf453a7436f6deb9c61abe --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/typing_utils.py @@ -0,0 +1,91 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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. +""" +Typing utilities: Utilities related to type checking and validation +""" + +from typing import Any, Dict, List, Set, Tuple, Type, Union, get_args, get_origin + + +def _is_valid_type(obj: Any, class_or_tuple: Union[Type, Tuple[Type, ...]]) -> bool: + """ + Checks if an object is an instance of any of the provided types. For collections, it checks if every element is of + the correct type as well. + """ + if not isinstance(class_or_tuple, tuple): + class_or_tuple = (class_or_tuple,) + + # Unpack unions + unpacked_class_or_tuple = [] + for t in class_or_tuple: + if get_origin(t) is Union: + unpacked_class_or_tuple.extend(get_args(t)) + else: + unpacked_class_or_tuple.append(t) + class_or_tuple = tuple(unpacked_class_or_tuple) + + if Any in class_or_tuple: + return True + + obj_type = type(obj) + # Classes with obj's type + class_or_tuple = {t for t in class_or_tuple if isinstance(obj, get_origin(t) or t)} + + # Singular types (e.g. int, ControlNet, ...) + # Untyped collections (e.g. List, but not List[int]) + elem_class_or_tuple = {get_args(t) for t in class_or_tuple} + if () in elem_class_or_tuple: + return True + # Typed lists or sets + elif obj_type in (list, set): + return any(all(_is_valid_type(x, t) for x in obj) for t in elem_class_or_tuple) + # Typed tuples + elif obj_type is tuple: + return any( + # Tuples with any length and single type (e.g. Tuple[int, ...]) + (len(t) == 2 and t[-1] is Ellipsis and all(_is_valid_type(x, t[0]) for x in obj)) + or + # Tuples with fixed length and any types (e.g. Tuple[int, str]) + (len(obj) == len(t) and all(_is_valid_type(x, tt) for x, tt in zip(obj, t))) + for t in elem_class_or_tuple + ) + # Typed dicts + elif obj_type is dict: + return any( + all(_is_valid_type(k, kt) and _is_valid_type(v, vt) for k, v in obj.items()) + for kt, vt in elem_class_or_tuple + ) + + else: + return False + + +def _get_detailed_type(obj: Any) -> Type: + """ + Gets a detailed type for an object, including nested types for collections. + """ + obj_type = type(obj) + + if obj_type in (list, set): + obj_origin_type = List if obj_type is list else Set + elems_type = Union[tuple({_get_detailed_type(x) for x in obj})] + return obj_origin_type[elems_type] + elif obj_type is tuple: + return Tuple[tuple(_get_detailed_type(x) for x in obj)] + elif obj_type is dict: + keys_type = Union[tuple({_get_detailed_type(k) for k in obj.keys()})] + values_type = Union[tuple({_get_detailed_type(k) for k in obj.values()})] + return Dict[keys_type, values_type] + else: + return obj_type diff --git a/venv/lib/python3.10/site-packages/diffusers/utils/versions.py b/venv/lib/python3.10/site-packages/diffusers/utils/versions.py new file mode 100644 index 0000000000000000000000000000000000000000..945a3977ce62a9a55307862193e4be6f12c3c17f --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/utils/versions.py @@ -0,0 +1,117 @@ +# Copyright 2020 The HuggingFace Team. All rights reserved. +# +# 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. +""" +Utilities for working with package versions +""" + +import importlib.metadata +import operator +import re +import sys +from typing import Optional + +from packaging import version + + +ops = { + "<": operator.lt, + "<=": operator.le, + "==": operator.eq, + "!=": operator.ne, + ">=": operator.ge, + ">": operator.gt, +} + + +def _compare_versions(op, got_ver, want_ver, requirement, pkg, hint): + if got_ver is None or want_ver is None: + raise ValueError( + f"Unable to compare versions for {requirement}: need={want_ver} found={got_ver}. This is unusual. Consider" + f" reinstalling {pkg}." + ) + if not ops[op](version.parse(got_ver), version.parse(want_ver)): + raise ImportError( + f"{requirement} is required for a normal functioning of this module, but found {pkg}=={got_ver}.{hint}" + ) + + +def require_version(requirement: str, hint: Optional[str] = None) -> None: + """ + Perform a runtime check of the dependency versions, using the exact same syntax used by pip. + + The installed module version comes from the *site-packages* dir via *importlib.metadata*. + + Args: + requirement (`str`): pip style definition, e.g., "tokenizers==0.9.4", "tqdm>=4.27", "numpy" + hint (`str`, *optional*): what suggestion to print in case of requirements not being met + + Example: + + ```python + require_version("pandas>1.1.2") + require_version("numpy>1.18.5", "this is important to have for whatever reason") + ```""" + + hint = f"\n{hint}" if hint is not None else "" + + # non-versioned check + if re.match(r"^[\w_\-\d]+$", requirement): + pkg, op, want_ver = requirement, None, None + else: + match = re.findall(r"^([^!=<>\s]+)([\s!=<>]{1,2}.+)", requirement) + if not match: + raise ValueError( + "requirement needs to be in the pip package format, .e.g., package_a==1.23, or package_b>=1.23, but" + f" got {requirement}" + ) + pkg, want_full = match[0] + want_range = want_full.split(",") # there could be multiple requirements + wanted = {} + for w in want_range: + match = re.findall(r"^([\s!=<>]{1,2})(.+)", w) + if not match: + raise ValueError( + "requirement needs to be in the pip package format, .e.g., package_a==1.23, or package_b>=1.23," + f" but got {requirement}" + ) + op, want_ver = match[0] + wanted[op] = want_ver + if op not in ops: + raise ValueError(f"{requirement}: need one of {list(ops.keys())}, but got {op}") + + # special case + if pkg == "python": + got_ver = ".".join([str(x) for x in sys.version_info[:3]]) + for op, want_ver in wanted.items(): + _compare_versions(op, got_ver, want_ver, requirement, pkg, hint) + return + + # check if any version is installed + try: + got_ver = importlib.metadata.version(pkg) + except importlib.metadata.PackageNotFoundError: + raise importlib.metadata.PackageNotFoundError( + f"The '{requirement}' distribution was not found and is required by this application. {hint}" + ) + + # check that the right version is installed if version number or a range was provided + if want_ver is not None: + for op, want_ver in wanted.items(): + _compare_versions(op, got_ver, want_ver, requirement, pkg, hint) + + +def require_version_core(requirement): + """require_version wrapper which emits a core-specific hint on failure""" + hint = "Try: pip install transformers -U or pip install -e '.[dev]' if you're working with git main" + return require_version(requirement, hint) diff --git a/venv/lib/python3.10/site-packages/diffusers/video_processor.py b/venv/lib/python3.10/site-packages/diffusers/video_processor.py new file mode 100644 index 0000000000000000000000000000000000000000..59b59b47d2c7d194874ab56753ca3efdbff90f5e --- /dev/null +++ b/venv/lib/python3.10/site-packages/diffusers/video_processor.py @@ -0,0 +1,113 @@ +# Copyright 2025 The HuggingFace Team. All rights reserved. +# +# 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 warnings +from typing import List, Optional, Union + +import numpy as np +import PIL +import torch + +from .image_processor import VaeImageProcessor, is_valid_image, is_valid_image_imagelist + + +class VideoProcessor(VaeImageProcessor): + r"""Simple video processor.""" + + def preprocess_video(self, video, height: Optional[int] = None, width: Optional[int] = None) -> torch.Tensor: + r""" + Preprocesses input video(s). + + Args: + video (`List[PIL.Image]`, `List[List[PIL.Image]]`, `torch.Tensor`, `np.array`, `List[torch.Tensor]`, `List[np.array]`): + The input video. It can be one of the following: + * List of the PIL images. + * List of list of PIL images. + * 4D Torch tensors (expected shape for each tensor `(num_frames, num_channels, height, width)`). + * 4D NumPy arrays (expected shape for each array `(num_frames, height, width, num_channels)`). + * List of 4D Torch tensors (expected shape for each tensor `(num_frames, num_channels, height, + width)`). + * List of 4D NumPy arrays (expected shape for each array `(num_frames, height, width, num_channels)`). + * 5D NumPy arrays: expected shape for each array `(batch_size, num_frames, height, width, + num_channels)`. + * 5D Torch tensors: expected shape for each array `(batch_size, num_frames, num_channels, height, + width)`. + height (`int`, *optional*, defaults to `None`): + The height in preprocessed frames of the video. If `None`, will use the `get_default_height_width()` to + get default height. + width (`int`, *optional*`, defaults to `None`): + The width in preprocessed frames of the video. If `None`, will use get_default_height_width()` to get + the default width. + """ + if isinstance(video, list) and isinstance(video[0], np.ndarray) and video[0].ndim == 5: + warnings.warn( + "Passing `video` as a list of 5d np.ndarray is deprecated." + "Please concatenate the list along the batch dimension and pass it as a single 5d np.ndarray", + FutureWarning, + ) + video = np.concatenate(video, axis=0) + if isinstance(video, list) and isinstance(video[0], torch.Tensor) and video[0].ndim == 5: + warnings.warn( + "Passing `video` as a list of 5d torch.Tensor is deprecated." + "Please concatenate the list along the batch dimension and pass it as a single 5d torch.Tensor", + FutureWarning, + ) + video = torch.cat(video, axis=0) + + # ensure the input is a list of videos: + # - if it is a batch of videos (5d torch.Tensor or np.ndarray), it is converted to a list of videos (a list of 4d torch.Tensor or np.ndarray) + # - if it is a single video, it is converted to a list of one video. + if isinstance(video, (np.ndarray, torch.Tensor)) and video.ndim == 5: + video = list(video) + elif isinstance(video, list) and is_valid_image(video[0]) or is_valid_image_imagelist(video): + video = [video] + elif isinstance(video, list) and is_valid_image_imagelist(video[0]): + video = video + else: + raise ValueError( + "Input is in incorrect format. Currently, we only support numpy.ndarray, torch.Tensor, PIL.Image.Image" + ) + + video = torch.stack([self.preprocess(img, height=height, width=width) for img in video], dim=0) + + # move the number of channels before the number of frames. + video = video.permute(0, 2, 1, 3, 4) + + return video + + def postprocess_video( + self, video: torch.Tensor, output_type: str = "np" + ) -> Union[np.ndarray, torch.Tensor, List[PIL.Image.Image]]: + r""" + Converts a video tensor to a list of frames for export. + + Args: + video (`torch.Tensor`): The video as a tensor. + output_type (`str`, defaults to `"np"`): Output type of the postprocessed `video` tensor. + """ + batch_size = video.shape[0] + outputs = [] + for batch_idx in range(batch_size): + batch_vid = video[batch_idx].permute(1, 0, 2, 3) + batch_output = self.postprocess(batch_vid, output_type) + outputs.append(batch_output) + + if output_type == "np": + outputs = np.stack(outputs) + elif output_type == "pt": + outputs = torch.stack(outputs) + elif not output_type == "pil": + raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil']") + + return outputs diff --git a/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/INSTALLER b/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/INSTALLER new file mode 100644 index 0000000000000000000000000000000000000000..a1b589e38a32041e49332e5e81c2d363dc418d68 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/INSTALLER @@ -0,0 +1 @@ +pip diff --git a/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/LICENSE b/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..76f67efdc6470081b512a8db5bf2b1d4962d9c3c --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2023 Georgi Gerganov + +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. diff --git a/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/METADATA b/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/METADATA new file mode 100644 index 0000000000000000000000000000000000000000..9e4fb97c351b77c21be3ba04cdfe8c6c6b420ee4 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/METADATA @@ -0,0 +1,126 @@ +Metadata-Version: 2.3 +Name: gguf +Version: 0.17.1 +Summary: Read and write ML models in GGUF for GGML +Keywords: ggml,gguf,llama.cpp +Author: GGML +Author-email: ggml@ggml.ai +Requires-Python: >=3.8 +Classifier: License :: OSI Approved :: MIT License +Classifier: Operating System :: OS Independent +Classifier: Programming Language :: Python :: 3 +Classifier: Programming Language :: Python :: 3.8 +Classifier: Programming Language :: Python :: 3.9 +Classifier: Programming Language :: Python :: 3.10 +Classifier: Programming Language :: Python :: 3.11 +Classifier: Programming Language :: Python :: 3.12 +Classifier: Programming Language :: Python :: 3.13 +Provides-Extra: gui +Requires-Dist: PySide6 (>=6.9,<7.0) ; (python_version >= "3.9" and python_version < "3.14") and (extra == "gui") +Requires-Dist: numpy (>=1.17) +Requires-Dist: pyyaml (>=5.1) +Requires-Dist: tqdm (>=4.27) +Project-URL: Homepage, https://ggml.ai +Project-URL: Repository, https://github.com/ggml-org/llama.cpp +Description-Content-Type: text/markdown + +## gguf + +This is a Python package for writing binary files in the [GGUF](https://github.com/ggml-org/ggml/pull/302) +(GGML Universal File) format. + +See [convert_hf_to_gguf.py](https://github.com/ggml-org/llama.cpp/blob/master/convert_hf_to_gguf.py) +as an example for its usage. + +## Installation +```sh +pip install gguf +``` + +Optionally, you can install gguf with the extra 'gui' to enable the visual GGUF editor. +```sh +pip install gguf[gui] +``` + +## API Examples/Simple Tools + +[examples/writer.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/examples/writer.py) — Generates `example.gguf` in the current directory to demonstrate generating a GGUF file. Note that this file cannot be used as a model. + +[examples/reader.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/examples/reader.py) — Extracts and displays key-value pairs and tensor details from a GGUF file in a readable format. + +[gguf/scripts/gguf_dump.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_dump.py) — Dumps a GGUF file's metadata to the console. + +[gguf/scripts/gguf_set_metadata.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_set_metadata.py) — Allows changing simple metadata values in a GGUF file by key. + +[gguf/scripts/gguf_convert_endian.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_convert_endian.py) — Allows converting the endianness of GGUF files. + +[gguf/scripts/gguf_new_metadata.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_new_metadata.py) — Copies a GGUF file with added/modified/removed metadata values. + +[gguf/scripts/gguf_editor_gui.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_editor_gui.py) — Allows for viewing, editing, adding, or removing metadata values within a GGUF file as well as viewing its tensors with a Qt interface. + +## Development +Maintainers who participate in development of this package are advised to install it in editable mode: + +```sh +cd /path/to/llama.cpp/gguf-py + +pip install --editable . +``` + +**Note**: This may require to upgrade your Pip installation, with a message saying that editable installation currently requires `setup.py`. +In this case, upgrade Pip to the latest: + +```sh +pip install --upgrade pip +``` + +## Automatic publishing with CI + +There's a GitHub workflow to make a release automatically upon creation of tags in a specified format. + +1. Bump the version in `pyproject.toml`. +2. Create a tag named `gguf-vx.x.x` where `x.x.x` is the semantic version number. + +```sh +git tag -a gguf-v1.0.0 -m "Version 1.0 release" +``` + +3. Push the tags. + +```sh +git push origin --tags +``` + +## Manual publishing +If you want to publish the package manually for any reason, you need to have `twine` and `build` installed: + +```sh +pip install build twine +``` + +Then, follow these steps to release a new version: + +1. Bump the version in `pyproject.toml`. +2. Build the package: + +```sh +python -m build +``` + +3. Upload the generated distribution archives: + +```sh +python -m twine upload dist/* +``` + +## Run Unit Tests + +From root of this repository you can run this command to run all the unit tests + +```bash +python -m unittest discover ./gguf-py -v +``` + +## TODO +- [ ] Include conversion scripts as command line entry points in this package. + diff --git a/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/RECORD b/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/RECORD new file mode 100644 index 0000000000000000000000000000000000000000..b83178ee6ac492fe40229230f0d0b4a2054ee83b --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/RECORD @@ -0,0 +1,47 @@ +../../../bin/gguf-convert-endian,sha256=dnXJg54diixDuUXImRlYLQCtDYdWr0aIMFVKOUEqHXA,249 +../../../bin/gguf-dump,sha256=dvhrFZ73eFsZ9qcrzDpNsCORWGARak19D3OPqHkiuTM,239 +../../../bin/gguf-editor-gui,sha256=39oEXapsbJzBsHp6UHzA7DXK6h5h6g2B08k8GI5i6xs,245 +../../../bin/gguf-new-metadata,sha256=-pNoLbvw09fEAZwAKKbeHecty6xCsfeE_22U8JKTrX8,247 +../../../bin/gguf-set-metadata,sha256=ooevAx7qxsR5j2cq3E4lXGLBjcV2niEdSl4k3UIhDag,247 +gguf-0.17.1.dist-info/INSTALLER,sha256=zuuue4knoyJ-UwPPXg8fezS7VCrXJQrAP7zeNuwvFQg,4 +gguf-0.17.1.dist-info/LICENSE,sha256=73jH5mWeNMeYGU8NNE6AfHIt5wy8oTWe9UdyZh4Ryjg,1072 +gguf-0.17.1.dist-info/METADATA,sha256=pa8_ce5ufoWKHq1zzh97He53EZjFtRFONFDWtl7mkAI,4348 +gguf-0.17.1.dist-info/RECORD,, +gguf-0.17.1.dist-info/REQUESTED,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0 +gguf-0.17.1.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88 +gguf-0.17.1.dist-info/entry_points.txt,sha256=TozYSFmMxpOaKE3brn9nWp-QkgM_sZ3a_uFDOXUCYig,273 +gguf/__init__.py,sha256=PM_AEEzX6ojGAodDt78_LIm19HRCXeA6IXpgcjINfC8,219 +gguf/__pycache__/__init__.cpython-310.pyc,, +gguf/__pycache__/constants.cpython-310.pyc,, +gguf/__pycache__/gguf.cpython-310.pyc,, +gguf/__pycache__/gguf_reader.cpython-310.pyc,, +gguf/__pycache__/gguf_writer.cpython-310.pyc,, +gguf/__pycache__/lazy.cpython-310.pyc,, +gguf/__pycache__/metadata.cpython-310.pyc,, +gguf/__pycache__/quants.cpython-310.pyc,, +gguf/__pycache__/tensor_mapping.cpython-310.pyc,, +gguf/__pycache__/utility.cpython-310.pyc,, +gguf/__pycache__/vocab.cpython-310.pyc,, +gguf/constants.py,sha256=-92nfNa8XjXbHsMV9Hf0iGHwcmoXzH2JJ73rAv73NPE,89508 +gguf/gguf.py,sha256=8MDu7a0JEXhLUv_tjhYqDrWubVNc41cFvBYZbkZZenI,478 +gguf/gguf_reader.py,sha256=6uI4vaLeRC2MJV-uUjgsRoUZ-Rdszi8mE_bEPwnk6QE,14828 +gguf/gguf_writer.py,sha256=UJXqFa3T2EOLfPZ6ej3vuyV7OBfKXXaszvHKeHIWmSI,43935 +gguf/lazy.py,sha256=Axy_plbHoC34nMeK_ORbbxEyEFm-Fz8BxFIeqeipSYA,9214 +gguf/metadata.py,sha256=v0kxoYfCk_yXZSnzlSO9AhQ6z_HOUpzs0c6ZV4BRSbA,33301 +gguf/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0 +gguf/quants.py,sha256=BLtCCqhHBtabaYaAp7EckyYOR5idmoKWhu_Hy3glotk,60771 +gguf/scripts/__pycache__/gguf_convert_endian.cpython-310.pyc,, +gguf/scripts/__pycache__/gguf_dump.cpython-310.pyc,, +gguf/scripts/__pycache__/gguf_editor_gui.cpython-310.pyc,, +gguf/scripts/__pycache__/gguf_hash.cpython-310.pyc,, +gguf/scripts/__pycache__/gguf_new_metadata.cpython-310.pyc,, +gguf/scripts/__pycache__/gguf_set_metadata.cpython-310.pyc,, +gguf/scripts/gguf_convert_endian.py,sha256=yzl_MAQ3jyn_9MmOWV1CksHqlohd7DmrG7REwas0rlo,7365 +gguf/scripts/gguf_dump.py,sha256=zDgZSSQLyO3S4YJsSUzdebDfwmdqQPN7_VtFZ5BkGAk,21785 +gguf/scripts/gguf_editor_gui.py,sha256=frdErSIB90N-sAvqUpbLfdDsaUGMMOWQ-0iumwzzm_M,64398 +gguf/scripts/gguf_hash.py,sha256=nyd8kzjRKnOFek5UaD19pNXeAVMXUfFEASZ8konkGX8,3725 +gguf/scripts/gguf_new_metadata.py,sha256=U_v5FgbH292x7bsi2dG4rbQcWc14nmAtZEWdLnbkIZs,9767 +gguf/scripts/gguf_set_metadata.py,sha256=yGEqcQlCimd-pVl23V7u1giJNN3vfvASRqW8em5YWzs,4145 +gguf/tensor_mapping.py,sha256=6pUDYgly0-yErLV8HSZFSnqyFKAp7SieditMEHMsYGI,55999 +gguf/utility.py,sha256=80rZ3MdGZ6bX0_yFvLoPTTOlxga8THcihF0q38y5h6M,10808 +gguf/vocab.py,sha256=Gr43idKqc4yBHQh20HfeW9oegYWLpM-L0ZKCVxPj4Lc,20788 diff --git a/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/REQUESTED b/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/REQUESTED new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/WHEEL b/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/WHEEL new file mode 100644 index 0000000000000000000000000000000000000000..9ed4d8facdffc976b12b8144fcda70b971ffea8f --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/WHEEL @@ -0,0 +1,4 @@ +Wheel-Version: 1.0 +Generator: poetry-core 2.1.3 +Root-Is-Purelib: true +Tag: py3-none-any diff --git a/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/entry_points.txt b/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/entry_points.txt new file mode 100644 index 0000000000000000000000000000000000000000..5502bee07257f9f8e862720f7253205ff217f4d3 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf-0.17.1.dist-info/entry_points.txt @@ -0,0 +1,7 @@ +[console_scripts] +gguf-convert-endian=gguf.scripts.gguf_convert_endian:main +gguf-dump=gguf.scripts.gguf_dump:main +gguf-editor-gui=gguf.scripts.gguf_editor_gui:main +gguf-new-metadata=gguf.scripts.gguf_new_metadata:main +gguf-set-metadata=gguf.scripts.gguf_set_metadata:main + diff --git a/venv/lib/python3.10/site-packages/gguf/__init__.py b/venv/lib/python3.10/site-packages/gguf/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..243defc4c1ca42d3713017d8902592f54ac849cd --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/__init__.py @@ -0,0 +1,9 @@ +from .constants import * +from .lazy import * +from .gguf_reader import * +from .gguf_writer import * +from .quants import * +from .tensor_mapping import * +from .vocab import * +from .utility import * +from .metadata import * diff --git a/venv/lib/python3.10/site-packages/gguf/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9c1eb627fd02faa46d38a69fb01ea9c1684e768c Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/__pycache__/constants.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/__pycache__/constants.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..eddedc5206d63f64c37fbe86ed2f28a220b7ce33 Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/__pycache__/constants.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/__pycache__/gguf.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/__pycache__/gguf.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..da2fb023e0613dff7d563a05dbbd913154244922 Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/__pycache__/gguf.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/__pycache__/gguf_reader.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/__pycache__/gguf_reader.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e9d3e53f196bec3f88d2a5eb411da12adc4b30c7 Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/__pycache__/gguf_reader.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/__pycache__/gguf_writer.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/__pycache__/gguf_writer.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b029077befaad30188833cc4595b54cb3cedf70f Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/__pycache__/gguf_writer.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/__pycache__/lazy.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/__pycache__/lazy.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b82d53ab649dac4fb61a5fb865d56c0b6fbc3f8f Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/__pycache__/lazy.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/__pycache__/metadata.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/__pycache__/metadata.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e8742f7e3fc44d44816514f59f778d69a9805329 Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/__pycache__/metadata.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/__pycache__/quants.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/__pycache__/quants.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fcee09ae4d3aa53c7c9382aa718dbe88cc6ff323 Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/__pycache__/quants.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/__pycache__/tensor_mapping.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/__pycache__/tensor_mapping.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ed2be677a6100f02c03fafb25207bf6a58d553bb Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/__pycache__/tensor_mapping.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/__pycache__/utility.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/__pycache__/utility.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..040620346acbdbb1650b5d61e015be5e5fc94166 Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/__pycache__/utility.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/__pycache__/vocab.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/__pycache__/vocab.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..217268c022806a809a6698c3d4a5af4915924fc7 Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/__pycache__/vocab.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/constants.py b/venv/lib/python3.10/site-packages/gguf/constants.py new file mode 100644 index 0000000000000000000000000000000000000000..834a1d5e1a97ed98f01196e0a0542afa95739b8c --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/constants.py @@ -0,0 +1,2438 @@ +from __future__ import annotations + +from enum import Enum, IntEnum, auto +from typing import Any + +# +# constants +# + +GGUF_MAGIC = 0x46554747 # "GGUF" +GGUF_VERSION = 3 +GGUF_DEFAULT_ALIGNMENT = 32 +GGML_QUANT_VERSION = 2 # GGML_QNT_VERSION from ggml.h + +# +# metadata keys +# + + +class Keys: + class General: + TYPE = "general.type" + ARCHITECTURE = "general.architecture" + QUANTIZATION_VERSION = "general.quantization_version" + ALIGNMENT = "general.alignment" + FILE_TYPE = "general.file_type" + + # Authorship Metadata + NAME = "general.name" + AUTHOR = "general.author" + VERSION = "general.version" + ORGANIZATION = "general.organization" + + FINETUNE = "general.finetune" + BASENAME = "general.basename" + + DESCRIPTION = "general.description" + QUANTIZED_BY = "general.quantized_by" + + SIZE_LABEL = "general.size_label" + + # Licensing details + LICENSE = "general.license" + LICENSE_NAME = "general.license.name" + LICENSE_LINK = "general.license.link" + + # Typically represents the converted GGUF repo (Unless native) + URL = "general.url" # Model Website/Paper + DOI = "general.doi" + UUID = "general.uuid" + REPO_URL = "general.repo_url" # Model Source Repository (git/svn/etc...) + + # Model Source during conversion + SOURCE_URL = "general.source.url" # Model Website/Paper + SOURCE_DOI = "general.source.doi" + SOURCE_UUID = "general.source.uuid" + SOURCE_REPO_URL = "general.source.repo_url" # Model Source Repository (git/svn/etc...) + + # Base Model Source. There can be more than one source if it's a merged + # model like with 'Mistral-7B-Merge-14-v0.1'. This will assist in + # tracing linage of models as it is finetuned or merged over time. + BASE_MODEL_COUNT = "general.base_model.count" + BASE_MODEL_NAME = "general.base_model.{id}.name" + BASE_MODEL_AUTHOR = "general.base_model.{id}.author" + BASE_MODEL_VERSION = "general.base_model.{id}.version" + BASE_MODEL_ORGANIZATION = "general.base_model.{id}.organization" + BASE_MODEL_DESCRIPTION = "general.base_model.{id}.description" + BASE_MODEL_URL = "general.base_model.{id}.url" # Model Website/Paper + BASE_MODEL_DOI = "general.base_model.{id}.doi" + BASE_MODEL_UUID = "general.base_model.{id}.uuid" + BASE_MODEL_REPO_URL = "general.base_model.{id}.repo_url" # Model Source Repository (git/svn/etc...) + + # Dataset Source + DATASET_COUNT = "general.dataset.count" + DATASET_NAME = "general.dataset.{id}.name" + DATASET_AUTHOR = "general.dataset.{id}.author" + DATASET_VERSION = "general.dataset.{id}.version" + DATASET_ORGANIZATION = "general.dataset.{id}.organization" + DATASET_DESCRIPTION = "general.dataset.{id}.description" + DATASET_URL = "general.dataset.{id}.url" # Model Website/Paper + DATASET_DOI = "general.dataset.{id}.doi" + DATASET_UUID = "general.dataset.{id}.uuid" + DATASET_REPO_URL = "general.dataset.{id}.repo_url" # Model Source Repository (git/svn/etc...) + + # Array based KV stores + TAGS = "general.tags" + LANGUAGES = "general.languages" + + class LLM: + VOCAB_SIZE = "{arch}.vocab_size" + CONTEXT_LENGTH = "{arch}.context_length" + EMBEDDING_LENGTH = "{arch}.embedding_length" + FEATURES_LENGTH = "{arch}.features_length" + BLOCK_COUNT = "{arch}.block_count" + LEADING_DENSE_BLOCK_COUNT = "{arch}.leading_dense_block_count" + FEED_FORWARD_LENGTH = "{arch}.feed_forward_length" + EXPERT_FEED_FORWARD_LENGTH = "{arch}.expert_feed_forward_length" + EXPERT_SHARED_FEED_FORWARD_LENGTH = "{arch}.expert_shared_feed_forward_length" + USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual" + TENSOR_DATA_LAYOUT = "{arch}.tensor_data_layout" + EXPERT_COUNT = "{arch}.expert_count" + EXPERT_USED_COUNT = "{arch}.expert_used_count" + EXPERT_SHARED_COUNT = "{arch}.expert_shared_count" + EXPERT_WEIGHTS_SCALE = "{arch}.expert_weights_scale" + EXPERT_WEIGHTS_NORM = "{arch}.expert_weights_norm" + EXPERT_GATING_FUNC = "{arch}.expert_gating_func" + MOE_EVERY_N_LAYERS = "{arch}.moe_every_n_layers" + POOLING_TYPE = "{arch}.pooling_type" + LOGIT_SCALE = "{arch}.logit_scale" + DECODER_START_TOKEN_ID = "{arch}.decoder_start_token_id" + ATTN_LOGIT_SOFTCAPPING = "{arch}.attn_logit_softcapping" + FINAL_LOGIT_SOFTCAPPING = "{arch}.final_logit_softcapping" + SWIN_NORM = "{arch}.swin_norm" + RESCALE_EVERY_N_LAYERS = "{arch}.rescale_every_n_layers" + TIME_MIX_EXTRA_DIM = "{arch}.time_mix_extra_dim" + TIME_DECAY_EXTRA_DIM = "{arch}.time_decay_extra_dim" + RESIDUAL_SCALE = "{arch}.residual_scale" + EMBEDDING_SCALE = "{arch}.embedding_scale" + TOKEN_SHIFT_COUNT = "{arch}.token_shift_count" + INTERLEAVE_MOE_LAYER_STEP = "{arch}.interleave_moe_layer_step" + + class Attention: + HEAD_COUNT = "{arch}.attention.head_count" + HEAD_COUNT_KV = "{arch}.attention.head_count_kv" + MAX_ALIBI_BIAS = "{arch}.attention.max_alibi_bias" + CLAMP_KQV = "{arch}.attention.clamp_kqv" + KEY_LENGTH = "{arch}.attention.key_length" + VALUE_LENGTH = "{arch}.attention.value_length" + LAYERNORM_EPS = "{arch}.attention.layer_norm_epsilon" + LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon" + GROUPNORM_EPS = "{arch}.attention.group_norm_epsilon" + GROUPNORM_GROUPS = "{arch}.attention.group_norm_groups" + CAUSAL = "{arch}.attention.causal" + Q_LORA_RANK = "{arch}.attention.q_lora_rank" + KV_LORA_RANK = "{arch}.attention.kv_lora_rank" + DECAY_LORA_RANK = "{arch}.attention.decay_lora_rank" + ICLR_LORA_RANK = "{arch}.attention.iclr_lora_rank" + VALUE_RESIDUAL_MIX_LORA_RANK = "{arch}.attention.value_residual_mix_lora_rank" + GATE_LORA_RANK = "{arch}.attention.gate_lora_rank" + REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count" + SLIDING_WINDOW = "{arch}.attention.sliding_window" + SCALE = "{arch}.attention.scale" + KEY_LENGTH_MLA = "{arch}.attention.key_length_mla" + VALUE_LENGTH_MLA = "{arch}.attention.value_length_mla" + + class Rope: + DIMENSION_COUNT = "{arch}.rope.dimension_count" + DIMENSION_SECTIONS = "{arch}.rope.dimension_sections" + FREQ_BASE = "{arch}.rope.freq_base" + SCALING_TYPE = "{arch}.rope.scaling.type" + SCALING_FACTOR = "{arch}.rope.scaling.factor" + SCALING_ATTN_FACTOR = "{arch}.rope.scaling.attn_factor" + SCALING_ORIG_CTX_LEN = "{arch}.rope.scaling.original_context_length" + SCALING_FINETUNED = "{arch}.rope.scaling.finetuned" + SCALING_YARN_LOG_MUL = "{arch}.rope.scaling.yarn_log_multiplier" + + class Split: + LLM_KV_SPLIT_NO = "split.no" + LLM_KV_SPLIT_COUNT = "split.count" + LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count" + + class SSM: + CONV_KERNEL = "{arch}.ssm.conv_kernel" + INNER_SIZE = "{arch}.ssm.inner_size" + STATE_SIZE = "{arch}.ssm.state_size" + TIME_STEP_RANK = "{arch}.ssm.time_step_rank" + DT_B_C_RMS = "{arch}.ssm.dt_b_c_rms" + + class WKV: + HEAD_SIZE = "{arch}.wkv.head_size" + + class PosNet: + EMBEDDING_LENGTH = "{arch}.posnet.embedding_length" + BLOCK_COUNT = "{arch}.posnet.block_count" + + class ConvNext: + EMBEDDING_LENGTH = "{arch}.convnext.embedding_length" + BLOCK_COUNT = "{arch}.convnext.block_count" + + class Classifier: + OUTPUT_LABELS = "{arch}.classifier.output_labels" + + class Tokenizer: + MODEL = "tokenizer.ggml.model" + PRE = "tokenizer.ggml.pre" + LIST = "tokenizer.ggml.tokens" + TOKEN_TYPE = "tokenizer.ggml.token_type" + TOKEN_TYPE_COUNT = "tokenizer.ggml.token_type_count" # for BERT-style token types + SCORES = "tokenizer.ggml.scores" + MERGES = "tokenizer.ggml.merges" + BOS_ID = "tokenizer.ggml.bos_token_id" + EOS_ID = "tokenizer.ggml.eos_token_id" + EOT_ID = "tokenizer.ggml.eot_token_id" + EOM_ID = "tokenizer.ggml.eom_token_id" + UNK_ID = "tokenizer.ggml.unknown_token_id" + SEP_ID = "tokenizer.ggml.seperator_token_id" + PAD_ID = "tokenizer.ggml.padding_token_id" + MASK_ID = "tokenizer.ggml.mask_token_id" + ADD_BOS = "tokenizer.ggml.add_bos_token" + ADD_EOS = "tokenizer.ggml.add_eos_token" + ADD_PREFIX = "tokenizer.ggml.add_space_prefix" + REMOVE_EXTRA_WS = "tokenizer.ggml.remove_extra_whitespaces" + PRECOMPILED_CHARSMAP = "tokenizer.ggml.precompiled_charsmap" + HF_JSON = "tokenizer.huggingface.json" + RWKV = "tokenizer.rwkv.world" + CHAT_TEMPLATE = "tokenizer.chat_template" + CHAT_TEMPLATE_N = "tokenizer.chat_template.{name}" + CHAT_TEMPLATES = "tokenizer.chat_templates" + # FIM/Infill special tokens constants + FIM_PRE_ID = "tokenizer.ggml.fim_pre_token_id" + FIM_SUF_ID = "tokenizer.ggml.fim_suf_token_id" + FIM_MID_ID = "tokenizer.ggml.fim_mid_token_id" + FIM_PAD_ID = "tokenizer.ggml.fim_pad_token_id" + FIM_REP_ID = "tokenizer.ggml.fim_rep_token_id" + FIM_SEP_ID = "tokenizer.ggml.fim_sep_token_id" + # deprecated: + PREFIX_ID = "tokenizer.ggml.prefix_token_id" + SUFFIX_ID = "tokenizer.ggml.suffix_token_id" + MIDDLE_ID = "tokenizer.ggml.middle_token_id" + + class Adapter: + TYPE = "adapter.type" + LORA_ALPHA = "adapter.lora.alpha" + + class Clip: + PROJECTOR_TYPE = "clip.projector_type" + HAS_VISION_ENCODER = "clip.has_vision_encoder" + HAS_AUDIO_ENCODER = "clip.has_audio_encoder" + HAS_LLAVA_PROJECTOR = "clip.has_llava_projector" + + class ClipVision: + IMAGE_SIZE = "clip.vision.image_size" + PATCH_SIZE = "clip.vision.patch_size" + EMBEDDING_LENGTH = "clip.vision.embedding_length" + FEED_FORWARD_LENGTH = "clip.vision.feed_forward_length" + PROJECTION_DIM = "clip.vision.projection_dim" + BLOCK_COUNT = "clip.vision.block_count" + IMAGE_MEAN = "clip.vision.image_mean" + IMAGE_STD = "clip.vision.image_std" + SPATIAL_MERGE_SIZE = "clip.vision.spatial_merge_size" + USE_GELU = "clip.use_gelu" + USE_SILU = "clip.use_silu" + N_WA_PATTERN = "clip.vision.n_wa_pattern" # used by qwen2.5vl + + class Attention: + HEAD_COUNT = "clip.vision.attention.head_count" + LAYERNORM_EPS = "clip.vision.attention.layer_norm_epsilon" + + class Projector: + SCALE_FACTOR = "clip.vision.projector.scale_factor" + + class ClipAudio: + NUM_MEL_BINS = "clip.audio.num_mel_bins" + EMBEDDING_LENGTH = "clip.audio.embedding_length" + FEED_FORWARD_LENGTH = "clip.audio.feed_forward_length" + PROJECTION_DIM = "clip.audio.projection_dim" + BLOCK_COUNT = "clip.audio.block_count" + + class Attention: + HEAD_COUNT = "clip.audio.attention.head_count" + LAYERNORM_EPS = "clip.audio.attention.layer_norm_epsilon" + + class Projector: + STACK_FACTOR = "clip.audio.projector.stack_factor" + +# +# recommended mapping of model tensor names for storage in gguf +# + + +class GGUFType: + MODEL = "model" + ADAPTER = "adapter" + MMPROJ = "mmproj" # dummy, unused for now + + +class MODEL_ARCH(IntEnum): + MMPROJ = auto() # dummy arch for clip.cpp + LLAMA = auto() + LLAMA4 = auto() + DECI = auto() + FALCON = auto() + BAICHUAN = auto() + GROK = auto() + GPT2 = auto() + GPTJ = auto() + GPTNEOX = auto() + MPT = auto() + STARCODER = auto() + REFACT = auto() + BERT = auto() + NOMIC_BERT = auto() + NOMIC_BERT_MOE = auto() + NEO_BERT = auto() + JINA_BERT_V2 = auto() + BLOOM = auto() + STABLELM = auto() + QWEN = auto() + QWEN2 = auto() + QWEN2MOE = auto() + QWEN2VL = auto() + QWEN3 = auto() + QWEN3MOE = auto() + PHI2 = auto() + PHI3 = auto() + PHIMOE = auto() + PLAMO = auto() + CODESHELL = auto() + ORION = auto() + INTERNLM2 = auto() + MINICPM = auto() + MINICPM3 = auto() + GEMMA = auto() + GEMMA2 = auto() + GEMMA3 = auto() + STARCODER2 = auto() + RWKV6 = auto() + RWKV6QWEN2 = auto() + RWKV7 = auto() + ARWKV7 = auto() + MAMBA = auto() + XVERSE = auto() + COMMAND_R = auto() + COHERE2 = auto() + DBRX = auto() + OLMO = auto() + OLMO2 = auto() + OLMOE = auto() + OPENELM = auto() + ARCTIC = auto() + DEEPSEEK = auto() + DEEPSEEK2 = auto() + CHATGLM = auto() + GLM4 = auto() + BITNET = auto() + T5 = auto() + T5ENCODER = auto() + JAIS = auto() + NEMOTRON = auto() + EXAONE = auto() + GRANITE = auto() + GRANITE_MOE = auto() + CHAMELEON = auto() + WAVTOKENIZER_DEC = auto() + PLM = auto() + BAILINGMOE = auto() + DOTS1 = auto() + ARCEE = auto() + + +class VISION_PROJECTOR_TYPE(IntEnum): + MLP = auto() + LDP = auto() + LDPV2 = auto() + RESAMPLER = auto() + GLM_EDGE = auto() + MERGER = auto() + GEMMA3 = auto() + + +class MODEL_TENSOR(IntEnum): + TOKEN_EMBD = auto() + TOKEN_EMBD_NORM = auto() + TOKEN_TYPES = auto() + POS_EMBD = auto() + OUTPUT = auto() + OUTPUT_NORM = auto() + ROPE_FREQS = auto() + ROPE_FACTORS_LONG = auto() + ROPE_FACTORS_SHORT = auto() + ATTN_Q = auto() + ATTN_K = auto() + ATTN_V = auto() + ATTN_QKV = auto() + ATTN_OUT = auto() + ATTN_NORM = auto() + ATTN_NORM_2 = auto() + ATTN_OUT_NORM = auto() + ATTN_POST_NORM = auto() + ATTN_ROT_EMBD = auto() + FFN_GATE_INP = auto() + FFN_GATE_INP_SHEXP = auto() + FFN_NORM = auto() + FFN_PRE_NORM = auto() + FFN_POST_NORM = auto() + FFN_GATE = auto() + FFN_DOWN = auto() + FFN_UP = auto() + FFN_ACT = auto() + FFN_NORM_EXP = auto() + FFN_GATE_EXP = auto() + FFN_DOWN_EXP = auto() + FFN_UP_EXP = auto() + FFN_GATE_SHEXP = auto() + FFN_DOWN_SHEXP = auto() + FFN_UP_SHEXP = auto() + FFN_EXP_PROBS_B = auto() + ATTN_Q_NORM = auto() + ATTN_K_NORM = auto() + LAYER_OUT_NORM = auto() + SSM_IN = auto() + SSM_CONV1D = auto() + SSM_X = auto() + SSM_DT = auto() + SSM_A = auto() + SSM_D = auto() + SSM_OUT = auto() + TIME_MIX_W0 = auto() + TIME_MIX_W1 = auto() + TIME_MIX_W2 = auto() + TIME_MIX_A0 = auto() + TIME_MIX_A1 = auto() + TIME_MIX_A2 = auto() + TIME_MIX_V0 = auto() + TIME_MIX_V1 = auto() + TIME_MIX_V2 = auto() + TIME_MIX_G1 = auto() + TIME_MIX_G2 = auto() + TIME_MIX_K_K = auto() + TIME_MIX_K_A = auto() + TIME_MIX_R_K = auto() + TIME_MIX_LERP_X = auto() + TIME_MIX_LERP_K = auto() + TIME_MIX_LERP_V = auto() + TIME_MIX_LERP_R = auto() + TIME_MIX_LERP_G = auto() + TIME_MIX_LERP_FUSED = auto() + TIME_MIX_LERP_W = auto() + TIME_MIX_FIRST = auto() + TIME_MIX_DECAY = auto() + TIME_MIX_DECAY_W1 = auto() + TIME_MIX_DECAY_W2 = auto() + TIME_MIX_KEY = auto() + TIME_MIX_VALUE = auto() + TIME_MIX_RECEPTANCE = auto() + TIME_MIX_GATE = auto() + TIME_MIX_LN = auto() + TIME_MIX_OUTPUT = auto() + CHANNEL_MIX_LERP_K = auto() + CHANNEL_MIX_LERP_R = auto() + CHANNEL_MIX_KEY = auto() + CHANNEL_MIX_RECEPTANCE = auto() + CHANNEL_MIX_VALUE = auto() + ATTN_Q_A = auto() + ATTN_Q_B = auto() + ATTN_KV_A_MQA = auto() + ATTN_KV_B = auto() + ATTN_K_B = auto() + ATTN_V_B = auto() + ATTN_Q_A_NORM = auto() + ATTN_KV_A_NORM = auto() + FFN_SUB_NORM = auto() + ATTN_SUB_NORM = auto() + DEC_ATTN_NORM = auto() + DEC_ATTN_Q = auto() + DEC_ATTN_K = auto() + DEC_ATTN_V = auto() + DEC_ATTN_OUT = auto() + DEC_ATTN_REL_B = auto() + DEC_CROSS_ATTN_NORM = auto() + DEC_CROSS_ATTN_Q = auto() + DEC_CROSS_ATTN_K = auto() + DEC_CROSS_ATTN_V = auto() + DEC_CROSS_ATTN_OUT = auto() + DEC_CROSS_ATTN_REL_B = auto() + DEC_FFN_NORM = auto() + DEC_FFN_GATE = auto() + DEC_FFN_DOWN = auto() + DEC_FFN_UP = auto() + DEC_OUTPUT_NORM = auto() + ENC_ATTN_NORM = auto() + ENC_ATTN_Q = auto() + ENC_ATTN_K = auto() + ENC_ATTN_V = auto() + ENC_ATTN_OUT = auto() + ENC_ATTN_REL_B = auto() + ENC_FFN_NORM = auto() + ENC_FFN_GATE = auto() + ENC_FFN_DOWN = auto() + ENC_FFN_UP = auto() + ENC_OUTPUT_NORM = auto() + CLS = auto() # classifier + CLS_OUT = auto() # classifier output projection + CONV1D = auto() + CONVNEXT_DW = auto() + CONVNEXT_NORM = auto() + CONVNEXT_PW1 = auto() + CONVNEXT_PW2 = auto() + CONVNEXT_GAMMA = auto() + POSNET_CONV1 = auto() + POSNET_CONV2 = auto() + POSNET_NORM = auto() + POSNET_NORM1 = auto() + POSNET_NORM2 = auto() + POSNET_ATTN_NORM = auto() + POSNET_ATTN_Q = auto() + POSNET_ATTN_K = auto() + POSNET_ATTN_V = auto() + POSNET_ATTN_OUT = auto() + # vision + V_MMPROJ = auto() + V_MMPROJ_FC = auto() + V_MMPROJ_MLP = auto() + V_MMPROJ_PEG = auto() + V_ENC_EMBD_CLS = auto() + V_ENC_EMBD_PATCH = auto() + V_ENC_EMBD_POS = auto() + V_ENC_INPUT_NORM = auto() + V_ENC_ATTN_Q = auto() + V_ENC_ATTN_Q_NORM = auto() + V_ENC_ATTN_K = auto() + V_ENC_ATTN_K_NORM = auto() + V_ENC_ATTN_V = auto() + V_ENC_ATTN_O = auto() + V_ENC_ATTN_O_NORM = auto() + V_ENC_POST_ATTN_NORM = auto() + V_ENC_FFN_UP = auto() + V_ENC_FFN_GATE = auto() + V_ENC_FFN_DOWN = auto() + V_LAYER_SCALE_1 = auto() + V_LAYER_SCALE_2 = auto() + V_PRE_NORM = auto() + V_POST_NORM = auto() + V_MM_INP_NORM = auto() + V_MM_INP_PROJ = auto() # gemma3 + V_MM_SOFT_EMB_NORM = auto() # gemma3 + V_RESMPL_POS_EMBD_K = auto() # minicpmv + V_RESMPL_ATTN_Q = auto() # minicpmv + V_RESMPL_ATTN_K = auto() # minicpmv + V_RESMPL_ATTN_V = auto() # minicpmv + V_RESMPL_ATTN_OUT = auto() # minicpmv + V_RESMPL_KV = auto() # minicpmv + V_RESMPL_KV_NORM = auto() # minicpmv + V_RESMPL_POST_NORM = auto() # minicpmv + V_RESMPL_Q_NORM = auto() # minicpmv + V_RESMPL_PROJ = auto() # minicpmv + V_RESMPL_QUERY = auto() # minicpmv + V_TOK_EMBD_IMG_BREAK = auto() # pixtral + V_MM_PATCH_MERGER = auto() # mistral small 3.1 + # audio (mtmd) + A_ENC_EMBD_POS = auto() + A_ENC_CONV1D = auto() + A_PRE_NORM = auto() + A_POST_NORM = auto() + A_ENC_ATTN_Q = auto() + A_ENC_ATTN_K = auto() + A_ENC_ATTN_V = auto() + A_ENC_INPUT_NORM = auto() + A_ENC_OUTPUT = auto() + A_ENC_OUTPUT_NORM = auto() + A_ENC_FFN_UP = auto() + A_ENC_FFN_GATE = auto() + A_ENC_FFN_DOWN = auto() + A_MMPROJ = auto() + A_MMPROJ_FC = auto() + A_MM_NORM_PRE = auto() + A_MM_NORM_MID = auto() + + +MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { + MODEL_ARCH.MMPROJ: "clip", # dummy arch for clip.cpp + MODEL_ARCH.LLAMA: "llama", + MODEL_ARCH.LLAMA4: "llama4", + MODEL_ARCH.DECI: "deci", + MODEL_ARCH.FALCON: "falcon", + MODEL_ARCH.BAICHUAN: "baichuan", + MODEL_ARCH.GROK: "grok", + MODEL_ARCH.GPT2: "gpt2", + MODEL_ARCH.GPTJ: "gptj", + MODEL_ARCH.GPTNEOX: "gptneox", + MODEL_ARCH.MPT: "mpt", + MODEL_ARCH.STARCODER: "starcoder", + MODEL_ARCH.REFACT: "refact", + MODEL_ARCH.BERT: "bert", + MODEL_ARCH.NOMIC_BERT: "nomic-bert", + MODEL_ARCH.NOMIC_BERT_MOE: "nomic-bert-moe", + MODEL_ARCH.NEO_BERT: "neo-bert", + MODEL_ARCH.JINA_BERT_V2: "jina-bert-v2", + MODEL_ARCH.BLOOM: "bloom", + MODEL_ARCH.STABLELM: "stablelm", + MODEL_ARCH.QWEN: "qwen", + MODEL_ARCH.QWEN2: "qwen2", + MODEL_ARCH.QWEN2MOE: "qwen2moe", + MODEL_ARCH.QWEN2VL: "qwen2vl", + MODEL_ARCH.QWEN3: "qwen3", + MODEL_ARCH.QWEN3MOE: "qwen3moe", + MODEL_ARCH.PHI2: "phi2", + MODEL_ARCH.PHI3: "phi3", + MODEL_ARCH.PHIMOE: "phimoe", + MODEL_ARCH.PLAMO: "plamo", + MODEL_ARCH.CODESHELL: "codeshell", + MODEL_ARCH.ORION: "orion", + MODEL_ARCH.INTERNLM2: "internlm2", + MODEL_ARCH.MINICPM: "minicpm", + MODEL_ARCH.MINICPM3: "minicpm3", + MODEL_ARCH.GEMMA: "gemma", + MODEL_ARCH.GEMMA2: "gemma2", + MODEL_ARCH.GEMMA3: "gemma3", + MODEL_ARCH.STARCODER2: "starcoder2", + MODEL_ARCH.RWKV6: "rwkv6", + MODEL_ARCH.RWKV6QWEN2: "rwkv6qwen2", + MODEL_ARCH.RWKV7: "rwkv7", + MODEL_ARCH.ARWKV7: "arwkv7", + MODEL_ARCH.MAMBA: "mamba", + MODEL_ARCH.XVERSE: "xverse", + MODEL_ARCH.COMMAND_R: "command-r", + MODEL_ARCH.COHERE2: "cohere2", + MODEL_ARCH.DBRX: "dbrx", + MODEL_ARCH.OLMO: "olmo", + MODEL_ARCH.OLMO2: "olmo2", + MODEL_ARCH.OLMOE: "olmoe", + MODEL_ARCH.OPENELM: "openelm", + MODEL_ARCH.ARCTIC: "arctic", + MODEL_ARCH.DEEPSEEK: "deepseek", + MODEL_ARCH.DEEPSEEK2: "deepseek2", + MODEL_ARCH.CHATGLM: "chatglm", + MODEL_ARCH.GLM4: "glm4", + MODEL_ARCH.BITNET: "bitnet", + MODEL_ARCH.T5: "t5", + MODEL_ARCH.T5ENCODER: "t5encoder", + MODEL_ARCH.JAIS: "jais", + MODEL_ARCH.NEMOTRON: "nemotron", + MODEL_ARCH.EXAONE: "exaone", + MODEL_ARCH.GRANITE: "granite", + MODEL_ARCH.GRANITE_MOE: "granitemoe", + MODEL_ARCH.CHAMELEON: "chameleon", + MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec", + MODEL_ARCH.PLM: "plm", + MODEL_ARCH.BAILINGMOE: "bailingmoe", + MODEL_ARCH.DOTS1: "dots1", + MODEL_ARCH.ARCEE: "arcee", +} + +VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = { + VISION_PROJECTOR_TYPE.MLP: "mlp", + VISION_PROJECTOR_TYPE.LDP: "ldp", + VISION_PROJECTOR_TYPE.LDPV2: "ldpv2", + VISION_PROJECTOR_TYPE.RESAMPLER: "resampler", + VISION_PROJECTOR_TYPE.GLM_EDGE: "adapter", + VISION_PROJECTOR_TYPE.MERGER: "qwen2vl_merger", + VISION_PROJECTOR_TYPE.GEMMA3: "gemma3", +} + +TENSOR_NAMES: dict[MODEL_TENSOR, str] = { + MODEL_TENSOR.TOKEN_EMBD: "token_embd", + MODEL_TENSOR.TOKEN_EMBD_NORM: "token_embd_norm", + MODEL_TENSOR.TOKEN_TYPES: "token_types", + MODEL_TENSOR.POS_EMBD: "position_embd", + MODEL_TENSOR.OUTPUT_NORM: "output_norm", + MODEL_TENSOR.OUTPUT: "output", + MODEL_TENSOR.ROPE_FREQS: "rope_freqs", + MODEL_TENSOR.ROPE_FACTORS_LONG: "rope_factors_long", + MODEL_TENSOR.ROPE_FACTORS_SHORT: "rope_factors_short", + MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm", + MODEL_TENSOR.ATTN_NORM_2: "blk.{bid}.attn_norm_2", + MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn_qkv", + MODEL_TENSOR.ATTN_Q: "blk.{bid}.attn_q", + MODEL_TENSOR.ATTN_K: "blk.{bid}.attn_k", + MODEL_TENSOR.ATTN_V: "blk.{bid}.attn_v", + MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output", + MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd", + MODEL_TENSOR.ATTN_Q_NORM: "blk.{bid}.attn_q_norm", + MODEL_TENSOR.ATTN_K_NORM: "blk.{bid}.attn_k_norm", + MODEL_TENSOR.ATTN_OUT_NORM: "blk.{bid}.attn_output_norm", + MODEL_TENSOR.ATTN_POST_NORM: "blk.{bid}.post_attention_norm", + MODEL_TENSOR.FFN_GATE_INP: "blk.{bid}.ffn_gate_inp", + MODEL_TENSOR.FFN_GATE_INP_SHEXP: "blk.{bid}.ffn_gate_inp_shexp", + MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm", + MODEL_TENSOR.FFN_PRE_NORM: "blk.{bid}.ffn_norm", + MODEL_TENSOR.FFN_POST_NORM: "blk.{bid}.post_ffw_norm", + MODEL_TENSOR.FFN_GATE: "blk.{bid}.ffn_gate", + MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down", + MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up", + MODEL_TENSOR.FFN_GATE_SHEXP: "blk.{bid}.ffn_gate_shexp", + MODEL_TENSOR.FFN_DOWN_SHEXP: "blk.{bid}.ffn_down_shexp", + MODEL_TENSOR.FFN_UP_SHEXP: "blk.{bid}.ffn_up_shexp", + MODEL_TENSOR.FFN_ACT: "blk.{bid}.ffn", + MODEL_TENSOR.FFN_NORM_EXP: "blk.{bid}.ffn_norm_exps", + MODEL_TENSOR.FFN_GATE_EXP: "blk.{bid}.ffn_gate_exps", + MODEL_TENSOR.FFN_DOWN_EXP: "blk.{bid}.ffn_down_exps", + MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up_exps", + MODEL_TENSOR.FFN_EXP_PROBS_B: "blk.{bid}.exp_probs_b", + MODEL_TENSOR.LAYER_OUT_NORM: "blk.{bid}.layer_output_norm", + MODEL_TENSOR.SSM_IN: "blk.{bid}.ssm_in", + MODEL_TENSOR.SSM_CONV1D: "blk.{bid}.ssm_conv1d", + MODEL_TENSOR.SSM_X: "blk.{bid}.ssm_x", + MODEL_TENSOR.SSM_DT: "blk.{bid}.ssm_dt", + MODEL_TENSOR.SSM_A: "blk.{bid}.ssm_a", + MODEL_TENSOR.SSM_D: "blk.{bid}.ssm_d", + MODEL_TENSOR.SSM_OUT: "blk.{bid}.ssm_out", + MODEL_TENSOR.TIME_MIX_W0: "blk.{bid}.time_mix_w0", + MODEL_TENSOR.TIME_MIX_W1: "blk.{bid}.time_mix_w1", + MODEL_TENSOR.TIME_MIX_W2: "blk.{bid}.time_mix_w2", + MODEL_TENSOR.TIME_MIX_A0: "blk.{bid}.time_mix_a0", + MODEL_TENSOR.TIME_MIX_A1: "blk.{bid}.time_mix_a1", + MODEL_TENSOR.TIME_MIX_A2: "blk.{bid}.time_mix_a2", + MODEL_TENSOR.TIME_MIX_V0: "blk.{bid}.time_mix_v0", + MODEL_TENSOR.TIME_MIX_V1: "blk.{bid}.time_mix_v1", + MODEL_TENSOR.TIME_MIX_V2: "blk.{bid}.time_mix_v2", + MODEL_TENSOR.TIME_MIX_G1: "blk.{bid}.time_mix_g1", + MODEL_TENSOR.TIME_MIX_G2: "blk.{bid}.time_mix_g2", + MODEL_TENSOR.TIME_MIX_K_K: "blk.{bid}.time_mix_k_k", + MODEL_TENSOR.TIME_MIX_K_A: "blk.{bid}.time_mix_k_a", + MODEL_TENSOR.TIME_MIX_R_K: "blk.{bid}.time_mix_r_k", + MODEL_TENSOR.TIME_MIX_LERP_X: "blk.{bid}.time_mix_lerp_x", + MODEL_TENSOR.TIME_MIX_LERP_K: "blk.{bid}.time_mix_lerp_k", + MODEL_TENSOR.TIME_MIX_LERP_V: "blk.{bid}.time_mix_lerp_v", + MODEL_TENSOR.TIME_MIX_LERP_R: "blk.{bid}.time_mix_lerp_r", + MODEL_TENSOR.TIME_MIX_LERP_G: "blk.{bid}.time_mix_lerp_g", + MODEL_TENSOR.TIME_MIX_LERP_FUSED: "blk.{bid}.time_mix_lerp_fused", + MODEL_TENSOR.TIME_MIX_LERP_W: "blk.{bid}.time_mix_lerp_w", + MODEL_TENSOR.TIME_MIX_FIRST: "blk.{bid}.time_mix_first", + MODEL_TENSOR.TIME_MIX_DECAY: "blk.{bid}.time_mix_decay", + MODEL_TENSOR.TIME_MIX_DECAY_W1: "blk.{bid}.time_mix_decay_w1", + MODEL_TENSOR.TIME_MIX_DECAY_W2: "blk.{bid}.time_mix_decay_w2", + MODEL_TENSOR.TIME_MIX_KEY: "blk.{bid}.time_mix_key", + MODEL_TENSOR.TIME_MIX_VALUE: "blk.{bid}.time_mix_value", + MODEL_TENSOR.TIME_MIX_RECEPTANCE: "blk.{bid}.time_mix_receptance", + MODEL_TENSOR.TIME_MIX_GATE: "blk.{bid}.time_mix_gate", + MODEL_TENSOR.TIME_MIX_LN: "blk.{bid}.time_mix_ln", + MODEL_TENSOR.TIME_MIX_OUTPUT: "blk.{bid}.time_mix_output", + MODEL_TENSOR.CHANNEL_MIX_LERP_K: "blk.{bid}.channel_mix_lerp_k", + MODEL_TENSOR.CHANNEL_MIX_LERP_R: "blk.{bid}.channel_mix_lerp_r", + MODEL_TENSOR.CHANNEL_MIX_KEY: "blk.{bid}.channel_mix_key", + MODEL_TENSOR.CHANNEL_MIX_RECEPTANCE: "blk.{bid}.channel_mix_receptance", + MODEL_TENSOR.CHANNEL_MIX_VALUE: "blk.{bid}.channel_mix_value", + MODEL_TENSOR.ATTN_Q_A: "blk.{bid}.attn_q_a", + MODEL_TENSOR.ATTN_Q_B: "blk.{bid}.attn_q_b", + MODEL_TENSOR.ATTN_KV_A_MQA: "blk.{bid}.attn_kv_a_mqa", + MODEL_TENSOR.ATTN_KV_B: "blk.{bid}.attn_kv_b", + MODEL_TENSOR.ATTN_K_B: "blk.{bid}.attn_k_b", + MODEL_TENSOR.ATTN_V_B: "blk.{bid}.attn_v_b", + MODEL_TENSOR.ATTN_Q_A_NORM: "blk.{bid}.attn_q_a_norm", + MODEL_TENSOR.ATTN_KV_A_NORM: "blk.{bid}.attn_kv_a_norm", + MODEL_TENSOR.ATTN_SUB_NORM: "blk.{bid}.attn_sub_norm", + MODEL_TENSOR.FFN_SUB_NORM: "blk.{bid}.ffn_sub_norm", + MODEL_TENSOR.DEC_ATTN_NORM: "dec.blk.{bid}.attn_norm", + MODEL_TENSOR.DEC_ATTN_Q: "dec.blk.{bid}.attn_q", + MODEL_TENSOR.DEC_ATTN_K: "dec.blk.{bid}.attn_k", + MODEL_TENSOR.DEC_ATTN_V: "dec.blk.{bid}.attn_v", + MODEL_TENSOR.DEC_ATTN_OUT: "dec.blk.{bid}.attn_o", + MODEL_TENSOR.DEC_ATTN_REL_B: "dec.blk.{bid}.attn_rel_b", + MODEL_TENSOR.DEC_CROSS_ATTN_NORM: "dec.blk.{bid}.cross_attn_norm", + MODEL_TENSOR.DEC_CROSS_ATTN_Q: "dec.blk.{bid}.cross_attn_q", + MODEL_TENSOR.DEC_CROSS_ATTN_K: "dec.blk.{bid}.cross_attn_k", + MODEL_TENSOR.DEC_CROSS_ATTN_V: "dec.blk.{bid}.cross_attn_v", + MODEL_TENSOR.DEC_CROSS_ATTN_OUT: "dec.blk.{bid}.cross_attn_o", + MODEL_TENSOR.DEC_CROSS_ATTN_REL_B: "dec.blk.{bid}.cross_attn_rel_b", + MODEL_TENSOR.DEC_FFN_NORM: "dec.blk.{bid}.ffn_norm", + MODEL_TENSOR.DEC_FFN_GATE: "dec.blk.{bid}.ffn_gate", + MODEL_TENSOR.DEC_FFN_DOWN: "dec.blk.{bid}.ffn_down", + MODEL_TENSOR.DEC_FFN_UP: "dec.blk.{bid}.ffn_up", + MODEL_TENSOR.DEC_OUTPUT_NORM: "dec.output_norm", + MODEL_TENSOR.ENC_ATTN_NORM: "enc.blk.{bid}.attn_norm", + MODEL_TENSOR.ENC_ATTN_Q: "enc.blk.{bid}.attn_q", + MODEL_TENSOR.ENC_ATTN_K: "enc.blk.{bid}.attn_k", + MODEL_TENSOR.ENC_ATTN_V: "enc.blk.{bid}.attn_v", + MODEL_TENSOR.ENC_ATTN_OUT: "enc.blk.{bid}.attn_o", + MODEL_TENSOR.ENC_ATTN_REL_B: "enc.blk.{bid}.attn_rel_b", + MODEL_TENSOR.ENC_FFN_NORM: "enc.blk.{bid}.ffn_norm", + MODEL_TENSOR.ENC_FFN_GATE: "enc.blk.{bid}.ffn_gate", + MODEL_TENSOR.ENC_FFN_DOWN: "enc.blk.{bid}.ffn_down", + MODEL_TENSOR.ENC_FFN_UP: "enc.blk.{bid}.ffn_up", + MODEL_TENSOR.ENC_OUTPUT_NORM: "enc.output_norm", + MODEL_TENSOR.CLS: "cls", + MODEL_TENSOR.CLS_OUT: "cls.output", + MODEL_TENSOR.CONV1D: "conv1d", + MODEL_TENSOR.CONVNEXT_DW: "convnext.{bid}.dw", + MODEL_TENSOR.CONVNEXT_NORM: "convnext.{bid}.norm", + MODEL_TENSOR.CONVNEXT_PW1: "convnext.{bid}.pw1", + MODEL_TENSOR.CONVNEXT_PW2: "convnext.{bid}.pw2", + MODEL_TENSOR.CONVNEXT_GAMMA: "convnext.{bid}.gamma", + MODEL_TENSOR.POSNET_CONV1: "posnet.{bid}.conv1", + MODEL_TENSOR.POSNET_CONV2: "posnet.{bid}.conv2", + MODEL_TENSOR.POSNET_NORM: "posnet.{bid}.norm", + MODEL_TENSOR.POSNET_NORM1: "posnet.{bid}.norm1", + MODEL_TENSOR.POSNET_NORM2: "posnet.{bid}.norm2", + MODEL_TENSOR.POSNET_ATTN_NORM: "posnet.{bid}.attn_norm", + MODEL_TENSOR.POSNET_ATTN_Q: "posnet.{bid}.attn_q", + MODEL_TENSOR.POSNET_ATTN_K: "posnet.{bid}.attn_k", + MODEL_TENSOR.POSNET_ATTN_V: "posnet.{bid}.attn_v", + MODEL_TENSOR.POSNET_ATTN_OUT: "posnet.{bid}.attn_output", + # vision + MODEL_TENSOR.V_MMPROJ: "mm.{bid}", + MODEL_TENSOR.V_MMPROJ_FC: "mm.model.fc", + MODEL_TENSOR.V_MMPROJ_MLP: "mm.model.mlp.{bid}", + MODEL_TENSOR.V_MMPROJ_PEG: "mm.model.peg.{bid}", + MODEL_TENSOR.V_ENC_EMBD_CLS: "v.class_embd", + MODEL_TENSOR.V_ENC_EMBD_PATCH: "v.patch_embd", + MODEL_TENSOR.V_ENC_EMBD_POS: "v.position_embd", + MODEL_TENSOR.V_ENC_ATTN_Q: "v.blk.{bid}.attn_q", + MODEL_TENSOR.V_ENC_ATTN_Q_NORM: "v.blk.{bid}.attn_q_norm", + MODEL_TENSOR.V_ENC_ATTN_K: "v.blk.{bid}.attn_k", + MODEL_TENSOR.V_ENC_ATTN_K_NORM: "v.blk.{bid}.attn_k_norm", + MODEL_TENSOR.V_ENC_ATTN_V: "v.blk.{bid}.attn_v", + MODEL_TENSOR.V_ENC_INPUT_NORM: "v.blk.{bid}.ln1", + MODEL_TENSOR.V_ENC_ATTN_O: "v.blk.{bid}.attn_out", + MODEL_TENSOR.V_ENC_ATTN_O_NORM: "v.blk.{bid}.attn_out_norm", + MODEL_TENSOR.V_ENC_POST_ATTN_NORM: "v.blk.{bid}.ln2", + MODEL_TENSOR.V_ENC_FFN_UP: "v.blk.{bid}.ffn_up", + MODEL_TENSOR.V_ENC_FFN_GATE: "v.blk.{bid}.ffn_gate", + MODEL_TENSOR.V_ENC_FFN_DOWN: "v.blk.{bid}.ffn_down", + MODEL_TENSOR.V_LAYER_SCALE_1: "v.blk.{bid}.ls1", + MODEL_TENSOR.V_LAYER_SCALE_2: "v.blk.{bid}.ls2", + MODEL_TENSOR.V_PRE_NORM: "v.pre_ln", + MODEL_TENSOR.V_POST_NORM: "v.post_ln", + MODEL_TENSOR.V_MM_INP_PROJ: "mm.input_projection", + MODEL_TENSOR.V_MM_INP_NORM: "mm.input_norm", + MODEL_TENSOR.V_MM_SOFT_EMB_NORM: "mm.soft_emb_norm", + MODEL_TENSOR.V_RESMPL_POS_EMBD_K: "resampler.pos_embd_k", + MODEL_TENSOR.V_RESMPL_ATTN_Q: "resampler.attn.q", + MODEL_TENSOR.V_RESMPL_ATTN_K: "resampler.attn.k", + MODEL_TENSOR.V_RESMPL_ATTN_V: "resampler.attn.v", + MODEL_TENSOR.V_RESMPL_ATTN_OUT: "resampler.attn.out", + MODEL_TENSOR.V_RESMPL_KV: "resampler.kv", + MODEL_TENSOR.V_RESMPL_KV_NORM: "resampler.ln_kv", + MODEL_TENSOR.V_RESMPL_POST_NORM: "resampler.ln_post", + MODEL_TENSOR.V_RESMPL_Q_NORM: "resampler.ln_q", + MODEL_TENSOR.V_RESMPL_PROJ: "resampler.proj", + MODEL_TENSOR.V_RESMPL_QUERY: "resampler.query", + MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK: "v.token_embd.img_break", # pixtral + MODEL_TENSOR.V_MM_PATCH_MERGER: "mm.patch_merger", # mistral small 3.1 + # audio (mtmd) + MODEL_TENSOR.A_ENC_EMBD_POS: "a.position_embd", + MODEL_TENSOR.A_ENC_CONV1D: "a.conv1d.{bid}", + MODEL_TENSOR.A_PRE_NORM: "a.pre_ln", + MODEL_TENSOR.A_POST_NORM: "a.post_ln", + MODEL_TENSOR.A_ENC_ATTN_Q: "a.blk.{bid}.attn_q", + MODEL_TENSOR.A_ENC_ATTN_K: "a.blk.{bid}.attn_k", + MODEL_TENSOR.A_ENC_ATTN_V: "a.blk.{bid}.attn_v", + MODEL_TENSOR.A_ENC_INPUT_NORM: "a.blk.{bid}.ln1", + MODEL_TENSOR.A_ENC_OUTPUT: "a.blk.{bid}.attn_out", + MODEL_TENSOR.A_ENC_OUTPUT_NORM: "a.blk.{bid}.ln2", + MODEL_TENSOR.A_ENC_FFN_UP: "a.blk.{bid}.ffn_up", + MODEL_TENSOR.A_ENC_FFN_GATE: "a.blk.{bid}.ffn_gate", + MODEL_TENSOR.A_ENC_FFN_DOWN: "a.blk.{bid}.ffn_down", + MODEL_TENSOR.A_MMPROJ: "mm.a.mlp.{bid}", + MODEL_TENSOR.A_MMPROJ_FC: "mm.a.fc", + MODEL_TENSOR.A_MM_NORM_PRE: "mm.a.norm_pre", + MODEL_TENSOR.A_MM_NORM_MID: "mm.a.norm_mid", +} + +MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { + MODEL_ARCH.MMPROJ: [ + MODEL_TENSOR.V_MMPROJ, + MODEL_TENSOR.V_MMPROJ_FC, + MODEL_TENSOR.V_MMPROJ_MLP, + MODEL_TENSOR.V_MMPROJ_PEG, + MODEL_TENSOR.V_ENC_EMBD_CLS, + MODEL_TENSOR.V_ENC_EMBD_PATCH, + MODEL_TENSOR.V_ENC_EMBD_POS, + MODEL_TENSOR.V_ENC_INPUT_NORM, + MODEL_TENSOR.V_ENC_ATTN_Q, + MODEL_TENSOR.V_ENC_ATTN_Q_NORM, + MODEL_TENSOR.V_ENC_ATTN_K, + MODEL_TENSOR.V_ENC_ATTN_K_NORM, + MODEL_TENSOR.V_ENC_ATTN_V, + MODEL_TENSOR.V_ENC_ATTN_O, + MODEL_TENSOR.V_ENC_ATTN_O_NORM, + MODEL_TENSOR.V_ENC_POST_ATTN_NORM, + MODEL_TENSOR.V_ENC_FFN_UP, + MODEL_TENSOR.V_ENC_FFN_GATE, + MODEL_TENSOR.V_ENC_FFN_DOWN, + MODEL_TENSOR.V_LAYER_SCALE_1, + MODEL_TENSOR.V_LAYER_SCALE_2, + MODEL_TENSOR.V_PRE_NORM, + MODEL_TENSOR.V_POST_NORM, + MODEL_TENSOR.V_MM_INP_PROJ, + MODEL_TENSOR.V_MM_INP_NORM, + MODEL_TENSOR.V_MM_SOFT_EMB_NORM, + MODEL_TENSOR.V_RESMPL_POS_EMBD_K, + MODEL_TENSOR.V_RESMPL_ATTN_Q, + MODEL_TENSOR.V_RESMPL_ATTN_K, + MODEL_TENSOR.V_RESMPL_ATTN_V, + MODEL_TENSOR.V_RESMPL_ATTN_OUT, + MODEL_TENSOR.V_RESMPL_KV, + MODEL_TENSOR.V_RESMPL_KV_NORM, + MODEL_TENSOR.V_RESMPL_POST_NORM, + MODEL_TENSOR.V_RESMPL_Q_NORM, + MODEL_TENSOR.V_RESMPL_PROJ, + MODEL_TENSOR.V_RESMPL_QUERY, + MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK, + MODEL_TENSOR.V_MM_PATCH_MERGER, + # audio + MODEL_TENSOR.A_ENC_EMBD_POS, + MODEL_TENSOR.A_ENC_CONV1D, + MODEL_TENSOR.A_PRE_NORM, + MODEL_TENSOR.A_POST_NORM, + MODEL_TENSOR.A_ENC_ATTN_Q, + MODEL_TENSOR.A_ENC_ATTN_K, + MODEL_TENSOR.A_ENC_ATTN_V, + MODEL_TENSOR.A_ENC_INPUT_NORM, + MODEL_TENSOR.A_ENC_OUTPUT, + MODEL_TENSOR.A_ENC_OUTPUT_NORM, + MODEL_TENSOR.A_ENC_FFN_UP, + MODEL_TENSOR.A_ENC_FFN_GATE, + MODEL_TENSOR.A_ENC_FFN_DOWN, + MODEL_TENSOR.A_MMPROJ, + MODEL_TENSOR.A_MMPROJ_FC, + MODEL_TENSOR.A_MM_NORM_PRE, + MODEL_TENSOR.A_MM_NORM_MID, + ], + MODEL_ARCH.LLAMA: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + ], + MODEL_ARCH.LLAMA4: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + ], + MODEL_ARCH.DECI: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + ], + MODEL_ARCH.GROK: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.ATTN_OUT_NORM, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.LAYER_OUT_NORM, + ], + MODEL_ARCH.GPTNEOX: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.FALCON: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_NORM_2, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.BAICHUAN: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.STARCODER: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.POS_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.BERT: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.TOKEN_TYPES, + MODEL_TENSOR.POS_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_OUT_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.LAYER_OUT_NORM, + MODEL_TENSOR.CLS, + MODEL_TENSOR.CLS_OUT, + ], + MODEL_ARCH.NOMIC_BERT: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.TOKEN_TYPES, + MODEL_TENSOR.POS_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_OUT_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.LAYER_OUT_NORM, + ], + MODEL_ARCH.NOMIC_BERT_MOE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.TOKEN_TYPES, + MODEL_TENSOR.POS_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_OUT_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.LAYER_OUT_NORM, + ], + MODEL_ARCH.NEO_BERT: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.ENC_OUTPUT_NORM, + MODEL_TENSOR.CLS, + MODEL_TENSOR.CLS_OUT, + ], + MODEL_ARCH.JINA_BERT_V2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.TOKEN_TYPES, + MODEL_TENSOR.ATTN_NORM_2, + MODEL_TENSOR.ATTN_OUT_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.LAYER_OUT_NORM, + MODEL_TENSOR.CLS, + ], + MODEL_ARCH.MPT: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_ACT, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.POS_EMBD, + ], + MODEL_ARCH.GPTJ: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.REFACT: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.BLOOM: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.STABLELM: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K_NORM, + ], + MODEL_ARCH.QWEN: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.QWEN2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.QWEN2VL: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.QWEN2MOE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_INP_SHEXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + ], + MODEL_ARCH.QWEN3: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.QWEN3MOE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + ], + MODEL_ARCH.PLAMO: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.GPT2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.POS_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.PHI2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.PHI3: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FACTORS_LONG, + MODEL_TENSOR.ROPE_FACTORS_SHORT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.PHIMOE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FACTORS_LONG, + MODEL_TENSOR.ROPE_FACTORS_SHORT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + ], + MODEL_ARCH.CODESHELL: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.POS_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.ORION: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.INTERNLM2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.MINICPM: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ROPE_FACTORS_LONG, + MODEL_TENSOR.ROPE_FACTORS_SHORT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + ], + MODEL_ARCH.MINICPM3: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FACTORS_LONG, + MODEL_TENSOR.ROPE_FACTORS_SHORT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q_A, + MODEL_TENSOR.ATTN_Q_B, + MODEL_TENSOR.ATTN_KV_A_MQA, + MODEL_TENSOR.ATTN_KV_B, + MODEL_TENSOR.ATTN_Q_A_NORM, + MODEL_TENSOR.ATTN_KV_A_NORM, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.GEMMA: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_NORM, + ], + MODEL_ARCH.GEMMA2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_POST_NORM, + MODEL_TENSOR.FFN_PRE_NORM, + MODEL_TENSOR.FFN_POST_NORM, + ], + MODEL_ARCH.GEMMA3: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_POST_NORM, + MODEL_TENSOR.FFN_PRE_NORM, + MODEL_TENSOR.FFN_POST_NORM, + ], + MODEL_ARCH.STARCODER2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.RWKV6: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_NORM_2, + MODEL_TENSOR.TIME_MIX_W1, + MODEL_TENSOR.TIME_MIX_W2, + MODEL_TENSOR.TIME_MIX_LERP_X, + MODEL_TENSOR.TIME_MIX_LERP_K, + MODEL_TENSOR.TIME_MIX_LERP_V, + MODEL_TENSOR.TIME_MIX_LERP_R, + MODEL_TENSOR.TIME_MIX_LERP_G, + MODEL_TENSOR.TIME_MIX_LERP_W, + MODEL_TENSOR.TIME_MIX_LERP_FUSED, + MODEL_TENSOR.TIME_MIX_FIRST, + MODEL_TENSOR.TIME_MIX_DECAY, + MODEL_TENSOR.TIME_MIX_DECAY_W1, + MODEL_TENSOR.TIME_MIX_DECAY_W2, + MODEL_TENSOR.TIME_MIX_KEY, + MODEL_TENSOR.TIME_MIX_VALUE, + MODEL_TENSOR.TIME_MIX_RECEPTANCE, + MODEL_TENSOR.TIME_MIX_GATE, + MODEL_TENSOR.TIME_MIX_LN, + MODEL_TENSOR.TIME_MIX_OUTPUT, + MODEL_TENSOR.CHANNEL_MIX_LERP_K, + MODEL_TENSOR.CHANNEL_MIX_LERP_R, + MODEL_TENSOR.CHANNEL_MIX_KEY, + MODEL_TENSOR.CHANNEL_MIX_RECEPTANCE, + MODEL_TENSOR.CHANNEL_MIX_VALUE, + ], + MODEL_ARCH.RWKV6QWEN2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.TIME_MIX_W1, + MODEL_TENSOR.TIME_MIX_W2, + MODEL_TENSOR.TIME_MIX_LERP_X, + MODEL_TENSOR.TIME_MIX_LERP_K, + MODEL_TENSOR.TIME_MIX_LERP_V, + MODEL_TENSOR.TIME_MIX_LERP_R, + MODEL_TENSOR.TIME_MIX_LERP_G, + MODEL_TENSOR.TIME_MIX_LERP_W, + MODEL_TENSOR.TIME_MIX_LERP_FUSED, + MODEL_TENSOR.TIME_MIX_FIRST, + MODEL_TENSOR.TIME_MIX_DECAY, + MODEL_TENSOR.TIME_MIX_DECAY_W1, + MODEL_TENSOR.TIME_MIX_DECAY_W2, + MODEL_TENSOR.TIME_MIX_KEY, + MODEL_TENSOR.TIME_MIX_VALUE, + MODEL_TENSOR.TIME_MIX_RECEPTANCE, + MODEL_TENSOR.TIME_MIX_GATE, + MODEL_TENSOR.TIME_MIX_LN, + MODEL_TENSOR.TIME_MIX_OUTPUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.RWKV7: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_NORM_2, + MODEL_TENSOR.TIME_MIX_LERP_FUSED, + MODEL_TENSOR.TIME_MIX_W0, + MODEL_TENSOR.TIME_MIX_W1, + MODEL_TENSOR.TIME_MIX_W2, + MODEL_TENSOR.TIME_MIX_A0, + MODEL_TENSOR.TIME_MIX_A1, + MODEL_TENSOR.TIME_MIX_A2, + MODEL_TENSOR.TIME_MIX_V0, + MODEL_TENSOR.TIME_MIX_V1, + MODEL_TENSOR.TIME_MIX_V2, + MODEL_TENSOR.TIME_MIX_G1, + MODEL_TENSOR.TIME_MIX_G2, + MODEL_TENSOR.TIME_MIX_K_K, + MODEL_TENSOR.TIME_MIX_K_A, + MODEL_TENSOR.TIME_MIX_R_K, + MODEL_TENSOR.TIME_MIX_KEY, + MODEL_TENSOR.TIME_MIX_VALUE, + MODEL_TENSOR.TIME_MIX_RECEPTANCE, + MODEL_TENSOR.TIME_MIX_LN, + MODEL_TENSOR.TIME_MIX_OUTPUT, + MODEL_TENSOR.CHANNEL_MIX_LERP_K, + MODEL_TENSOR.CHANNEL_MIX_KEY, + MODEL_TENSOR.CHANNEL_MIX_VALUE, + ], + MODEL_ARCH.ARWKV7: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.TIME_MIX_LERP_FUSED, + MODEL_TENSOR.TIME_MIX_W0, + MODEL_TENSOR.TIME_MIX_W1, + MODEL_TENSOR.TIME_MIX_W2, + MODEL_TENSOR.TIME_MIX_A0, + MODEL_TENSOR.TIME_MIX_A1, + MODEL_TENSOR.TIME_MIX_A2, + MODEL_TENSOR.TIME_MIX_V0, + MODEL_TENSOR.TIME_MIX_V1, + MODEL_TENSOR.TIME_MIX_V2, + MODEL_TENSOR.TIME_MIX_G1, + MODEL_TENSOR.TIME_MIX_G2, + MODEL_TENSOR.TIME_MIX_K_K, + MODEL_TENSOR.TIME_MIX_K_A, + MODEL_TENSOR.TIME_MIX_R_K, + MODEL_TENSOR.TIME_MIX_KEY, + MODEL_TENSOR.TIME_MIX_VALUE, + MODEL_TENSOR.TIME_MIX_RECEPTANCE, + MODEL_TENSOR.TIME_MIX_LN, + MODEL_TENSOR.TIME_MIX_OUTPUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.MAMBA: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.SSM_IN, + MODEL_TENSOR.SSM_CONV1D, + MODEL_TENSOR.SSM_X, + MODEL_TENSOR.SSM_DT, + MODEL_TENSOR.SSM_A, + MODEL_TENSOR.SSM_D, + MODEL_TENSOR.SSM_OUT, + ], + MODEL_ARCH.XVERSE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.COMMAND_R: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_Q_NORM, + ], + MODEL_ARCH.COHERE2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.DBRX: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_OUT_NORM, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + ], + MODEL_ARCH.OLMO: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.OLMO2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_POST_NORM, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.FFN_POST_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.OLMOE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + ], + MODEL_ARCH.OPENELM: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.ARCTIC: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_NORM_EXP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + ], + MODEL_ARCH.DEEPSEEK: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + ], + MODEL_ARCH.DEEPSEEK2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_A, + MODEL_TENSOR.ATTN_Q_B, + MODEL_TENSOR.ATTN_KV_A_MQA, + MODEL_TENSOR.ATTN_KV_B, + MODEL_TENSOR.ATTN_K_B, + MODEL_TENSOR.ATTN_V_B, + MODEL_TENSOR.ATTN_Q_A_NORM, + MODEL_TENSOR.ATTN_KV_A_NORM, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + MODEL_TENSOR.FFN_EXP_PROBS_B, + ], + MODEL_ARCH.PLM: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_KV_A_MQA, + MODEL_TENSOR.ATTN_KV_A_NORM, + MODEL_TENSOR.ATTN_KV_B, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_DOWN, + ], + MODEL_ARCH.CHATGLM : [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.GLM4 : [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.ATTN_POST_NORM, + MODEL_TENSOR.FFN_POST_NORM, + ], + MODEL_ARCH.BITNET: [ + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.ATTN_SUB_NORM, + MODEL_TENSOR.FFN_SUB_NORM, + ], + MODEL_ARCH.T5: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.DEC_ATTN_NORM, + MODEL_TENSOR.DEC_ATTN_Q, + MODEL_TENSOR.DEC_ATTN_K, + MODEL_TENSOR.DEC_ATTN_V, + MODEL_TENSOR.DEC_ATTN_OUT, + MODEL_TENSOR.DEC_ATTN_REL_B, + MODEL_TENSOR.DEC_CROSS_ATTN_NORM, + MODEL_TENSOR.DEC_CROSS_ATTN_Q, + MODEL_TENSOR.DEC_CROSS_ATTN_K, + MODEL_TENSOR.DEC_CROSS_ATTN_V, + MODEL_TENSOR.DEC_CROSS_ATTN_OUT, + MODEL_TENSOR.DEC_CROSS_ATTN_REL_B, + MODEL_TENSOR.DEC_FFN_NORM, + MODEL_TENSOR.DEC_FFN_GATE, + MODEL_TENSOR.DEC_FFN_DOWN, + MODEL_TENSOR.DEC_FFN_UP, + MODEL_TENSOR.DEC_OUTPUT_NORM, + MODEL_TENSOR.ENC_ATTN_NORM, + MODEL_TENSOR.ENC_ATTN_Q, + MODEL_TENSOR.ENC_ATTN_K, + MODEL_TENSOR.ENC_ATTN_V, + MODEL_TENSOR.ENC_ATTN_OUT, + MODEL_TENSOR.ENC_ATTN_REL_B, + MODEL_TENSOR.ENC_FFN_NORM, + MODEL_TENSOR.ENC_FFN_GATE, + MODEL_TENSOR.ENC_FFN_DOWN, + MODEL_TENSOR.ENC_FFN_UP, + MODEL_TENSOR.ENC_OUTPUT_NORM, + ], + MODEL_ARCH.T5ENCODER: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ENC_ATTN_NORM, + MODEL_TENSOR.ENC_ATTN_Q, + MODEL_TENSOR.ENC_ATTN_K, + MODEL_TENSOR.ENC_ATTN_V, + MODEL_TENSOR.ENC_ATTN_OUT, + MODEL_TENSOR.ENC_ATTN_REL_B, + MODEL_TENSOR.ENC_FFN_NORM, + MODEL_TENSOR.ENC_FFN_GATE, + MODEL_TENSOR.ENC_FFN_DOWN, + MODEL_TENSOR.ENC_FFN_UP, + MODEL_TENSOR.ENC_OUTPUT_NORM, + ], + MODEL_ARCH.JAIS: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.NEMOTRON: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.EXAONE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.GRANITE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.GRANITE_MOE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + ], + MODEL_ARCH.CHAMELEON: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.WAVTOKENIZER_DEC: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.CONV1D, + MODEL_TENSOR.CONVNEXT_DW, + MODEL_TENSOR.CONVNEXT_NORM, + MODEL_TENSOR.CONVNEXT_PW1, + MODEL_TENSOR.CONVNEXT_PW2, + MODEL_TENSOR.CONVNEXT_GAMMA, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.POSNET_CONV1, + MODEL_TENSOR.POSNET_CONV2, + MODEL_TENSOR.POSNET_NORM, + MODEL_TENSOR.POSNET_NORM1, + MODEL_TENSOR.POSNET_NORM2, + MODEL_TENSOR.POSNET_ATTN_NORM, + MODEL_TENSOR.POSNET_ATTN_Q, + MODEL_TENSOR.POSNET_ATTN_K, + MODEL_TENSOR.POSNET_ATTN_V, + MODEL_TENSOR.POSNET_ATTN_OUT, + ], + MODEL_ARCH.BAILINGMOE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + ], + MODEL_ARCH.DOTS1: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_EXP_PROBS_B, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_UP_SHEXP, + ], + MODEL_ARCH.ARCEE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + # TODO +} + +# tensors that will not be serialized +MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { + MODEL_ARCH.LLAMA: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], + MODEL_ARCH.DECI: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], + MODEL_ARCH.BAICHUAN: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], + MODEL_ARCH.QWEN: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], + MODEL_ARCH.CODESHELL: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], + MODEL_ARCH.ORION: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], + MODEL_ARCH.STARCODER2: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], + MODEL_ARCH.XVERSE: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], + MODEL_ARCH.DEEPSEEK: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], + MODEL_ARCH.DEEPSEEK2: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], + MODEL_ARCH.CHATGLM: [ + MODEL_TENSOR.ROPE_FREQS, + ], + MODEL_ARCH.NEMOTRON: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], + MODEL_ARCH.BAILINGMOE: [ + MODEL_TENSOR.ROPE_FREQS, + ], +} + +# +# types +# + + +class TokenType(IntEnum): + NORMAL = 1 + UNKNOWN = 2 + CONTROL = 3 + USER_DEFINED = 4 + UNUSED = 5 + BYTE = 6 + + +class RopeScalingType(Enum): + NONE = 'none' + LINEAR = 'linear' + YARN = 'yarn' + LONGROPE = 'longrope' + + +class PoolingType(IntEnum): + NONE = 0 + MEAN = 1 + CLS = 2 + LAST = 3 + RANK = 4 + + +class GGMLQuantizationType(IntEnum): + F32 = 0 + F16 = 1 + Q4_0 = 2 + Q4_1 = 3 + Q5_0 = 6 + Q5_1 = 7 + Q8_0 = 8 + Q8_1 = 9 + Q2_K = 10 + Q3_K = 11 + Q4_K = 12 + Q5_K = 13 + Q6_K = 14 + Q8_K = 15 + IQ2_XXS = 16 + IQ2_XS = 17 + IQ3_XXS = 18 + IQ1_S = 19 + IQ4_NL = 20 + IQ3_S = 21 + IQ2_S = 22 + IQ4_XS = 23 + I8 = 24 + I16 = 25 + I32 = 26 + I64 = 27 + F64 = 28 + IQ1_M = 29 + BF16 = 30 + TQ1_0 = 34 + TQ2_0 = 35 + + +class ExpertGatingFuncType(IntEnum): + SOFTMAX = 1 + SIGMOID = 2 + + +# TODO: add GGMLFileType from ggml_ftype in ggml.h + + +# from llama_ftype in llama.h +# ALL VALUES SHOULD BE THE SAME HERE AS THEY ARE OVER THERE. +class LlamaFileType(IntEnum): + ALL_F32 = 0 + MOSTLY_F16 = 1 # except 1d tensors + MOSTLY_Q4_0 = 2 # except 1d tensors + MOSTLY_Q4_1 = 3 # except 1d tensors + # MOSTLY_Q4_1_SOME_F16 = 4 # tok_embeddings.weight and output.weight are F16 + # MOSTLY_Q4_2 = 5 # support has been removed + # MOSTLY_Q4_3 = 6 # support has been removed + MOSTLY_Q8_0 = 7 # except 1d tensors + MOSTLY_Q5_0 = 8 # except 1d tensors + MOSTLY_Q5_1 = 9 # except 1d tensors + MOSTLY_Q2_K = 10 # except 1d tensors + MOSTLY_Q3_K_S = 11 # except 1d tensors + MOSTLY_Q3_K_M = 12 # except 1d tensors + MOSTLY_Q3_K_L = 13 # except 1d tensors + MOSTLY_Q4_K_S = 14 # except 1d tensors + MOSTLY_Q4_K_M = 15 # except 1d tensors + MOSTLY_Q5_K_S = 16 # except 1d tensors + MOSTLY_Q5_K_M = 17 # except 1d tensors + MOSTLY_Q6_K = 18 # except 1d tensors + MOSTLY_IQ2_XXS = 19 # except 1d tensors + MOSTLY_IQ2_XS = 20 # except 1d tensors + MOSTLY_Q2_K_S = 21 # except 1d tensors + MOSTLY_IQ3_XS = 22 # except 1d tensors + MOSTLY_IQ3_XXS = 23 # except 1d tensors + MOSTLY_IQ1_S = 24 # except 1d tensors + MOSTLY_IQ4_NL = 25 # except 1d tensors + MOSTLY_IQ3_S = 26 # except 1d tensors + MOSTLY_IQ3_M = 27 # except 1d tensors + MOSTLY_IQ2_S = 28 # except 1d tensors + MOSTLY_IQ2_M = 29 # except 1d tensors + MOSTLY_IQ4_XS = 30 # except 1d tensors + MOSTLY_IQ1_M = 31 # except 1d tensors + MOSTLY_BF16 = 32 # except 1d tensors + # MOSTLY_Q4_0_4_4 = 33 # removed from gguf files, use Q4_0 and runtime repack + # MOSTLY_Q4_0_4_8 = 34 # removed from gguf files, use Q4_0 and runtime repack + # MOSTLY_Q4_0_8_8 = 35 # removed from gguf files, use Q4_0 and runtime repack + MOSTLY_TQ1_0 = 36 # except 1d tensors + MOSTLY_TQ2_0 = 37 # except 1d tensors + + GUESSED = 1024 # not specified in the model file + + +class GGUFEndian(IntEnum): + LITTLE = 0 + BIG = 1 + + +class GGUFValueType(IntEnum): + UINT8 = 0 + INT8 = 1 + UINT16 = 2 + INT16 = 3 + UINT32 = 4 + INT32 = 5 + FLOAT32 = 6 + BOOL = 7 + STRING = 8 + ARRAY = 9 + UINT64 = 10 + INT64 = 11 + FLOAT64 = 12 + + @staticmethod + def get_type(val: Any) -> GGUFValueType: + if isinstance(val, (str, bytes, bytearray)): + return GGUFValueType.STRING + elif isinstance(val, list): + return GGUFValueType.ARRAY + elif isinstance(val, float): + return GGUFValueType.FLOAT32 + elif isinstance(val, bool): + return GGUFValueType.BOOL + elif isinstance(val, int): + return GGUFValueType.INT32 + # TODO: need help with 64-bit types in Python + else: + raise ValueError(f"Unknown type: {type(val)}") + + +class VisionProjectorType: + GEMMA3 = "gemma3" + IDEFICS3 = "idefics3" + PIXTRAL = "pixtral" + LLAMA4 = "llama4" + QWEN2VL = "qwen2vl_merger" + QWEN25VL = "qwen2.5vl_merger" + ULTRAVOX = "ultravox" + INTERNVL = "internvl" + QWEN2A = "qwen2a" # audio + QWEN25O = "qwen2.5o" # omni + + +# Items here are (block size, type size) +QK_K = 256 +GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = { + GGMLQuantizationType.F32: (1, 4), + GGMLQuantizationType.F16: (1, 2), + GGMLQuantizationType.Q4_0: (32, 2 + 16), + GGMLQuantizationType.Q4_1: (32, 2 + 2 + 16), + GGMLQuantizationType.Q5_0: (32, 2 + 4 + 16), + GGMLQuantizationType.Q5_1: (32, 2 + 2 + 4 + 16), + GGMLQuantizationType.Q8_0: (32, 2 + 32), + GGMLQuantizationType.Q8_1: (32, 4 + 4 + 32), + GGMLQuantizationType.Q2_K: (256, 2 + 2 + QK_K // 16 + QK_K // 4), + GGMLQuantizationType.Q3_K: (256, 2 + QK_K // 4 + QK_K // 8 + 12), + GGMLQuantizationType.Q4_K: (256, 2 + 2 + QK_K // 2 + 12), + GGMLQuantizationType.Q5_K: (256, 2 + 2 + QK_K // 2 + QK_K // 8 + 12), + GGMLQuantizationType.Q6_K: (256, 2 + QK_K // 2 + QK_K // 4 + QK_K // 16), + GGMLQuantizationType.Q8_K: (256, 4 + QK_K + QK_K // 8), + GGMLQuantizationType.IQ2_XXS: (256, 2 + QK_K // 4), + GGMLQuantizationType.IQ2_XS: (256, 2 + QK_K // 4 + QK_K // 32), + GGMLQuantizationType.IQ3_XXS: (256, 2 + QK_K // 4 + QK_K // 8), + GGMLQuantizationType.IQ1_S: (256, 2 + QK_K // 8 + QK_K // 16), + GGMLQuantizationType.IQ4_NL: (32, 2 + 16), + GGMLQuantizationType.IQ3_S: (256, 2 + QK_K // 4 + QK_K // 8 + QK_K // 32 + 4), + GGMLQuantizationType.IQ2_S: (256, 2 + QK_K // 4 + QK_K // 16), + GGMLQuantizationType.IQ4_XS: (256, 2 + 2 + QK_K // 2 + QK_K // 64), + GGMLQuantizationType.I8: (1, 1), + GGMLQuantizationType.I16: (1, 2), + GGMLQuantizationType.I32: (1, 4), + GGMLQuantizationType.I64: (1, 8), + GGMLQuantizationType.F64: (1, 8), + GGMLQuantizationType.IQ1_M: (256, QK_K // 8 + QK_K // 16 + QK_K // 32), + GGMLQuantizationType.BF16: (1, 2), + GGMLQuantizationType.TQ1_0: (256, 2 + 4 * 13), + GGMLQuantizationType.TQ2_0: (256, 2 + 64), +} + + +# Aliases for backward compatibility. + +# general +KEY_GENERAL_ARCHITECTURE = Keys.General.ARCHITECTURE +KEY_GENERAL_QUANTIZATION_VERSION = Keys.General.QUANTIZATION_VERSION +KEY_GENERAL_ALIGNMENT = Keys.General.ALIGNMENT +KEY_GENERAL_NAME = Keys.General.NAME +KEY_GENERAL_AUTHOR = Keys.General.AUTHOR +KEY_GENERAL_URL = Keys.General.URL +KEY_GENERAL_DESCRIPTION = Keys.General.DESCRIPTION +KEY_GENERAL_LICENSE = Keys.General.LICENSE +KEY_GENERAL_SOURCE_URL = Keys.General.SOURCE_URL +KEY_GENERAL_FILE_TYPE = Keys.General.FILE_TYPE + +# LLM +KEY_VOCAB_SIZE = Keys.LLM.VOCAB_SIZE +KEY_CONTEXT_LENGTH = Keys.LLM.CONTEXT_LENGTH +KEY_EMBEDDING_LENGTH = Keys.LLM.EMBEDDING_LENGTH +KEY_BLOCK_COUNT = Keys.LLM.BLOCK_COUNT +KEY_FEED_FORWARD_LENGTH = Keys.LLM.FEED_FORWARD_LENGTH +KEY_USE_PARALLEL_RESIDUAL = Keys.LLM.USE_PARALLEL_RESIDUAL +KEY_TENSOR_DATA_LAYOUT = Keys.LLM.TENSOR_DATA_LAYOUT + +# attention +KEY_ATTENTION_HEAD_COUNT = Keys.Attention.HEAD_COUNT +KEY_ATTENTION_HEAD_COUNT_KV = Keys.Attention.HEAD_COUNT_KV +KEY_ATTENTION_MAX_ALIBI_BIAS = Keys.Attention.MAX_ALIBI_BIAS +KEY_ATTENTION_CLAMP_KQV = Keys.Attention.CLAMP_KQV +KEY_ATTENTION_LAYERNORM_EPS = Keys.Attention.LAYERNORM_EPS +KEY_ATTENTION_LAYERNORM_RMS_EPS = Keys.Attention.LAYERNORM_RMS_EPS + +# RoPE +KEY_ROPE_DIMENSION_COUNT = Keys.Rope.DIMENSION_COUNT +KEY_ROPE_FREQ_BASE = Keys.Rope.FREQ_BASE +KEY_ROPE_SCALING_TYPE = Keys.Rope.SCALING_TYPE +KEY_ROPE_SCALING_FACTOR = Keys.Rope.SCALING_FACTOR +KEY_ROPE_SCALING_ORIG_CTX_LEN = Keys.Rope.SCALING_ORIG_CTX_LEN +KEY_ROPE_SCALING_FINETUNED = Keys.Rope.SCALING_FINETUNED + +# SSM +KEY_SSM_CONV_KERNEL = Keys.SSM.CONV_KERNEL +KEY_SSM_INNER_SIZE = Keys.SSM.INNER_SIZE +KEY_SSM_STATE_SIZE = Keys.SSM.STATE_SIZE +KEY_SSM_TIME_STEP_RANK = Keys.SSM.TIME_STEP_RANK +KEY_SSM_DT_B_C_RMS = Keys.SSM.DT_B_C_RMS + +# tokenization +KEY_TOKENIZER_MODEL = Keys.Tokenizer.MODEL +KEY_TOKENIZER_PRE = Keys.Tokenizer.PRE +KEY_TOKENIZER_LIST = Keys.Tokenizer.LIST +KEY_TOKENIZER_TOKEN_TYPE = Keys.Tokenizer.TOKEN_TYPE +KEY_TOKENIZER_SCORES = Keys.Tokenizer.SCORES +KEY_TOKENIZER_MERGES = Keys.Tokenizer.MERGES +KEY_TOKENIZER_BOS_ID = Keys.Tokenizer.BOS_ID +KEY_TOKENIZER_EOS_ID = Keys.Tokenizer.EOS_ID +KEY_TOKENIZER_EOT_ID = Keys.Tokenizer.EOT_ID +KEY_TOKENIZER_EOM_ID = Keys.Tokenizer.EOM_ID +KEY_TOKENIZER_UNK_ID = Keys.Tokenizer.UNK_ID +KEY_TOKENIZER_SEP_ID = Keys.Tokenizer.SEP_ID +KEY_TOKENIZER_PAD_ID = Keys.Tokenizer.PAD_ID +KEY_TOKENIZER_MASK_ID = Keys.Tokenizer.MASK_ID +KEY_TOKENIZER_HF_JSON = Keys.Tokenizer.HF_JSON +KEY_TOKENIZER_RWKV = Keys.Tokenizer.RWKV + +KEY_TOKENIZER_FIM_PRE_ID = Keys.Tokenizer.FIM_PRE_ID +KEY_TOKENIZER_FIM_SUF_ID = Keys.Tokenizer.FIM_SUF_ID +KEY_TOKENIZER_FIM_MID_ID = Keys.Tokenizer.FIM_MID_ID +KEY_TOKENIZER_FIM_PAD_ID = Keys.Tokenizer.FIM_PAD_ID +KEY_TOKENIZER_FIM_REP_ID = Keys.Tokenizer.FIM_REP_ID +KEY_TOKENIZER_FIM_SEP_ID = Keys.Tokenizer.FIM_SEP_ID + +# deprecated +KEY_TOKENIZER_PREFIX_ID = Keys.Tokenizer.PREFIX_ID +KEY_TOKENIZER_SUFFIX_ID = Keys.Tokenizer.SUFFIX_ID +KEY_TOKENIZER_MIDDLE_ID = Keys.Tokenizer.MIDDLE_ID diff --git a/venv/lib/python3.10/site-packages/gguf/gguf.py b/venv/lib/python3.10/site-packages/gguf/gguf.py new file mode 100644 index 0000000000000000000000000000000000000000..651a81eb828248728f854c85c1a437b52892f275 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/gguf.py @@ -0,0 +1,15 @@ +# This file left for compatibility. If you want to use the GGUF API from Python +# then don't import gguf/gguf.py directly. If you're looking for examples, see the +# examples/ directory for gguf-py + +import importlib +import sys +from pathlib import Path + +sys.path.insert(0, str(Path(__file__).parent.parent)) + +# Compatibility for people trying to import gguf/gguf.py directly instead of as a package. +importlib.invalidate_caches() +import gguf # noqa: E402 + +importlib.reload(gguf) diff --git a/venv/lib/python3.10/site-packages/gguf/gguf_reader.py b/venv/lib/python3.10/site-packages/gguf/gguf_reader.py new file mode 100644 index 0000000000000000000000000000000000000000..d87e8f72321b3f776cf214e117e21cdde9e1e18b --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/gguf_reader.py @@ -0,0 +1,367 @@ +# +# GGUF file reading/modification support. For API usage information, +# please see the files scripts/ for some fairly simple examples. +# +from __future__ import annotations + +import logging +import os +import sys +from collections import OrderedDict +from typing import Any, Literal, NamedTuple, TypeVar, Union + +import numpy as np +import numpy.typing as npt + +from .quants import quant_shape_to_byte_shape + +if __name__ == "__main__": + from pathlib import Path + + # Allow running file in package as a script. + sys.path.insert(0, str(Path(__file__).parent.parent)) + +from gguf.constants import ( + GGML_QUANT_SIZES, + GGUF_DEFAULT_ALIGNMENT, + GGUF_MAGIC, + GGUF_VERSION, + GGMLQuantizationType, + GGUFValueType, + GGUFEndian, +) + +logger = logging.getLogger(__name__) + +READER_SUPPORTED_VERSIONS = [2, GGUF_VERSION] + + +class ReaderField(NamedTuple): + # Offset to start of this field. + offset: int + + # Name of the field (not necessarily from file data). + name: str + + # Data parts. Some types have multiple components, such as strings + # that consist of a length followed by the string data. + parts: list[npt.NDArray[Any]] = [] + + # Indexes into parts that we can call the actual data. For example + # an array of strings will be populated with indexes to the actual + # string data. + data: list[int] = [-1] + + types: list[GGUFValueType] = [] + + def contents(self, index_or_slice: int | slice = slice(None)) -> Any: + if self.types: + to_string = lambda x: str(x.tobytes(), encoding='utf-8') # noqa: E731 + main_type = self.types[0] + + if main_type == GGUFValueType.ARRAY: + sub_type = self.types[-1] + + if sub_type == GGUFValueType.STRING: + indices = self.data[index_or_slice] + + if isinstance(index_or_slice, int): + return to_string(self.parts[indices]) # type: ignore + else: + return [to_string(self.parts[idx]) for idx in indices] # type: ignore + else: + # FIXME: When/if _get_field_parts() support multi-dimensional arrays, this must do so too + + # Check if it's unsafe to perform slice optimization on data + # if any(True for idx in self.data if len(self.parts[idx]) != 1): + # optim_slice = slice(None) + # else: + # optim_slice = index_or_slice + # index_or_slice = slice(None) + + # if isinstance(optim_slice, int): + # return self.parts[self.data[optim_slice]].tolist()[0] + # else: + # return [pv for idx in self.data[optim_slice] for pv in self.parts[idx].tolist()][index_or_slice] + + if isinstance(index_or_slice, int): + return self.parts[self.data[index_or_slice]].tolist()[0] + else: + return [pv for idx in self.data[index_or_slice] for pv in self.parts[idx].tolist()] + + if main_type == GGUFValueType.STRING: + return to_string(self.parts[-1]) + else: + return self.parts[-1].tolist()[0] + + return None + + +class ReaderTensor(NamedTuple): + name: str + tensor_type: GGMLQuantizationType + shape: npt.NDArray[np.uint32] + n_elements: int + n_bytes: int + data_offset: int + data: npt.NDArray[Any] + field: ReaderField + + +class GGUFReader: + # I - same as host, S - swapped + byte_order: Literal['I', 'S'] = 'I' + alignment: int = GGUF_DEFAULT_ALIGNMENT + data_offset: int + + # Note: Internal helper, API may change. + gguf_scalar_to_np: dict[GGUFValueType, type[np.generic]] = { + GGUFValueType.UINT8: np.uint8, + GGUFValueType.INT8: np.int8, + GGUFValueType.UINT16: np.uint16, + GGUFValueType.INT16: np.int16, + GGUFValueType.UINT32: np.uint32, + GGUFValueType.INT32: np.int32, + GGUFValueType.FLOAT32: np.float32, + GGUFValueType.UINT64: np.uint64, + GGUFValueType.INT64: np.int64, + GGUFValueType.FLOAT64: np.float64, + GGUFValueType.BOOL: np.bool_, + } + + def __init__(self, path: os.PathLike[str] | str, mode: Literal['r', 'r+', 'c'] = 'r'): + self.data = np.memmap(path, mode = mode) + offs = 0 + + # Check for GGUF magic + if self._get(offs, np.uint32, override_order = '<')[0] != GGUF_MAGIC: + raise ValueError('GGUF magic invalid') + offs += 4 + + # Check GGUF version + temp_version = self._get(offs, np.uint32) + if temp_version[0] & 65535 == 0: + # If we get 0 here that means it's (probably) a GGUF file created for + # the opposite byte order of the machine this script is running on. + self.byte_order = 'S' + temp_version = temp_version.view(temp_version.dtype.newbyteorder(self.byte_order)) + version = temp_version[0] + if version not in READER_SUPPORTED_VERSIONS: + raise ValueError(f'Sorry, file appears to be version {version} which we cannot handle') + if sys.byteorder == "little": + # Host is little endian + host_endian = GGUFEndian.LITTLE + swapped_endian = GGUFEndian.BIG + else: + # Sorry PDP or other weird systems that don't use BE or LE. + host_endian = GGUFEndian.BIG + swapped_endian = GGUFEndian.LITTLE + self.endianess = swapped_endian if self.byte_order == "S" else host_endian + self.fields: OrderedDict[str, ReaderField] = OrderedDict() + self.tensors: list[ReaderTensor] = [] + offs += self._push_field(ReaderField(offs, 'GGUF.version', [temp_version], [0], [GGUFValueType.UINT32])) + + # Check tensor count and kv count + temp_counts = self._get(offs, np.uint64, 2) + offs += self._push_field(ReaderField(offs, 'GGUF.tensor_count', [temp_counts[:1]], [0], [GGUFValueType.UINT64])) + offs += self._push_field(ReaderField(offs, 'GGUF.kv_count', [temp_counts[1:]], [0], [GGUFValueType.UINT64])) + tensor_count, kv_count = temp_counts + offs = self._build_fields(offs, kv_count) + + # Build Tensor Info Fields + offs, tensors_fields = self._build_tensor_info(offs, tensor_count) + new_align = self.fields.get('general.alignment') + if new_align is not None: + if new_align.types != [GGUFValueType.UINT32]: + raise ValueError('Bad type for general.alignment field') + self.alignment = new_align.parts[-1][0] + padding = offs % self.alignment + if padding != 0: + offs += self.alignment - padding + self.data_offset = offs + self._build_tensors(offs, tensors_fields) + + _DT = TypeVar('_DT', bound = npt.DTypeLike) + + # Fetch a key/value metadata field by key. + def get_field(self, key: str) -> Union[ReaderField, None]: + return self.fields.get(key, None) + + # Fetch a tensor from the list by index. + def get_tensor(self, idx: int) -> ReaderTensor: + return self.tensors[idx] + + def _get( + self, offset: int, dtype: npt.DTypeLike, count: int = 1, override_order: None | Literal['I', 'S', '<'] = None, + ) -> npt.NDArray[Any]: + count = int(count) + itemsize = int(np.empty([], dtype = dtype).itemsize) + end_offs = offset + itemsize * count + arr = self.data[offset:end_offs].view(dtype=dtype)[:count] + return arr.view(arr.dtype.newbyteorder(self.byte_order if override_order is None else override_order)) + + def _push_field(self, field: ReaderField, skip_sum: bool = False) -> int: + if field.name in self.fields: + # TODO: add option to generate error on duplicate keys + # raise KeyError(f'Duplicate {field.name} already in list at offset {field.offset}') + + logger.warning(f'Duplicate key {field.name} at offset {field.offset}') + self.fields[field.name + '_{}'.format(field.offset)] = field + else: + self.fields[field.name] = field + return 0 if skip_sum else sum(int(part.nbytes) for part in field.parts) + + def _get_str(self, offset: int) -> tuple[npt.NDArray[np.uint64], npt.NDArray[np.uint8]]: + slen = self._get(offset, np.uint64) + return slen, self._get(offset + 8, np.uint8, slen[0]) + + def _get_field_parts( + self, orig_offs: int, raw_type: int, + ) -> tuple[int, list[npt.NDArray[Any]], list[int], list[GGUFValueType]]: + offs = orig_offs + types: list[GGUFValueType] = [] + gtype = GGUFValueType(raw_type) + types.append(gtype) + # Handle strings. + if gtype == GGUFValueType.STRING: + sparts: list[npt.NDArray[Any]] = list(self._get_str(offs)) + size = sum(int(part.nbytes) for part in sparts) + return size, sparts, [1], types + # Check if it's a simple scalar type. + nptype = self.gguf_scalar_to_np.get(gtype) + if nptype is not None: + val = self._get(offs, nptype) + return int(val.nbytes), [val], [0], types + # Handle arrays. + if gtype == GGUFValueType.ARRAY: + raw_itype = self._get(offs, np.uint32) + offs += int(raw_itype.nbytes) + alen = self._get(offs, np.uint64) + offs += int(alen.nbytes) + aparts: list[npt.NDArray[Any]] = [raw_itype, alen] + data_idxs: list[int] = [] + # FIXME: Handle multi-dimensional arrays properly instead of flattening + for idx in range(alen[0]): + curr_size, curr_parts, curr_idxs, curr_types = self._get_field_parts(offs, raw_itype[0]) + if idx == 0: + types += curr_types + idxs_offs = len(aparts) + aparts += curr_parts + data_idxs += (idx + idxs_offs for idx in curr_idxs) + offs += curr_size + return offs - orig_offs, aparts, data_idxs, types + # We can't deal with this one. + raise ValueError(f'Unknown/unhandled field type {gtype}') + + def _get_tensor_info_field(self, orig_offs: int) -> ReaderField: + offs = orig_offs + + # Get Tensor Name + name_len, name_data = self._get_str(offs) + offs += int(name_len.nbytes + name_data.nbytes) + + # Get Tensor Dimensions Count + n_dims = self._get(offs, np.uint32) + offs += int(n_dims.nbytes) + + # Get Tensor Dimension Array + dims = self._get(offs, np.uint64, n_dims[0]) + offs += int(dims.nbytes) + + # Get Tensor Encoding Scheme Type + raw_dtype = self._get(offs, np.uint32) + offs += int(raw_dtype.nbytes) + + # Get Tensor Offset + offset_tensor = self._get(offs, np.uint64) + offs += int(offset_tensor.nbytes) + + return ReaderField( + orig_offs, + str(bytes(name_data), encoding = 'utf-8'), + [name_len, name_data, n_dims, dims, raw_dtype, offset_tensor], + [1, 3, 4, 5], + ) + + def _build_fields(self, offs: int, count: int) -> int: + for _ in range(count): + orig_offs = offs + kv_klen, kv_kdata = self._get_str(offs) + offs += int(kv_klen.nbytes + kv_kdata.nbytes) + raw_kv_type = self._get(offs, np.uint32) + offs += int(raw_kv_type.nbytes) + parts: list[npt.NDArray[Any]] = [kv_klen, kv_kdata, raw_kv_type] + idxs_offs = len(parts) + field_size, field_parts, field_idxs, field_types = self._get_field_parts(offs, raw_kv_type[0]) + parts += field_parts + self._push_field(ReaderField( + orig_offs, + str(bytes(kv_kdata), encoding = 'utf-8'), + parts, + [idx + idxs_offs for idx in field_idxs], + field_types, + ), skip_sum = True) + offs += field_size + return offs + + def _build_tensor_info(self, offs: int, count: int) -> tuple[int, list[ReaderField]]: + tensor_fields = [] + for _ in range(count): + field = self._get_tensor_info_field(offs) + offs += sum(int(part.nbytes) for part in field.parts) + tensor_fields.append(field) + return offs, tensor_fields + + def _build_tensors(self, start_offs: int, fields: list[ReaderField]) -> None: + tensors = [] + tensor_names = set() # keep track of name to prevent duplicated tensors + for field in fields: + _name_len, name_data, _n_dims, dims, raw_dtype, offset_tensor = field.parts + # check if there's any tensor having same name already in the list + tensor_name = str(bytes(name_data), encoding = 'utf-8') + if tensor_name in tensor_names: + raise ValueError(f'Found duplicated tensor with name {tensor_name}') + tensor_names.add(tensor_name) + ggml_type = GGMLQuantizationType(raw_dtype[0]) + n_elems = int(np.prod(dims)) + np_dims = tuple(reversed(dims.tolist())) + block_size, type_size = GGML_QUANT_SIZES[ggml_type] + n_bytes = n_elems * type_size // block_size + data_offs = int(start_offs + offset_tensor[0]) + item_type: npt.DTypeLike + if ggml_type == GGMLQuantizationType.F16: + item_count = n_elems + item_type = np.float16 + elif ggml_type == GGMLQuantizationType.F32: + item_count = n_elems + item_type = np.float32 + elif ggml_type == GGMLQuantizationType.F64: + item_count = n_elems + item_type = np.float64 + elif ggml_type == GGMLQuantizationType.I8: + item_count = n_elems + item_type = np.int8 + elif ggml_type == GGMLQuantizationType.I16: + item_count = n_elems + item_type = np.int16 + elif ggml_type == GGMLQuantizationType.I32: + item_count = n_elems + item_type = np.int32 + elif ggml_type == GGMLQuantizationType.I64: + item_count = n_elems + item_type = np.int64 + else: + item_count = n_bytes + item_type = np.uint8 + np_dims = quant_shape_to_byte_shape(np_dims, ggml_type) + tensors.append(ReaderTensor( + name = tensor_name, + tensor_type = ggml_type, + shape = dims, + n_elements = n_elems, + n_bytes = n_bytes, + data_offset = data_offs, + data = self._get(data_offs, item_type, item_count).reshape(np_dims), + field = field, + )) + self.tensors = tensors diff --git a/venv/lib/python3.10/site-packages/gguf/gguf_writer.py b/venv/lib/python3.10/site-packages/gguf/gguf_writer.py new file mode 100644 index 0000000000000000000000000000000000000000..54ca0c33fd3368daa6a19a15023b792da88951d2 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/gguf_writer.py @@ -0,0 +1,1076 @@ +from __future__ import annotations + +import logging +import os +import shutil +import struct +import tempfile +from dataclasses import dataclass +from enum import Enum, auto +from math import prod +from pathlib import Path +from io import BufferedWriter +from typing import IO, Any, Sequence, Mapping +from string import ascii_letters, digits + +import numpy as np + +from .constants import ( + GGUF_DEFAULT_ALIGNMENT, + GGUF_MAGIC, + GGUF_VERSION, + GGMLQuantizationType, + GGUFEndian, + GGUFValueType, + Keys, + RopeScalingType, + PoolingType, + TokenType, + ExpertGatingFuncType, +) + +from .quants import quant_shape_from_byte_shape + +logger = logging.getLogger(__name__) + + +SHARD_NAME_FORMAT = "{:s}-{:05d}-of-{:05d}.gguf" + + +@dataclass +class TensorInfo: + shape: Sequence[int] + dtype: GGMLQuantizationType + nbytes: int + tensor: np.ndarray[Any, Any] | None = None + + +@dataclass +class GGUFValue: + value: Any + type: GGUFValueType + sub_type: GGUFValueType | None = None + + +class WriterState(Enum): + NO_FILE = auto() + EMPTY = auto() + HEADER = auto() + KV_DATA = auto() + TI_DATA = auto() + WEIGHTS = auto() + + +class GGUFWriter: + fout: list[BufferedWriter] | None + path: Path | None + temp_file: tempfile.SpooledTemporaryFile[bytes] | None + tensors: list[dict[str, TensorInfo]] + kv_data: list[dict[str, GGUFValue]] + state: WriterState + _simple_value_packing = { + GGUFValueType.UINT8: "B", + GGUFValueType.INT8: "b", + GGUFValueType.UINT16: "H", + GGUFValueType.INT16: "h", + GGUFValueType.UINT32: "I", + GGUFValueType.INT32: "i", + GGUFValueType.FLOAT32: "f", + GGUFValueType.UINT64: "Q", + GGUFValueType.INT64: "q", + GGUFValueType.FLOAT64: "d", + GGUFValueType.BOOL: "?", + } + + def __init__( + self, path: os.PathLike[str] | str | None, arch: str, use_temp_file: bool = False, endianess: GGUFEndian = GGUFEndian.LITTLE, + split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False + ): + self.fout = None + self.path = Path(path) if path else None + self.arch = arch + self.endianess = endianess + self.data_alignment = GGUF_DEFAULT_ALIGNMENT + self.use_temp_file = use_temp_file + self.temp_file = None + self.tensors = [{}] + self.kv_data = [{}] + self.split_max_tensors = split_max_tensors + self.split_max_size = split_max_size + self.dry_run = dry_run + self.small_first_shard = small_first_shard + logger.info("gguf: This GGUF file is for {0} Endian only".format( + "Big" if self.endianess == GGUFEndian.BIG else "Little", + )) + self.state = WriterState.NO_FILE + + if self.small_first_shard: + self.tensors.append({}) + + self.add_architecture() + + def get_total_parameter_count(self) -> tuple[int, int, int, int]: + total_params = 0 + shared_params = 0 + expert_params = 0 + + expert_sum = 0 + n_expert_tensors = 0 + + last_lora_a: tuple[str, TensorInfo] | None = None + + for tensors in self.tensors: + for name, info in tensors.items(): + + shape = info.shape + + if name.endswith(".lora_a"): + last_lora_a = (name, info) + continue + elif name.endswith(".lora_b"): + if last_lora_a is None or last_lora_a[0] != name[:-1] + "a": + # Bail when the LoRA pair can't be found trivially + logger.warning("can't measure LoRA size correctly, tensor order is unusual") + return 0, 0, 0, 0 + else: + shape = (*shape[:-1], last_lora_a[1].shape[-1]) + + size = prod(shape) + + if "_exps." in name: + expert_params += (size // shape[-3]) + expert_sum += shape[-3] + n_expert_tensors += 1 + else: + shared_params += size + + total_params += size + + # Hopefully this should work even for variable-expert-count models + expert_count = (expert_sum // n_expert_tensors) if n_expert_tensors > 0 else 0 + + # Negate the total to signal it's likely not exact + if last_lora_a is not None: + total_params = -total_params + + # NOTE: keep the output in the same order as accepted by 'size_label' in gguf-py/gguf/utility.py + return total_params, shared_params, expert_params, expert_count + + def format_shard_names(self, path: Path) -> list[Path]: + if len(self.tensors) == 1: + return [path] + return [path.with_name(SHARD_NAME_FORMAT.format(path.stem, i + 1, len(self.tensors))) for i in range(len(self.tensors))] + + def open_output_file(self, path: Path | None = None) -> None: + if self.state is WriterState.EMPTY and self.fout is not None and (path is None or path == self.path): + # allow calling this multiple times as long as the path is the same + return + + if self.state is not WriterState.NO_FILE: + raise ValueError(f'Expected output file to be not yet opened, got {self.state}') + + if path is not None: + self.path = path + + if self.path is not None: + filenames = self.print_plan() + self.fout = [open(filename, "wb") for filename in filenames] + self.state = WriterState.EMPTY + + def print_plan(self) -> list[Path]: + logger.info("Writing the following files:") + assert self.path is not None + filenames = self.format_shard_names(self.path) + assert len(filenames) == len(self.tensors) + for name, tensors in zip(filenames, self.tensors): + logger.info(f"{name}: n_tensors = {len(tensors)}, total_size = {GGUFWriter.format_n_bytes_to_str(sum(ti.nbytes for ti in tensors.values()))}") + + if self.dry_run: + logger.info("Dry run, not writing files") + for name in filenames: + print(name) # noqa: NP100 + exit() + + return filenames + + def add_shard_kv_data(self) -> None: + if len(self.tensors) == 1: + return + + total_tensors = sum(len(t) for t in self.tensors) + assert self.fout is not None + total_splits = len(self.fout) + self.kv_data.extend({} for _ in range(len(self.kv_data), total_splits)) + for i, kv_data in enumerate(self.kv_data): + kv_data[Keys.Split.LLM_KV_SPLIT_NO] = GGUFValue(i, GGUFValueType.UINT16) + kv_data[Keys.Split.LLM_KV_SPLIT_COUNT] = GGUFValue(total_splits, GGUFValueType.UINT16) + kv_data[Keys.Split.LLM_KV_SPLIT_TENSORS_COUNT] = GGUFValue(total_tensors, GGUFValueType.INT32) + + def write_header_to_file(self, path: Path | None = None) -> None: + if len(self.tensors) == 1 and (self.split_max_tensors != 0 or self.split_max_size != 0): + logger.warning("Model fails split requirements, not splitting") + + self.open_output_file(path) + + if self.state is not WriterState.EMPTY: + raise ValueError(f'Expected output file to be empty, got {self.state}') + + assert self.fout is not None + assert len(self.fout) == len(self.tensors) + assert len(self.kv_data) == 1 + + self.add_shard_kv_data() + + for fout, tensors, kv_data in zip(self.fout, self.tensors, self.kv_data): + fout.write(self._pack(" None: + if self.state is not WriterState.HEADER: + raise ValueError(f'Expected output file to contain the header, got {self.state}') + assert self.fout is not None + + for fout, kv_data in zip(self.fout, self.kv_data): + kv_bytes = bytearray() + + for key, val in kv_data.items(): + kv_bytes += self._pack_val(key, GGUFValueType.STRING, add_vtype=False) + kv_bytes += self._pack_val(val.value, val.type, add_vtype=True, sub_type=val.sub_type) + + fout.write(kv_bytes) + + self.flush() + self.state = WriterState.KV_DATA + + def write_ti_data_to_file(self) -> None: + if self.state is not WriterState.KV_DATA: + raise ValueError(f'Expected output file to contain KV data, got {self.state}') + assert self.fout is not None + + for fout, tensors in zip(self.fout, self.tensors): + ti_data = bytearray() + offset_tensor = 0 + + for name, ti in tensors.items(): + ti_data += self._pack_val(name, GGUFValueType.STRING, add_vtype=False) + n_dims = len(ti.shape) + ti_data += self._pack("I", n_dims) + for j in range(n_dims): + ti_data += self._pack("Q", ti.shape[n_dims - 1 - j]) + ti_data += self._pack("I", ti.dtype) + ti_data += self._pack("Q", offset_tensor) + offset_tensor += GGUFWriter.ggml_pad(ti.nbytes, self.data_alignment) + + fout.write(ti_data) + fout.flush() + self.state = WriterState.TI_DATA + + def add_key_value(self, key: str, val: Any, vtype: GGUFValueType, sub_type: GGUFValueType | None = None) -> None: + if any(key in kv_data for kv_data in self.kv_data): + logger.warning(f'Duplicated key name {key!r}, overwriting it with new value {val!r} of type {vtype.name}') + + self.kv_data[0][key] = GGUFValue(value=val, type=vtype, sub_type=sub_type) + + def add_uint8(self, key: str, val: int) -> None: + self.add_key_value(key,val, GGUFValueType.UINT8) + + def add_int8(self, key: str, val: int) -> None: + self.add_key_value(key, val, GGUFValueType.INT8) + + def add_uint16(self, key: str, val: int) -> None: + self.add_key_value(key, val, GGUFValueType.UINT16) + + def add_int16(self, key: str, val: int) -> None: + self.add_key_value(key, val, GGUFValueType.INT16) + + def add_uint32(self, key: str, val: int) -> None: + self.add_key_value(key, val, GGUFValueType.UINT32) + + def add_int32(self, key: str, val: int) -> None: + self.add_key_value(key, val, GGUFValueType.INT32) + + def add_float32(self, key: str, val: float) -> None: + self.add_key_value(key, val, GGUFValueType.FLOAT32) + + def add_uint64(self, key: str, val: int) -> None: + self.add_key_value(key, val, GGUFValueType.UINT64) + + def add_int64(self, key: str, val: int) -> None: + self.add_key_value(key, val, GGUFValueType.INT64) + + def add_float64(self, key: str, val: float) -> None: + self.add_key_value(key, val, GGUFValueType.FLOAT64) + + def add_bool(self, key: str, val: bool) -> None: + self.add_key_value(key, val, GGUFValueType.BOOL) + + def add_string(self, key: str, val: str) -> None: + if not val: + return + self.add_key_value(key, val, GGUFValueType.STRING) + + def add_array(self, key: str, val: Sequence[Any]) -> None: + if len(val) == 0: + return + self.add_key_value(key, val, GGUFValueType.ARRAY) + + @staticmethod + def ggml_pad(x: int, n: int) -> int: + return ((x + n - 1) // n) * n + + def add_tensor_info( + self, name: str, tensor_shape: Sequence[int], tensor_dtype: np.dtype, + tensor_nbytes: int, raw_dtype: GGMLQuantizationType | None = None, + ) -> None: + if self.state is not WriterState.NO_FILE: + raise ValueError(f'Expected output file to be not yet opened, got {self.state}') + + if any(name in tensors for tensors in self.tensors): + raise ValueError(f'Duplicated tensor name {name!r}') + + if raw_dtype is None: + if tensor_dtype == np.float16: + dtype = GGMLQuantizationType.F16 + elif tensor_dtype == np.float32: + dtype = GGMLQuantizationType.F32 + elif tensor_dtype == np.float64: + dtype = GGMLQuantizationType.F64 + elif tensor_dtype == np.int8: + dtype = GGMLQuantizationType.I8 + elif tensor_dtype == np.int16: + dtype = GGMLQuantizationType.I16 + elif tensor_dtype == np.int32: + dtype = GGMLQuantizationType.I32 + elif tensor_dtype == np.int64: + dtype = GGMLQuantizationType.I64 + else: + raise ValueError("Only F16, F32, F64, I8, I16, I32, I64 tensors are supported for now") + else: + dtype = raw_dtype + if tensor_dtype == np.uint8: + tensor_shape = quant_shape_from_byte_shape(tensor_shape, raw_dtype) + + # make sure there is at least one tensor before splitting + if len(self.tensors[-1]) > 0: + if ( # split when over tensor limit + self.split_max_tensors != 0 + and len(self.tensors[-1]) >= self.split_max_tensors + ) or ( # split when over size limit + self.split_max_size != 0 + and sum(ti.nbytes for ti in self.tensors[-1].values()) + tensor_nbytes > self.split_max_size + ): + self.tensors.append({}) + + self.tensors[-1][name] = TensorInfo(shape=tensor_shape, dtype=dtype, nbytes=tensor_nbytes) + + def add_tensor( + self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None, + raw_dtype: GGMLQuantizationType | None = None, + ) -> None: + if self.endianess == GGUFEndian.BIG: + tensor.byteswap(inplace=True) + if self.use_temp_file and self.temp_file is None: + fp = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256 * 1024 * 1024) + fp.seek(0) + self.temp_file = fp + + shape: Sequence[int] = raw_shape if raw_shape is not None else tensor.shape + self.add_tensor_info(name, shape, tensor.dtype, tensor.nbytes, raw_dtype=raw_dtype) + + if self.temp_file is None: + self.tensors[-1][name].tensor = tensor + return + + tensor.tofile(self.temp_file) + self.write_padding(self.temp_file, tensor.nbytes) + + def write_padding(self, fp: IO[bytes], n: int, align: int | None = None) -> None: + pad = GGUFWriter.ggml_pad(n, align if align is not None else self.data_alignment) - n + if pad != 0: + fp.write(bytes([0] * pad)) + + def write_tensor_data(self, tensor: np.ndarray[Any, Any]) -> None: + if self.state is not WriterState.TI_DATA and self.state is not WriterState.WEIGHTS: + raise ValueError(f'Expected output file to contain tensor info or weights, got {self.state}') + assert self.fout is not None + + if self.endianess == GGUFEndian.BIG: + tensor.byteswap(inplace=True) + + file_id = -1 + for i, tensors in enumerate(self.tensors): + if len(tensors) > 0: + file_id = i + break + + fout = self.fout[file_id] + + # pop the first tensor info + # TODO: cleaner way to get the first key + first_tensor_name = [name for name, _ in zip(self.tensors[file_id].keys(), range(1))][0] + ti = self.tensors[file_id].pop(first_tensor_name) + assert ti.nbytes == tensor.nbytes + + self.write_padding(fout, fout.tell()) + tensor.tofile(fout) + self.write_padding(fout, tensor.nbytes) + + self.state = WriterState.WEIGHTS + + def write_tensors_to_file(self, *, progress: bool = False) -> None: + self.write_ti_data_to_file() + + assert self.fout is not None + + for fout in self.fout: + self.write_padding(fout, fout.tell()) + + if self.temp_file is None: + shard_bar = None + bar = None + + if progress: + from tqdm import tqdm + + total_bytes = sum(ti.nbytes for t in self.tensors for ti in t.values()) + + if len(self.fout) > 1: + shard_bar = tqdm(desc=f"Shard (0/{len(self.fout)})", total=None, unit="byte", unit_scale=True) + bar = tqdm(desc="Writing", total=total_bytes, unit="byte", unit_scale=True) + + for i, (fout, tensors) in enumerate(zip(self.fout, self.tensors)): + if shard_bar is not None: + shard_bar.set_description(f"Shard ({i + 1}/{len(self.fout)})") + total = sum(ti.nbytes for ti in tensors.values()) + shard_bar.reset(total=(total if total > 0 else None)) + + # relying on the fact that Python dicts preserve insertion order (since 3.7) + for ti in tensors.values(): + assert ti.tensor is not None # can only iterate once over the tensors + assert ti.tensor.nbytes == ti.nbytes + ti.tensor.tofile(fout) + if shard_bar is not None: + shard_bar.update(ti.nbytes) + if bar is not None: + bar.update(ti.nbytes) + self.write_padding(fout, ti.nbytes) + ti.tensor = None + else: + self.temp_file.seek(0) + + shutil.copyfileobj(self.temp_file, self.fout[0 if not self.small_first_shard else 1]) + self.flush() + self.temp_file.close() + + self.state = WriterState.WEIGHTS + + def flush(self) -> None: + assert self.fout is not None + for fout in self.fout: + fout.flush() + + def close(self) -> None: + if self.fout is not None: + for fout in self.fout: + fout.close() + self.fout = None + + def add_type(self, type_name: str) -> None: + self.add_string(Keys.General.TYPE, type_name) + + def add_architecture(self) -> None: + self.add_string(Keys.General.ARCHITECTURE, self.arch) + + def add_quantization_version(self, quantization_version: int) -> None: + self.add_uint32(Keys.General.QUANTIZATION_VERSION, quantization_version) + + def add_custom_alignment(self, alignment: int) -> None: + self.data_alignment = alignment + self.add_uint32(Keys.General.ALIGNMENT, alignment) + + def add_file_type(self, ftype: int) -> None: + self.add_uint32(Keys.General.FILE_TYPE, ftype) + + def add_name(self, name: str) -> None: + self.add_string(Keys.General.NAME, name) + + def add_author(self, author: str) -> None: + self.add_string(Keys.General.AUTHOR, author) + + def add_version(self, version: str) -> None: + self.add_string(Keys.General.VERSION, version) + + def add_organization(self, organization: str) -> None: + self.add_string(Keys.General.ORGANIZATION, organization) + + def add_finetune(self, finetune: str) -> None: + self.add_string(Keys.General.FINETUNE, finetune) + + def add_basename(self, basename: str) -> None: + self.add_string(Keys.General.BASENAME, basename) + + def add_description(self, description: str) -> None: + self.add_string(Keys.General.DESCRIPTION, description) + + def add_quantized_by(self, quantized: str) -> None: + self.add_string(Keys.General.QUANTIZED_BY, quantized) + + def add_size_label(self, size_label: str) -> None: + self.add_string(Keys.General.SIZE_LABEL, size_label) + + def add_license(self, license: str) -> None: + self.add_string(Keys.General.LICENSE, license) + + def add_license_name(self, license: str) -> None: + self.add_string(Keys.General.LICENSE_NAME, license) + + def add_license_link(self, license: str) -> None: + self.add_string(Keys.General.LICENSE_LINK, license) + + def add_url(self, url: str) -> None: + self.add_string(Keys.General.URL, url) + + def add_doi(self, doi: str) -> None: + self.add_string(Keys.General.DOI, doi) + + def add_uuid(self, uuid: str) -> None: + self.add_string(Keys.General.UUID, uuid) + + def add_repo_url(self, repo_url: str) -> None: + self.add_string(Keys.General.REPO_URL, repo_url) + + def add_source_url(self, url: str) -> None: + self.add_string(Keys.General.SOURCE_URL, url) + + def add_source_doi(self, doi: str) -> None: + self.add_string(Keys.General.SOURCE_DOI, doi) + + def add_source_uuid(self, uuid: str) -> None: + self.add_string(Keys.General.SOURCE_UUID, uuid) + + def add_source_repo_url(self, repo_url: str) -> None: + self.add_string(Keys.General.SOURCE_REPO_URL, repo_url) + + def add_base_model_count(self, source_count: int) -> None: + self.add_uint32(Keys.General.BASE_MODEL_COUNT, source_count) + + def add_base_model_name(self, source_id: int, name: str) -> None: + self.add_string(Keys.General.BASE_MODEL_NAME.format(id=source_id), name) + + def add_base_model_author(self, source_id: int, author: str) -> None: + self.add_string(Keys.General.BASE_MODEL_AUTHOR.format(id=source_id), author) + + def add_base_model_version(self, source_id: int, version: str) -> None: + self.add_string(Keys.General.BASE_MODEL_VERSION.format(id=source_id), version) + + def add_base_model_organization(self, source_id: int, organization: str) -> None: + self.add_string(Keys.General.BASE_MODEL_ORGANIZATION.format(id=source_id), organization) + + def add_base_model_description(self, source_id: int, description: str) -> None: + self.add_string(Keys.General.BASE_MODEL_DESCRIPTION.format(id=source_id), description) + + def add_base_model_url(self, source_id: int, url: str) -> None: + self.add_string(Keys.General.BASE_MODEL_URL.format(id=source_id), url) + + def add_base_model_doi(self, source_id: int, doi: str) -> None: + self.add_string(Keys.General.BASE_MODEL_DOI.format(id=source_id), doi) + + def add_base_model_uuid(self, source_id: int, uuid: str) -> None: + self.add_string(Keys.General.BASE_MODEL_UUID.format(id=source_id), uuid) + + def add_base_model_repo_url(self, source_id: int, repo_url: str) -> None: + self.add_string(Keys.General.BASE_MODEL_REPO_URL.format(id=source_id), repo_url) + + def add_dataset_count(self, source_count: int) -> None: + self.add_uint32(Keys.General.DATASET_COUNT, source_count) + + def add_dataset_name(self, source_id: int, name: str) -> None: + self.add_string(Keys.General.DATASET_NAME.format(id=source_id), name) + + def add_dataset_author(self, source_id: int, author: str) -> None: + self.add_string(Keys.General.DATASET_AUTHOR.format(id=source_id), author) + + def add_dataset_version(self, source_id: int, version: str) -> None: + self.add_string(Keys.General.DATASET_VERSION.format(id=source_id), version) + + def add_dataset_organization(self, source_id: int, organization: str) -> None: + self.add_string(Keys.General.DATASET_ORGANIZATION.format(id=source_id), organization) + + def add_dataset_description(self, source_id: int, description: str) -> None: + self.add_string(Keys.General.DATASET_DESCRIPTION.format(id=source_id), description) + + def add_dataset_url(self, source_id: int, url: str) -> None: + self.add_string(Keys.General.DATASET_URL.format(id=source_id), url) + + def add_dataset_doi(self, source_id: int, doi: str) -> None: + self.add_string(Keys.General.DATASET_DOI.format(id=source_id), doi) + + def add_dataset_uuid(self, source_id: int, uuid: str) -> None: + self.add_string(Keys.General.DATASET_UUID.format(id=source_id), uuid) + + def add_dataset_repo_url(self, source_id: int, repo_url: str) -> None: + self.add_string(Keys.General.DATASET_REPO_URL.format(id=source_id), repo_url) + + def add_tags(self, tags: Sequence[str]) -> None: + self.add_array(Keys.General.TAGS, tags) + + def add_languages(self, languages: Sequence[str]) -> None: + self.add_array(Keys.General.LANGUAGES, languages) + + def add_tensor_data_layout(self, layout: str) -> None: + self.add_string(Keys.LLM.TENSOR_DATA_LAYOUT.format(arch=self.arch), layout) + + def add_vocab_size(self, size: int) -> None: + self.add_uint32(Keys.LLM.VOCAB_SIZE.format(arch=self.arch), size) + + def add_context_length(self, length: int) -> None: + self.add_uint32(Keys.LLM.CONTEXT_LENGTH.format(arch=self.arch), length) + + def add_embedding_length(self, length: int) -> None: + self.add_uint32(Keys.LLM.EMBEDDING_LENGTH.format(arch=self.arch), length) + + def add_features_length(self, length: int) -> None: + self.add_uint32(Keys.LLM.FEATURES_LENGTH.format(arch=self.arch), length) + + def add_posnet_embedding_length(self, length: int) -> None: + self.add_uint32(Keys.PosNet.EMBEDDING_LENGTH.format(arch=self.arch), length) + + def add_posnet_block_count(self, length: int) -> None: + self.add_uint32(Keys.PosNet.BLOCK_COUNT.format(arch=self.arch), length) + + def add_convnext_embedding_length(self, length: int) -> None: + self.add_uint32(Keys.ConvNext.EMBEDDING_LENGTH.format(arch=self.arch), length) + + def add_convnext_block_count(self, length: int) -> None: + self.add_uint32(Keys.ConvNext.BLOCK_COUNT.format(arch=self.arch), length) + + def add_block_count(self, length: int) -> None: + self.add_uint32(Keys.LLM.BLOCK_COUNT.format(arch=self.arch), length) + + def add_leading_dense_block_count(self, length: int) -> None: + self.add_uint32(Keys.LLM.LEADING_DENSE_BLOCK_COUNT.format(arch=self.arch), length) + + def add_feed_forward_length(self, length: int | Sequence[int]) -> None: + if isinstance(length, int): + self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length) + else: + self.add_array(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length) + + def add_expert_feed_forward_length(self, length: int) -> None: + self.add_uint32(Keys.LLM.EXPERT_FEED_FORWARD_LENGTH.format(arch=self.arch), length) + + def add_expert_shared_feed_forward_length(self, length: int) -> None: + self.add_uint32(Keys.LLM.EXPERT_SHARED_FEED_FORWARD_LENGTH.format(arch=self.arch), length) + + def add_parallel_residual(self, use: bool) -> None: + self.add_bool(Keys.LLM.USE_PARALLEL_RESIDUAL.format(arch=self.arch), use) + + def add_decoder_start_token_id(self, id: int) -> None: + self.add_uint32(Keys.LLM.DECODER_START_TOKEN_ID.format(arch=self.arch), id) + + def add_head_count(self, count: int | Sequence[int]) -> None: + if isinstance(count, int): + self.add_uint32(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count) + else: + self.add_array(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count) + + def add_head_count_kv(self, count: int | Sequence[int]) -> None: + if isinstance(count, int): + self.add_uint32(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count) + else: + self.add_array(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count) + + def add_key_length(self, length: int) -> None: + self.add_uint32(Keys.Attention.KEY_LENGTH.format(arch=self.arch), length) + + def add_value_length(self, length: int) -> None: + self.add_uint32(Keys.Attention.VALUE_LENGTH.format(arch=self.arch), length) + + def add_key_length_mla(self, length: int) -> None: + self.add_uint32(Keys.Attention.KEY_LENGTH_MLA.format(arch=self.arch), length) + + def add_value_length_mla(self, length: int) -> None: + self.add_uint32(Keys.Attention.VALUE_LENGTH_MLA.format(arch=self.arch), length) + + def add_max_alibi_bias(self, bias: float) -> None: + self.add_float32(Keys.Attention.MAX_ALIBI_BIAS.format(arch=self.arch), bias) + + def add_clamp_kqv(self, value: float) -> None: + self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value) + + def add_logit_scale(self, value: float) -> None: + self.add_float32(Keys.LLM.LOGIT_SCALE.format(arch=self.arch), value) + + def add_attn_logit_softcapping(self, value: float) -> None: + self.add_float32(Keys.LLM.ATTN_LOGIT_SOFTCAPPING.format(arch=self.arch), value) + + def add_final_logit_softcapping(self, value: float) -> None: + self.add_float32(Keys.LLM.FINAL_LOGIT_SOFTCAPPING.format(arch=self.arch), value) + + def add_expert_count(self, count: int) -> None: + self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count) + + def add_expert_used_count(self, count: int) -> None: + self.add_uint32(Keys.LLM.EXPERT_USED_COUNT.format(arch=self.arch), count) + + def add_expert_shared_count(self, count: int) -> None: + self.add_uint32(Keys.LLM.EXPERT_SHARED_COUNT.format(arch=self.arch), count) + + def add_expert_weights_scale(self, value: float) -> None: + self.add_float32(Keys.LLM.EXPERT_WEIGHTS_SCALE.format(arch=self.arch), value) + + def add_expert_weights_norm(self, value: bool) -> None: + self.add_bool(Keys.LLM.EXPERT_WEIGHTS_NORM.format(arch=self.arch), value) + + def add_expert_gating_func(self, value: ExpertGatingFuncType) -> None: + self.add_uint32(Keys.LLM.EXPERT_GATING_FUNC.format(arch=self.arch), value.value) + + def add_moe_every_n_layers(self, value: int) -> None: + self.add_uint32(Keys.LLM.MOE_EVERY_N_LAYERS.format(arch=self.arch), value) + + def add_swin_norm(self, value: bool) -> None: + self.add_bool(Keys.LLM.SWIN_NORM.format(arch=self.arch), value) + + def add_rescale_every_n_layers(self, count: int) -> None: + self.add_uint32(Keys.LLM.RESCALE_EVERY_N_LAYERS.format(arch=self.arch), count) + + def add_time_mix_extra_dim(self, dim: int) -> None: + self.add_uint32(Keys.LLM.TIME_MIX_EXTRA_DIM.format(arch=self.arch), dim) + + def add_time_decay_extra_dim(self, dim: int) -> None: + self.add_uint32(Keys.LLM.TIME_DECAY_EXTRA_DIM.format(arch=self.arch), dim) + + def add_residual_scale(self, value: float) -> None: + self.add_float32(Keys.LLM.RESIDUAL_SCALE.format(arch=self.arch), value) + + def add_embedding_scale(self, value: float) -> None: + self.add_float32(Keys.LLM.EMBEDDING_SCALE.format(arch=self.arch), value) + + def add_wkv_head_size(self, size: int) -> None: + self.add_uint32(Keys.WKV.HEAD_SIZE.format(arch=self.arch), size) + + def add_token_shift_count(self, count: int) -> None: + self.add_uint32(Keys.LLM.TOKEN_SHIFT_COUNT.format(arch=self.arch), count) + + def add_interleave_moe_layer_step(self, value: int) -> None: + self.add_uint32(Keys.LLM.INTERLEAVE_MOE_LAYER_STEP.format(arch=self.arch), value) + + def add_layer_norm_eps(self, value: float) -> None: + self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value) + + def add_layer_norm_rms_eps(self, value: float) -> None: + self.add_float32(Keys.Attention.LAYERNORM_RMS_EPS.format(arch=self.arch), value) + + def add_group_norm_eps(self, value: float) -> None: + self.add_float32(Keys.Attention.GROUPNORM_EPS.format(arch=self.arch), value) + + def add_group_norm_groups(self, value: int) -> None: + self.add_uint32(Keys.Attention.GROUPNORM_GROUPS.format(arch=self.arch), value) + + def add_causal_attention(self, value: bool) -> None: + self.add_bool(Keys.Attention.CAUSAL.format(arch=self.arch), value) + + def add_q_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.Q_LORA_RANK.format(arch=self.arch), length) + + def add_kv_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.KV_LORA_RANK.format(arch=self.arch), length) + + def add_decay_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.DECAY_LORA_RANK.format(arch=self.arch), length) + + def add_iclr_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.ICLR_LORA_RANK.format(arch=self.arch), length) + + def add_value_residual_mix_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.VALUE_RESIDUAL_MIX_LORA_RANK.format(arch=self.arch), length) + + def add_gate_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.GATE_LORA_RANK.format(arch=self.arch), length) + + def add_relative_attn_buckets_count(self, value: int) -> None: + self.add_uint32(Keys.Attention.REL_BUCKETS_COUNT.format(arch=self.arch), value) + + def add_sliding_window(self, value: int) -> None: + self.add_uint32(Keys.Attention.SLIDING_WINDOW.format(arch=self.arch), value) + + def add_attention_scale(self, value: float) -> None: + self.add_float32(Keys.Attention.SCALE.format(arch=self.arch), value) + + def add_pooling_type(self, value: PoolingType) -> None: + self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value) + + def add_rope_dimension_count(self, count: int) -> None: + self.add_uint32(Keys.Rope.DIMENSION_COUNT.format(arch=self.arch), count) + + def add_rope_dimension_sections(self, dims: Sequence[int]) -> None: + self.add_array(Keys.Rope.DIMENSION_SECTIONS.format(arch=self.arch), dims) + + def add_rope_freq_base(self, value: float) -> None: + self.add_float32(Keys.Rope.FREQ_BASE.format(arch=self.arch), value) + + def add_rope_scaling_type(self, value: RopeScalingType) -> None: + self.add_string(Keys.Rope.SCALING_TYPE.format(arch=self.arch), value.value) + + def add_rope_scaling_factor(self, value: float) -> None: + self.add_float32(Keys.Rope.SCALING_FACTOR.format(arch=self.arch), value) + + def add_rope_scaling_attn_factors(self, value: float) -> None: + self.add_float32(Keys.Rope.SCALING_ATTN_FACTOR.format(arch=self.arch), value) + + def add_rope_scaling_orig_ctx_len(self, value: int) -> None: + self.add_uint32(Keys.Rope.SCALING_ORIG_CTX_LEN.format(arch=self.arch), value) + + def add_rope_scaling_finetuned(self, value: bool) -> None: + self.add_bool(Keys.Rope.SCALING_FINETUNED.format(arch=self.arch), value) + + def add_rope_scaling_yarn_log_mul(self, value: float) -> None: + self.add_float32(Keys.Rope.SCALING_YARN_LOG_MUL.format(arch=self.arch), value) + + def add_ssm_conv_kernel(self, value: int) -> None: + self.add_uint32(Keys.SSM.CONV_KERNEL.format(arch=self.arch), value) + + def add_ssm_inner_size(self, value: int) -> None: + self.add_uint32(Keys.SSM.INNER_SIZE.format(arch=self.arch), value) + + def add_ssm_state_size(self, value: int) -> None: + self.add_uint32(Keys.SSM.STATE_SIZE.format(arch=self.arch), value) + + def add_ssm_time_step_rank(self, value: int) -> None: + self.add_uint32(Keys.SSM.TIME_STEP_RANK.format(arch=self.arch), value) + + def add_ssm_dt_b_c_rms(self, value: bool) -> None: + self.add_bool(Keys.SSM.DT_B_C_RMS.format(arch=self.arch), value) + + def add_tokenizer_model(self, model: str) -> None: + self.add_string(Keys.Tokenizer.MODEL, model) + + def add_tokenizer_pre(self, pre: str) -> None: + self.add_string(Keys.Tokenizer.PRE, pre) + + def add_token_list(self, tokens: Sequence[str] | Sequence[bytes] | Sequence[bytearray]) -> None: + self.add_array(Keys.Tokenizer.LIST, tokens) + + def add_token_merges(self, merges: Sequence[str] | Sequence[bytes] | Sequence[bytearray]) -> None: + self.add_array(Keys.Tokenizer.MERGES, merges) + + def add_token_types(self, types: Sequence[TokenType] | Sequence[int]) -> None: + self.add_array(Keys.Tokenizer.TOKEN_TYPE, types) + + def add_token_type_count(self, value: int) -> None: + self.add_uint32(Keys.Tokenizer.TOKEN_TYPE_COUNT, value) + + def add_token_scores(self, scores: Sequence[float]) -> None: + self.add_array(Keys.Tokenizer.SCORES, scores) + + def add_bos_token_id(self, id: int) -> None: + self.add_uint32(Keys.Tokenizer.BOS_ID, id) + + def add_eos_token_id(self, id: int) -> None: + self.add_uint32(Keys.Tokenizer.EOS_ID, id) + + def add_unk_token_id(self, id: int) -> None: + self.add_uint32(Keys.Tokenizer.UNK_ID, id) + + def add_sep_token_id(self, id: int) -> None: + self.add_uint32(Keys.Tokenizer.SEP_ID, id) + + def add_pad_token_id(self, id: int) -> None: + self.add_uint32(Keys.Tokenizer.PAD_ID, id) + + def add_mask_token_id(self, id: int) -> None: + self.add_uint32(Keys.Tokenizer.MASK_ID, id) + + def add_add_bos_token(self, value: bool) -> None: + self.add_bool(Keys.Tokenizer.ADD_BOS, value) + + def add_add_eos_token(self, value: bool) -> None: + self.add_bool(Keys.Tokenizer.ADD_EOS, value) + + def add_add_space_prefix(self, value: bool) -> None: + self.add_bool(Keys.Tokenizer.ADD_PREFIX, value) + + def add_remove_extra_whitespaces(self, value: bool) -> None: + self.add_bool(Keys.Tokenizer.REMOVE_EXTRA_WS, value) + + def add_precompiled_charsmap(self, charsmap: bytes) -> None: + self.add_array(Keys.Tokenizer.PRECOMPILED_CHARSMAP, charsmap) + + def add_chat_template(self, value: str | Sequence[Mapping[str, str]]) -> None: + if not isinstance(value, str): + template_default = None + template_names = set() + + for choice in value: + name = choice.get('name', '') + template = choice.get('template') + + # Allowing non-alphanumerical characters in template name is probably not a good idea, so filter it + name = ''.join((c if c in ascii_letters + digits else '_' for c in name)) + + if name and template is not None: + if name == 'default': + template_default = template + else: + template_names.add(name) + self.add_string(Keys.Tokenizer.CHAT_TEMPLATE_N.format(name=name), template) + + if template_names: + self.add_array(Keys.Tokenizer.CHAT_TEMPLATES, list(template_names)) + + if template_default is None: + return + + value = template_default + + self.add_string(Keys.Tokenizer.CHAT_TEMPLATE, value) + + def add_eot_token_id(self, id: int) -> None: + self.add_uint32(Keys.Tokenizer.EOT_ID, id) + + def add_eom_token_id(self, id: int) -> None: + self.add_uint32(Keys.Tokenizer.EOM_ID, id) + + def add_classifier_output_labels(self, labels: Sequence[str]) -> None: + self.add_array(Keys.Classifier.OUTPUT_LABELS.format(arch=self.arch), labels) + + # for vision models + + def add_clip_has_vision_encoder(self, value: bool) -> None: + self.add_bool(Keys.Clip.HAS_VISION_ENCODER, value) + + def add_clip_has_audio_encoder(self, value: bool) -> None: + self.add_bool(Keys.Clip.HAS_AUDIO_ENCODER, value) + + def add_clip_projector_type(self, value: str) -> None: + self.add_string(Keys.Clip.PROJECTOR_TYPE, value) + + def add_vision_projection_dim(self, value: int) -> None: + self.add_uint32(Keys.ClipVision.PROJECTION_DIM, value) + + def add_vision_patch_size(self, value: int) -> None: + self.add_uint32(Keys.ClipVision.PATCH_SIZE, value) + + def add_vision_embedding_length(self, value: int) -> None: + self.add_uint32(Keys.ClipVision.EMBEDDING_LENGTH, value) + + def add_vision_feed_forward_length(self, value: int) -> None: + self.add_uint32(Keys.ClipVision.FEED_FORWARD_LENGTH, value) + + def add_vision_block_count(self, value: int) -> None: + self.add_uint32(Keys.ClipVision.BLOCK_COUNT, value) + + def add_vision_head_count(self, value: int) -> None: + self.add_uint32(Keys.ClipVision.Attention.HEAD_COUNT, value) + + def add_vision_attention_layernorm_eps(self, value: float) -> None: + self.add_float32(Keys.ClipVision.Attention.LAYERNORM_EPS, value) + + def add_vision_image_size(self, value: int) -> None: + self.add_uint32(Keys.ClipVision.IMAGE_SIZE, value) + + def add_vision_image_mean(self, values: Sequence[float]) -> None: + self.add_array(Keys.ClipVision.IMAGE_MEAN, values) + + def add_vision_image_std(self, values: Sequence[float]) -> None: + self.add_array(Keys.ClipVision.IMAGE_STD, values) + + def add_vision_spatial_merge_size(self, value: int) -> None: + self.add_uint32(Keys.ClipVision.SPATIAL_MERGE_SIZE, value) + + def add_vision_use_gelu(self, value: bool) -> None: + self.add_bool(Keys.ClipVision.USE_GELU, value) + + def add_vision_use_silu(self, value: bool) -> None: + self.add_bool(Keys.ClipVision.USE_SILU, value) + + def add_vision_projector_scale_factor(self, value: int) -> None: + self.add_uint32(Keys.ClipVision.Projector.SCALE_FACTOR, value) + + def add_vision_n_wa_pattern(self, value: int) -> None: + self.add_uint32(Keys.ClipVision.N_WA_PATTERN, value) + + # audio models + + def add_audio_projection_dim(self, value: int) -> None: + self.add_uint32(Keys.ClipAudio.PROJECTION_DIM, value) + + def add_audio_embedding_length(self, value: int) -> None: + self.add_uint32(Keys.ClipAudio.EMBEDDING_LENGTH, value) + + def add_audio_feed_forward_length(self, value: int) -> None: + self.add_uint32(Keys.ClipAudio.FEED_FORWARD_LENGTH, value) + + def add_audio_block_count(self, value: int) -> None: + self.add_uint32(Keys.ClipAudio.BLOCK_COUNT, value) + + def add_audio_head_count(self, value: int) -> None: + self.add_uint32(Keys.ClipAudio.Attention.HEAD_COUNT, value) + + def add_audio_attention_layernorm_eps(self, value: float) -> None: + self.add_float32(Keys.ClipAudio.Attention.LAYERNORM_EPS, value) + + def add_audio_num_mel_bins(self, value: int) -> None: + self.add_uint32(Keys.ClipAudio.NUM_MEL_BINS, value) + + def add_audio_stack_factor(self, value: int) -> None: + self.add_uint32(Keys.ClipAudio.Projector.STACK_FACTOR, value) + + def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes: + pack_prefix = '' + if not skip_pack_prefix: + pack_prefix = '<' if self.endianess == GGUFEndian.LITTLE else '>' + return struct.pack(f'{pack_prefix}{fmt}', value) + + def _pack_val(self, val: Any, vtype: GGUFValueType, add_vtype: bool, sub_type: GGUFValueType | None = None) -> bytes: + kv_data = bytearray() + + if add_vtype: + kv_data += self._pack("I", vtype) + + pack_fmt = self._simple_value_packing.get(vtype) + if pack_fmt is not None: + kv_data += self._pack(pack_fmt, val, skip_pack_prefix = vtype == GGUFValueType.BOOL) + elif vtype == GGUFValueType.STRING: + encoded_val = val.encode("utf-8") if isinstance(val, str) else val + kv_data += self._pack("Q", len(encoded_val)) + kv_data += encoded_val + elif vtype == GGUFValueType.ARRAY: + + if not isinstance(val, Sequence): + raise ValueError("Invalid GGUF metadata array, expecting sequence") + + if len(val) == 0: + raise ValueError("Invalid GGUF metadata array. Empty array") + + if sub_type is not None: + ltype = sub_type + elif isinstance(val, bytes): + ltype = GGUFValueType.UINT8 + else: + ltype = GGUFValueType.get_type(val[0]) + if not all(GGUFValueType.get_type(i) is ltype for i in val[1:]): + raise ValueError("All items in a GGUF array should be of the same type") + kv_data += self._pack("I", ltype) + kv_data += self._pack("Q", len(val)) + for item in val: + kv_data += self._pack_val(item, ltype, add_vtype=False) + else: + raise ValueError("Invalid GGUF metadata value type or value") + + return kv_data + + @staticmethod + def format_n_bytes_to_str(num: int) -> str: + if num == 0: + return "negligible - metadata only" + fnum = float(num) + for unit in ("", "K", "M", "G"): + if abs(fnum) < 1000.0: + return f"{fnum:3.1f}{unit}" + fnum /= 1000.0 + return f"{fnum:.1f}T - over 1TB, split recommended" diff --git a/venv/lib/python3.10/site-packages/gguf/lazy.py b/venv/lib/python3.10/site-packages/gguf/lazy.py new file mode 100644 index 0000000000000000000000000000000000000000..f9bcadae0224bac5f922ce6565dd1e3883a923e1 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/lazy.py @@ -0,0 +1,223 @@ +from __future__ import annotations +from abc import ABC, ABCMeta, abstractmethod + +import logging +from typing import Any, Callable + +import numpy as np +from numpy.typing import DTypeLike + + +logger = logging.getLogger(__name__) + + +class LazyMeta(ABCMeta): + + def __new__(cls, name: str, bases: tuple[type, ...], namespace: dict[str, Any], **kwargs): + def __getattr__(self, name: str) -> Any: + meta_attr = getattr(self._meta, name) + if callable(meta_attr): + return type(self)._wrap_fn( + (lambda s, *args, **kwargs: getattr(s, name)(*args, **kwargs)), + use_self=self, + ) + elif isinstance(meta_attr, self._tensor_type): + # e.g. self.T with torch.Tensor should still be wrapped + return type(self)._wrap_fn(lambda s: getattr(s, name))(self) + else: + # no need to wrap non-tensor properties, + # and they likely don't depend on the actual contents of the tensor + return meta_attr + + namespace["__getattr__"] = __getattr__ + + # need to make a builder for the wrapped wrapper to copy the name, + # or else it fails with very cryptic error messages, + # because somehow the same string would end up in every closures + def mk_wrap(op_name: str, *, meta_noop: bool = False): + # need to wrap the wrapper to get self + def wrapped_special_op(self, *args, **kwargs): + return type(self)._wrap_fn( + getattr(type(self)._tensor_type, op_name), + meta_noop=meta_noop, + )(self, *args, **kwargs) + return wrapped_special_op + + # special methods bypass __getattr__, so they need to be added manually + # ref: https://docs.python.org/3/reference/datamodel.html#special-lookup + # NOTE: doing this from a metaclass is very convenient + # TODO: make this even more comprehensive + for binary_op in ( + "lt", "le", "eq", "ne", "ge", "gt", "not" + "abs", "add", "and", "floordiv", "invert", "lshift", "mod", "mul", "matmul", + "neg", "or", "pos", "pow", "rshift", "sub", "truediv", "xor", + "iadd", "iand", "ifloordiv", "ilshift", "imod", "imul", "ior", "irshift", "isub", "ixor", + "radd", "rand", "rfloordiv", "rmul", "ror", "rpow", "rsub", "rtruediv", "rxor", + ): + attr_name = f"__{binary_op}__" + # the result of these operators usually has the same shape and dtype as the input, + # so evaluation on the meta tensor can be skipped. + namespace[attr_name] = mk_wrap(attr_name, meta_noop=True) + + for special_op in ( + "getitem", "setitem", "len", + ): + attr_name = f"__{special_op}__" + namespace[attr_name] = mk_wrap(attr_name, meta_noop=False) + + return super().__new__(cls, name, bases, namespace, **kwargs) + + +# Tree of lazy tensors +class LazyBase(ABC, metaclass=LazyMeta): + _tensor_type: type + _meta: Any + _data: Any | None + _args: tuple + _kwargs: dict[str, Any] + _func: Callable[[Any], Any] | None + + def __init__(self, *, meta: Any, data: Any | None = None, args: tuple = (), kwargs: dict[str, Any] | None = None, func: Callable[[Any], Any] | None = None): + super().__init__() + self._meta = meta + self._data = data + self._args = args + self._kwargs = kwargs if kwargs is not None else {} + self._func = func + assert self._func is not None or self._data is not None + + def __init_subclass__(cls) -> None: + if "_tensor_type" not in cls.__dict__: + raise TypeError(f"property '_tensor_type' must be defined for {cls!r}") + return super().__init_subclass__() + + @staticmethod + def _recurse_apply(o: Any, fn: Callable[[Any], Any]) -> Any: + # TODO: dict and set + if isinstance(o, (list, tuple)): + L = [] + for item in o: + L.append(LazyBase._recurse_apply(item, fn)) + if isinstance(o, tuple): + L = tuple(L) + return L + elif isinstance(o, LazyBase): + return fn(o) + else: + return o + + @classmethod + def _wrap_fn(cls, fn: Callable, *, use_self: LazyBase | None = None, meta_noop: bool | DTypeLike | tuple[DTypeLike, Callable[[tuple[int, ...]], tuple[int, ...]]] = False) -> Callable[[Any], Any]: + def wrapped_fn(*args, **kwargs): + if kwargs is None: + kwargs = {} + args = ((use_self,) if use_self is not None else ()) + args + + meta_args = LazyBase._recurse_apply(args, lambda t: t._meta) + # TODO: maybe handle tensors in kwargs too + + if isinstance(meta_noop, bool) and not meta_noop: + try: + res = fn(*meta_args, **kwargs) + except NotImplementedError: + # running some operations on PyTorch's Meta tensors can cause this exception + res = None + else: + # some operators don't need to actually run on the meta tensors + assert len(args) > 0 + res = args[0] + assert isinstance(res, cls) + res = res._meta + # allow operations to override the dtype and shape + if meta_noop is not True: + if isinstance(meta_noop, tuple): + dtype, shape = meta_noop + assert callable(shape) + res = cls.meta_with_dtype_and_shape(dtype, shape(res.shape)) + else: + res = cls.meta_with_dtype_and_shape(meta_noop, res.shape) + + if isinstance(res, cls._tensor_type): + return cls(meta=cls.eager_to_meta(res), args=args, kwargs=kwargs, func=fn) + elif isinstance(res, tuple) and all(isinstance(t, cls._tensor_type) for t in res): + # share the evaluation between lazy tuple elements + shared_args: list = [args, None] + + def eager_tuple_element(a: list[Any], i: int = 0, /, **kw) -> LazyBase: + assert len(a) == 2 + if a[1] is None: + a[1] = fn(*a[0], **kw) + return a[1][i] + return tuple(cls(meta=cls.eager_to_meta(res[i]), args=(shared_args, i), kwargs=kwargs, func=eager_tuple_element) for i in range(len(res))) + else: + del res # not needed + # non-tensor return likely relies on the contents of the args + # (e.g. the result of torch.equal) + eager_args = cls.to_eager(args) + return fn(*eager_args, **kwargs) + return wrapped_fn + + @classmethod + def to_eager(cls, t: Any) -> Any: + def simple_to_eager(_t: LazyBase) -> Any: + if _t._data is not None: + return _t._data + + # NOTE: there's a recursion limit in Python (usually 1000) + + assert _t._func is not None + _t._args = cls._recurse_apply(_t._args, simple_to_eager) + _t._data = _t._func(*_t._args, **_t._kwargs) + # sanity check + assert _t._data is not None + assert _t._data.dtype == _t._meta.dtype + assert _t._data.shape == _t._meta.shape + + return _t._data + + # recurse into lists and/or tuples, keeping their structure + return cls._recurse_apply(t, simple_to_eager) + + @classmethod + def eager_to_meta(cls, t: Any) -> Any: + return cls.meta_with_dtype_and_shape(t.dtype, t.shape) + + # must be overridden, meta tensor init is backend-specific + @classmethod + @abstractmethod + def meta_with_dtype_and_shape(cls, dtype: Any, shape: Any) -> Any: pass + + @classmethod + def from_eager(cls, t: Any) -> Any: + if type(t) is cls: + # already lazy + return t + elif isinstance(t, cls._tensor_type): + return cls(meta=cls.eager_to_meta(t), data=t) + else: + return TypeError(f"{type(t)!r} is not compatible with {cls._tensor_type!r}") + + +class LazyNumpyTensor(LazyBase): + _tensor_type = np.ndarray + + shape: tuple[int, ...] # Makes the type checker happy in quants.py + + @classmethod + def meta_with_dtype_and_shape(cls, dtype: DTypeLike, shape: tuple[int, ...]) -> np.ndarray[Any, Any]: + # The initial idea was to use np.nan as the fill value, + # but non-float types like np.int16 can't use that. + # So zero it is. + cheat = np.zeros(1, dtype) + return np.lib.stride_tricks.as_strided(cheat, shape, (0 for _ in shape)) + + def astype(self, dtype, *args, **kwargs): + meta = type(self).meta_with_dtype_and_shape(dtype, self._meta.shape) + full_args = (self, dtype,) + args + return type(self)(meta=meta, args=full_args, kwargs=kwargs, func=(lambda a, *args, **kwargs: a.astype(*args, **kwargs))) + + def tofile(self, *args, **kwargs): + eager = LazyNumpyTensor.to_eager(self) + return eager.tofile(*args, **kwargs) + + # TODO: __array_function__ diff --git a/venv/lib/python3.10/site-packages/gguf/metadata.py b/venv/lib/python3.10/site-packages/gguf/metadata.py new file mode 100644 index 0000000000000000000000000000000000000000..e807f434689de669816504921f54809cac15fc71 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/metadata.py @@ -0,0 +1,642 @@ +from __future__ import annotations + +import re +import json +import yaml +import logging +from pathlib import Path +from typing import Any, Literal, Optional +from dataclasses import dataclass + +from .constants import Keys + +import gguf + +logger = logging.getLogger("metadata") + + +@dataclass +class Metadata: + # Authorship Metadata to be written to GGUF KV Store + name: Optional[str] = None + author: Optional[str] = None + version: Optional[str] = None + organization: Optional[str] = None + finetune: Optional[str] = None + basename: Optional[str] = None + description: Optional[str] = None + quantized_by: Optional[str] = None + size_label: Optional[str] = None + url: Optional[str] = None + doi: Optional[str] = None + uuid: Optional[str] = None + repo_url: Optional[str] = None + source_url: Optional[str] = None + source_doi: Optional[str] = None + source_uuid: Optional[str] = None + source_repo_url: Optional[str] = None + license: Optional[str] = None + license_name: Optional[str] = None + license_link: Optional[str] = None + base_models: Optional[list[dict]] = None + tags: Optional[list[str]] = None + languages: Optional[list[str]] = None + datasets: Optional[list[dict]] = None + + @staticmethod + def load(metadata_override_path: Optional[Path] = None, model_path: Optional[Path] = None, model_name: Optional[str] = None, total_params: int = 0) -> Metadata: + # This grabs as many contextual authorship metadata as possible from the model repository + # making any conversion as required to match the gguf kv store metadata format + # as well as giving users the ability to override any authorship metadata that may be incorrect + + # Create a new Metadata instance + metadata = Metadata() + + model_card = Metadata.load_model_card(model_path) + hf_params = Metadata.load_hf_parameters(model_path) + # TODO: load adapter_config.json when possible, it usually contains the base model of the LoRA adapter + + # heuristics + metadata = Metadata.apply_metadata_heuristic(metadata, model_card, hf_params, model_path, total_params) + + # Metadata Override File Provided + # This is based on LLM_KV_NAMES mapping in llama.cpp + metadata_override = Metadata.load_metadata_override(metadata_override_path) + + metadata.name = metadata_override.get(Keys.General.NAME, metadata.name) + metadata.author = metadata_override.get(Keys.General.AUTHOR, metadata.author) + metadata.version = metadata_override.get(Keys.General.VERSION, metadata.version) + metadata.organization = metadata_override.get(Keys.General.ORGANIZATION, metadata.organization) + + metadata.finetune = metadata_override.get(Keys.General.FINETUNE, metadata.finetune) + metadata.basename = metadata_override.get(Keys.General.BASENAME, metadata.basename) + + metadata.description = metadata_override.get(Keys.General.DESCRIPTION, metadata.description) + metadata.quantized_by = metadata_override.get(Keys.General.QUANTIZED_BY, metadata.quantized_by) + + metadata.size_label = metadata_override.get(Keys.General.SIZE_LABEL, metadata.size_label) + metadata.license_name = metadata_override.get(Keys.General.LICENSE_NAME, metadata.license_name) + metadata.license_link = metadata_override.get(Keys.General.LICENSE_LINK, metadata.license_link) + + metadata.url = metadata_override.get(Keys.General.URL, metadata.url) + metadata.doi = metadata_override.get(Keys.General.DOI, metadata.doi) + metadata.uuid = metadata_override.get(Keys.General.UUID, metadata.uuid) + metadata.repo_url = metadata_override.get(Keys.General.REPO_URL, metadata.repo_url) + + metadata.source_url = metadata_override.get(Keys.General.SOURCE_URL, metadata.source_url) + metadata.source_doi = metadata_override.get(Keys.General.SOURCE_DOI, metadata.source_doi) + metadata.source_uuid = metadata_override.get(Keys.General.SOURCE_UUID, metadata.source_uuid) + metadata.source_repo_url = metadata_override.get(Keys.General.SOURCE_REPO_URL, metadata.source_repo_url) + + # Base Models is received here as an array of models + metadata.base_models = metadata_override.get("general.base_models", metadata.base_models) + + # Datasets is received here as an array of datasets + metadata.datasets = metadata_override.get("general.datasets", metadata.datasets) + + metadata.tags = metadata_override.get(Keys.General.TAGS, metadata.tags) + metadata.languages = metadata_override.get(Keys.General.LANGUAGES, metadata.languages) + + # Direct Metadata Override (via direct cli argument) + if model_name is not None: + metadata.name = model_name + + return metadata + + @staticmethod + def load_metadata_override(metadata_override_path: Optional[Path] = None) -> dict[str, Any]: + if metadata_override_path is None or not metadata_override_path.is_file(): + return {} + + with open(metadata_override_path, "r", encoding="utf-8") as f: + return json.load(f) + + @staticmethod + def load_model_card(model_path: Optional[Path] = None) -> dict[str, Any]: + if model_path is None or not model_path.is_dir(): + return {} + + model_card_path = model_path / "README.md" + + if not model_card_path.is_file(): + return {} + + # The model card metadata is assumed to always be in YAML (frontmatter) + # ref: https://github.com/huggingface/transformers/blob/a5c642fe7a1f25d3bdcd76991443ba6ff7ee34b2/src/transformers/modelcard.py#L468-L473 + yaml_content: str = "" + with open(model_card_path, "r", encoding="utf-8") as f: + content = f.read() + lines = content.splitlines() + lines_yaml = [] + if len(lines) == 0: + # Empty file + return {} + if len(lines) > 0 and lines[0] != "---": + # No frontmatter + return {} + for line in lines[1:]: + if line == "---": + break # End of frontmatter + else: + lines_yaml.append(line) + yaml_content = "\n".join(lines_yaml) + "\n" + + # Quick hack to fix the Norway problem + # https://hitchdev.com/strictyaml/why/implicit-typing-removed/ + yaml_content = yaml_content.replace("- no\n", "- \"no\"\n") + + if yaml_content: + data = yaml.safe_load(yaml_content) + if isinstance(data, dict): + return data + else: + logger.error(f"while reading YAML model card frontmatter, data is {type(data)} instead of dict") + return {} + else: + return {} + + @staticmethod + def load_hf_parameters(model_path: Optional[Path] = None) -> dict[str, Any]: + if model_path is None or not model_path.is_dir(): + return {} + + config_path = model_path / "config.json" + + if not config_path.is_file(): + return {} + + with open(config_path, "r", encoding="utf-8") as f: + return json.load(f) + + @staticmethod + def id_to_title(string): + # Convert capitalization into title form unless acronym or version number + return ' '.join([w.title() if w.islower() and not re.match(r'^(v\d+(?:\.\d+)*|\d.*)$', w) else w for w in string.strip().replace('-', ' ').split()]) + + @staticmethod + def get_model_id_components(model_id: Optional[str] = None, total_params: int = 0) -> tuple[str | None, str | None, str | None, str | None, str | None, str | None]: + # Huggingface often store model id as '/' + # so let's parse it and apply some heuristics if possible for model name components + + if model_id is None: + # model ID missing + return None, None, None, None, None, None + + if ' ' in model_id: + # model ID is actually a normal human sentence + # which means its most likely a normal model name only + # not part of the hugging face naming standard, but whatever + return model_id, None, None, None, None, None + + if '/' in model_id: + # model ID (huggingface style) + org_component, model_full_name_component = model_id.split('/', 1) + else: + # model ID but missing org components + org_component, model_full_name_component = None, model_id + + # Check if we erroneously matched against './' or '../' etc... + if org_component is not None and len(org_component) > 0 and org_component[0] == '.': + org_component = None + + name_parts: list[str] = model_full_name_component.split('-') + + # Remove empty parts + for i in reversed(range(len(name_parts))): + if len(name_parts[i]) == 0: + del name_parts[i] + + name_types: list[ + set[Literal["basename", "size_label", "finetune", "version", "type"]] + ] = [set() for _ in name_parts] + + # Annotate the name + for i, part in enumerate(name_parts): + # Version + if re.fullmatch(r'(v|iter)?\d+([.]\d+)*', part, re.IGNORECASE): + name_types[i].add("version") + # Quant type (should not be there for base models, but still annotated) + elif re.fullmatch(r'i?q\d(_\w)*|b?fp?(16|32)', part, re.IGNORECASE): + name_types[i].add("type") + name_parts[i] = part.upper() + # Model size + elif i > 0 and re.fullmatch(r'(([A]|\d+[x])?\d+([._]\d+)?[KMBT][\d]?|small|mini|medium|large|x?xl)', part, re.IGNORECASE): + part = part.replace("_", ".") + # Handle weird bloom-7b1 notation + if part[-1].isdecimal(): + part = part[:-2] + "." + part[-1] + part[-2] + # Normalize the size suffixes + if len(part) > 1 and part[-2].isdecimal(): + if part[-1] in "kmbt": + part = part[:-1] + part[-1].upper() + if total_params != 0: + try: + label_params = float(part[:-1]) * pow(1000, " KMBT".find(part[-1])) + # Only use it as a size label if it's close or bigger than the model size + # Note that LoRA adapters don't necessarily include all layers, + # so this is why bigger label sizes are accepted. + # Do not use the size label when it's smaller than 1/8 of the model size + if (total_params < 0 and label_params < abs(total_params) // 8) or ( + # Check both directions when the current model isn't a LoRA adapter + total_params > 0 and abs(label_params - total_params) > 7 * total_params // 8 + ): + # Likely a context length + name_types[i].add("finetune") + # Lowercase the size when it's a context length + part = part[:-1] + part[-1].lower() + except ValueError: + # Failed to convert the size label to float, use it anyway + pass + if len(name_types[i]) == 0: + name_types[i].add("size_label") + name_parts[i] = part + # Some easy to recognize finetune names + elif i > 0 and re.fullmatch(r'chat|instruct|vision|lora', part, re.IGNORECASE): + if total_params < 0 and part.lower() == "lora": + # ignore redundant "lora" in the finetune part when the output is a lora adapter + name_types[i].add("type") + else: + name_types[i].add("finetune") + + # Ignore word-based size labels when there is at least a number-based one present + # TODO: should word-based size labels always be removed instead? + if any(c.isdecimal() for n, t in zip(name_parts, name_types) if "size_label" in t for c in n): + for n, t in zip(name_parts, name_types): + if "size_label" in t: + if all(c.isalpha() for c in n): + t.remove("size_label") + + at_start = True + # Find the basename through the annotated name + for part, t in zip(name_parts, name_types): + if at_start and ((len(t) == 0 and part[0].isalpha()) or "version" in t): + t.add("basename") + else: + if at_start: + at_start = False + if len(t) == 0: + t.add("finetune") + + # Remove the basename annotation from trailing version + for part, t in zip(reversed(name_parts), reversed(name_types)): + if "basename" in t and len(t) > 1: + t.remove("basename") + else: + break + + basename = "-".join(n for n, t in zip(name_parts, name_types) if "basename" in t) or None + # Deduplicate size labels using order-preserving 'dict' ('set' seems to sort the keys) + size_label = "-".join(dict.fromkeys(s for s, t in zip(name_parts, name_types) if "size_label" in t).keys()) or None + finetune = "-".join(f for f, t in zip(name_parts, name_types) if "finetune" in t) or None + # TODO: should the basename version always be excluded? + # NOTE: multiple finetune versions are joined together + version = "-".join(v for v, t, in zip(name_parts, name_types) if "version" in t and "basename" not in t) or None + + if size_label is None and finetune is None and version is None: + # Too ambiguous, output nothing + basename = None + + return model_full_name_component, org_component, basename, finetune, version, size_label + + @staticmethod + def apply_metadata_heuristic(metadata: Metadata, model_card: Optional[dict] = None, hf_params: Optional[dict] = None, model_path: Optional[Path] = None, total_params: int = 0) -> Metadata: + # Reference Model Card Metadata: https://github.com/huggingface/hub-docs/blob/main/modelcard.md?plain=1 + + # Model Card Heuristics + ######################## + if model_card is not None: + + def use_model_card_metadata(metadata_key: str, model_card_key: str): + if model_card_key in model_card and getattr(metadata, metadata_key, None) is None: + setattr(metadata, metadata_key, model_card.get(model_card_key)) + + def use_array_model_card_metadata(metadata_key: str, model_card_key: str): + # Note: Will append rather than replace if already exist + tags_value = model_card.get(model_card_key, None) + if tags_value is None: + return + + current_value = getattr(metadata, metadata_key, None) + if current_value is None: + current_value = [] + + if isinstance(tags_value, str): + current_value.append(tags_value) + elif isinstance(tags_value, list): + current_value.extend(tags_value) + + setattr(metadata, metadata_key, current_value) + + # LLAMA.cpp's direct internal convention + # (Definitely not part of hugging face formal/informal standard) + ######################################### + use_model_card_metadata("name", "name") + use_model_card_metadata("author", "author") + use_model_card_metadata("version", "version") + use_model_card_metadata("organization", "organization") + use_model_card_metadata("description", "description") + use_model_card_metadata("finetune", "finetune") + use_model_card_metadata("basename", "basename") + use_model_card_metadata("size_label", "size_label") + use_model_card_metadata("source_url", "url") + use_model_card_metadata("source_doi", "doi") + use_model_card_metadata("source_uuid", "uuid") + use_model_card_metadata("source_repo_url", "repo_url") + + # LLAMA.cpp's huggingface style convention + # (Definitely not part of hugging face formal/informal standard... but with model_ appended to match their style) + ########################################### + use_model_card_metadata("name", "model_name") + use_model_card_metadata("author", "model_author") + use_model_card_metadata("version", "model_version") + use_model_card_metadata("organization", "model_organization") + use_model_card_metadata("description", "model_description") + use_model_card_metadata("finetune", "model_finetune") + use_model_card_metadata("basename", "model_basename") + use_model_card_metadata("size_label", "model_size_label") + use_model_card_metadata("source_url", "model_url") + use_model_card_metadata("source_doi", "model_doi") + use_model_card_metadata("source_uuid", "model_uuid") + use_model_card_metadata("source_repo_url", "model_repo_url") + + # Hugging Face Direct Convention + ################################# + + # Not part of huggingface model card standard but notice some model creator using it + # such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF' + use_model_card_metadata("name", "model_name") + use_model_card_metadata("author", "model_creator") + use_model_card_metadata("basename", "model_type") + + if "base_model" in model_card or "base_models" in model_card or "base_model_sources" in model_card: + # This represents the parent models that this is based on + # Example: stabilityai/stable-diffusion-xl-base-1.0. Can also be a list (for merges) + # Example of merges: https://huggingface.co/EmbeddedLLM/Mistral-7B-Merge-14-v0.1/blob/main/README.md + metadata_base_models = [] + base_model_value = model_card.get("base_model", model_card.get("base_models", model_card.get("base_model_sources", None))) + + if base_model_value is not None: + if isinstance(base_model_value, str): + metadata_base_models.append(base_model_value) + elif isinstance(base_model_value, list): + metadata_base_models.extend(base_model_value) + + if metadata.base_models is None: + metadata.base_models = [] + + for model_id in metadata_base_models: + # NOTE: model size of base model is assumed to be similar to the size of the current model + base_model = {} + if isinstance(model_id, str): + if model_id.startswith("http://") or model_id.startswith("https://") or model_id.startswith("ssh://"): + base_model["repo_url"] = model_id + + # Check if Hugging Face ID is present in URL + if "huggingface.co" in model_id: + match = re.match(r"https?://huggingface.co/([^/]+/[^/]+)$", model_id) + if match: + model_id_component = match.group(1) + model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id_component, total_params) + + # Populate model dictionary with extracted components + if model_full_name_component is not None: + base_model["name"] = Metadata.id_to_title(model_full_name_component) + if org_component is not None: + base_model["organization"] = Metadata.id_to_title(org_component) + if version is not None: + base_model["version"] = version + + else: + # Likely a Hugging Face ID + model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id, total_params) + + # Populate model dictionary with extracted components + if model_full_name_component is not None: + base_model["name"] = Metadata.id_to_title(model_full_name_component) + if org_component is not None: + base_model["organization"] = Metadata.id_to_title(org_component) + if version is not None: + base_model["version"] = version + if org_component is not None and model_full_name_component is not None: + base_model["repo_url"] = f"https://huggingface.co/{org_component}/{model_full_name_component}" + + elif isinstance(model_id, dict): + base_model = model_id + + else: + logger.error(f"base model entry '{str(model_id)}' not in a known format") + + metadata.base_models.append(base_model) + + if "datasets" in model_card or "dataset" in model_card or "dataset_sources" in model_card: + # This represents the datasets that this was trained from + metadata_datasets = [] + dataset_value = model_card.get("datasets", model_card.get("dataset", model_card.get("dataset_sources", None))) + + if dataset_value is not None: + if isinstance(dataset_value, str): + metadata_datasets.append(dataset_value) + elif isinstance(dataset_value, list): + metadata_datasets.extend(dataset_value) + + if metadata.datasets is None: + metadata.datasets = [] + + for dataset_id in metadata_datasets: + # NOTE: model size of base model is assumed to be similar to the size of the current model + dataset = {} + if isinstance(dataset_id, str): + if dataset_id.startswith(("http://", "https://", "ssh://")): + dataset["repo_url"] = dataset_id + + # Check if Hugging Face ID is present in URL + if "huggingface.co" in dataset_id: + match = re.match(r"https?://huggingface.co/([^/]+/[^/]+)$", dataset_id) + if match: + dataset_id_component = match.group(1) + dataset_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(dataset_id_component, total_params) + + # Populate dataset dictionary with extracted components + if dataset_name_component is not None: + dataset["name"] = Metadata.id_to_title(dataset_name_component) + if org_component is not None: + dataset["organization"] = Metadata.id_to_title(org_component) + if version is not None: + dataset["version"] = version + + else: + # Likely a Hugging Face ID + dataset_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(dataset_id, total_params) + + # Populate dataset dictionary with extracted components + if dataset_name_component is not None: + dataset["name"] = Metadata.id_to_title(dataset_name_component) + if org_component is not None: + dataset["organization"] = Metadata.id_to_title(org_component) + if version is not None: + dataset["version"] = version + if org_component is not None and dataset_name_component is not None: + dataset["repo_url"] = f"https://huggingface.co/{org_component}/{dataset_name_component}" + + elif isinstance(dataset_id, dict): + dataset = dataset_id + + else: + logger.error(f"dataset entry '{str(dataset_id)}' not in a known format") + + metadata.datasets.append(dataset) + + use_model_card_metadata("license", "license") + use_model_card_metadata("license_name", "license_name") + use_model_card_metadata("license_link", "license_link") + + use_array_model_card_metadata("tags", "tags") + use_array_model_card_metadata("tags", "pipeline_tag") + + use_array_model_card_metadata("languages", "languages") + use_array_model_card_metadata("languages", "language") + + # Hugging Face Parameter Heuristics + #################################### + + if hf_params is not None: + + hf_name_or_path = hf_params.get("_name_or_path") + if hf_name_or_path is not None and hf_name_or_path.count('/') <= 1: + # Use _name_or_path only if its actually a model name and not some computer path + # e.g. 'meta-llama/Llama-2-7b-hf' + model_id = hf_name_or_path + model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id, total_params) + if metadata.name is None and model_full_name_component is not None: + metadata.name = Metadata.id_to_title(model_full_name_component) + if metadata.organization is None and org_component is not None: + metadata.organization = Metadata.id_to_title(org_component) + if metadata.basename is None and basename is not None: + metadata.basename = basename + if metadata.finetune is None and finetune is not None: + metadata.finetune = finetune + if metadata.version is None and version is not None: + metadata.version = version + if metadata.size_label is None and size_label is not None: + metadata.size_label = size_label + + # Directory Folder Name Fallback Heuristics + ############################################ + if model_path is not None: + model_id = model_path.name + model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id, total_params) + if metadata.name is None and model_full_name_component is not None: + metadata.name = Metadata.id_to_title(model_full_name_component) + if metadata.organization is None and org_component is not None: + metadata.organization = Metadata.id_to_title(org_component) + if metadata.basename is None and basename is not None: + metadata.basename = basename + if metadata.finetune is None and finetune is not None: + metadata.finetune = finetune + if metadata.version is None and version is not None: + metadata.version = version + if metadata.size_label is None and size_label is not None: + metadata.size_label = size_label + + return metadata + + def set_gguf_meta_model(self, gguf_writer: gguf.GGUFWriter): + assert self.name is not None + gguf_writer.add_name(self.name) + + if self.author is not None: + gguf_writer.add_author(self.author) + if self.version is not None: + gguf_writer.add_version(self.version) + if self.organization is not None: + gguf_writer.add_organization(self.organization) + + if self.finetune is not None: + gguf_writer.add_finetune(self.finetune) + if self.basename is not None: + gguf_writer.add_basename(self.basename) + + if self.description is not None: + gguf_writer.add_description(self.description) + if self.quantized_by is not None: + gguf_writer.add_quantized_by(self.quantized_by) + + if self.size_label is not None: + gguf_writer.add_size_label(self.size_label) + + if self.license is not None: + if isinstance(self.license, list): + gguf_writer.add_license(",".join(self.license)) + else: + gguf_writer.add_license(self.license) + if self.license_name is not None: + gguf_writer.add_license_name(self.license_name) + if self.license_link is not None: + gguf_writer.add_license_link(self.license_link) + + if self.url is not None: + gguf_writer.add_url(self.url) + if self.doi is not None: + gguf_writer.add_doi(self.doi) + if self.uuid is not None: + gguf_writer.add_uuid(self.uuid) + if self.repo_url is not None: + gguf_writer.add_repo_url(self.repo_url) + + if self.source_url is not None: + gguf_writer.add_source_url(self.source_url) + if self.source_doi is not None: + gguf_writer.add_source_doi(self.source_doi) + if self.source_uuid is not None: + gguf_writer.add_source_uuid(self.source_uuid) + if self.source_repo_url is not None: + gguf_writer.add_source_repo_url(self.source_repo_url) + + if self.base_models is not None: + gguf_writer.add_base_model_count(len(self.base_models)) + for key, base_model_entry in enumerate(self.base_models): + if "name" in base_model_entry: + gguf_writer.add_base_model_name(key, base_model_entry["name"]) + if "author" in base_model_entry: + gguf_writer.add_base_model_author(key, base_model_entry["author"]) + if "version" in base_model_entry: + gguf_writer.add_base_model_version(key, base_model_entry["version"]) + if "organization" in base_model_entry: + gguf_writer.add_base_model_organization(key, base_model_entry["organization"]) + if "description" in base_model_entry: + gguf_writer.add_base_model_description(key, base_model_entry["description"]) + if "url" in base_model_entry: + gguf_writer.add_base_model_url(key, base_model_entry["url"]) + if "doi" in base_model_entry: + gguf_writer.add_base_model_doi(key, base_model_entry["doi"]) + if "uuid" in base_model_entry: + gguf_writer.add_base_model_uuid(key, base_model_entry["uuid"]) + if "repo_url" in base_model_entry: + gguf_writer.add_base_model_repo_url(key, base_model_entry["repo_url"]) + + if self.datasets is not None: + gguf_writer.add_dataset_count(len(self.datasets)) + for key, dataset_entry in enumerate(self.datasets): + if "name" in dataset_entry: + gguf_writer.add_dataset_name(key, dataset_entry["name"]) + if "author" in dataset_entry: + gguf_writer.add_dataset_author(key, dataset_entry["author"]) + if "version" in dataset_entry: + gguf_writer.add_dataset_version(key, dataset_entry["version"]) + if "organization" in dataset_entry: + gguf_writer.add_dataset_organization(key, dataset_entry["organization"]) + if "description" in dataset_entry: + gguf_writer.add_dataset_description(key, dataset_entry["description"]) + if "url" in dataset_entry: + gguf_writer.add_dataset_url(key, dataset_entry["url"]) + if "doi" in dataset_entry: + gguf_writer.add_dataset_doi(key, dataset_entry["doi"]) + if "uuid" in dataset_entry: + gguf_writer.add_dataset_uuid(key, dataset_entry["uuid"]) + if "repo_url" in dataset_entry: + gguf_writer.add_dataset_repo_url(key, dataset_entry["repo_url"]) + + if self.tags is not None: + gguf_writer.add_tags(self.tags) + if self.languages is not None: + gguf_writer.add_languages(self.languages) diff --git a/venv/lib/python3.10/site-packages/gguf/py.typed b/venv/lib/python3.10/site-packages/gguf/py.typed new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/venv/lib/python3.10/site-packages/gguf/quants.py b/venv/lib/python3.10/site-packages/gguf/quants.py new file mode 100644 index 0000000000000000000000000000000000000000..3c8ba82e19d3d9e984ba39caf5cf865b0ee8e72a --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/quants.py @@ -0,0 +1,1269 @@ +from __future__ import annotations +from abc import ABC, abstractmethod +from typing import Any, Callable, Sequence +from math import log2, ceil + +from numpy.typing import DTypeLike + +from .constants import GGML_QUANT_SIZES, GGMLQuantizationType, QK_K +from .lazy import LazyNumpyTensor + +import numpy as np + + +def quant_shape_to_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType) -> tuple[int, ...]: + block_size, type_size = GGML_QUANT_SIZES[quant_type] + if shape[-1] % block_size != 0: + raise ValueError(f"Quantized tensor row size ({shape[-1]}) is not a multiple of {quant_type.name} block size ({block_size})") + return (*shape[:-1], shape[-1] // block_size * type_size) + + +def quant_shape_from_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType) -> tuple[int, ...]: + block_size, type_size = GGML_QUANT_SIZES[quant_type] + if shape[-1] % type_size != 0: + raise ValueError(f"Quantized tensor bytes per row ({shape[-1]}) is not a multiple of {quant_type.name} type size ({type_size})") + return (*shape[:-1], shape[-1] // type_size * block_size) + + +# This is faster than np.vectorize and np.apply_along_axis because it works on more than one row at a time +def _apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.ndarray, otype: DTypeLike, oshape: tuple[int, ...]) -> np.ndarray: + rows = arr.reshape((-1, arr.shape[-1])) + osize = 1 + for dim in oshape: + osize *= dim + out = np.empty(shape=osize, dtype=otype) + # compute over groups of 16 rows (arbitrary, but seems good for performance) + n_groups = (rows.shape[0] // 16) or 1 + np.concatenate([func(group).ravel() for group in np.array_split(rows, n_groups)], axis=0, out=out) + return out.reshape(oshape) + + +# round away from zero +# ref: https://stackoverflow.com/a/59143326/22827863 +def np_roundf(n: np.ndarray) -> np.ndarray: + a = abs(n) + floored = np.floor(a) + b = floored + np.floor(2 * (a - floored)) + return np.sign(n) * b + + +class QuantError(Exception): ... + + +_type_traits: dict[GGMLQuantizationType, type[__Quant]] = {} + + +def quantize(data: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray: + if qtype == GGMLQuantizationType.F32: + return data.astype(np.float32, copy=False) + elif qtype == GGMLQuantizationType.F16: + return data.astype(np.float16, copy=False) + elif (q := _type_traits.get(qtype)) is not None: + return q.quantize(data) + else: + raise NotImplementedError(f"Quantization for {qtype.name} is not yet implemented") + + +def dequantize(data: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray: + if qtype == GGMLQuantizationType.F32: + return data.view(np.float32) + elif qtype == GGMLQuantizationType.F16: + return data.view(np.float16).astype(np.float32) + elif (q := _type_traits.get(qtype)) is not None: + return q.dequantize(data) + else: + raise NotImplementedError(f"Dequantization for {qtype.name} is not yet implemented") + + +class __Quant(ABC): + qtype: GGMLQuantizationType + block_size: int + type_size: int + + grid: np.ndarray[Any, np.dtype[np.float32]] | None = None + grid_shape: tuple[int, int] = (0, 0) + grid_map: tuple[int | float, ...] = () + grid_hex: bytes | None = None + + def __init__(self): + return TypeError("Quant conversion classes can't have instances") + + def __init_subclass__(cls, qtype: GGMLQuantizationType) -> None: + cls.qtype = qtype + cls.block_size, cls.type_size = GGML_QUANT_SIZES[qtype] + cls.__quantize_lazy = LazyNumpyTensor._wrap_fn( + cls.__quantize_array, + meta_noop=(np.uint8, cls.__shape_to_bytes) + ) + cls.__dequantize_lazy = LazyNumpyTensor._wrap_fn( + cls.__dequantize_array, + meta_noop=(np.float32, cls.__shape_from_bytes) + ) + assert qtype not in _type_traits + _type_traits[qtype] = cls + + @classmethod + def init_grid(cls): + if cls.grid is not None or cls.grid_hex is None: + return + + bits_per_elem = ceil(log2(len(cls.grid_map))) + assert bits_per_elem != 0, cls.qtype.name + elems_per_byte = 8 // bits_per_elem + + grid = np.frombuffer(cls.grid_hex, dtype=np.uint8) + # decode hexadecimal chars from grid + grid = grid.reshape((-1, 2)) + grid = (np.where(grid > 0x40, grid + 9, grid) & 0x0F) << np.array([4, 0], dtype=np.uint8).reshape((1, 2)) + grid = grid[..., 0] | grid[..., 1] + # unpack the grid values + grid = grid.reshape((-1, 1)) >> np.array([i for i in range(0, 8, 8 // elems_per_byte)], dtype=np.uint8).reshape((1, elems_per_byte)) + grid = (grid & ((1 << bits_per_elem) - 1)).reshape((-1, 1)) + grid_map = np.array(cls.grid_map, dtype=np.float32).reshape((1, -1)) + grid = np.take_along_axis(grid_map, grid, axis=-1) + cls.grid = grid.reshape((1, 1, *cls.grid_shape)) + + @classmethod + @abstractmethod + def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + raise NotImplementedError + + @classmethod + @abstractmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + raise NotImplementedError + + @classmethod + def quantize_rows(cls, rows: np.ndarray) -> np.ndarray: + rows = rows.astype(np.float32, copy=False) + shape = rows.shape + n_blocks = rows.size // cls.block_size + blocks = rows.reshape((n_blocks, cls.block_size)) + blocks = cls.quantize_blocks(blocks) + assert blocks.dtype == np.uint8 + assert blocks.shape[-1] == cls.type_size + return blocks.reshape(cls.__shape_to_bytes(shape)) + + @classmethod + def dequantize_rows(cls, rows: np.ndarray) -> np.ndarray: + rows = rows.view(np.uint8) + shape = rows.shape + n_blocks = rows.size // cls.type_size + blocks = rows.reshape((n_blocks, cls.type_size)) + blocks = cls.dequantize_blocks(blocks) + assert blocks.dtype == np.float32 + assert blocks.shape[-1] == cls.block_size + return blocks.reshape(cls.__shape_from_bytes(shape)) + + @classmethod + def __shape_to_bytes(cls, shape: Sequence[int]): + return quant_shape_to_byte_shape(shape, cls.qtype) + + @classmethod + def __shape_from_bytes(cls, shape: Sequence[int]): + return quant_shape_from_byte_shape(shape, cls.qtype) + + @classmethod + def __quantize_array(cls, array: np.ndarray) -> np.ndarray: + return _apply_over_grouped_rows(cls.quantize_rows, arr=array, otype=np.uint8, oshape=cls.__shape_to_bytes(array.shape)) + + @classmethod + def __dequantize_array(cls, array: np.ndarray) -> np.ndarray: + cls.init_grid() + return _apply_over_grouped_rows(cls.dequantize_rows, arr=array, otype=np.float32, oshape=cls.__shape_from_bytes(array.shape)) + + @classmethod + def __quantize_lazy(cls, lazy_tensor: LazyNumpyTensor, /) -> Any: + pass + + @classmethod + def __dequantize_lazy(cls, lazy_tensor: LazyNumpyTensor, /) -> Any: + pass + + @classmethod + def can_quantize(cls, tensor: np.ndarray | LazyNumpyTensor) -> bool: + return tensor.shape[-1] % cls.block_size == 0 + + @classmethod + def quantize(cls, tensor: np.ndarray | LazyNumpyTensor) -> np.ndarray: + if not cls.can_quantize(tensor): + raise QuantError(f"Can't quantize tensor with shape {tensor.shape} to {cls.qtype.name}") + if isinstance(tensor, LazyNumpyTensor): + return cls.__quantize_lazy(tensor) + else: + return cls.__quantize_array(tensor) + + @classmethod + def dequantize(cls, tensor: np.ndarray | LazyNumpyTensor) -> np.ndarray: + if isinstance(tensor, LazyNumpyTensor): + return cls.__dequantize_lazy(tensor) + else: + return cls.__dequantize_array(tensor) + + +class BF16(__Quant, qtype=GGMLQuantizationType.BF16): + @classmethod + # same as ggml_compute_fp32_to_bf16 in ggml-impl.h + def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n = blocks.view(np.uint32) + # force nan to quiet + n = np.where((n & 0x7fffffff) > 0x7f800000, (n & np.uint32(0xffff0000)) | np.uint32(64 << 16), n) + # round to nearest even + n = (np.uint64(n) + (0x7fff + ((n >> 16) & 1))) >> 16 + return n.astype(np.uint16).view(np.uint8) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + return (blocks.view(np.int16).astype(np.int32) << 16).view(np.float32) + + +class Q4_0(__Quant, qtype=GGMLQuantizationType.Q4_0): + @classmethod + def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + imax = abs(blocks).argmax(axis=-1, keepdims=True) + max = np.take_along_axis(blocks, imax, axis=-1) + + d = max / -8 + with np.errstate(divide="ignore"): + id = np.where(d == 0, 0, 1 / d) + # FIXME: Q4_0's reference rounding is cursed and depends on FMA + qs = np.trunc((np.float64(blocks) * np.float64(id)) + np.float64(8.5), dtype=np.float32).astype(np.uint8).clip(0, 15) + + qs = qs.reshape((n_blocks, 2, cls.block_size // 2)) + qs = qs[..., 0, :] | (qs[..., 1, :] << np.uint8(4)) + + d = d.astype(np.float16).view(np.uint8) + + return np.concatenate([d, qs], axis=-1) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d, qs = np.hsplit(blocks, [2]) + + d = d.view(np.float16).astype(np.float32) + + qs = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1)) + qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1)).astype(np.int8) - np.int8(8) + + return (d * qs.astype(np.float32)) + + +class Q4_1(__Quant, qtype=GGMLQuantizationType.Q4_1): + @classmethod + def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + max = blocks.max(axis=-1, keepdims=True) + min = blocks.min(axis=-1, keepdims=True) + + d = (max - min) / 15 + with np.errstate(divide="ignore"): + id = np.where(d == 0, 0, 1 / d) + qs = np.trunc((blocks - min) * id + np.float32(0.5), dtype=np.float32).astype(np.uint8).clip(0, 15) + + qs = qs.reshape((n_blocks, 2, cls.block_size // 2)) + qs = qs[..., 0, :] | (qs[..., 1, :] << np.uint8(4)) + + d = d.astype(np.float16).view(np.uint8) + m = min.astype(np.float16).view(np.uint8) + + return np.concatenate([d, m, qs], axis=-1) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d, rest = np.hsplit(blocks, [2]) + m, qs = np.hsplit(rest, [2]) + + d = d.view(np.float16).astype(np.float32) + m = m.view(np.float16).astype(np.float32) + + qs = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1)) + qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1)).astype(np.float32) + + return (d * qs) + m + + +class Q5_0(__Quant, qtype=GGMLQuantizationType.Q5_0): + @classmethod + def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + imax = abs(blocks).argmax(axis=-1, keepdims=True) + max = np.take_along_axis(blocks, imax, axis=-1) + + d = max / -16 + with np.errstate(divide="ignore"): + id = np.where(d == 0, 0, 1 / d) + # FIXME: Q5_0's reference rounding is cursed and depends on FMA + q = np.trunc((np.float64(blocks) * np.float64(id)) + np.float64(16.5), dtype=np.float32).astype(np.uint8).clip(0, 31) + + qs = q.reshape((n_blocks, 2, cls.block_size // 2)) + qs = (qs[..., 0, :] & np.uint8(0x0F)) | (qs[..., 1, :] << np.uint8(4)) + + qh = np.packbits(q.reshape((n_blocks, 1, 32)) >> np.uint8(4), axis=-1, bitorder="little").reshape(n_blocks, 4) + + d = d.astype(np.float16).view(np.uint8) + + return np.concatenate([d, qh, qs], axis=-1) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d, rest = np.hsplit(blocks, [2]) + qh, qs = np.hsplit(rest, [4]) + + d = d.view(np.float16).astype(np.float32) + qh = qh.view(np.uint32) + + qh = qh.reshape((n_blocks, 1)) >> np.array([i for i in range(32)], dtype=np.uint32).reshape((1, 32)) + ql = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1)) + qh = (qh & np.uint32(0x01)).astype(np.uint8) + ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1)) + + qs = (ql | (qh << np.uint8(4))).astype(np.int8) - np.int8(16) + + return (d * qs.astype(np.float32)) + + +class Q5_1(__Quant, qtype=GGMLQuantizationType.Q5_1): + @classmethod + def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + max = blocks.max(axis=-1, keepdims=True) + min = blocks.min(axis=-1, keepdims=True) + + d = (max - min) / 31 + with np.errstate(divide="ignore"): + id = np.where(d == 0, 0, 1 / d) + q = np.trunc((blocks - min) * id + np.float32(0.5), dtype=np.float32).astype(np.uint8).clip(0, 31) + + qs = q.reshape((n_blocks, 2, cls.block_size // 2)) + qs = (qs[..., 0, :] & np.uint8(0x0F)) | (qs[..., 1, :] << np.uint8(4)) + + qh = np.packbits(q.reshape((n_blocks, 1, 32)) >> np.uint8(4), axis=-1, bitorder="little").reshape(n_blocks, 4) + + d = d.astype(np.float16).view(np.uint8) + m = min.astype(np.float16).view(np.uint8) + + return np.concatenate([d, m, qh, qs], axis=-1) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d, rest = np.hsplit(blocks, [2]) + m, rest = np.hsplit(rest, [2]) + qh, qs = np.hsplit(rest, [4]) + + d = d.view(np.float16).astype(np.float32) + m = m.view(np.float16).astype(np.float32) + qh = qh.view(np.uint32) + + qh = qh.reshape((n_blocks, 1)) >> np.array([i for i in range(32)], dtype=np.uint32).reshape((1, 32)) + ql = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1)) + qh = (qh & np.uint32(0x01)).astype(np.uint8) + ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1)) + + qs = (ql | (qh << np.uint8(4))).astype(np.float32) + + return (d * qs) + m + + +class Q8_0(__Quant, qtype=GGMLQuantizationType.Q8_0): + @classmethod + # Implementation of Q8_0 with bit-exact same results as reference implementation in ggml-quants.c + def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + + d = abs(blocks).max(axis=1, keepdims=True) / 127 + with np.errstate(divide="ignore"): + id = np.where(d == 0, 0, 1 / d) + qs = np_roundf(blocks * id) + + # (n_blocks, 2) + d = d.astype(np.float16).view(np.uint8) + # (n_blocks, block_size) + qs = qs.astype(np.int8).view(np.uint8) + + return np.concatenate([d, qs], axis=1) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + d, x = np.split(blocks, [2], axis=1) + d = d.view(np.float16).astype(np.float32) + x = x.view(np.int8).astype(np.float32) + + return (x * d) + + +class Q2_K(__Quant, qtype=GGMLQuantizationType.Q2_K): + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + scales, rest = np.hsplit(blocks, [QK_K // 16]) + qs, rest = np.hsplit(rest, [QK_K // 4]) + d, dmin = np.hsplit(rest, [2]) + + d = d.view(np.float16).astype(np.float32) + dmin = dmin.view(np.float16).astype(np.float32) + + # (n_blocks, 16, 1) + dl = (d * (scales & 0xF).astype(np.float32)).reshape((n_blocks, QK_K // 16, 1)) + ml = (dmin * (scales >> 4).astype(np.float32)).reshape((n_blocks, QK_K // 16, 1)) + + shift = np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1)) + + qs = (qs.reshape((n_blocks, -1, 1, 32)) >> shift) & np.uint8(3) + + qs = qs.reshape((n_blocks, QK_K // 16, 16)).astype(np.float32) + + qs = dl * qs - ml + + return qs.reshape((n_blocks, -1)) + + +class Q3_K(__Quant, qtype=GGMLQuantizationType.Q3_K): + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + hmask, rest = np.hsplit(blocks, [QK_K // 8]) + qs, rest = np.hsplit(rest, [QK_K // 4]) + scales, d = np.hsplit(rest, [12]) + + d = d.view(np.float16).astype(np.float32) + + # The scales are packed at 6-bit each in this pattern: + # 0: IIIIAAAA + # 1: JJJJBBBB + # 2: KKKKCCCC + # 3: LLLLDDDD + # 4: MMMMEEEE + # 5: NNNNFFFF + # 6: OOOOGGGG + # 7: PPPPHHHH + # 8: MMIIEEAA + # 9: NNJJFFBB + # 10: OOKKGGCC + # 11: PPLLHHDD + lscales, hscales = np.hsplit(scales, [8]) + lscales = lscales.reshape((n_blocks, 1, 8)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 2, 1)) + lscales = lscales.reshape((n_blocks, 16)) + hscales = hscales.reshape((n_blocks, 1, 4)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 4, 1)) + hscales = hscales.reshape((n_blocks, 16)) + scales = (lscales & np.uint8(0x0F)) | ((hscales & np.uint8(0x03)) << np.uint8(4)) + scales = (scales.astype(np.int8) - np.int8(32)).astype(np.float32) + + dl = (d * scales).reshape((n_blocks, 16, 1)) + + ql = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1)) + qh = hmask.reshape(n_blocks, -1, 1, 32) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8, 1)) + ql = ql.reshape((n_blocks, 16, QK_K // 16)) & np.uint8(3) + qh = (qh.reshape((n_blocks, 16, QK_K // 16)) & np.uint8(1)) + qh = qh ^ np.uint8(1) # strangely, the offset is zero when the bitmask is 1 + q = (ql.astype(np.int8) - (qh << np.uint8(2)).astype(np.int8)).astype(np.float32) + + return (dl * q).reshape((n_blocks, QK_K)) + + +class Q4_K(__Quant, qtype=GGMLQuantizationType.Q4_K): + K_SCALE_SIZE = 12 + + @staticmethod + def get_scale_min(scales: np.ndarray) -> tuple[np.ndarray, np.ndarray]: + n_blocks = scales.shape[0] + scales = scales.view(np.uint8) + ### Unpacking the following: ### + # 0 EEAAAAAA + # 1 FFBBBBBB + # 2 GGCCCCCC + # 3 HHDDDDDD + # 4 eeaaaaaa + # 5 ffbbbbbb + # 6 ggcccccc + # 7 hhdddddd + # 8 eeeeEEEE + # 9 ffffFFFF + # 10 ggggGGGG + # 11 hhhhHHHH + scales = scales.reshape((n_blocks, 3, 4)) + d, m, m_d = np.split(scales, 3, axis=-2) + + sc = np.concatenate([d & 0x3F, (m_d & 0x0F) | ((d >> 2) & 0x30)], axis=-1) + min = np.concatenate([m & 0x3F, (m_d >> 4) | ((m >> 2) & 0x30)], axis=-1) + + return (sc.reshape((n_blocks, 8)), min.reshape((n_blocks, 8))) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d, rest = np.hsplit(blocks, [2]) + dmin, rest = np.hsplit(rest, [2]) + scales, qs = np.hsplit(rest, [cls.K_SCALE_SIZE]) + + d = d.view(np.float16).astype(np.float32) + dmin = dmin.view(np.float16).astype(np.float32) + + sc, m = Q4_K.get_scale_min(scales) + + d = (d * sc.astype(np.float32)).reshape((n_blocks, -1, 1)) + dm = (dmin * m.astype(np.float32)).reshape((n_blocks, -1, 1)) + + qs = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1)) + qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1, 32)).astype(np.float32) + + return (d * qs - dm).reshape((n_blocks, QK_K)) + + +class Q5_K(__Quant, qtype=GGMLQuantizationType.Q5_K): + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d, rest = np.hsplit(blocks, [2]) + dmin, rest = np.hsplit(rest, [2]) + scales, rest = np.hsplit(rest, [Q4_K.K_SCALE_SIZE]) + qh, qs = np.hsplit(rest, [QK_K // 8]) + + d = d.view(np.float16).astype(np.float32) + dmin = dmin.view(np.float16).astype(np.float32) + + sc, m = Q4_K.get_scale_min(scales) + + d = (d * sc.astype(np.float32)).reshape((n_blocks, -1, 1)) + dm = (dmin * m.astype(np.float32)).reshape((n_blocks, -1, 1)) + + ql = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1)) + qh = qh.reshape((n_blocks, -1, 1, 32)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8, 1)) + ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1, 32)) + qh = (qh & np.uint8(0x01)).reshape((n_blocks, -1, 32)) + q = (ql | (qh << np.uint8(4))).astype(np.float32) + + return (d * q - dm).reshape((n_blocks, QK_K)) + + +class Q6_K(__Quant, qtype=GGMLQuantizationType.Q6_K): + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + ql, rest = np.hsplit(blocks, [QK_K // 2]) + qh, rest = np.hsplit(rest, [QK_K // 4]) + scales, d = np.hsplit(rest, [QK_K // 16]) + + scales = scales.view(np.int8).astype(np.float32) + d = d.view(np.float16).astype(np.float32) + d = (d * scales).reshape((n_blocks, QK_K // 16, 1)) + + ql = ql.reshape((n_blocks, -1, 1, 64)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1)) + ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1, 32)) + qh = qh.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1)) + qh = (qh & np.uint8(0x03)).reshape((n_blocks, -1, 32)) + q = (ql | (qh << np.uint8(4))).astype(np.int8) - np.int8(32) + q = q.reshape((n_blocks, QK_K // 16, -1)).astype(np.float32) + + return (d * q).reshape((n_blocks, QK_K)) + + +class TQ1_0(__Quant, qtype=GGMLQuantizationType.TQ1_0): + @classmethod + def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d = abs(blocks).max(axis=-1, keepdims=True) + with np.errstate(divide="ignore"): + id = np.where(d == 0, 0, 1 / d) + qs = np_roundf(blocks * id) + qs = (qs.astype(np.int8) + np.int8(1)).astype(np.uint8) + + qs0, qs1, qh = qs[..., :(32 * 5)], qs[..., (32 * 5):(48 * 5)], qs[..., (48 * 5):] + qs0 = qs0.reshape((n_blocks, -1, 5, 32)) * np.array([81, 27, 9, 3, 1], dtype=np.uint8).reshape((1, 1, 5, 1)) + qs0 = np.sum(qs0, axis=-2).reshape((n_blocks, -1)) + qs1 = qs1.reshape((n_blocks, -1, 5, 16)) * np.array([81, 27, 9, 3, 1], dtype=np.uint8).reshape((1, 1, 5, 1)) + qs1 = np.sum(qs1, axis=-2).reshape((n_blocks, -1)) + qh = qh.reshape((n_blocks, -1, 4, 4)) * np.array([81, 27, 9, 3], dtype=np.uint8).reshape((1, 1, 4, 1)) + qh = np.sum(qh, axis=-2).reshape((n_blocks, -1)) + qs = np.concatenate([qs0, qs1, qh], axis=-1) + qs = (qs.astype(np.uint16) * 256 + (243 - 1)) // 243 + + qs = qs.astype(np.uint8) + d = d.astype(np.float16).view(np.uint8) + + return np.concatenate([qs, d], axis=-1) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + qs, rest = np.hsplit(blocks, [(QK_K - 4 * QK_K // 64) // 5]) + qh, d = np.hsplit(rest, [QK_K // 64]) + + d = d.view(np.float16).astype(np.float32) + + qs0, qs1 = qs[..., :32], qs[..., 32:] + qs0 = qs0.reshape((n_blocks, -1, 1, 32)) * np.array([1, 3, 9, 27, 81], dtype=np.uint8).reshape((1, 1, 5, 1)) + qs0 = qs0.reshape((n_blocks, -1)) + qs1 = qs1.reshape((n_blocks, -1, 1, 16)) * np.array([1, 3, 9, 27, 81], dtype=np.uint8).reshape((1, 1, 5, 1)) + qs1 = qs1.reshape((n_blocks, -1)) + qh = qh.reshape((n_blocks, -1, 1, 4)) * np.array([1, 3, 9, 27], dtype=np.uint8).reshape((1, 1, 4, 1)) + qh = qh.reshape((n_blocks, -1)) + qs = np.concatenate([qs0, qs1, qh], axis=-1) + qs = ((qs.astype(np.uint16) * 3) >> 8).astype(np.int8) - np.int8(1) + + return (d * qs.astype(np.float32)) + + +class TQ2_0(__Quant, qtype=GGMLQuantizationType.TQ2_0): + @classmethod + def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d = abs(blocks).max(axis=-1, keepdims=True) + with np.errstate(divide="ignore"): + id = np.where(d == 0, 0, 1 / d) + qs = np_roundf(blocks * id) + qs = (qs.astype(np.int8) + np.int8(1)).astype(np.uint8) + + qs = qs.reshape((n_blocks, -1, 4, 32)) << np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1)) + qs = qs[..., 0, :] | qs[..., 1, :] | qs[..., 2, :] | qs[..., 3, :] + qs = qs.reshape((n_blocks, -1)) + + d = d.astype(np.float16).view(np.uint8) + + return np.concatenate([qs, d], axis=-1) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + qs, d = np.hsplit(blocks, [QK_K // 4]) + + d = d.view(np.float16).astype(np.float32) + + qs = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1)) + qs = (qs & 0x03).reshape((n_blocks, -1)).astype(np.int8) - np.int8(1) + + return (d * qs.astype(np.float32)) + + +class IQ2_XXS(__Quant, qtype=GGMLQuantizationType.IQ2_XXS): + ksigns: bytes = ( + b"\x00\x81\x82\x03\x84\x05\x06\x87\x88\x09\x0a\x8b\x0c\x8d\x8e\x0f" + b"\x90\x11\x12\x93\x14\x95\x96\x17\x18\x99\x9a\x1b\x9c\x1d\x1e\x9f" + b"\xa0\x21\x22\xa3\x24\xa5\xa6\x27\x28\xa9\xaa\x2b\xac\x2d\x2e\xaf" + b"\x30\xb1\xb2\x33\xb4\x35\x36\xb7\xb8\x39\x3a\xbb\x3c\xbd\xbe\x3f" + b"\xc0\x41\x42\xc3\x44\xc5\xc6\x47\x48\xc9\xca\x4b\xcc\x4d\x4e\xcf" + b"\x50\xd1\xd2\x53\xd4\x55\x56\xd7\xd8\x59\x5a\xdb\x5c\xdd\xde\x5f" + b"\x60\xe1\xe2\x63\xe4\x65\x66\xe7\xe8\x69\x6a\xeb\x6c\xed\xee\x6f" + b"\xf0\x71\x72\xf3\x74\xf5\xf6\x77\x78\xf9\xfa\x7b\xfc\x7d\x7e\xff" + ) + + # iq2xxs_grid, but with each byte of the original packed in 2 bits, + # by mapping 0x08 to 0, 0x19 to 1, and 0x2b to 2. + grid_shape = (256, 8) + grid_map = (0x08, 0x19, 0x2b) + grid_hex = ( + b"00000200050008000a00110014002000220028002a0041004400500058006100" + b"6400800082008a00a20001010401100115014001840198010002020222028202" + b"010404041004210424044004420448046004810484049004a404000502050805" + b"200546056905800591050906100640068406a406000805080808140828084108" + b"440850085208880804094009020a140a01100410101021104010601084109010" + b"951000110811201150115a118011241245120014081420142514491480141815" + b"6215001616160118041810184018811800190519a019511a002002200a204420" + b"6120802082202921482100220222012404241024402456240025412564259026" + b"082820289428442a014004401040184021402440404048405640604081408440" + b"9040004120416141804185410142104248425642684200440844204480449944" + b"124524450046014804481048404845480049584961498249454a904a00500850" + b"1150195020508050885004514251a4519152905492540a550156545600581158" + b"195864584059085a046010604060686000615561186260620064056410651265" + b"84654268008002800a8041808280048118814081118201840484108415844084" + b"608400854685948509864086608602880489118a0490109024904090a1901691" + b"8091459200942294449451958198209902a050a085a009a100a218a450a804a9" + ) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d, qs = np.hsplit(blocks, [2]) + + d = d.view(np.float16).astype(np.float32) + + qs = qs.view(np.uint32).reshape(n_blocks, -1, 2) + + db = d * (np.float32(0.5) + (qs[..., 1] >> 28).astype(np.float32)) * np.float32(0.25) + db = db.reshape((n_blocks, -1, 1, 1)) + + # get the sign indices and unpack the bits + signs = qs[..., 1].reshape((n_blocks, -1, 1)) >> np.array([0, 7, 14, 21], dtype=np.uint32).reshape((1, 1, 4)) + ksigns = np.frombuffer(cls.ksigns, dtype=np.uint8).reshape((1, 1, 1, 128)) + signs = (signs & np.uint32(0x7F)).reshape((n_blocks, -1, 4, 1)) + signs = np.take_along_axis(ksigns, signs, axis=-1) + signs = signs.reshape((n_blocks, -1, 4, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 1, 8)) + signs = signs & np.uint8(0x01) + signs = np.where(signs == 0, np.float32(1), np.float32(-1)) + signs = signs.reshape((n_blocks, -1, 4, 8)) + + assert cls.grid is not None + grid = np.take_along_axis(cls.grid, qs[..., 0].copy().view(np.uint8).reshape((n_blocks, -1, 1, 1)), axis=-2) + grid = grid.reshape((n_blocks, -1, 4, 8)) + + return (db * grid * signs).reshape((n_blocks, -1)) + + +class IQ2_XS(__Quant, qtype=GGMLQuantizationType.IQ2_XS): + # iq2xs_grid, but with each byte of the original packed in 2 bits, + # by mapping 0x08 to 0, 0x19 to 1, and 0x2b to 2. + grid_shape = (512, 8) + grid_map = (0x08, 0x19, 0x2b) + grid_hex = ( + b"00000200050008000a0011001400160019002000220025002800410044004600" + b"49005000520055005800610064008000820085008800910094009900a0000101" + b"04010601090110011201150118011a0121012401400142014501480151015401" + b"6001680181018401900100020202050208021102140220024102440250025502" + b"80028a0201040404060409041004120415041804210424044004420445044804" + b"5104540456046004810484049004000502050505080511051405200541054405" + b"500561058005010604061006260640064206840600080208050808080a081108" + b"14082008250841084408500858088008a008aa08010904091009400981098909" + b"000a200a280a960aa00a01100410061009101010121015101810211024104010" + b"4210451048105110541060106a10811084109010001102110511081111111411" + b"2011411144115011801194119611011204120612101240126012001402140514" + b"0814111414142014411444144914501464148014011504151015401500161416" + b"49160118041810181218401854188618001905196619511aa91a002002200520" + b"08200a201120142020204120442050208020a020012104211021402148216521" + b"002222228022a82201240424102429244024002541255225992501261a26a626" + b"002808280a28202855288828a22868299029082a202a822a882a8a2a01400440" + b"0640094010401240154018402140244040404240454048404a40514054406040" + b"6540814084409040004102410541084111411441204141414441504180418541" + b"a241014204421042124229424042004402440544084411441444194420444144" + b"4444504480449444014504451045244540459a4500460a464446504601480448" + b"1048404845485448624800491149444950496949044a00500250055008501150" + b"145020502850415044505050805001510451105115514051425100524452aa52" + b"0154045410542154405460548154a154005508558055885521566856a1560058" + b"14584158505899581a5940594259855a0160046010604060546062608660a960" + b"006124624a62926200641664106540654565a46501686a682569066a546a626a" + b"00800280058008801180148020802a8041804480508080808280a880aa800181" + b"0481068110814081518159810082208280828282a082a8820184048410841284" + b"158440846084898400854485a58518866a860088088825885a8880888288a888" + b"0689228a808a888a968aa88a0190049010904090569084900091229164915692" + b"89920094059444945094589429959095929541965198a6984999159a609a00a0" + b"02a008a00aa020a02aa0a0a051a159a1a6a100a202a208a22aa280a2a0a240a4" + b"95a465a698a60aa820a822a828a8a0a8a8a804a984a986a928aa2aaa91aaaaaa" + ) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d, rest = np.hsplit(blocks, [2]) + qs, scales = np.hsplit(rest, [2 * QK_K // 8]) + + d = d.view(np.float16).astype(np.float32) + qs = qs.view(np.uint16) + + scales = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2)) + scales = (scales & 0x0F).reshape((n_blocks, -1)) + db = d * (np.float32(0.5) + scales) * np.float32(0.25) + db = db.reshape((n_blocks, -1, 1, 1)) + + # get the sign indices and unpack the bits + signs = np.frombuffer(IQ2_XXS.ksigns, dtype=np.uint8).reshape(1, 1, 128) + signs = np.take_along_axis(signs, (qs >> 9).reshape((n_blocks, -1, 1)), axis=-1) + signs = signs.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8)) + signs = signs & np.uint8(0x01) + signs = np.where(signs == 0, np.float32(1), np.float32(-1)) + signs = signs.reshape((n_blocks, -1, 2, 8)) + + assert cls.grid is not None + grid = np.take_along_axis(cls.grid, (qs & np.uint16(511)).reshape((n_blocks, -1, 1, 1)), axis=-2) + grid = grid.reshape((n_blocks, -1, 2, 8)) + + return (db * grid * signs).reshape((n_blocks, -1)) + + +class IQ2_S(__Quant, qtype=GGMLQuantizationType.IQ2_S): + # iq2s_grid, but with each byte of the original packed in 2 bits, + # by mapping 0x08 to 0, 0x19 to 1, and 0x2b to 2. + grid_shape = (1024, 8) + grid_map = (0x08, 0x19, 0x2b) + grid_hex = ( + b"00000200050008000a0011001400160019002000220025002800410044004600" + b"490050005200550058006100640066006900800082008500880091009400a000" + b"a500aa0001010401060109011001120115011801210124014001420145014801" + b"510154015601590160016501680181018401900192019501a101a40100020202" + b"050208021102140220022a02410244024602490250025502800285028a029402" + b"a202010404040604090410041204150418042104240426042904400442044504" + b"48044a0451045404560459046004620465048104840486048904900495049804" + b"a104a40400050205050508050a05110514051605190520052505280541054405" + b"46054905500552055505580561056405800582058505880591059405a0050106" + b"0406060609061006150640064506480651065406600681068406900600080208" + b"050808081108140816081908200825082a084108440846084908500852085508" + b"580861086408800885089408aa08010904091009120915091809210940094509" + b"480951095409600981099009000a110a140a220a280a2a0a500a990a01100410" + b"0610091010101210151018102110241026104010421045104810511054105610" + b"59106010621065106810811084108610901095109810a110a410001102110511" + b"08110a1111111411161119112011221125112811411144114611491150115211" + b"5511581161116411801182118511881191119411011204120912101215122112" + b"2412401245125112541281128412901200140214051408141114141416141914" + b"2014251428144114441446144914501452145514581461146414801482148514" + b"881491149414a014011504150615091510151215151518152115241540154215" + b"4515481551155415601581158415901500160516081611161416201641164416" + b"50168016aa160118041806180918101815181818211840184218451848185118" + b"541860188118841800190219051908191119141920194119441950196919a219" + b"041a101a401a561a00200220052008201120142016201920202025202a204120" + b"4420502052205520642080208a209420aa200121042110211221152121214021" + b"4221452151215421602181218421902100220a22222228222a22442250228822" + b"8a22a82201240424062409241024152418242124242440244224452448245124" + b"5424602481248424902400250525082511251425202541254425502566258025" + b"0126042610264026592600280528112814284128442850288a28aa2801290429" + b"102995290a2a222a642a882a8a2a014004400640094010401240154018401a40" + b"21402440264040404240454048404a4051405440564059406040624065408140" + b"8440904095409840a140a4400041024105410841114114411641194120412241" + b"2541414144414641494150415241554158416141644180418241854188419141" + b"9441a04101420442104212421542184224424042454248425142544260428142" + b"844200440244054408440a441144144416441944204422442544284441444444" + b"46444944504452445544584461446444804482448544884491449444a0440145" + b"0445064509451045124515451845214524454045424545454845514554456045" + b"6a4581458445904500460246054608461146144620464146444650468046a546" + b"0148044809481048124815481848214824484048424845484848514854486048" + b"84489048004902490549084911491449204941494449504980499649014a044a" + b"104a404a00500250055008501150145016501950205022502550285041504450" + b"4650495050505250555058506150645080508250855088509150945001510451" + b"0651095110511251155118512151245140514251455148515151545160518151" + b"8451905100520552085211521452205241524452505269528052015404540654" + b"0954105412541554185421542454405442544554485451545454605481548454" + b"9054005502550555085511551455205541554455505580550156045610562656" + b"405600580258055808581158145820584158445850585a588058015904591059" + b"4059005a195a855aa85a01600460066010601260156018602160246040604560" + b"4860516054606060846090600061026105610861116114612061416144615061" + b"806199610462106240625662a162006405640864116414642064416444645064" + b"806401650465106540654a656865926500669466016804681068656898680069" + b"2a69426aa16a0080028005800880118014801980208025804180448050805280" + b"5580588061808080858091809480018104810981108112811581188121812481" + b"408142814581488151815481818184819081a981008205820a82118214824182" + b"4482508201840484068409841084128415841884218440844284458448845184" + b"5484608481848484908400850285058508851185148520854185448550858085" + b"8a85018604861086298640860088058811881488418844885088a28801890489" + b"40896589228a588a5a8a828aa28a019004900990109012901590189024904090" + b"4290459048905190549060908190849090900091059111911491419144915091" + b"5a910192049210924092a6920094029405940894119414942094419444945094" + b"8094969401950495109540959895a19500964696649601980498109826984098" + b"a998009949995299909a00a005a00aa014a022a02aa041a044a050a0a2a0aaa0" + b"40a165a102a20aa222a228a22aa282a288a28aa2a8a201a404a410a440a489a4" + b"a4a400a519a551a60aa828a8a2a854a986a908aa0aaa20aa22aa28aa88aaaaaa" + ) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d, rest = np.hsplit(blocks, [2]) + qs, rest = np.hsplit(rest, [QK_K // 8]) + signs, rest = np.hsplit(rest, [QK_K // 8]) + qh, scales = np.hsplit(rest, [QK_K // 32]) + + d = d.view(np.float16).astype(np.float32) + + scales = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2)) + scales = (scales & 0x0F).reshape((n_blocks, -1)) + db = d * (np.float32(0.5) + scales) * np.float32(0.25) + db = db.reshape((n_blocks, -1, 1, 1)) + + # unpack the sign bits + signs = signs.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8)) + signs = signs & np.uint8(0x01) + signs = np.where(signs == 0, np.float32(1), np.float32(-1)) + signs = signs.reshape((n_blocks, -1, 2, 8)) + + qh = qh.reshape((n_blocks, -1, 1)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4)) + qs = qs.astype(np.uint16) | ((qh & 0x03).astype(np.uint16) << 8).reshape((n_blocks, -1)) + + assert cls.grid is not None + grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2) + grid = grid.reshape((n_blocks, -1, 2, 8)) + + return (db * grid * signs).reshape((n_blocks, -1)) + + +class IQ3_XXS(__Quant, qtype=GGMLQuantizationType.IQ3_XXS): + grid_shape = (256, 4) + grid_map = (0x04, 0x0c, 0x14, 0x1c, 0x24, 0x2c, 0x34, 0x3e) + grid_hex = ( + b"0000020004001100130017002000220031004200730075000101030110011201" + b"2101250130013201410154017001000202020402110220022202310233023702" + b"5102570275020103070310031203250370031304370444045704730475040105" + b"0705320552053506640610071407160743076107011003101010121021102310" + b"3010321034104710501000110211111120112211011203121012121221123012" + b"7212001302132013311346136613011405145014201524154615711505162217" + b"4017002002201120132020202220262031204220012103210521102112212121" + b"3021632167217021002202221122172220222222372240225522012310231423" + b"7023742335245324032527254125742501270327162745270130103012302130" + b"2330503065307230003102312031313144314631013203321032253252327232" + b"1133333330344734723400350635223555351436363663363337603704401740" + b"3540374053405740744120423742404260426642074345430444514464442545" + b"4345704505471047124730471250415070500051065126515551145232527252" + b"0253535310542354275472540255315550562457425724604460466064602161" + b"6161176264623063366344640565526533660367216703700570077010703270" + b"5270267140711272457252720073157333736073217441740075027524753076" + ) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d, rest = np.hsplit(blocks, [2]) + qs, scales = np.hsplit(rest, [QK_K // 4]) + + d = d.view(np.float16).astype(np.float32) + scales = scales.view(np.uint32) + + db = d * (np.float32(0.5) + (scales >> 28).astype(np.float32)) * np.float32(0.5) + db = db.reshape((n_blocks, -1, 1, 1)) + + # get the sign indices and unpack the bits + signs = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 7, 14, 21], dtype=np.uint32).reshape((1, 1, 4)) + ksigns = np.frombuffer(IQ2_XXS.ksigns, dtype=np.uint8).reshape((1, 1, 1, 128)) + signs = (signs & np.uint32(0x7F)).reshape((n_blocks, -1, 4, 1)) + signs = np.take_along_axis(ksigns, signs, axis=-1) + signs = signs.reshape((n_blocks, -1, 4, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 1, 8)) + signs = signs & np.uint8(0x01) + signs = np.where(signs == 0, np.float32(1), np.float32(-1)) + signs = signs.reshape((n_blocks, -1, 4, 8)) + + assert cls.grid is not None + grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2) + grid = grid.reshape((n_blocks, -1, 4, 8)) + + return (db * grid * signs).reshape((n_blocks, -1)) + + +class IQ3_S(__Quant, qtype=GGMLQuantizationType.IQ3_S): + grid_shape = (512, 4) + grid_map = (0x01, 0x03, 0x05, 0x07, 0x09, 0x0b, 0x0d, 0x0f) + grid_hex = ( + b"0000010002000500070010001100120014001600200021002500330040004200" + b"4500470051005300600062007100740077000001010102010401100111011501" + b"2001230127013101350144016101650172010002010205020702100213021602" + b"2102250230023402420245024702510253027002730203031103150320032203" + b"3103330336034403500352036703710375030004130417042104240432044004" + b"4304510470040205040520052205260533054105450547056605730506061106" + b"1306310652067106000702070407200722072607330750075407001001100210" + b"0410101011101310151017102010221031103410361054105610611072100011" + b"0111031106111011141121113011331141115011521170117611001212121512" + b"1712201224123212401243125512601272120113041307131013131321132713" + b"3013341341136213701303140514121414143114331442144614501454140115" + b"1015131521153015321551152016241627164416461601170317101712172117" + b"3517411762177017002001200320052007201020122014201620212023202720" + b"3020322041204320452050205220672070207320752000210221102113211721" + b"2221252131213421422151210122042207222122232230223722412253225722" + b"7122742200230223052311232223242331233323422350236623012407242024" + b"2324322435244124722475240425112522253725402553257025002602260726" + b"2126552661260527112726273027432750270230113013301530173022303130" + b"3330353042304430473051306330713001310331053114312131233140316031" + b"7231763100321232203232323432503201331033143321332333273330334133" + b"4333473355337333033411341634223431345234603464340135103512352535" + b"3235443556357335163641360137033720372237353700400440124020402440" + b"2740324041405040704002410741114113412241304135414341514155410142" + b"0342104215422142334240425742624270420443114313432043224331433543" + b"0044024424443744404471440545074521456245134634466046104715473047" + b"4347514702501050145022504050445047505250665074500151035105511251" + b"2151325172510052115223523052365253520253075310532753445351536553" + b"7353015404542054325446541255265551555355425602570457225711601360" + b"1560316033606060006120612761646112623462426255626262706200631463" + b"2163406325644364626400650365346560650566406611671367007004700770" + b"2070227036704070547062700271117124714371457101720472107216722172" + b"3072517202733273357353730174057413742074507422754275027631760077" + ) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d, rest = np.hsplit(blocks, [2]) + qs, rest = np.hsplit(rest, [QK_K // 4]) + qh, rest = np.hsplit(rest, [QK_K // 32]) + signs, scales = np.hsplit(rest, [QK_K // 8]) + + d = d.view(np.float16).astype(np.float32) + + scales = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2)) + scales = (scales & 0x0F).reshape((n_blocks, -1)) + db = d * (1 + 2 * scales) + db = db.reshape((n_blocks, -1, 1, 1)) + + # unpack the sign bits + signs = signs.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8)) + signs = signs & np.uint8(0x01) + signs = np.where(signs == 0, np.float32(1), np.float32(-1)) + signs = signs.reshape((n_blocks, -1, 4, 8)) + + qh = qh.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8) + qh = (qh & 0x01).astype(np.uint16).reshape((n_blocks, -1)) + qs = qs.astype(np.uint16) | (qh << 8) + + assert cls.grid is not None + grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2) + grid = grid.reshape((n_blocks, -1, 4, 8)) + + return (db * grid * signs).reshape((n_blocks, -1)) + + +class IQ1_S(__Quant, qtype=GGMLQuantizationType.IQ1_S): + # iq1s_grid, with each byte packed into 2 bits + # -1, 0, 1 <=> 0, 1, 2 + grid_shape = (2048, 8) + grid_map = (-1, 0, 1) + grid_hex = ( + b"00000200050008000a00110015002000220028002a0045005100540056006500" + b"8000820088008a009500a000a200a800aa000401050111011401160119011a01" + b"2501410146014901520155015a0161016401660168018501910194019601a501" + b"0002020208020a0215022002220228022a024502510259026402690280028202" + b"88028a02910295029902a002a202a802aa021104140416042504410449045504" + b"5a046404650491049904a5040105040505050605150518051a05290540054505" + b"4a0550055105540555055605590560056205650568056a058105910595059805" + b"9a05a105a405a505a605a9051406190641064406500652065506580660066106" + b"6606690685069106940699060008020808080a0815082008220828082a084508" + b"5108560865088008820888088a089508a008a208a808aa080509110914091909" + b"2409250941095009510955096109640969099109940996099909a509000a020a" + b"080a0a0a150a200a220a280a2a0a450a510a590a610a650a800a820a850a880a" + b"8a0a950aa00aa20aa80aaa0a1010111014101910241025104110441050105510" + b"58106110641065106910911094109610a110a510011104110611091110111211" + b"1511181121112411291145114a11501151115211541155115611591160116511" + b"841192119511a111a41111121412161225124012461249125212551258125a12" + b"641266128512911294129612a512011406140914141415141814191421142614" + b"41144514461448144a1451145414551456145914621465146814841489149014" + b"94149514981499149a14a114a414a514a914021505150a151115141515151615" + b"191520152215251528152a154115441545154615511552155415551556155915" + b"5a1561156415651566156915801582158415851588158a159015911594159515" + b"961599159a15a015a215a51501160416051606161516161618161a1621162616" + b"401642164416451648164a165116551656165816591661166416651668166916" + b"6a1686168a1692169516a416a916111816182518411844184618491850185518" + b"58185a1860186118641866186918851891189418a5181019121915191a192119" + b"25194219441945194819511954195519561959195a19601965196a1989199119" + b"921995199819a119a619a919091a161a241a261a441a461a491a501a521a551a" + b"581a611a661a691a851a911a961a9a1a0020022008200a201520202022202520" + b"28202a20452051205920612065208020822088208a209520a020a220a520a820" + b"aa2005211121142119212521422144214921552158215a216121642165216621" + b"8521902196219921a521012208220a22112215222022222228222a2245225122" + b"562259226522812288228a2291229522a022a222a822aa220524142416241924" + b"252444244524462449245224552458245a2466248524912494249924a124a524" + b"0925152521252925402545254825512554255525592562256525682589259025" + b"9425952598259a25a125a425a625a92505261026122619262526412649265526" + b"6026612669268426862690269a260028022808280a2815282028222828282a28" + b"45285128542865288028822888288a28a028a228a828aa280929112914291929" + b"2529462949295229552961296429662969298529902996299929a429a529002a" + b"022a082a0a2a202a222a282a2a2a452a512a562a592a652a802a822a882a8a2a" + b"952aa02aa22aa82aaa2a054011401640254049405240554058405a4061406440" + b"664094409940a140a6400041014104410641094112411541164118411a412141" + b"26412941454148414a41514154415541564159415a41654168416a4181418441" + b"8641904192419541a041a141a241054211421442164225424142524255425a42" + b"6442694289429442a5420144154419442944454448444a445144544455445644" + b"61446244654468446a44814486448944904492449544a044a144a94401450245" + b"05450a4511451445154516451945204525452a45414544454545464549455045" + b"5145544555455645584559456145644565456645694582458445854588459145" + b"94459545964599459a45a545a845aa450146054609461446154618461a462146" + b"2446294640464246454648465046514652465546564659466246654668468146" + b"85468a4694469546a146a446a6460548114815481a4825484248494850485548" + b"5848614864486648694885489148944896489948a5480149054906490a491049" + b"144915491849214924492649404945494a495149524954495549564959496049" + b"6249654966496a49864989499249954996499849a149a449a649a949164a444a" + b"464a494a554a584a5a4a644a694a944aa54a0150045005500650095012501550" + b"1a50215024502950405045504850515054505550565059506550685086508950" + b"95509850a050a150a650a9500551085109510a51115114511551165118511951" + b"20512551265128512a5141514451455146514951505151515251545155515651" + b"585159515a51615164516551665169518251855191519451955196519951a051" + b"a551aa5101520652125215521a5221522452425245524a525152545255525652" + b"595262526552855290529252955299529a52a452045405541154145415541654" + b"185419542154255428542a54415444544554465449544a545054515454545554" + b"5654585459545a54615462546454655466546954805488548a54915494549554" + b"96549954a154a454a554aa540155025504550555065509551055115512551455" + b"1555165519551a55215524552555265529554055415542554455455546554855" + b"4955505551555255545555555655585559555a55605561556455655566556855" + b"69556a5581558455855589558a559055915594559555965598559955a155a455" + b"a555a655a9550056015602560456065608560956115614561556185619562056" + b"2156225624562556265628562956415645564656485649564a56505651565256" + b"545655565656585659565a566156645665566956825685568656885689568a56" + b"915695569a56a256a556a656a856a95604580558065809581058155818582158" + b"2a58455848584a58515854585558565858585958605862586458655882588958" + b"9058925895589858a158a9580159025905590a59115914591559165919592559" + b"41594459455946594959505951595259545955595659585959595a5961596459" + b"655966596959815985598959915994599559965998599959a559045a085a155a" + b"1a5a205a255a265a295a455a485a495a515a555a565a585a595a625a655a685a" + b"6a5a815a8a5a925a955a965a985a9a5aa15a0560146016601960256044605060" + b"5560566058605a60616064606660696081609660a56001610461066109611261" + b"15612161226126612961456149615161556156615961656166616a6184618a61" + b"92619561a161a661a96111621662196240624162466255625662586260628562" + b"91629662a56211641264156416641a6421642664296440644264456448644a64" + b"516454645564566459645a646064626465648464856489649064926494649564" + b"966498649a64a164a464a964056508650a651165156516651965446545654665" + b"496550655165546555655665596561656465656566656965866589658a659165" + b"9565966599659a65a265a565a665a86502660966156620662666286629664066" + b"456648664a66516654665566566658665a666066656668668066826685668a66" + b"9466966698669966a066a466a666aa661668196825684168526855685a686168" + b"6968856891689868a66801690469106915692169246926692969406941694569" + b"4669486951695469556956695969606965696a69826984698a699569a169a469" + b"a569a969116a166a186a416a446a496a506a556a586a5a6a646a656a696a866a" + b"946a986a9a6aa66a0080028008800a802080228028802a804580508051805480" + b"5680598065808080828088808a809580a080a280a880aa800581118114811681" + b"1981258141814481498150815281558156815881598164816681698185818981" + b"948196819981a5810082028208820a8215822082228228822a82518254825982" + b"65828082828288828a829582a082a282a882aa82148419844184448451845584" + b"5a846184648469849484998401850985128515851a8526852985408541854585" + b"4885518554855585568559855a856585668568856a8581858485868589859085" + b"928595859885a68511861686198625864186448649864a865086558659865a86" + b"618666866a86858691869a86a4860088028808880a8815882088228828882a88" + b"41884588518854885988658869888088828888888a889588a088a288a888aa88" + b"05890689118914891689258941894489468949895089528955895a8961896489" + b"858996899989a589008a028a088a0a8a158a208a228a288a2a8a458a518a548a" + b"568a808a828a888a8a8a958aa08aa28aa88aaa8a059011901690189019902590" + b"419046904990559058905a9069906a9085909190949096909990a59001910491" + b"069109911091159118911a912191249126912991409145915091519154915591" + b"569159916291659184918691929195919891a191a491a691a991059211921492" + b"19922592449246924992509252925592589266926992859294929692a9920194" + b"04940694109415941894269440944a9451945494559456945894599460946194" + b"62946594849486949294949495949894a194a9940095059508950a9510951195" + b"14951595169519952195259529952a9541954495459546954995509551955295" + b"549555955695589559955a956195649565956695699581958595889591959295" + b"94959595969599959a95a095a295a595a895aa95019604961096159619962096" + b"2696299645964896499651965296559656965996659668968296849689968a96" + b"929694969596a496a696a9960598169819982598419846985098529855985698" + b"5a98649865988598919896989998a59804990699099910991299159918991a99" + b"209921992499269940994299459948994a995199549955995699599962996599" + b"66996a99819984999099929995999a99a199a699059a159a259a449a469a499a" + b"509a559a589a619a859a919a949a959a969a00a002a008a00aa015a020a022a0" + b"28a02aa045a051a054a056a059a080a082a088a08aa095a0a0a0a2a0a8a0aaa0" + b"05a109a111a114a116a119a11aa146a149a151a155a158a15aa161a164a185a1" + b"90a192a196a199a102a208a20aa210a219a222a228a22aa245a251a256a259a2" + b"65a280a282a288a28aa295a2a0a2a2a2a8a2aaa219a425a441a444a450a454a4" + b"55a458a45aa461a465a466a468a469a485a406a509a510a512a515a518a526a5" + b"29a542a545a551a554a555a556a559a565a56aa581a584a585a586a589a592a5" + b"95a598a505a611a616a61aa621a625a644a646a64aa652a655a656a658a660a6" + b"62a686a690a695a696a699a6a1a6a4a6a6a600a802a808a80aa820a822a828a8" + b"2aa851a854a856a859a880a882a888a88aa895a8a0a8a2a8a8a8aaa805a914a9" + b"19a921a925a941a950a955a95aa961a966a969a990a996a900aa02aa08aa0aaa" + b"20aa22aa28aa2aaa51aa54aa56aa80aa82aa88aa8aaa95aaa0aaa2aaa8aaaaaa" + ) + + delta = np.float32(0.125) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d, rest = np.hsplit(blocks, [2]) + qs, qh = np.hsplit(rest, [QK_K // 8]) + + d = d.view(np.float16).astype(np.float32) + qh = qh.view(np.uint16) + + dl = d * (2 * ((qh >> 12) & 7) + 1) + dl = dl.reshape((n_blocks, -1, 1, 1)) + delta = np.where((qh & np.uint16(0x8000)) == 0, cls.delta, -cls.delta) + delta = delta.reshape((n_blocks, -1, 1, 1)) + + qh = qh.reshape((n_blocks, -1, 1)) >> np.array([0, 3, 6, 9], dtype=np.uint16).reshape((1, 1, 4)) + qs = qs.astype(np.uint16) | ((qh & 7) << 8).reshape((n_blocks, -1)) + + assert cls.grid is not None + grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2) + grid = grid.reshape((n_blocks, -1, 4, 8)) + + return (dl * (grid + delta)).reshape((n_blocks, -1)) + + +class IQ1_M(__Quant, qtype=GGMLQuantizationType.IQ1_M): + grid_shape = IQ1_S.grid_shape + grid_map = IQ1_S.grid_map + grid_hex = IQ1_S.grid_hex + + delta = IQ1_S.delta + + # Okay *this* type is weird. It's the only one which stores the f16 scales in multiple parts. + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + qs, rest = np.hsplit(blocks, [QK_K // 8]) + qh, scales = np.hsplit(rest, [QK_K // 16]) + + # The f16 scale is packed across multiple bytes + scales = scales.view(np.uint16) + d = (scales.reshape((n_blocks, 4)) & np.uint16(0xF000)) >> np.array([12, 8, 4, 0], dtype=np.uint16).reshape((1, 4)) + d = d[..., 0] | d[..., 1] | d[..., 2] | d[..., 3] + d = d.view(np.float16).astype(np.float32).reshape((n_blocks, 1)) + + scales = scales.reshape(n_blocks, -1, 1) >> np.array([0, 3, 6, 9], dtype=np.uint16).reshape((1, 1, 4)) + scales = (scales & 0x07).reshape((n_blocks, -1)) + dl = d * (2 * scales + 1) + dl = dl.reshape((n_blocks, -1, 2, 1, 1)) + + qh = qh.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2)) + qs = qs.astype(np.uint16) | ((qh & 0x07).astype(np.uint16) << 8).reshape((n_blocks, -1)) + + delta = np.where(qh & 0x08 == 0, cls.delta, -cls.delta) + delta = delta.reshape((n_blocks, -1, 2, 2, 1)) + + assert cls.grid is not None + grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2) + grid = grid.reshape((n_blocks, -1, 2, 2, 8)) + + return (dl * (grid + delta)).reshape((n_blocks, -1)) + + +class IQ4_NL(__Quant, qtype=GGMLQuantizationType.IQ4_NL): + kvalues = (-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113) + + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d, qs = np.hsplit(blocks, [2]) + + d = d.view(np.float16).astype(np.float32) + + qs = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1)) + + qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1, 1)) + + kvalues = np.array(cls.kvalues, dtype=np.int8).reshape(1, 1, 16) + qs = np.take_along_axis(kvalues, qs, axis=-1).astype(np.float32).reshape((n_blocks, -1)) + + return (d * qs) + + +class IQ4_XS(__Quant, qtype=GGMLQuantizationType.IQ4_XS): + @classmethod + def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray: + n_blocks = blocks.shape[0] + + d, rest = np.hsplit(blocks, [2]) + scales_h, rest = np.hsplit(rest, [2]) + scales_l, qs = np.hsplit(rest, [QK_K // 64]) + + d = d.view(np.float16).astype(np.float32) + scales_h = scales_h.view(np.uint16) + + scales_l = scales_l.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2)) + scales_h = scales_h.reshape((n_blocks, 1, -1)) >> np.array([2 * i for i in range(QK_K // 32)], dtype=np.uint16).reshape((1, -1, 1)) + scales_l = scales_l.reshape((n_blocks, -1)) & np.uint8(0x0F) + scales_h = scales_h.reshape((n_blocks, -1)).astype(np.uint8) & np.uint8(0x03) + + scales = (scales_l | (scales_h << np.uint8(4))).astype(np.int8) - np.int8(32) + dl = (d * scales.astype(np.float32)).reshape((n_blocks, -1, 1)) + + qs = qs.reshape((n_blocks, -1, 1, 16)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1)) + qs = qs.reshape((n_blocks, -1, 32, 1)) & np.uint8(0x0F) + + kvalues = np.array(IQ4_NL.kvalues, dtype=np.int8).reshape((1, 1, 1, -1)) + qs = np.take_along_axis(kvalues, qs, axis=-1).astype(np.float32).reshape((n_blocks, -1, 32)) + + return (dl * qs).reshape((n_blocks, -1)) diff --git a/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_convert_endian.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_convert_endian.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c838a8224124af5a86e71a3882cc28c4cc220650 Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_convert_endian.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_dump.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_dump.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..47835b43e03d1aae30f35079f083814724b154e1 Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_dump.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_editor_gui.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_editor_gui.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bf28e52f2ed73dc8696afbbc570eec2392ed9774 Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_editor_gui.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_hash.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_hash.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4d7856368336b0a06817bd058d8c755c6daa67e5 Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_hash.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_new_metadata.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_new_metadata.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c9013d9dc44b7679f60ef4c7769150ce7d2c9ea8 Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_new_metadata.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_set_metadata.cpython-310.pyc b/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_set_metadata.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a8d41e04fe5365260d3a95d5e2dd371b0e0685f3 Binary files /dev/null and b/venv/lib/python3.10/site-packages/gguf/scripts/__pycache__/gguf_set_metadata.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/gguf/scripts/gguf_convert_endian.py b/venv/lib/python3.10/site-packages/gguf/scripts/gguf_convert_endian.py new file mode 100644 index 0000000000000000000000000000000000000000..0e0febaa791786fb9d5c713478377e5d54642df8 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/scripts/gguf_convert_endian.py @@ -0,0 +1,182 @@ +#!/usr/bin/env python3 +from __future__ import annotations + +import logging +import argparse +import os +import sys +from tqdm import tqdm +from pathlib import Path + +import numpy as np + +# Necessary to load the local gguf package +if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent.parent)) + +import gguf + +logger = logging.getLogger("gguf-convert-endian") + + +def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None: + file_endian = reader.endianess.name + if reader.byte_order == 'S': + host_endian = 'BIG' if file_endian == 'LITTLE' else 'LITTLE' + else: + host_endian = file_endian + order = host_endian if args.order == "native" else args.order.upper() + logger.info(f"* Host is {host_endian} endian, GGUF file seems to be {file_endian} endian") + if file_endian == order: + logger.info(f"* File is already {order} endian. Nothing to do.") + sys.exit(0) + logger.info("* Checking tensors for conversion compatibility") + for tensor in reader.tensors: + if tensor.tensor_type not in ( + gguf.GGMLQuantizationType.F32, + gguf.GGMLQuantizationType.F16, + gguf.GGMLQuantizationType.Q8_0, + gguf.GGMLQuantizationType.Q4_K, + gguf.GGMLQuantizationType.Q6_K, + ): + raise ValueError(f"Cannot handle type {tensor.tensor_type.name} for tensor {repr(tensor.name)}") + logger.info(f"* Preparing to convert from {file_endian} to {order}") + if args.dry_run: + return + logger.warning("*** Warning *** Warning *** Warning **") + logger.warning("* This conversion process may damage the file. Ensure you have a backup.") + if order != host_endian: + logger.warning("* Requested endian differs from host, you will not be able to load the model on this machine.") + logger.warning("* The file will be modified immediately, so if conversion fails or is interrupted") + logger.warning("* the file will be corrupted. Enter exactly YES if you are positive you want to proceed:") + response = input("YES, I am sure> ") + if response != "YES": + logger.warning("You didn't enter YES. Okay then, see ya!") + sys.exit(0) + logger.info(f"* Converting fields ({len(reader.fields)})") + for idx, field in enumerate(reader.fields.values()): + logger.info(f"- {idx:4}: Converting field {repr(field.name)}, part count: {len(field.parts)}") + for part in field.parts: + part.byteswap(inplace=True) + logger.info(f"* Converting tensors ({len(reader.tensors)})") + + for idx, tensor in enumerate(pbar := tqdm(reader.tensors, desc="Converting tensor")): + log_message = ( + f"Converting tensor {repr(tensor.name)}, " + f"type={tensor.tensor_type.name}, " + f"elements={tensor.n_elements} " + ) + + # Byte-swap each part of the tensor's field + for part in tensor.field.parts: + part.byteswap(inplace=True) + + # Byte-swap tensor data if necessary + if tensor.tensor_type == gguf.GGMLQuantizationType.Q8_0: + # Handle Q8_0 tensor blocks (block_q8_0) + # Specific handling of block_q8_0 is required. + # Each block_q8_0 consists of an f16 delta (scaling factor) followed by 32 int8 quantizations. + + block_size = 34 # 34 bytes = + 32 * + + n_blocks = len(tensor.data) // block_size + for block_num in (inner_pbar := tqdm(range(n_blocks), desc="Byte-swapping Blocks", leave=False)): + block_offs = block_num * block_size + + # Byte-Swap f16 sized delta field + delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16) + delta.byteswap(inplace=True) + + # Byte-Swap Q8 weights + if block_num % 100000 == 0: + inner_pbar.set_description(f"Byte-swapping Blocks [{(n_blocks - block_num) // n_blocks}]") + + elif tensor.tensor_type == gguf.GGMLQuantizationType.Q4_K: + # Handle Q4_K tensor blocks (block_q4_k) + # Specific handling of block_q4_k is required. + # Each block_q4_k consists of 2 f16 values followed by 140 int8 values. + + # first flatten structure + newshape = 1 + for i in tensor.data.shape: + newshape *= i + + tensor.data.resize(newshape) + + block_size = 144 + n_blocks = len(tensor.data) // block_size + for block_num in (inner_pbar := tqdm(range(n_blocks), desc="Byte-swapping Blocks", leave=False)): + block_offs = block_num * block_size + + # Byte-Swap f16 sized fields + delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16) + delta.byteswap(inplace=True) + + delta = tensor.data[block_offs + 2:block_offs + 4].view(dtype=np.uint16) + delta.byteswap(inplace=True) + + # Byte-Swap + if block_num % 100000 == 0: + inner_pbar.set_description(f"Byte-swapping Blocks [{(n_blocks - block_num) // n_blocks}]") + + elif tensor.tensor_type == gguf.GGMLQuantizationType.Q6_K: + # Handle Q6_K tensor blocks (block_q6_k) + # Specific handling of block_q6_k is required. + # Each block_q6_k consists of 208 int8 values followed by 1 f16 value. + + # first flatten structure + newshape = 1 + for i in tensor.data.shape: + newshape *= i + + tensor.data.resize(newshape) + + block_size = 210 + n_blocks = len(tensor.data) // block_size + for block_num in (inner_pbar := tqdm(range(n_blocks), desc="Byte-swapping Blocks", leave=False)): + block_offs = block_num * block_size + + # Byte-Swap f16 sized field + delta = tensor.data[block_offs + 208:block_offs + 210].view(dtype=np.uint16) + delta.byteswap(inplace=True) + + # Byte-Swap + if block_num % 100000 == 0: + inner_pbar.set_description(f"Byte-swapping Blocks [{(n_blocks - block_num) // n_blocks}]") + + else: + # Handle other tensor types + tensor.data.byteswap(inplace=True) + + pbar.set_description(log_message) + + logger.info("* Completion") + + +def main() -> None: + parser = argparse.ArgumentParser(description="Convert GGUF file byte order") + parser.add_argument( + "model", type=str, + help="GGUF format model filename", + ) + parser.add_argument( + "order", type=str, choices=['big', 'little', 'native'], + help="Requested byte order", + ) + parser.add_argument( + "--dry-run", action="store_true", + help="Don't actually change anything", + ) + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") + + args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"]) + + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + + logger.info(f'* Loading: {args.model}') + reader = gguf.GGUFReader(args.model, 'r' if args.dry_run else 'r+') + convert_byteorder(reader, args) + + +if __name__ == "__main__": + main() diff --git a/venv/lib/python3.10/site-packages/gguf/scripts/gguf_dump.py b/venv/lib/python3.10/site-packages/gguf/scripts/gguf_dump.py new file mode 100644 index 0000000000000000000000000000000000000000..e282892d645c7c31a1d5e00f383dcaaa5890a127 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/scripts/gguf_dump.py @@ -0,0 +1,454 @@ +#!/usr/bin/env python3 +from __future__ import annotations + +import logging +import argparse +import os +import re +import sys +from pathlib import Path +from typing import Any + +# Necessary to load the local gguf package +if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent.parent)) + +from gguf import GGUFReader, GGUFValueType, ReaderTensor # noqa: E402 + +logger = logging.getLogger("gguf-dump") + + +def get_file_host_endian(reader: GGUFReader) -> tuple[str, str]: + file_endian = reader.endianess.name + if reader.byte_order == 'S': + host_endian = 'BIG' if file_endian == 'LITTLE' else 'LITTLE' + else: + host_endian = file_endian + return (host_endian, file_endian) + + +# For more information about what field.parts and field.data represent, +# please see the comments in the modify_gguf.py example. +def dump_metadata(reader: GGUFReader, args: argparse.Namespace) -> None: + host_endian, file_endian = get_file_host_endian(reader) + print(f'* File is {file_endian} endian, script is running on a {host_endian} endian host.') # noqa: NP100 + print(f'* Dumping {len(reader.fields)} key/value pair(s)') # noqa: NP100 + for n, field in enumerate(reader.fields.values(), 1): + if not field.types: + pretty_type = 'N/A' + elif field.types[0] == GGUFValueType.ARRAY: + nest_count = len(field.types) - 1 + pretty_type = '[' * nest_count + str(field.types[-1].name) + ']' * nest_count + else: + pretty_type = str(field.types[-1].name) + + log_message = f' {n:5}: {pretty_type:10} | {len(field.data):8} | {field.name}' + if field.types: + curr_type = field.types[0] + if curr_type == GGUFValueType.STRING: + content = field.contents() + if len(content) > 60: + content = content[:57] + '...' + log_message += ' = {0}'.format(repr(content)) + elif curr_type in reader.gguf_scalar_to_np: + log_message += ' = {0}'.format(field.contents()) + else: + content = repr(field.contents(slice(6))) + if len(field.data) > 6: + content = content[:-1] + ', ...]' + log_message += ' = {0}'.format(content) + print(log_message) # noqa: NP100 + if args.no_tensors: + return + print(f'* Dumping {len(reader.tensors)} tensor(s)') # noqa: NP100 + for n, tensor in enumerate(reader.tensors, 1): + prettydims = ', '.join('{0:5}'.format(d) for d in list(tensor.shape) + [1] * (4 - len(tensor.shape))) + print(f' {n:5}: {tensor.n_elements:10} | {prettydims} | {tensor.tensor_type.name:7} | {tensor.name}') # noqa: NP100 + + +def dump_metadata_json(reader: GGUFReader, args: argparse.Namespace) -> None: + import json + host_endian, file_endian = get_file_host_endian(reader) + metadata: dict[str, Any] = {} + tensors: dict[str, Any] = {} + result = { + "filename": args.model, + "endian": file_endian, + "metadata": metadata, + "tensors": tensors, + } + for idx, field in enumerate(reader.fields.values()): + curr: dict[str, Any] = { + "index": idx, + "type": field.types[0].name if field.types else 'UNKNOWN', + "offset": field.offset, + } + metadata[field.name] = curr + if field.types[:1] == [GGUFValueType.ARRAY]: + curr["array_types"] = [t.name for t in field.types][1:] + if not args.json_array: + continue + curr["value"] = field.contents() + else: + curr["value"] = field.contents() + if not args.no_tensors: + for idx, tensor in enumerate(reader.tensors): + tensors[tensor.name] = { + "index": idx, + "shape": tensor.shape.tolist(), + "type": tensor.tensor_type.name, + "offset": tensor.field.offset, + } + json.dump(result, sys.stdout) + + +def markdown_table_with_alignment_support(header_map: list[dict[str, str]], data: list[dict[str, Any]]): + # JSON to Markdown table formatting: https://stackoverflow.com/a/72983854/2850957 + + # Alignment Utility Function + def strAlign(padding: int, alignMode: str | None, strVal: str): + if alignMode == 'center': + return strVal.center(padding) + elif alignMode == 'right': + return strVal.rjust(padding - 1) + ' ' + elif alignMode == 'left': + return ' ' + strVal.ljust(padding - 1) + else: # default left + return ' ' + strVal.ljust(padding - 1) + + def dashAlign(padding: int, alignMode: str | None): + if alignMode == 'center': + return ':' + '-' * (padding - 2) + ':' + elif alignMode == 'right': + return '-' * (padding - 1) + ':' + elif alignMode == 'left': + return ':' + '-' * (padding - 1) + else: # default left + return '-' * (padding) + + # Calculate Padding For Each Column Based On Header and Data Length + rowsPadding = {} + for index, columnEntry in enumerate(header_map): + padCount = max([len(str(v)) for d in data for k, v in d.items() if k == columnEntry['key_name']], default=0) + 2 + headerPadCount = len(columnEntry['header_name']) + 2 + rowsPadding[index] = headerPadCount if padCount <= headerPadCount else padCount + + # Render Markdown Header + rows = [] + rows.append('|'.join(strAlign(rowsPadding[index], columnEntry.get('align'), str(columnEntry['header_name'])) for index, columnEntry in enumerate(header_map))) + rows.append('|'.join(dashAlign(rowsPadding[index], columnEntry.get('align')) for index, columnEntry in enumerate(header_map))) + + # Render Tabular Data + for item in data: + rows.append('|'.join(strAlign(rowsPadding[index], columnEntry.get('align'), str(item[columnEntry['key_name']])) for index, columnEntry in enumerate(header_map))) + + # Convert Tabular String Rows Into String + tableString = "" + for row in rows: + tableString += f'|{row}|\n' + + return tableString + + +def element_count_rounded_notation(count: int) -> str: + if count > 1e15 : + # Quadrillion + scaled_amount = count * 1e-15 + scale_suffix = "Q" + elif count > 1e12 : + # Trillions + scaled_amount = count * 1e-12 + scale_suffix = "T" + elif count > 1e9 : + # Billions + scaled_amount = count * 1e-9 + scale_suffix = "B" + elif count > 1e6 : + # Millions + scaled_amount = count * 1e-6 + scale_suffix = "M" + elif count > 1e3 : + # Thousands + scaled_amount = count * 1e-3 + scale_suffix = "K" + else: + # Under Thousands + scaled_amount = count + scale_suffix = "" + return f"{'~' if count > 1e3 else ''}{round(scaled_amount)}{scale_suffix}" + + +def translate_tensor_name(name): + words = name.split(".") + + # Source: https://github.com/ggml-org/ggml/blob/master/docs/gguf.md#standardized-tensor-names + abbreviation_dictionary = { + 'token_embd': 'Token embedding', + 'pos_embd': 'Position embedding', + 'output_norm': 'Output normalization', + 'output': 'Output', + 'attn_norm': 'Attention normalization', + 'attn_norm_2': 'Attention normalization', + 'attn_qkv': 'Attention query-key-value', + 'attn_q': 'Attention query', + 'attn_k': 'Attention key', + 'attn_v': 'Attention value', + 'attn_output': 'Attention output', + 'ffn_norm': 'Feed-forward network normalization', + 'ffn_up': 'Feed-forward network "up"', + 'ffn_gate': 'Feed-forward network "gate"', + 'ffn_down': 'Feed-forward network "down"', + 'ffn_gate_inp': 'Expert-routing layer for the Feed-forward network in Mixture of Expert models', + 'ffn_gate_exp': 'Feed-forward network "gate" layer per expert in Mixture of Expert models', + 'ffn_down_exp': 'Feed-forward network "down" layer per expert in Mixture of Expert models', + 'ffn_up_exp': 'Feed-forward network "up" layer per expert in Mixture of Expert models', + 'ssm_in': 'State space model input projections', + 'ssm_conv1d': 'State space model rolling/shift', + 'ssm_x': 'State space model selective parametrization', + 'ssm_a': 'State space model state compression', + 'ssm_d': 'State space model skip connection', + 'ssm_dt': 'State space model time step', + 'ssm_out': 'State space model output projection', + 'blk': 'Block', + 'enc': 'Encoder', + 'dec': 'Decoder', + } + + expanded_words = [] + for word in words: + word_norm = word.strip().lower() + if word_norm in abbreviation_dictionary: + expanded_words.append(abbreviation_dictionary[word_norm].title()) + else: + expanded_words.append(word.title()) + + return ' '.join(expanded_words) + + +def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None: + host_endian, file_endian = get_file_host_endian(reader) + markdown_content = "" + markdown_content += f'# {args.model} - GGUF Internal File Dump\n\n' + markdown_content += f'- Endian: {file_endian} endian\n' + markdown_content += '\n' + markdown_content += '## Key Value Metadata Store\n\n' + markdown_content += f'There are {len(reader.fields)} key-value pairs in this file\n' + markdown_content += '\n' + + kv_dump_table: list[dict[str, str | int]] = [] + for n, field in enumerate(reader.fields.values(), 1): + if not field.types: + pretty_type = 'N/A' + elif field.types[0] == GGUFValueType.ARRAY: + nest_count = len(field.types) - 1 + pretty_type = '[' * nest_count + str(field.types[-1].name) + ']' * nest_count + else: + pretty_type = str(field.types[-1].name) + + def escape_markdown_inline_code(value_string): + # Find the longest contiguous sequence of backticks in the string then + # wrap string with appropriate number of backticks required to escape it + max_backticks = max((len(match.group(0)) for match in re.finditer(r'`+', value_string)), default=0) + inline_code_marker = '`' * (max_backticks + 1) + + # If the string starts or ends with a backtick, add a space at the beginning and end + if value_string.startswith('`') or value_string.endswith('`'): + value_string = f" {value_string} " + + return f"{inline_code_marker}{value_string}{inline_code_marker}" + + total_elements = len(field.data) + value = "" + if len(field.types) == 1: + curr_type = field.types[0] + if curr_type == GGUFValueType.STRING: + truncate_length = 60 + value_string = str(bytes(field.parts[-1]), encoding='utf-8') + if len(value_string) > truncate_length: + head = escape_markdown_inline_code(value_string[:truncate_length // 2]) + tail = escape_markdown_inline_code(value_string[-truncate_length // 2:]) + value = "{head}...{tail}".format(head=head, tail=tail) + else: + value = escape_markdown_inline_code(value_string) + elif curr_type in reader.gguf_scalar_to_np: + value = str(field.parts[-1][0]) + else: + if field.types[0] == GGUFValueType.ARRAY: + curr_type = field.types[1] + array_elements = [] + + if curr_type == GGUFValueType.STRING: + render_element = min(5, total_elements) + for element_pos in range(render_element): + truncate_length = 30 + value_string = str(bytes(field.parts[-1 - (total_elements - element_pos - 1) * 2]), encoding='utf-8') + if len(value_string) > truncate_length: + head = escape_markdown_inline_code(value_string[:truncate_length // 2]) + tail = escape_markdown_inline_code(value_string[-truncate_length // 2:]) + value = "{head}...{tail}".format(head=head, tail=tail) + else: + value = escape_markdown_inline_code(value_string) + array_elements.append(value) + + elif curr_type in reader.gguf_scalar_to_np: + render_element = min(7, total_elements) + for element_pos in range(render_element): + array_elements.append(str(field.parts[-1 - (total_elements - element_pos - 1)][0])) + + value = f'[ {", ".join(array_elements).strip()}{", ..." if total_elements > len(array_elements) else ""} ]' + + kv_dump_table.append({"n":n, "pretty_type":pretty_type, "total_elements":total_elements, "field_name":field.name, "value":value}) + + kv_dump_table_header_map = [ + {'key_name':'n', 'header_name':'POS', 'align':'right'}, + {'key_name':'pretty_type', 'header_name':'TYPE', 'align':'left'}, + {'key_name':'total_elements', 'header_name':'Count', 'align':'right'}, + {'key_name':'field_name', 'header_name':'Key', 'align':'left'}, + {'key_name':'value', 'header_name':'Value', 'align':'left'}, + ] + + markdown_content += markdown_table_with_alignment_support(kv_dump_table_header_map, kv_dump_table) + + markdown_content += "\n" + + if not args.no_tensors: + # Group tensors by their prefix and maintain order + tensor_prefix_order: list[str] = [] + tensor_name_to_key: dict[str, int] = {} + tensor_groups: dict[str, list[ReaderTensor]] = {} + total_elements = sum(tensor.n_elements for tensor in reader.tensors) + + # Parsing Tensors Record + for key, tensor in enumerate(reader.tensors): + tensor_components = tensor.name.split('.') + + # Classify Tensor Group + tensor_group_name = "base" + if tensor_components[0] == 'blk': + tensor_group_name = f"{tensor_components[0]}.{tensor_components[1]}" + elif tensor_components[0] in ['enc', 'dec'] and tensor_components[1] == 'blk': + tensor_group_name = f"{tensor_components[0]}.{tensor_components[1]}.{tensor_components[2]}" + elif tensor_components[0] in ['enc', 'dec']: + tensor_group_name = f"{tensor_components[0]}" + + # Check if new Tensor Group + if tensor_group_name not in tensor_groups: + tensor_groups[tensor_group_name] = [] + tensor_prefix_order.append(tensor_group_name) + + # Record Tensor and Tensor Position + tensor_groups[tensor_group_name].append(tensor) + tensor_name_to_key[tensor.name] = key + + # Tensors Mapping Dump + markdown_content += f'## Tensors Overview {element_count_rounded_notation(total_elements)} Elements\n\n' + markdown_content += f'Total number of elements in all tensors: {total_elements} Elements\n' + markdown_content += '\n' + + for group in tensor_prefix_order: + tensors = tensor_groups[group] + group_elements = sum(tensor.n_elements for tensor in tensors) + markdown_content += f"- [{translate_tensor_name(group)} Tensor Group - {element_count_rounded_notation(group_elements)} Elements](#{group.replace('.', '_')})\n" + + markdown_content += "\n" + + markdown_content += "### Tensor Data Offset\n" + markdown_content += '\n' + markdown_content += 'This table contains the offset and data segment relative to start of file\n' + markdown_content += '\n' + + tensor_mapping_table: list[dict[str, str | int]] = [] + for key, tensor in enumerate(reader.tensors): + data_offset_pretty = '{0:#16x}'.format(tensor.data_offset) + data_size_pretty = '{0:#16x}'.format(tensor.n_bytes) + tensor_mapping_table.append({"t_id":key, "layer_name":tensor.name, "data_offset":data_offset_pretty, "data_size":data_size_pretty}) + + tensors_mapping_table_header_map = [ + {'key_name':'t_id', 'header_name':'T_ID', 'align':'right'}, + {'key_name':'layer_name', 'header_name':'Tensor Layer Name', 'align':'left'}, + {'key_name':'data_offset', 'header_name':'Data Offset (B)', 'align':'right'}, + {'key_name':'data_size', 'header_name':'Data Size (B)', 'align':'right'}, + ] + + markdown_content += markdown_table_with_alignment_support(tensors_mapping_table_header_map, tensor_mapping_table) + markdown_content += "\n" + + for group in tensor_prefix_order: + tensors = tensor_groups[group] + group_elements = sum(tensor.n_elements for tensor in tensors) + group_percentage = group_elements / total_elements * 100 + markdown_content += f"### {translate_tensor_name(group)} Tensor Group : {element_count_rounded_notation(group_elements)} Elements\n\n" + + # Precalculate column sizing for visual consistency + prettify_element_est_count_size: int = 1 + prettify_element_count_size: int = 1 + prettify_dimension_max_widths: dict[int, int] = {} + for tensor in tensors: + prettify_element_est_count_size = max(prettify_element_est_count_size, len(str(element_count_rounded_notation(tensor.n_elements)))) + prettify_element_count_size = max(prettify_element_count_size, len(str(tensor.n_elements))) + for i, dimension_size in enumerate(list(tensor.shape) + [1] * (4 - len(tensor.shape))): + prettify_dimension_max_widths[i] = max(prettify_dimension_max_widths.get(i,1), len(str(dimension_size))) + + # Generate Tensor Layer Table Content + tensor_dump_table: list[dict[str, str | int]] = [] + for tensor in tensors: + human_friendly_name = translate_tensor_name(tensor.name.replace(".weight", ".(W)").replace(".bias", ".(B)")) + pretty_dimension = ' x '.join(f'{str(d):>{prettify_dimension_max_widths[i]}}' for i, d in enumerate(list(tensor.shape) + [1] * (4 - len(tensor.shape)))) + element_count_est = f"({element_count_rounded_notation(tensor.n_elements):>{prettify_element_est_count_size}})" + element_count_string = f"{element_count_est} {tensor.n_elements:>{prettify_element_count_size}}" + type_name_string = f"{tensor.tensor_type.name}" + tensor_dump_table.append({"t_id":tensor_name_to_key[tensor.name], "layer_name":tensor.name, "human_layer_name":human_friendly_name, "element_count":element_count_string, "pretty_dimension":pretty_dimension, "tensor_type":type_name_string}) + + tensor_dump_table_header_map = [ + {'key_name':'t_id', 'header_name':'T_ID', 'align':'right'}, + {'key_name':'layer_name', 'header_name':'Tensor Layer Name', 'align':'left'}, + {'key_name':'human_layer_name', 'header_name':'Human Friendly Tensor Layer Name', 'align':'left'}, + {'key_name':'element_count', 'header_name':'Elements', 'align':'left'}, + {'key_name':'pretty_dimension', 'header_name':'Shape', 'align':'left'}, + {'key_name':'tensor_type', 'header_name':'Type', 'align':'left'}, + ] + + markdown_content += markdown_table_with_alignment_support(tensor_dump_table_header_map, tensor_dump_table) + + markdown_content += "\n" + markdown_content += f"- Total elements in {group}: ({element_count_rounded_notation(group_elements):>4}) {group_elements}\n" + markdown_content += f"- Percentage of total elements: {group_percentage:.2f}%\n" + markdown_content += "\n\n" + + print(markdown_content) # noqa: NP100 + + +def main() -> None: + parser = argparse.ArgumentParser(description="Dump GGUF file metadata") + parser.add_argument("model", type=str, help="GGUF format model filename") + parser.add_argument("--no-tensors", action="store_true", help="Don't dump tensor metadata") + parser.add_argument("--json", action="store_true", help="Produce JSON output") + parser.add_argument("--json-array", action="store_true", help="Include full array values in JSON output (long)") + parser.add_argument("--data-offset", action="store_true", help="Start of data offset") + parser.add_argument("--data-alignment", action="store_true", help="Data alignment applied globally to data field") + parser.add_argument("--markdown", action="store_true", help="Produce markdown output") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") + + args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"]) + + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + + if not args.json and not args.markdown and not args.data_offset and not args.data_alignment: + logger.info(f'* Loading: {args.model}') + + reader = GGUFReader(args.model, 'r') + + if args.json: + dump_metadata_json(reader, args) + elif args.markdown: + dump_markdown_metadata(reader, args) + elif args.data_offset: + print(reader.data_offset) # noqa: NP100 + elif args.data_alignment: + print(reader.alignment) # noqa: NP100 + else: + dump_metadata(reader, args) + + +if __name__ == '__main__': + main() diff --git a/venv/lib/python3.10/site-packages/gguf/scripts/gguf_editor_gui.py b/venv/lib/python3.10/site-packages/gguf/scripts/gguf_editor_gui.py new file mode 100644 index 0000000000000000000000000000000000000000..05f4db0f8cdc89049a7e8719fcaa1fe741391413 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/scripts/gguf_editor_gui.py @@ -0,0 +1,1621 @@ +#!/usr/bin/env python3 +from __future__ import annotations + +import logging +import argparse +import os +import sys +import numpy +import enum +from pathlib import Path +from typing import Any, Optional, Tuple, Type +import warnings + +import numpy as np +from PySide6.QtWidgets import ( + QApplication, QMainWindow, QWidget, QVBoxLayout, QHBoxLayout, + QPushButton, QLabel, QLineEdit, QFileDialog, QTableWidget, + QTableWidgetItem, QComboBox, QMessageBox, QTabWidget, + QTextEdit, QFormLayout, + QHeaderView, QDialog, QDialogButtonBox +) +from PySide6.QtCore import Qt + +# Necessary to load the local gguf package +if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent.parent)) + +import gguf +from gguf import GGUFReader, GGUFWriter, GGUFValueType, ReaderField +from gguf.constants import TokenType, RopeScalingType, PoolingType, GGMLQuantizationType + +logger = logging.getLogger("gguf-editor-gui") + +# Map of key names to enum types for automatic enum interpretation +KEY_TO_ENUM_TYPE = { + gguf.Keys.Tokenizer.TOKEN_TYPE: TokenType, + gguf.Keys.Rope.SCALING_TYPE: RopeScalingType, + gguf.Keys.LLM.POOLING_TYPE: PoolingType, + gguf.Keys.General.FILE_TYPE: GGMLQuantizationType, +} + +# Define the tokenizer keys that should be edited together +TOKENIZER_LINKED_KEYS = [ + gguf.Keys.Tokenizer.LIST, + gguf.Keys.Tokenizer.TOKEN_TYPE, + gguf.Keys.Tokenizer.SCORES +] + + +class TokenizerEditorDialog(QDialog): + def __init__(self, tokens, token_types, scores, parent=None): + super().__init__(parent) + self.setWindowTitle("Edit Tokenizer Data") + self.resize(900, 600) + + self.tokens = tokens.copy() if tokens else [] + self.token_types = token_types.copy() if token_types else [] + self.scores = scores.copy() if scores else [] + + # Ensure all arrays have the same length + max_len = max(len(self.tokens), len(self.token_types), len(self.scores)) + if len(self.tokens) < max_len: + self.tokens.extend([""] * (max_len - len(self.tokens))) + if len(self.token_types) < max_len: + self.token_types.extend([0] * (max_len - len(self.token_types))) + if len(self.scores) < max_len: + self.scores.extend([0.0] * (max_len - len(self.scores))) + + layout = QVBoxLayout(self) + + # Add filter controls + filter_layout = QHBoxLayout() + filter_layout.addWidget(QLabel("Filter:")) + self.filter_edit = QLineEdit() + self.filter_edit.setPlaceholderText("Type to filter tokens...") + self.filter_edit.textChanged.connect(self.apply_filter) + filter_layout.addWidget(self.filter_edit) + + # Add page controls + self.page_size = 100 # Show 100 items per page + self.current_page = 0 + self.total_pages = max(1, (len(self.tokens) + self.page_size - 1) // self.page_size) + + self.page_label = QLabel(f"Page 1 of {self.total_pages}") + filter_layout.addWidget(self.page_label) + + prev_page = QPushButton("Previous") + prev_page.clicked.connect(self.previous_page) + filter_layout.addWidget(prev_page) + + next_page = QPushButton("Next") + next_page.clicked.connect(self.next_page) + filter_layout.addWidget(next_page) + + layout.addLayout(filter_layout) + + # Tokenizer data table + self.tokens_table = QTableWidget() + self.tokens_table.setColumnCount(4) + self.tokens_table.setHorizontalHeaderLabels(["Index", "Token", "Type", "Score"]) + self.tokens_table.horizontalHeader().setSectionResizeMode(0, QHeaderView.ResizeMode.ResizeToContents) + self.tokens_table.horizontalHeader().setSectionResizeMode(1, QHeaderView.ResizeMode.Stretch) + self.tokens_table.horizontalHeader().setSectionResizeMode(2, QHeaderView.ResizeMode.ResizeToContents) + self.tokens_table.horizontalHeader().setSectionResizeMode(3, QHeaderView.ResizeMode.ResizeToContents) + + layout.addWidget(self.tokens_table) + + # Controls + controls_layout = QHBoxLayout() + + add_button = QPushButton("Add Token") + add_button.clicked.connect(self.add_token) + controls_layout.addWidget(add_button) + + remove_button = QPushButton("Remove Selected") + remove_button.clicked.connect(self.remove_selected) + controls_layout.addWidget(remove_button) + + controls_layout.addStretch() + + layout.addLayout(controls_layout) + + # Buttons + buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel) + buttons.accepted.connect(self.accept) + buttons.rejected.connect(self.reject) + layout.addWidget(buttons) + + # Initialize the filtered values + self.filtered_indices = list(range(len(self.tokens))) + + # Load data for the first page + self.load_page() + + def apply_filter(self): + """Filter the tokens based on the search text.""" + filter_text = self.filter_edit.text().lower() + + if not filter_text: + # No filter, show all values + self.filtered_indices = list(range(len(self.tokens))) + else: + # Apply filter + self.filtered_indices = [] + for i, token in enumerate(self.tokens): + if filter_text in str(token).lower(): + self.filtered_indices.append(i) + + # Reset to first page and reload + self.total_pages = max(1, (len(self.filtered_indices) + self.page_size - 1) // self.page_size) + self.current_page = 0 + self.page_label.setText(f"Page 1 of {self.total_pages}") + self.load_page() + + def previous_page(self): + """Go to the previous page of results.""" + if self.current_page > 0: + self.current_page -= 1 + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + self.load_page() + + def next_page(self): + """Go to the next page of results.""" + if self.current_page < self.total_pages - 1: + self.current_page += 1 + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + self.load_page() + + def load_page(self): + """Load the current page of tokenizer data.""" + self.tokens_table.setRowCount(0) # Clear the table + + # Calculate start and end indices for the current page + start_idx = self.current_page * self.page_size + end_idx = min(start_idx + self.page_size, len(self.filtered_indices)) + + # Pre-allocate rows for better performance + self.tokens_table.setRowCount(end_idx - start_idx) + + for row, i in enumerate(range(start_idx, end_idx)): + orig_idx = self.filtered_indices[i] + + # Index + index_item = QTableWidgetItem(str(orig_idx)) + index_item.setData(Qt.ItemDataRole.UserRole, orig_idx) # Store original index + index_item.setFlags(index_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.tokens_table.setItem(row, 0, index_item) + + # Token + token_item = QTableWidgetItem(str(self.tokens[orig_idx])) + self.tokens_table.setItem(row, 1, token_item) + + # Token Type + token_type = self.token_types[orig_idx] if orig_idx < len(self.token_types) else 0 + try: + enum_val = TokenType(token_type) + display_text = f"{enum_val.name} ({token_type})" + except (ValueError, KeyError): + display_text = f"Unknown ({token_type})" + + type_item = QTableWidgetItem(display_text) + type_item.setData(Qt.ItemDataRole.UserRole, token_type) + + # Make type cell editable with a double-click handler + type_item.setFlags(type_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.tokens_table.setItem(row, 2, type_item) + + # Score + score = self.scores[orig_idx] if orig_idx < len(self.scores) else 0.0 + score_item = QTableWidgetItem(str(score)) + self.tokens_table.setItem(row, 3, score_item) + + # Connect double-click handler for token type cells + self.tokens_table.cellDoubleClicked.connect(self.handle_cell_double_click) + + def handle_cell_double_click(self, row, column): + """Handle double-click on a cell, specifically for token type editing.""" + if column == 2: # Token Type column + orig_item = self.tokens_table.item(row, 0) + if orig_item: + orig_idx = orig_item.data(Qt.ItemDataRole.UserRole) + self.edit_token_type(row, orig_idx) + + def edit_token_type(self, row, orig_idx): + """Edit a token type using a dialog with a dropdown of all enum options.""" + current_value = self.token_types[orig_idx] if orig_idx < len(self.token_types) else 0 + + # Create a dialog with enum options + dialog = QDialog(self) + dialog.setWindowTitle("Select Token Type") + layout = QVBoxLayout(dialog) + + combo = QComboBox() + for enum_val in TokenType: + combo.addItem(f"{enum_val.name} ({enum_val.value})", enum_val.value) + + # Set current value + try: + if isinstance(current_value, int): + enum_val = TokenType(current_value) + combo.setCurrentText(f"{enum_val.name} ({current_value})") + except (ValueError, KeyError): + pass + + layout.addWidget(combo) + + buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel) + buttons.accepted.connect(dialog.accept) + buttons.rejected.connect(dialog.reject) + layout.addWidget(buttons) + + if dialog.exec() == QDialog.DialogCode.Accepted: + # Get the selected value + new_value = combo.currentData() + enum_val = TokenType(new_value) + display_text = f"{enum_val.name} ({new_value})" + + # Update the display + type_item = self.tokens_table.item(row, 2) + if type_item: + type_item.setText(display_text) + type_item.setData(Qt.ItemDataRole.UserRole, new_value) + + # Update the actual value + self.token_types[orig_idx] = new_value + + def add_token(self): + """Add a new token to the end of the list.""" + # Add to the end of the arrays + self.tokens.append("") + self.token_types.append(0) # Default to normal token + self.scores.append(0.0) + + orig_idx = len(self.tokens) - 1 + + # Add to filtered indices if it matches the current filter + filter_text = self.filter_edit.text().lower() + if not filter_text or filter_text in "": + self.filtered_indices.append(orig_idx) + + # Update pagination + self.total_pages = max(1, (len(self.filtered_indices) + self.page_size - 1) // self.page_size) + + # Go to the last page to show the new item + self.current_page = self.total_pages - 1 + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + + # Reload the page + self.load_page() + + def remove_selected(self): + """Remove selected tokens from all arrays.""" + selected_rows = [] + for item in self.tokens_table.selectedItems(): + row = item.row() + if row not in selected_rows: + selected_rows.append(row) + + if not selected_rows: + return + + # Get original indices in descending order to avoid index shifting + orig_indices = [] + for row in selected_rows: + orig_item = self.tokens_table.item(row, 0) + if orig_item: + orig_indices.append(orig_item.data(Qt.ItemDataRole.UserRole)) + orig_indices.sort(reverse=True) + + # Remove from all arrays + for idx in orig_indices: + if idx < len(self.tokens): + del self.tokens[idx] + if idx < len(self.token_types): + del self.token_types[idx] + if idx < len(self.scores): + del self.scores[idx] + + # Rebuild filtered_indices + self.filtered_indices = [] + filter_text = self.filter_edit.text().lower() + + for i, token in enumerate(self.tokens): + if not filter_text or filter_text in str(token).lower(): + self.filtered_indices.append(i) + + # Update pagination + self.total_pages = max(1, (len(self.filtered_indices) + self.page_size - 1) // self.page_size) + self.current_page = min(self.current_page, self.total_pages - 1) + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + + # Reload the page + self.load_page() + + def get_data(self): + """Return the edited tokenizer data.""" + return self.tokens, self.token_types, self.scores + + +class ArrayEditorDialog(QDialog): + def __init__(self, array_values, element_type, key=None, parent=None): + super().__init__(parent) + self.setWindowTitle("Edit Array Values") + self.resize(700, 500) + + self.array_values = array_values + self.element_type = element_type + self.key = key + + # Get enum type for this array if applicable + self.enum_type = None + if key in KEY_TO_ENUM_TYPE and element_type == GGUFValueType.INT32: + self.enum_type = KEY_TO_ENUM_TYPE[key] + + layout = QVBoxLayout(self) + + # Add enum type information if applicable + if self.enum_type is not None: + enum_info_layout = QHBoxLayout() + enum_label = QLabel(f"Editing {self.enum_type.__name__} values:") + enum_info_layout.addWidget(enum_label) + + # Add a legend for the enum values + enum_values = ", ".join([f"{e.name}={e.value}" for e in self.enum_type]) + enum_values_label = QLabel(f"Available values: {enum_values}") + enum_values_label.setWordWrap(True) + enum_info_layout.addWidget(enum_values_label, 1) + + layout.addLayout(enum_info_layout) + + # Add search/filter controls + filter_layout = QHBoxLayout() + filter_layout.addWidget(QLabel("Filter:")) + self.filter_edit = QLineEdit() + self.filter_edit.setPlaceholderText("Type to filter values...") + self.filter_edit.textChanged.connect(self.apply_filter) + filter_layout.addWidget(self.filter_edit) + + # Add page controls for large arrays + self.page_size = 100 # Show 100 items per page + self.current_page = 0 + self.total_pages = max(1, (len(array_values) + self.page_size - 1) // self.page_size) + + self.page_label = QLabel(f"Page 1 of {self.total_pages}") + filter_layout.addWidget(self.page_label) + + prev_page = QPushButton("Previous") + prev_page.clicked.connect(self.previous_page) + filter_layout.addWidget(prev_page) + + next_page = QPushButton("Next") + next_page.clicked.connect(self.next_page) + filter_layout.addWidget(next_page) + + layout.addLayout(filter_layout) + + # Array items table + self.items_table = QTableWidget() + + # Set up columns based on whether we have an enum type + if self.enum_type is not None: + self.items_table.setColumnCount(3) + self.items_table.setHorizontalHeaderLabels(["Index", "Value", "Actions"]) + self.items_table.horizontalHeader().setSectionResizeMode(0, QHeaderView.ResizeMode.ResizeToContents) + self.items_table.horizontalHeader().setSectionResizeMode(1, QHeaderView.ResizeMode.Stretch) + self.items_table.horizontalHeader().setSectionResizeMode(2, QHeaderView.ResizeMode.ResizeToContents) + else: + self.items_table.setColumnCount(2) + self.items_table.setHorizontalHeaderLabels(["Index", "Value"]) + self.items_table.horizontalHeader().setSectionResizeMode(0, QHeaderView.ResizeMode.ResizeToContents) + self.items_table.horizontalHeader().setSectionResizeMode(1, QHeaderView.ResizeMode.Stretch) + + layout.addWidget(self.items_table) + + # Controls + controls_layout = QHBoxLayout() + + add_button = QPushButton("Add Item") + add_button.clicked.connect(self.add_item) + controls_layout.addWidget(add_button) + + remove_button = QPushButton("Remove Selected") + remove_button.clicked.connect(self.remove_selected) + controls_layout.addWidget(remove_button) + + # Add bulk edit button for enum arrays + if self.enum_type is not None: + bulk_edit_button = QPushButton("Bulk Edit Selected") + bulk_edit_button.clicked.connect(self.bulk_edit_selected) + controls_layout.addWidget(bulk_edit_button) + + controls_layout.addStretch() + + layout.addLayout(controls_layout) + + # Buttons + buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel) + buttons.accepted.connect(self.accept) + buttons.rejected.connect(self.reject) + layout.addWidget(buttons) + + # Initialize the filtered values + self.filtered_indices = list(range(len(self.array_values))) + + # Load array values for the first page + self.load_page() + + def apply_filter(self): + """Filter the array values based on the search text.""" + filter_text = self.filter_edit.text().lower() + + if not filter_text: + # No filter, show all values + self.filtered_indices = list(range(len(self.array_values))) + else: + # Apply filter + self.filtered_indices = [] + for i, value in enumerate(self.array_values): + # For enum values, search in both name and value + if self.enum_type is not None and isinstance(value, int): + try: + enum_val = self.enum_type(value) + display_text = f"{enum_val.name} ({value})".lower() + if filter_text in display_text: + self.filtered_indices.append(i) + except (ValueError, KeyError): + # If not a valid enum value, just check the raw value + if filter_text in str(value).lower(): + self.filtered_indices.append(i) + else: + # For non-enum values, just check the string representation + if filter_text in str(value).lower(): + self.filtered_indices.append(i) + + # Reset to first page and reload + self.total_pages = max(1, (len(self.filtered_indices) + self.page_size - 1) // self.page_size) + self.current_page = 0 + self.page_label.setText(f"Page 1 of {self.total_pages}") + self.load_page() + + def previous_page(self): + """Go to the previous page of results.""" + if self.current_page > 0: + self.current_page -= 1 + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + self.load_page() + + def next_page(self): + """Go to the next page of results.""" + if self.current_page < self.total_pages - 1: + self.current_page += 1 + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + self.load_page() + + def load_page(self): + """Load the current page of array values.""" + self.items_table.setRowCount(0) # Clear the table + + # Calculate start and end indices for the current page + start_idx = self.current_page * self.page_size + end_idx = min(start_idx + self.page_size, len(self.filtered_indices)) + + # Pre-allocate rows for better performance + self.items_table.setRowCount(end_idx - start_idx) + + for row, i in enumerate(range(start_idx, end_idx)): + orig_idx = self.filtered_indices[i] + value = self.array_values[orig_idx] + + # Index + index_item = QTableWidgetItem(str(orig_idx)) + index_item.setData(Qt.ItemDataRole.UserRole, orig_idx) # Store original index + index_item.setFlags(index_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.items_table.setItem(row, 0, index_item) + + # Value + if self.enum_type is not None: + # Display enum value and name + try: + if isinstance(value, (int, numpy.signedinteger)): + enum_val = self.enum_type(value) + display_text = f"{enum_val.name} ({value})" + else: + display_text = str(value) + except (ValueError, KeyError): + display_text = f"Unknown ({value})" + + # Store the enum value in the item + value_item = QTableWidgetItem(display_text) + value_item.setData(Qt.ItemDataRole.UserRole, value) + value_item.setFlags(value_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.items_table.setItem(row, 1, value_item) + + # Add an edit button in a separate column + edit_button = QPushButton("Edit") + edit_button.setProperty("row", row) + edit_button.clicked.connect(self.edit_array_enum_value) + + # Create a widget to hold the button + button_widget = QWidget() + button_layout = QHBoxLayout(button_widget) + button_layout.setContentsMargins(2, 2, 2, 2) + button_layout.addWidget(edit_button) + button_layout.addStretch() + + self.items_table.setCellWidget(row, 2, button_widget) + else: + value_item = QTableWidgetItem(str(value)) + self.items_table.setItem(row, 1, value_item) + + def edit_array_enum_value(self): + """Handle editing an enum value in the array editor.""" + button = self.sender() + row = button.property("row") + + # Get the original index from the table item + orig_item = self.items_table.item(row, 0) + new_item = self.items_table.item(row, 1) + if orig_item and new_item and self.enum_type and self.edit_enum_value(row, self.enum_type): + orig_idx = orig_item.data(Qt.ItemDataRole.UserRole) + new_value = new_item.data(Qt.ItemDataRole.UserRole) + # Update the stored value in the array + if isinstance(new_value, (int, float, str, bool)): + self.array_values[orig_idx] = new_value + + def bulk_edit_selected(self): + """Edit multiple enum values at once.""" + if not self.enum_type: + return + + selected_rows = set() + for item in self.items_table.selectedItems(): + selected_rows.add(item.row()) + + if not selected_rows: + QMessageBox.information(self, "No Selection", "Please select at least one row to edit.") + return + + # Create a dialog with enum options + dialog = QDialog(self) + dialog.setWindowTitle(f"Bulk Edit {self.enum_type.__name__} Values") + layout = QVBoxLayout(dialog) + + layout.addWidget(QLabel(f"Set {len(selected_rows)} selected items to:")) + + combo = QComboBox() + for enum_val in self.enum_type: + combo.addItem(f"{enum_val.name} ({enum_val.value})", enum_val.value) + + layout.addWidget(combo) + + buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel) + buttons.accepted.connect(dialog.accept) + buttons.rejected.connect(dialog.reject) + layout.addWidget(buttons) + + if dialog.exec() == QDialog.DialogCode.Accepted: + # Get the selected value + new_value = combo.currentData() + enum_val = self.enum_type(new_value) + display_text = f"{enum_val.name} ({new_value})" + + # Update all selected rows + for row in selected_rows: + orig_item = self.items_table.item(row, 0) + new_item = self.items_table.item(row, 1) + if orig_item and new_item: + orig_idx = orig_item.data(Qt.ItemDataRole.UserRole) + self.array_values[orig_idx] = new_value + + # Update the display + new_item.setText(display_text) + new_item.setData(Qt.ItemDataRole.UserRole, new_value) + + def add_item(self): + # Add to the end of the array + orig_idx = len(self.array_values) + + # Add default value based on type + if self.enum_type is not None: + # Default to first enum value + default_value = list(self.enum_type)[0].value + self.array_values.append(default_value) + else: + if self.element_type == GGUFValueType.STRING: + self.array_values.append("") + else: + self.array_values.append(0) + + # Add to filtered indices if it matches the current filter + self.filtered_indices.append(orig_idx) + + # Update pagination + self.total_pages = max(1, (len(self.filtered_indices) + self.page_size - 1) // self.page_size) + + # Go to the last page to show the new item + self.current_page = self.total_pages - 1 + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + + # Reload the page + self.load_page() + + def remove_selected(self): + selected_rows = [] + for item in self.items_table.selectedItems(): + row = item.row() + if row not in selected_rows: + selected_rows.append(row) + + if not selected_rows: + return + + # Get original indices in descending order to avoid index shifting + orig_indices = list() + for row in selected_rows: + orig_item = self.items_table.item(row, 0) + if orig_item: + orig_indices.append(orig_item.data(Qt.ItemDataRole.UserRole)) + orig_indices.sort(reverse=True) + + # Remove from array_values + for idx in orig_indices: + del self.array_values[idx] + + # Rebuild filtered_indices + self.filtered_indices = [] + filter_text = self.filter_edit.text().lower() + + for i, value in enumerate(self.array_values): + if not filter_text: + self.filtered_indices.append(i) + else: + # Apply filter + if self.enum_type is not None and isinstance(value, int): + try: + enum_val = self.enum_type(value) + display_text = f"{enum_val.name} ({value})".lower() + if filter_text in display_text: + self.filtered_indices.append(i) + except (ValueError, KeyError): + if filter_text in str(value).lower(): + self.filtered_indices.append(i) + else: + if filter_text in str(value).lower(): + self.filtered_indices.append(i) + + # Update pagination + self.total_pages = max(1, (len(self.filtered_indices) + self.page_size - 1) // self.page_size) + self.current_page = min(self.current_page, self.total_pages - 1) + self.page_label.setText(f"Page {self.current_page + 1} of {self.total_pages}") + + # Reload the page + self.load_page() + + def edit_enum_value(self, row: int, enum_type: Type[enum.Enum]): + """Edit an enum value using a dialog with a dropdown of all enum options.""" + # Get the original index from the table item + orig_item = self.items_table.item(row, 0) + if orig_item: + orig_idx = orig_item.data(Qt.ItemDataRole.UserRole) + else: + return + current_value = self.array_values[orig_idx] + + # Create a dialog with enum options + dialog = QDialog(self) + dialog.setWindowTitle(f"Select {enum_type.__name__} Value") + layout = QVBoxLayout(dialog) + + # Add description + description = QLabel(f"Select a {enum_type.__name__} value:") + layout.addWidget(description) + + # Use a combo box for quick selection + combo = QComboBox() + for enum_val in enum_type: + combo.addItem(f"{enum_val.name} ({enum_val.value})", enum_val.value) + + # Set current value + try: + if isinstance(current_value, int): + enum_val = enum_type(current_value) + combo.setCurrentText(f"{enum_val.name} ({current_value})") + except (ValueError, KeyError): + pass + + layout.addWidget(combo) + + buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel) + buttons.accepted.connect(dialog.accept) + buttons.rejected.connect(dialog.reject) + layout.addWidget(buttons) + + if dialog.exec() == QDialog.DialogCode.Accepted: + # Update the value display and stored data + new_value = combo.currentData() + enum_val = enum_type(new_value) + display_text = f"{enum_val.name} ({new_value})" + + new_item = self.items_table.item(row, 1) + if new_item: + new_item.setText(display_text) + new_item.setData(Qt.ItemDataRole.UserRole, new_value) + + # Update the actual array value + self.array_values[orig_idx] = new_value + return True + return False + + def get_array_values(self): + # The array_values list is kept up-to-date as edits are made + return self.array_values + + +class AddMetadataDialog(QDialog): + def __init__(self, parent=None): + super().__init__(parent) + self.setWindowTitle("Add Metadata") + self.resize(400, 200) + + layout = QVBoxLayout(self) + + form_layout = QFormLayout() + + self.key_edit = QLineEdit() + form_layout.addRow("Key:", self.key_edit) + + self.type_combo = QComboBox() + for value_type in GGUFValueType: + if value_type != GGUFValueType.ARRAY: # Skip array type for simplicity + self.type_combo.addItem(value_type.name, value_type) + form_layout.addRow("Type:", self.type_combo) + + self.value_edit = QTextEdit() + form_layout.addRow("Value:", self.value_edit) + + layout.addLayout(form_layout) + + buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel) + buttons.accepted.connect(self.accept) + buttons.rejected.connect(self.reject) + layout.addWidget(buttons) + + def get_data(self) -> Tuple[str, GGUFValueType, Any]: + key = self.key_edit.text() + value_type = self.type_combo.currentData() + value_text = self.value_edit.toPlainText() + + # Convert value based on type + if value_type == GGUFValueType.UINT8: + value = np.uint8(int(value_text)) + elif value_type == GGUFValueType.INT8: + value = np.int8(int(value_text)) + elif value_type == GGUFValueType.UINT16: + value = np.uint16(int(value_text)) + elif value_type == GGUFValueType.INT16: + value = np.int16(int(value_text)) + elif value_type == GGUFValueType.UINT32: + value = np.uint32(int(value_text)) + elif value_type == GGUFValueType.INT32: + value = np.int32(int(value_text)) + elif value_type == GGUFValueType.FLOAT32: + value = np.float32(float(value_text)) + elif value_type == GGUFValueType.BOOL: + value = value_text.lower() in ('true', 'yes', '1') + elif value_type == GGUFValueType.STRING: + value = value_text + else: + value = value_text + + return key, value_type, value + + +class GGUFEditorWindow(QMainWindow): + def __init__(self): + super().__init__() + + self.setWindowTitle("GGUF Editor") + self.resize(1000, 800) + + self.current_file = None + self.reader = None + self.modified = False + self.metadata_changes = {} # Store changes to apply when saving + self.metadata_to_remove = set() # Store keys to remove when saving + self.on_metadata_changed_is_connected = False + + self.setup_ui() + + def setup_ui(self): + central_widget = QWidget() + self.setCentralWidget(central_widget) + + main_layout = QVBoxLayout(central_widget) + + # File controls + file_layout = QHBoxLayout() + + self.file_path_edit = QLineEdit() + self.file_path_edit.setReadOnly(True) + file_layout.addWidget(self.file_path_edit) + + open_button = QPushButton("Open GGUF") + open_button.clicked.connect(self.open_file) + file_layout.addWidget(open_button) + + save_button = QPushButton("Save As...") + save_button.clicked.connect(self.save_file) + file_layout.addWidget(save_button) + + main_layout.addLayout(file_layout) + + # Tabs for different views + self.tabs = QTabWidget() + + # Metadata tab + self.metadata_tab = QWidget() + metadata_layout = QVBoxLayout(self.metadata_tab) + + # Metadata table + self.metadata_table = QTableWidget() + self.metadata_table.setColumnCount(4) + self.metadata_table.setHorizontalHeaderLabels(["Key", "Type", "Value", "Actions"]) + self.metadata_table.horizontalHeader().setSectionResizeMode(0, QHeaderView.ResizeMode.Stretch) + self.metadata_table.horizontalHeader().setSectionResizeMode(1, QHeaderView.ResizeMode.ResizeToContents) + self.metadata_table.horizontalHeader().setSectionResizeMode(2, QHeaderView.ResizeMode.Stretch) + self.metadata_table.horizontalHeader().setSectionResizeMode(3, QHeaderView.ResizeMode.ResizeToContents) + metadata_layout.addWidget(self.metadata_table) + + # Metadata controls + metadata_controls = QHBoxLayout() + + add_metadata_button = QPushButton("Add Metadata") + add_metadata_button.clicked.connect(self.add_metadata) + metadata_controls.addWidget(add_metadata_button) + + metadata_controls.addStretch() + + metadata_layout.addLayout(metadata_controls) + + # Tensors tab + self.tensors_tab = QWidget() + tensors_layout = QVBoxLayout(self.tensors_tab) + + self.tensors_table = QTableWidget() + self.tensors_table.setColumnCount(5) + self.tensors_table.setHorizontalHeaderLabels(["Name", "Type", "Shape", "Elements", "Size (bytes)"]) + self.tensors_table.horizontalHeader().setSectionResizeMode(0, QHeaderView.ResizeMode.Stretch) + self.tensors_table.horizontalHeader().setSectionResizeMode(1, QHeaderView.ResizeMode.ResizeToContents) + self.tensors_table.horizontalHeader().setSectionResizeMode(2, QHeaderView.ResizeMode.ResizeToContents) + self.tensors_table.horizontalHeader().setSectionResizeMode(3, QHeaderView.ResizeMode.ResizeToContents) + self.tensors_table.horizontalHeader().setSectionResizeMode(4, QHeaderView.ResizeMode.ResizeToContents) + tensors_layout.addWidget(self.tensors_table) + + # Add tabs to tab widget + self.tabs.addTab(self.metadata_tab, "Metadata") + self.tabs.addTab(self.tensors_tab, "Tensors") + + main_layout.addWidget(self.tabs) + + # Status bar + self.statusBar().showMessage("Ready") + + def load_file(self, file_path): + """Load a GGUF file by path""" + try: + self.statusBar().showMessage(f"Loading {file_path}...") + QApplication.processEvents() + + self.reader = GGUFReader(file_path, 'r') + self.current_file = file_path + self.file_path_edit.setText(file_path) + + self.load_metadata() + self.load_tensors() + + self.metadata_changes = {} + self.metadata_to_remove = set() + self.modified = False + + self.statusBar().showMessage(f"Loaded {file_path}") + return True + except Exception as e: + QMessageBox.critical(self, "Error", f"Failed to open file: {str(e)}") + self.statusBar().showMessage("Error loading file") + return False + + def open_file(self): + file_path, _ = QFileDialog.getOpenFileName( + self, "Open GGUF File", "", "GGUF Files (*.gguf);;All Files (*)" + ) + + if not file_path: + return + + self.load_file(file_path) + + def load_metadata(self): + self.metadata_table.setRowCount(0) + + if not self.reader: + return + + # Disconnect to prevent triggering during loading + if self.on_metadata_changed_is_connected: + with warnings.catch_warnings(): + warnings.filterwarnings('ignore') + self.metadata_table.itemChanged.disconnect(self.on_metadata_changed) + self.on_metadata_changed_is_connected = False + + for i, (key, field) in enumerate(self.reader.fields.items()): + self.metadata_table.insertRow(i) + + # Key + key_item = QTableWidgetItem(key) + key_item.setFlags(key_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.metadata_table.setItem(i, 0, key_item) + + # Type + if not field.types: + type_str = "N/A" + elif field.types[0] == GGUFValueType.ARRAY: + nest_count = len(field.types) - 1 + element_type = field.types[-1].name + # Check if this is an enum array + enum_type = self.get_enum_for_key(key) + if enum_type is not None and field.types[-1] == GGUFValueType.INT32: + element_type = enum_type.__name__ + type_str = '[' * nest_count + element_type + ']' * nest_count + else: + type_str = str(field.types[0].name) + # Check if this is an enum field + enum_type = self.get_enum_for_key(key) + if enum_type is not None and field.types[0] == GGUFValueType.INT32: + type_str = enum_type.__name__ + + type_item = QTableWidgetItem(type_str) + type_item.setFlags(type_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.metadata_table.setItem(i, 1, type_item) + + # Value + value_str = self.format_field_value(field) + value_item = QTableWidgetItem(value_str) + + # Make only simple values editable + if len(field.types) == 1 and field.types[0] != GGUFValueType.ARRAY: + value_item.setFlags(value_item.flags() | Qt.ItemFlag.ItemIsEditable) + else: + value_item.setFlags(value_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + + self.metadata_table.setItem(i, 2, value_item) + + # Actions + actions_widget = QWidget() + actions_layout = QHBoxLayout(actions_widget) + actions_layout.setContentsMargins(2, 2, 2, 2) + + # Add Edit button for arrays and enum fields + if field.types and field.types[0] == GGUFValueType.ARRAY: + edit_button = QPushButton("Edit") + edit_button.setProperty("row", i) + edit_button.setProperty("key", key) + edit_button.clicked.connect(self.edit_array_metadata) + actions_layout.addWidget(edit_button) + + # Add special label for tokenizer linked fields + if key in TOKENIZER_LINKED_KEYS: + edit_button.setText("Edit Tokenizer") + edit_button.setToolTip("Edit all tokenizer data together") + elif len(field.types) == 1 and self.get_enum_for_key(key) is not None: + edit_button = QPushButton("Edit") + edit_button.setProperty("row", i) + edit_button.setProperty("key", key) + edit_button.clicked.connect(self.edit_metadata_enum) + actions_layout.addWidget(edit_button) + + remove_button = QPushButton("Remove") + remove_button.setProperty("row", i) + remove_button.setProperty("key", key) + remove_button.clicked.connect(self.remove_metadata) + actions_layout.addWidget(remove_button) + + self.metadata_table.setCellWidget(i, 3, actions_widget) + + # Reconnect after loading + self.metadata_table.itemChanged.connect(self.on_metadata_changed) + self.on_metadata_changed_is_connected = True + + def extract_array_values(self, field: ReaderField) -> list: + """Extract all values from an array field.""" + if not field.types or field.types[0] != GGUFValueType.ARRAY: + return [] + + curr_type = field.types[1] + array_values = [] + total_elements = len(field.data) + + if curr_type == GGUFValueType.STRING: + for element_pos in range(total_elements): + value_string = str(bytes(field.parts[-1 - (total_elements - element_pos - 1) * 2]), encoding='utf-8') + array_values.append(value_string) + elif self.reader and curr_type in self.reader.gguf_scalar_to_np: + for element_pos in range(total_elements): + array_values.append(field.parts[-1 - (total_elements - element_pos - 1)][0]) + + return array_values + + def get_enum_for_key(self, key: str) -> Optional[Type[enum.Enum]]: + """Get the enum type for a given key if it exists.""" + return KEY_TO_ENUM_TYPE.get(key) + + def format_enum_value(self, value: Any, enum_type: Type[enum.Enum]) -> str: + """Format a value as an enum if possible.""" + try: + if isinstance(value, (int, str)): + enum_value = enum_type(value) + return f"{enum_value.name} ({value})" + except (ValueError, KeyError): + pass + return str(value) + + def format_field_value(self, field: ReaderField) -> str: + if not field.types: + return "N/A" + + if len(field.types) == 1: + curr_type = field.types[0] + if curr_type == GGUFValueType.STRING: + return str(bytes(field.parts[-1]), encoding='utf-8') + elif self.reader and curr_type in self.reader.gguf_scalar_to_np: + value = field.parts[-1][0] + # Check if this field has an enum type + enum_type = self.get_enum_for_key(field.name) + if enum_type is not None: + return self.format_enum_value(value, enum_type) + return str(value) + + if field.types[0] == GGUFValueType.ARRAY: + array_values = self.extract_array_values(field) + render_element = min(5, len(array_values)) + + # Get enum type for this array if applicable + enum_type = self.get_enum_for_key(field.name) + + if enum_type is not None: + array_elements = [] + for i in range(render_element): + array_elements.append(self.format_enum_value(array_values[i], enum_type)) + else: + array_elements = [str(array_values[i]) for i in range(render_element)] + + return f"[ {', '.join(array_elements).strip()}{', ...' if len(array_values) > len(array_elements) else ''} ]" + + return "Complex value" + + def load_tensors(self): + self.tensors_table.setRowCount(0) + + if not self.reader: + return + + for i, tensor in enumerate(self.reader.tensors): + self.tensors_table.insertRow(i) + + # Name + name_item = QTableWidgetItem(tensor.name) + name_item.setFlags(name_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.tensors_table.setItem(i, 0, name_item) + + # Type + type_item = QTableWidgetItem(tensor.tensor_type.name) + type_item.setFlags(type_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.tensors_table.setItem(i, 1, type_item) + + # Shape + shape_str = " × ".join(str(d) for d in tensor.shape) + shape_item = QTableWidgetItem(shape_str) + shape_item.setFlags(shape_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.tensors_table.setItem(i, 2, shape_item) + + # Elements + elements_item = QTableWidgetItem(str(tensor.n_elements)) + elements_item.setFlags(elements_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.tensors_table.setItem(i, 3, elements_item) + + # Size + size_item = QTableWidgetItem(f"{tensor.n_bytes:,}") + size_item.setFlags(size_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.tensors_table.setItem(i, 4, size_item) + + def on_metadata_changed(self, item): + if item.column() != 2: # Only handle value column changes + return + + row = item.row() + orig_item = self.metadata_table.item(row, 0) + key = None + if orig_item: + key = orig_item.text() + new_value = item.text() + + field = None + if self.reader and key: + field = self.reader.get_field(key) + if not field or not field.types or not key: + return + + value_type = field.types[0] + + # Check if this is an enum field + enum_type = self.get_enum_for_key(key) + if enum_type is not None and value_type == GGUFValueType.INT32: + # Try to parse the enum value from the text + try: + # Check if it's a name + try: + enum_val = enum_type[new_value] + converted_value = enum_val.value + except (KeyError, AttributeError): + # Check if it's a number or "NAME (value)" format + if '(' in new_value and ')' in new_value: + # Extract the value from "NAME (value)" format + value_part = new_value.split('(')[1].split(')')[0].strip() + converted_value = int(value_part) + else: + # Try to convert directly to int + converted_value = int(new_value) + + # Validate that it's a valid enum value + enum_type(converted_value) + + # Store the change + self.metadata_changes[key] = (value_type, converted_value) + self.modified = True + + # Update display with formatted enum value + formatted_value = self.format_enum_value(converted_value, enum_type) + item.setText(formatted_value) + + self.statusBar().showMessage(f"Changed {key} to {formatted_value}") + return + except (ValueError, KeyError) as e: + QMessageBox.warning( + self, + f"Invalid Enum Value ({e})", + f"'{new_value}' is not a valid {enum_type.__name__} value.\n" + f"Valid values are: {', '.join(v.name for v in enum_type)}") + + # Revert to original value + original_value = self.format_field_value(field) + item.setText(original_value) + return + + try: + # Convert the string value to the appropriate type + if value_type == GGUFValueType.UINT8: + converted_value = np.uint8(int(new_value)) + elif value_type == GGUFValueType.INT8: + converted_value = np.int8(int(new_value)) + elif value_type == GGUFValueType.UINT16: + converted_value = np.uint16(int(new_value)) + elif value_type == GGUFValueType.INT16: + converted_value = np.int16(int(new_value)) + elif value_type == GGUFValueType.UINT32: + converted_value = np.uint32(int(new_value)) + elif value_type == GGUFValueType.INT32: + converted_value = np.int32(int(new_value)) + elif value_type == GGUFValueType.FLOAT32: + converted_value = np.float32(float(new_value)) + elif value_type == GGUFValueType.BOOL: + converted_value = new_value.lower() in ('true', 'yes', '1') + elif value_type == GGUFValueType.STRING: + converted_value = new_value + else: + # Unsupported type for editing + return + + # Store the change + self.metadata_changes[key] = (value_type, converted_value) + self.modified = True + + self.statusBar().showMessage(f"Changed {key} to {new_value}") + except ValueError: + QMessageBox.warning(self, "Invalid Value", f"The value '{new_value}' is not valid for type {value_type.name}") + + # Revert to original value + original_value = self.format_field_value(field) + item.setText(original_value) + + def remove_metadata(self): + button = self.sender() + key = button.property("key") + row = button.property("row") + + reply = QMessageBox.question( + self, "Confirm Removal", + f"Are you sure you want to remove the metadata key '{key}'?", + QMessageBox.StandardButton.Yes | QMessageBox.StandardButton.No, QMessageBox.StandardButton.No + ) + + if reply == QMessageBox.StandardButton.Yes: + self.metadata_table.removeRow(row) + self.metadata_to_remove.add(key) + + # If we previously had changes for this key, remove them + if key in self.metadata_changes: + del self.metadata_changes[key] + + self.modified = True + self.statusBar().showMessage(f"Marked {key} for removal") + + def edit_metadata_enum(self): + """Edit an enum metadata field.""" + button = self.sender() + key = button.property("key") + row = button.property("row") + + field = None + if self.reader: + field = self.reader.get_field(key) + if not field or not field.types: + return + + enum_type = self.get_enum_for_key(key) + if enum_type is None: + return + + # Get current value + current_value = field.contents() + + # Create a dialog with enum options + dialog = QDialog(self) + dialog.setWindowTitle(f"Select {enum_type.__name__} Value") + layout = QVBoxLayout(dialog) + + combo = QComboBox() + for enum_val in enum_type: + combo.addItem(f"{enum_val.name} ({enum_val.value})", enum_val.value) + + # Set current value + try: + if isinstance(current_value, (int, str)): + enum_val = enum_type(current_value) + combo.setCurrentText(f"{enum_val.name} ({current_value})") + except (ValueError, KeyError): + pass + + layout.addWidget(combo) + + buttons = QDialogButtonBox(QDialogButtonBox.StandardButton.Ok | QDialogButtonBox.StandardButton.Cancel) + buttons.accepted.connect(dialog.accept) + buttons.rejected.connect(dialog.reject) + layout.addWidget(buttons) + + if dialog.exec() == QDialog.DialogCode.Accepted: + # Get the selected value + new_value = combo.currentData() + enum_val = enum_type(new_value) + + # Store the change + self.metadata_changes[key] = (field.types[0], new_value) + self.modified = True + + # Update display + display_text = f"{enum_val.name} ({new_value})" + target_item = self.metadata_table.item(row, 2) + if target_item: + target_item.setText(display_text) + + self.statusBar().showMessage(f"Changed {key} to {display_text}") + + def edit_array_metadata(self): + button = self.sender() + key = button.property("key") + row = button.property("row") + + # Check if this is one of the linked tokenizer keys + if key in TOKENIZER_LINKED_KEYS: + self.edit_tokenizer_metadata(key) + return + + field = None + if self.reader: + field = self.reader.get_field(key) + if not field or not field.types or field.types[0] != GGUFValueType.ARRAY: + return + + # Get array element type + element_type = field.types[1] + + # Extract array values + array_values = self.extract_array_values(field) + + # Open array editor dialog + dialog = ArrayEditorDialog(array_values, element_type, key, self) + if dialog.exec() == QDialog.DialogCode.Accepted: + new_values = dialog.get_array_values() + + # Store the change + self.metadata_changes[key] = (GGUFValueType.ARRAY, (element_type, new_values)) + self.modified = True + + # Update display + enum_type = self.get_enum_for_key(key) + if enum_type is not None and element_type == GGUFValueType.INT32: + value_str = f"[ {', '.join(self.format_enum_value(v, enum_type) for v in new_values[:5])}{', ...' if len(new_values) > 5 else ''} ]" + else: + value_str = f"[ {', '.join(str(v) for v in new_values[:5])}{', ...' if len(new_values) > 5 else ''} ]" + target_item = self.metadata_table.item(row, 2) + if target_item: + target_item.setText(value_str) + + self.statusBar().showMessage(f"Updated array values for {key}") + + def edit_tokenizer_metadata(self, trigger_key): + """Edit the linked tokenizer metadata arrays together.""" + if not self.reader: + return + + # Get all three fields + tokens_field = self.reader.get_field(gguf.Keys.Tokenizer.LIST) + token_types_field = self.reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE) + scores_field = self.reader.get_field(gguf.Keys.Tokenizer.SCORES) + + # Extract values from each field + tokens = self.extract_array_values(tokens_field) if tokens_field else [] + token_types = self.extract_array_values(token_types_field) if token_types_field else [] + scores = self.extract_array_values(scores_field) if scores_field else [] + + # Apply any pending changes + if gguf.Keys.Tokenizer.LIST in self.metadata_changes: + _, (_, tokens) = self.metadata_changes[gguf.Keys.Tokenizer.LIST] + if gguf.Keys.Tokenizer.TOKEN_TYPE in self.metadata_changes: + _, (_, token_types) = self.metadata_changes[gguf.Keys.Tokenizer.TOKEN_TYPE] + if gguf.Keys.Tokenizer.SCORES in self.metadata_changes: + _, (_, scores) = self.metadata_changes[gguf.Keys.Tokenizer.SCORES] + + # Open the tokenizer editor dialog + dialog = TokenizerEditorDialog(tokens, token_types, scores, self) + if dialog.exec() == QDialog.DialogCode.Accepted: + new_tokens, new_token_types, new_scores = dialog.get_data() + + # Store changes for all three arrays + if tokens_field: + self.metadata_changes[gguf.Keys.Tokenizer.LIST] = ( + GGUFValueType.ARRAY, + (tokens_field.types[1], new_tokens) + ) + + if token_types_field: + self.metadata_changes[gguf.Keys.Tokenizer.TOKEN_TYPE] = ( + GGUFValueType.ARRAY, + (token_types_field.types[1], new_token_types) + ) + + if scores_field: + self.metadata_changes[gguf.Keys.Tokenizer.SCORES] = ( + GGUFValueType.ARRAY, + (scores_field.types[1], new_scores) + ) + + self.modified = True + + # Update display for all three fields + self.update_tokenizer_display(gguf.Keys.Tokenizer.LIST, new_tokens) + self.update_tokenizer_display(gguf.Keys.Tokenizer.TOKEN_TYPE, new_token_types) + self.update_tokenizer_display(gguf.Keys.Tokenizer.SCORES, new_scores) + + self.statusBar().showMessage("Updated tokenizer data") + + def update_tokenizer_display(self, key, values): + """Update the display of a tokenizer field in the metadata table.""" + for row in range(self.metadata_table.rowCount()): + key_item = self.metadata_table.item(row, 0) + if key_item and key_item.text() == key: + value_str = f"[ {', '.join(str(v) for v in values[:5])}{', ...' if len(values) > 5 else ''} ]" + value_item = self.metadata_table.item(row, 2) + if value_item: + value_item.setText(value_str) + break + + def add_metadata(self): + dialog = AddMetadataDialog(self) + if dialog.exec() == QDialog.DialogCode.Accepted: + key, value_type, value = dialog.get_data() + + if not key: + QMessageBox.warning(self, "Invalid Key", "Key cannot be empty") + return + + # Check if key already exists + for row in range(self.metadata_table.rowCount()): + orig_item = self.metadata_table.item(row, 0) + if orig_item and orig_item.text() == key: + QMessageBox.warning(self, "Duplicate Key", f"Key '{key}' already exists") + return + + # Add to table + row = self.metadata_table.rowCount() + self.metadata_table.insertRow(row) + + # Key + key_item = QTableWidgetItem(key) + key_item.setFlags(key_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.metadata_table.setItem(row, 0, key_item) + + # Type + type_item = QTableWidgetItem(value_type.name) + type_item.setFlags(type_item.flags() & ~Qt.ItemFlag.ItemIsEditable) + self.metadata_table.setItem(row, 1, type_item) + + # Value + value_item = QTableWidgetItem(str(value)) + value_item.setFlags(value_item.flags() | Qt.ItemFlag.ItemIsEditable) + self.metadata_table.setItem(row, 2, value_item) + + # Actions + actions_widget = QWidget() + actions_layout = QHBoxLayout(actions_widget) + actions_layout.setContentsMargins(2, 2, 2, 2) + + remove_button = QPushButton("Remove") + remove_button.setProperty("row", row) + remove_button.setProperty("key", key) + remove_button.clicked.connect(self.remove_metadata) + actions_layout.addWidget(remove_button) + + self.metadata_table.setCellWidget(row, 3, actions_widget) + + # Store the change + self.metadata_changes[key] = (value_type, value) + self.modified = True + + self.statusBar().showMessage(f"Added new metadata key {key}") + + def save_file(self): + if not self.reader: + QMessageBox.warning(self, "No File Open", "Please open a GGUF file first") + return + + if not self.modified and not self.metadata_changes and not self.metadata_to_remove: + QMessageBox.information(self, "No Changes", "No changes to save") + return + + file_path, _ = QFileDialog.getSaveFileName( + self, "Save GGUF File As", "", "GGUF Files (*.gguf);;All Files (*)" + ) + + if not file_path: + return + + try: + self.statusBar().showMessage(f"Saving to {file_path}...") + QApplication.processEvents() + + # Get architecture and endianness from the original file + arch = 'unknown' + field = self.reader.get_field(gguf.Keys.General.ARCHITECTURE) + if field: + arch = field.contents() + + # Create writer + writer = GGUFWriter(file_path, arch=arch, endianess=self.reader.endianess) + + # Get alignment if present + alignment = None + field = self.reader.get_field(gguf.Keys.General.ALIGNMENT) + if field: + alignment = field.contents() + if alignment is not None: + writer.data_alignment = alignment + + # Copy metadata with changes + for field in self.reader.fields.values(): + # Skip virtual fields and fields written by GGUFWriter + if field.name == gguf.Keys.General.ARCHITECTURE or field.name.startswith('GGUF.'): + continue + + # Skip fields marked for removal + if field.name in self.metadata_to_remove: + continue + + # Apply changes if any + sub_type = None + if field.name in self.metadata_changes: + value_type, value = self.metadata_changes[field.name] + if value_type == GGUFValueType.ARRAY: + # Handle array values + sub_type, value = value + else: + # Copy original value + value = field.contents() + value_type = field.types[0] + if value_type == GGUFValueType.ARRAY: + sub_type = field.types[-1] + + if value is not None: + writer.add_key_value(field.name, value, value_type, sub_type=sub_type) + + # Add new metadata + for key, (value_type, value) in self.metadata_changes.items(): + # Skip if the key already existed (we handled it above) + if self.reader.get_field(key) is not None: + continue + + sub_type = None + if value_type == GGUFValueType.ARRAY: + # Handle array values + sub_type, value = value + + writer.add_key_value(key, value, value_type, sub_type=sub_type) + + # Add tensors (including data) + for tensor in self.reader.tensors: + writer.add_tensor(tensor.name, tensor.data, raw_shape=tensor.data.shape, raw_dtype=tensor.tensor_type) + + # Write header and metadata + writer.open_output_file(Path(file_path)) + writer.write_header_to_file() + writer.write_kv_data_to_file() + + # Write tensor data using the optimized method + writer.write_tensors_to_file(progress=False) + + writer.close() + + self.statusBar().showMessage(f"Saved to {file_path}") + + # Ask if user wants to open the new file + reply = QMessageBox.question( + self, "Open Saved File", + "Would you like to open the newly saved file?", + QMessageBox.StandardButton.Yes | QMessageBox.StandardButton.No, QMessageBox.StandardButton.Yes + ) + + if reply == QMessageBox.StandardButton.Yes: + self.reader = GGUFReader(file_path, 'r') + self.current_file = file_path + self.file_path_edit.setText(file_path) + + self.load_metadata() + self.load_tensors() + + self.metadata_changes = {} + self.metadata_to_remove = set() + self.modified = False + + except Exception as e: + QMessageBox.critical(self, "Error", f"Failed to save file: {str(e)}") + self.statusBar().showMessage("Error saving file") + + +def main() -> None: + parser = argparse.ArgumentParser(description="GUI GGUF Editor") + parser.add_argument("model_path", nargs="?", help="path to GGUF model file to load at startup") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") + + args = parser.parse_args() + + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + + app = QApplication(sys.argv) + window = GGUFEditorWindow() + window.show() + + # Load model if specified + if args.model_path: + if os.path.isfile(args.model_path) and args.model_path.endswith('.gguf'): + window.load_file(args.model_path) + else: + logger.error(f"Invalid model path: {args.model_path}") + QMessageBox.warning( + window, + "Invalid Model Path", + f"The specified file does not exist or is not a GGUF file: {args.model_path}") + + sys.exit(app.exec()) + + +if __name__ == '__main__': + main() diff --git a/venv/lib/python3.10/site-packages/gguf/scripts/gguf_hash.py b/venv/lib/python3.10/site-packages/gguf/scripts/gguf_hash.py new file mode 100644 index 0000000000000000000000000000000000000000..3ef98992197e987d8aefc8fb5f5bf9c787e9d06a --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/scripts/gguf_hash.py @@ -0,0 +1,102 @@ +#!/usr/bin/env python3 +from __future__ import annotations + +import uuid +import hashlib + +import logging +import argparse +import os +import sys +from pathlib import Path + +from tqdm import tqdm + +# Necessary to load the local gguf package +if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent.parent)) + +from gguf import GGUFReader # noqa: E402 + + +logger = logging.getLogger("gguf-hash") + +# UUID_NAMESPACE_LLAMA_CPP = uuid.uuid5(uuid.NAMESPACE_URL, 'en.wikipedia.org/wiki/Llama.cpp') +UUID_NAMESPACE_LLAMA_CPP = uuid.UUID('ef001206-dadc-5f6d-a15f-3359e577d4e5') + + +# For more information about what field.parts and field.data represent, +# please see the comments in the modify_gguf.py example. +def gguf_hash(reader: GGUFReader, filename: str, disable_progress_bar: bool, no_layer: bool) -> None: + sha1 = hashlib.sha1() + sha256 = hashlib.sha256() + uuidv5_sha1 = hashlib.sha1() + uuidv5_sha1.update(UUID_NAMESPACE_LLAMA_CPP.bytes) + + # Total Weight Calculation For Progress Bar + total_weights = 0 + for n, tensor in enumerate(reader.tensors, 1): + + # We don't need these + if tensor.name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")): + continue + + # Calculate Tensor Volume + sum_weights_in_tensor = 1 + for dim in tensor.shape: + sum_weights_in_tensor *= dim + total_weights += sum_weights_in_tensor + + # Hash Progress Bar + bar = tqdm(desc="Hashing", total=total_weights, unit="weights", unit_scale=True, disable=disable_progress_bar) + + # Hashing Process + for tensor in reader.tensors: + + # We don't need these + if tensor.name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")): + continue + + # Progressbar + sum_weights_in_tensor = 1 + for dim in tensor.shape: + sum_weights_in_tensor *= dim + bar.update(sum_weights_in_tensor) + + if not no_layer: + + sha1_layer = hashlib.sha1() + sha1_layer.update(tensor.data.data) + print("sha1 {0} {1}:{2}".format(sha1_layer.hexdigest(), filename, tensor.name)) # noqa: NP100 + + sha256_layer = hashlib.sha256() + sha256_layer.update(tensor.data.data) + print("sha256 {0} {1}:{2}".format(sha256_layer.hexdigest(), filename, tensor.name)) # noqa: NP100 + + sha1.update(tensor.data.data) + sha256.update(tensor.data.data) + uuidv5_sha1.update(tensor.data.data) + + # Flush Hash Progress Bar + bar.close() + + # Display Hash Output + print("sha1 {0} {1}".format(sha1.hexdigest(), filename)) # noqa: NP100 + print("sha256 {0} {1}".format(sha256.hexdigest(), filename)) # noqa: NP100 + print("uuid {0} {1}".format(uuid.UUID(bytes=uuidv5_sha1.digest()[:16], version=5), filename)) # noqa: NP100 + + +def main() -> None: + parser = argparse.ArgumentParser(description="Dump GGUF file metadata") + parser.add_argument("model", type=str, help="GGUF format model filename") + parser.add_argument("--no-layer", action="store_true", help="exclude per layer hash") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") + parser.add_argument("--progressbar", action="store_true", help="enable progressbar") + args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"]) + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + reader = GGUFReader(args.model, 'r') + gguf_hash(reader, args.model, not args.progressbar, args.no_layer) + + +if __name__ == '__main__': + main() diff --git a/venv/lib/python3.10/site-packages/gguf/scripts/gguf_new_metadata.py b/venv/lib/python3.10/site-packages/gguf/scripts/gguf_new_metadata.py new file mode 100644 index 0000000000000000000000000000000000000000..63f2300348ed0ffad3bb9a4be6aa7aae6b827350 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/scripts/gguf_new_metadata.py @@ -0,0 +1,210 @@ +#!/usr/bin/env python3 +from __future__ import annotations + +import logging +import argparse +import os +import sys +import json +from pathlib import Path + +from tqdm import tqdm +from typing import Any, Sequence, NamedTuple + +# Necessary to load the local gguf package +if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent.parent)) + +import gguf + +logger = logging.getLogger("gguf-new-metadata") + + +class MetadataDetails(NamedTuple): + type: gguf.GGUFValueType + value: Any + description: str = '' + sub_type: gguf.GGUFValueType | None = None + + +def get_field_data(reader: gguf.GGUFReader, key: str) -> Any: + field = reader.get_field(key) + + return field.contents() if field else None + + +def find_token(token_list: Sequence[int], token: str) -> Sequence[int]: + token_ids = [index for index, value in enumerate(token_list) if value == token] + + if len(token_ids) == 0: + raise LookupError(f'Unable to find "{token}" in token list!') + + return token_ids + + +def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new_metadata: dict[str, MetadataDetails], remove_metadata: Sequence[str]) -> None: + for field in reader.fields.values(): + # Suppress virtual fields and fields written by GGUFWriter + if field.name == gguf.Keys.General.ARCHITECTURE or field.name.startswith('GGUF.'): + logger.debug(f'Suppressing {field.name}') + continue + + # Skip old chat templates if we have new ones + if field.name.startswith(gguf.Keys.Tokenizer.CHAT_TEMPLATE) and gguf.Keys.Tokenizer.CHAT_TEMPLATE in new_metadata: + logger.debug(f'Skipping {field.name}') + continue + + if field.name in remove_metadata: + logger.debug(f'Removing {field.name}') + continue + + val_type = field.types[0] + sub_type = field.types[-1] if val_type == gguf.GGUFValueType.ARRAY else None + old_val = MetadataDetails(val_type, field.contents(), sub_type=sub_type) + val = new_metadata.get(field.name, old_val) + + if field.name in new_metadata: + logger.debug(f'Modifying {field.name}: "{old_val.value}" -> "{val.value}" {val.description}') + del new_metadata[field.name] + elif val.value is not None: + logger.debug(f'Copying {field.name}') + + if val.value is not None: + writer.add_key_value(field.name, val.value, val.type, sub_type=sub_type if val.sub_type is None else val.sub_type) + + if gguf.Keys.Tokenizer.CHAT_TEMPLATE in new_metadata: + logger.debug('Adding chat template(s)') + writer.add_chat_template(new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE].value) + del new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] + + for key, val in new_metadata.items(): + logger.debug(f'Adding {key}: "{val.value}" {val.description}') + writer.add_key_value(key, val.value, val.type) + + total_bytes = 0 + + for tensor in reader.tensors: + total_bytes += tensor.n_bytes + writer.add_tensor_info(tensor.name, tensor.data.shape, tensor.data.dtype, tensor.data.nbytes, tensor.tensor_type) + + bar = tqdm(desc="Writing", total=total_bytes, unit="byte", unit_scale=True) + + writer.write_header_to_file() + writer.write_kv_data_to_file() + writer.write_ti_data_to_file() + + for tensor in reader.tensors: + writer.write_tensor_data(tensor.data) + bar.update(tensor.n_bytes) + + writer.close() + + +def main() -> None: + tokenizer_metadata = (getattr(gguf.Keys.Tokenizer, n) for n in gguf.Keys.Tokenizer.__dict__.keys() if not n.startswith('_')) + token_names = dict((n.split('.')[-1][:-len('_token_id')], n) for n in tokenizer_metadata if n.endswith('_token_id')) + + parser = argparse.ArgumentParser(description="Make a copy of a GGUF file with new metadata") + parser.add_argument("input", type=Path, help="GGUF format model input filename") + parser.add_argument("output", type=Path, help="GGUF format model output filename") + parser.add_argument("--general-name", type=str, help="The models general.name", metavar='"name"') + parser.add_argument("--general-description", type=str, help="The models general.description", metavar='"Description ..."') + parser.add_argument("--chat-template", type=str, help="Chat template string (or JSON string containing templates)", metavar='"{% ... %} ..."') + parser.add_argument("--chat-template-config", type=Path, help="Config file containing chat template(s)", metavar='tokenizer_config.json') + parser.add_argument("--pre-tokenizer", type=str, help="The models tokenizer.ggml.pre", metavar='"pre tokenizer"') + parser.add_argument("--remove-metadata", action="append", type=str, help="Remove metadata (by key name) from output model", metavar='general.url') + parser.add_argument("--special-token", action="append", type=str, help="Special token by value", nargs=2, metavar=(' | '.join(token_names.keys()), '""')) + parser.add_argument("--special-token-by-id", action="append", type=str, help="Special token by id", nargs=2, metavar=(' | '.join(token_names.keys()), '0')) + parser.add_argument("--force", action="store_true", help="Bypass warnings without confirmation") + parser.add_argument("--verbose", action="store_true", help="Increase output verbosity") + args = parser.parse_args(None if len(sys.argv) > 2 else ["--help"]) + + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + + new_metadata = {} + remove_metadata = args.remove_metadata or [] + + if args.general_name: + new_metadata[gguf.Keys.General.NAME] = MetadataDetails(gguf.GGUFValueType.STRING, args.general_name) + + if args.general_description: + new_metadata[gguf.Keys.General.DESCRIPTION] = MetadataDetails(gguf.GGUFValueType.STRING, args.general_description) + + if args.chat_template: + new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, json.loads(args.chat_template) if args.chat_template.startswith('[') else args.chat_template) + + if args.chat_template_config: + with open(args.chat_template_config, 'r') as fp: + config = json.load(fp) + template = config.get('chat_template') + if template: + new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, template) + + if args.pre_tokenizer: + new_metadata[gguf.Keys.Tokenizer.PRE] = MetadataDetails(gguf.GGUFValueType.STRING, args.pre_tokenizer) + + if remove_metadata: + logger.warning('*** Warning *** Warning *** Warning **') + logger.warning('* Most metadata is required for a fully functional GGUF file,') + logger.warning('* removing crucial metadata may result in a corrupt output file!') + + if not args.force: + logger.warning('* Enter exactly YES if you are positive you want to proceed:') + response = input('YES, I am sure> ') + if response != 'YES': + logger.info("You didn't enter YES. Okay then, see ya!") + sys.exit(0) + + logger.info(f'* Loading: {args.input}') + reader = gguf.GGUFReader(args.input, 'r') + + arch = get_field_data(reader, gguf.Keys.General.ARCHITECTURE) + + token_list = get_field_data(reader, gguf.Keys.Tokenizer.LIST) or [] + + for name, token in args.special_token or []: + if name not in token_names: + logger.warning(f'Unknown special token "{name}", ignoring...') + else: + ids = find_token(token_list, token) + new_metadata[token_names[name]] = MetadataDetails(gguf.GGUFValueType.UINT32, ids[0], f'= {token}') + + if len(ids) > 1: + logger.warning(f'Multiple "{token}" tokens found, choosing ID {ids[0]}, use --special-token-by-id if you want another:') + logger.warning(', '.join(str(i) for i in ids)) + + for name, id_string in args.special_token_by_id or []: + if name not in token_names: + logger.warning(f'Unknown special token "{name}", ignoring...') + elif not id_string.isdecimal(): + raise LookupError(f'Token ID "{id_string}" is not a valid ID!') + else: + id_int = int(id_string) + + if id_int >= 0 and id_int < len(token_list): + new_metadata[token_names[name]] = MetadataDetails(gguf.GGUFValueType.UINT32, id_int, f'= {token_list[id_int]}') + else: + raise LookupError(f'Token ID {id_int} is not within token list!') + + if os.path.isfile(args.output) and not args.force: + logger.warning('*** Warning *** Warning *** Warning **') + logger.warning(f'* The "{args.output}" GGUF file already exists, it will be overwritten!') + logger.warning('* Enter exactly YES if you are positive you want to proceed:') + response = input('YES, I am sure> ') + if response != 'YES': + logger.info("You didn't enter YES. Okay then, see ya!") + sys.exit(0) + + logger.info(f'* Writing: {args.output}') + writer = gguf.GGUFWriter(args.output, arch=arch, endianess=reader.endianess) + + alignment = get_field_data(reader, gguf.Keys.General.ALIGNMENT) + if alignment is not None: + logger.debug(f'Setting custom alignment: {alignment}') + writer.data_alignment = alignment + + copy_with_new_metadata(reader, writer, new_metadata, remove_metadata) + + +if __name__ == '__main__': + main() diff --git a/venv/lib/python3.10/site-packages/gguf/scripts/gguf_set_metadata.py b/venv/lib/python3.10/site-packages/gguf/scripts/gguf_set_metadata.py new file mode 100644 index 0000000000000000000000000000000000000000..f5809c35c887086f8055e828719041f8ae3ea749 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/scripts/gguf_set_metadata.py @@ -0,0 +1,95 @@ +#!/usr/bin/env python3 +import logging +import argparse +import os +import sys +from pathlib import Path + +# Necessary to load the local gguf package +if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent.parent)) + +from gguf import GGUFReader # noqa: E402 + +logger = logging.getLogger("gguf-set-metadata") + + +def minimal_example(filename: str) -> None: + reader = GGUFReader(filename, 'r+') + field = reader.fields['tokenizer.ggml.bos_token_id'] + if field is None: + return + part_index = field.data[0] + field.parts[part_index][0] = 2 # Set tokenizer.ggml.bos_token_id to 2 + # + # So what's this field.data thing? It's helpful because field.parts contains + # _every_ part of the GGUF field. For example, tokenizer.ggml.bos_token_id consists + # of: + # + # Part index 0: Key length (27) + # Part index 1: Key data ("tokenizer.ggml.bos_token_id") + # Part index 2: Field type (4, the id for GGUFValueType.UINT32) + # Part index 3: Field value + # + # Note also that each part is an NDArray slice, so even a part that + # is only a single value like the key length will be a NDArray of + # the key length type (numpy.uint32). + # + # The .data attribute in the Field is a list of relevant part indexes + # and doesn't contain internal GGUF details like the key length part. + # In this case, .data will be [3] - just the part index of the + # field value itself. + + +def set_metadata(reader: GGUFReader, args: argparse.Namespace) -> None: + field = reader.get_field(args.key) + if field is None: + logger.error(f'! Field {repr(args.key)} not found') + sys.exit(1) + # Note that field.types is a list of types. This is because the GGUF + # format supports arrays. For example, an array of UINT32 would + # look like [GGUFValueType.ARRAY, GGUFValueType.UINT32] + handler = reader.gguf_scalar_to_np.get(field.types[0]) if field.types else None + if handler is None: + logger.error(f'! This tool only supports changing simple values, {repr(args.key)} has unsupported type {field.types}') + sys.exit(1) + current_value = field.parts[field.data[0]][0] + new_value = handler(args.value) + logger.info(f'* Preparing to change field {repr(args.key)} from {current_value} to {new_value}') + if current_value == new_value: + logger.info(f'- Key {repr(args.key)} already set to requested value {current_value}') + sys.exit(0) + if args.dry_run: + sys.exit(0) + if not args.force: + logger.warning('*** Warning *** Warning *** Warning **') + logger.warning('* Changing fields in a GGUF file can make it unusable. Proceed at your own risk.') + logger.warning('* Enter exactly YES if you are positive you want to proceed:') + response = input('YES, I am sure> ') + if response != 'YES': + logger.info("You didn't enter YES. Okay then, see ya!") + sys.exit(0) + field.parts[field.data[0]][0] = new_value + logger.info('* Field changed. Successful completion.') + + +def main() -> None: + parser = argparse.ArgumentParser(description="Set a simple value in GGUF file metadata") + parser.add_argument("model", type=str, help="GGUF format model filename") + parser.add_argument("key", type=str, help="Metadata key to set") + parser.add_argument("value", type=str, help="Metadata value to set") + parser.add_argument("--dry-run", action="store_true", help="Don't actually change anything") + parser.add_argument("--force", action="store_true", help="Change the field without confirmation") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") + + args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"]) + + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + + logger.info(f'* Loading: {args.model}') + reader = GGUFReader(args.model, 'r' if args.dry_run else 'r+') + set_metadata(reader, args) + + +if __name__ == '__main__': + main() diff --git a/venv/lib/python3.10/site-packages/gguf/tensor_mapping.py b/venv/lib/python3.10/site-packages/gguf/tensor_mapping.py new file mode 100644 index 0000000000000000000000000000000000000000..79f044d2a5945236b613ac3733ccc2b60ebd9ba4 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/tensor_mapping.py @@ -0,0 +1,1280 @@ +from __future__ import annotations + +from typing import Sequence + +from .constants import MODEL_ARCH, MODEL_TENSOR, MODEL_TENSORS, TENSOR_NAMES + + +class TensorNameMap: + mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = { + # Token embeddings + MODEL_TENSOR.TOKEN_EMBD: ( + "gpt_neox.embed_in", # gptneox + "transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx jais exaone + "transformer.word_embeddings", # falcon + "word_embeddings", # bloom + "model.embed_tokens", # llama-hf nemotron olmoe olmo2 rwkv6qwen2 glm4-0414 + "tok_embeddings", # llama-pth + "embeddings.word_embeddings", # bert nomic-bert + "language_model.embedding.word_embeddings", # persimmon + "wte", # gpt2 + "transformer.embd.wte", # phi2 + "model.tok_embeddings", # internlm2 + "model.embedding", # mamba-qbert + "backbone.embedding", # mamba + "backbone.embeddings", # mamba-hf + "transformer.in_out_embed", # Grok + "embedding.word_embeddings", # chatglm + "transformer.token_embeddings", # openelm + "shared", # t5 + "rwkv.embeddings", # rwkv6 + "model.embeddings", # rwkv7 + "model.word_embeddings", # bailingmoe + "language_model.model.embed_tokens", # llama4 + "encoder", # neobert + ), + + # Token type embeddings + MODEL_TENSOR.TOKEN_TYPES: ( + "embeddings.token_type_embeddings", # bert nomic-bert + ), + + # Normalization of token embeddings + MODEL_TENSOR.TOKEN_EMBD_NORM: ( + "word_embeddings_layernorm", # bloom + "embeddings.LayerNorm", # bert + "emb_ln", # nomic-bert + "transformer.norm", # openelm + "rwkv.blocks.0.pre_ln", # rwkv + "rwkv.blocks.0.pre_ln", # rwkv6 + "model.pre_ln", # rwkv7 + "model.layers.0.pre_norm", # rwkv7 + "backbone.norm", # wavtokenizer + ), + + # Position embeddings + MODEL_TENSOR.POS_EMBD: ( + "transformer.wpe", # gpt2 + "embeddings.position_embeddings", # bert + "wpe", # gpt2 + ), + + # Output + MODEL_TENSOR.OUTPUT: ( + "embed_out", # gptneox + "lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe olmo2 phimoe + "output", # llama-pth bloom internlm2 + "word_embeddings_for_head", # persimmon + "lm_head.linear", # phi2 + "output_layer", # chatglm + "head", # rwkv + "head.out", # wavtokenizer + "lm_head", # llama4 + ), + + # Output norm + MODEL_TENSOR.OUTPUT_NORM: ( + "gpt_neox.final_layer_norm", # gptneox + "transformer.ln_f", # gpt2 gpt-j falcon jais exaone + "model.norm", # llama-hf baichuan internlm2 olmoe olmo2 phimoe + "norm", # llama-pth + "transformer.norm_f", # mpt dbrx + "ln_f", # refact bloom qwen gpt2 + "language_model.encoder.final_layernorm", # persimmon + "model.final_layernorm", # persimmon + "lm_head.ln", # phi2 + "model.norm_f", # mamba-qbert + "backbone.norm_f", # mamba + "transformer.rms_norm", # Grok + "encoder.final_layernorm", # chatglm + "transformer.norm", # openelm + "model.norm", # nemotron + "rwkv.ln_out", # rwkv6 + "model.ln_out", # rwkv7 + "backbone.final_layer_norm", # wavtokenizer + "model.norm", # llama4 + ), + + # Rope frequencies + MODEL_TENSOR.ROPE_FREQS: ( + "rope.freqs", # llama-pth + "rotary_pos_emb.inv_freq", # chatglm + ), + + MODEL_TENSOR.ROPE_FACTORS_LONG: (), + MODEL_TENSOR.ROPE_FACTORS_SHORT: (), + + MODEL_TENSOR.CONV1D: ( + "backbone.embed", # roberta + ), + } + + block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = { + # Attention norm + MODEL_TENSOR.ATTN_NORM: ( + "gpt_neox.layers.{bid}.input_layernorm", # gptneox + "transformer.h.{bid}.ln_1", # gpt2 gpt-j refact qwen jais exaone + "transformer.blocks.{bid}.norm_1", # mpt + "transformer.h.{bid}.input_layernorm", # falcon7b + "h.{bid}.input_layernorm", # bloom + "transformer.h.{bid}.ln_mlp", # falcon40b + "model.layers.{bid}.input_layernorm", # llama-hf nemotron olmoe phimoe + "layers.{bid}.attention_norm", # llama-pth + "language_model.encoder.layers.{bid}.input_layernorm", # persimmon + "model.layers.{bid}.ln1", # yi + "h.{bid}.ln_1", # gpt2 + "transformer.h.{bid}.ln", # phi2 + "model.layers.layers.{bid}.norm", # plamo + "model.layers.{bid}.attention_norm", # internlm2 + "model.layers.{bid}.norm", # mamba-qbert + "backbone.layers.{bid}.norm", # mamba + "transformer.decoder_layer.{bid}.rms_norm", # Grok + "transformer.blocks.{bid}.norm_attn_norm.norm_1", # dbrx + "encoder.layers.{bid}.input_layernorm", # chatglm + "transformer.layers.{bid}.attn_norm", # openelm + "rwkv.blocks.{bid}.ln1", # rwkv6 + "model.layers.{bid}.ln1", # rwkv7 + "model.layers.{bid}.input_layernorm", # llama4 + "transformer_encoder.{bid}.attention_norm", # neobert + ), + + # Attention norm 2 + MODEL_TENSOR.ATTN_NORM_2: ( + "transformer.h.{bid}.ln_attn", # falcon40b + "encoder.layer.{bid}.layer_norm_1", # jina-v2-code + "rwkv.blocks.{bid}.ln2", # rwkv6 + "model.layers.{bid}.ln2", # rwkv7 + ), + + # Attention query-key-value + MODEL_TENSOR.ATTN_QKV: ( + "gpt_neox.layers.{bid}.attention.query_key_value", # gptneox + "transformer.h.{bid}.attn.c_attn", # gpt2 qwen jais + "transformer.blocks.{bid}.attn.Wqkv", # mpt + "transformer.blocks.{bid}.norm_attn_norm.attn.Wqkv", # dbrx + "transformer.h.{bid}.self_attention.query_key_value", # falcon + "h.{bid}.self_attention.query_key_value", # bloom + "language_model.encoder.layers.{bid}.self_attention.query_key_value", # persimmon + "model.layers.{bid}.self_attn.query_key_value", # persimmon + "h.{bid}.attn.c_attn", # gpt2 + "transformer.h.{bid}.mixer.Wqkv", # phi2 + "encoder.layers.{bid}.attn.Wqkv", # nomic-bert + "encoder.layers.{bid}.mixer.Wqkv", # jina + "model.layers.{bid}.self_attn.qkv_proj", # phi3 + "encoder.layers.{bid}.self_attention.query_key_value", # chatglm + "transformer.layers.{bid}.attn.qkv_proj", # openelm + "transformer_encoder.{bid}.qkv", # neobert + ), + + # Attention query + MODEL_TENSOR.ATTN_Q: ( + "model.layers.{bid}.self_attn.q_proj", # llama-hf nemotron olmoe olmo2 phimoe + "model.layers.{bid}.self_attn.q_proj_no_perm", # llama-custom + "layers.{bid}.attention.wq", # llama-pth + "encoder.layer.{bid}.attention.self.query", # bert + "transformer.layer.{bid}.attention.q_lin", # distillbert + "transformer.h.{bid}.attn.q_proj", # gpt-j + "model.layers.layers.{bid}.self_attn.q_proj", # plamo + "model.layers.{bid}.attention.wq", # internlm2 + "transformer.decoder_layer.{bid}.multi_head_attention.query",# Grok + "transformer.h.{bid}.attn.attention.q_proj", # exaone + "model.layers.{bid}.self_attn.q_proj", # llama4 + ), + + # Attention key + MODEL_TENSOR.ATTN_K: ( + "model.layers.{bid}.self_attn.k_proj", # llama-hf nemotron olmoe olmo2 phimoe + "model.layers.{bid}.self_attn.k_proj_no_perm", # llama-custom + "layers.{bid}.attention.wk", # llama-pth + "encoder.layer.{bid}.attention.self.key", # bert + "transformer.layer.{bid}.attention.k_lin", # distillbert + "transformer.h.{bid}.attn.k_proj", # gpt-j + "transformer.h.{bid}.attn.k", # refact + "model.layers.layers.{bid}.self_attn.k_proj", # plamo + "model.layers.{bid}.attention.wk", # internlm2 + "transformer.decoder_layer.{bid}.multi_head_attention.key",# Grok + "transformer.h.{bid}.attn.attention.k_proj", # exaone + "model.layers.{bid}.self_attn.k_proj", # llama4 + ), + + # Attention value + MODEL_TENSOR.ATTN_V: ( + "model.layers.{bid}.self_attn.v_proj", # llama-hf nemotron olmoe olmo2 phimoe + "layers.{bid}.attention.wv", # llama-pth + "encoder.layer.{bid}.attention.self.value", # bert + "transformer.layer.{bid}.attention.v_lin", # distillbert + "transformer.h.{bid}.attn.v_proj", # gpt-j + "transformer.h.{bid}.attn.v", # refact + "model.layers.layers.{bid}.self_attn.v_proj", # plamo + "model.layers.{bid}.attention.wv", # internlm2 + "transformer.decoder_layer.{bid}.multi_head_attention.value",# Grok + "transformer.h.{bid}.attn.attention.v_proj", # exaone + "model.layers.{bid}.self_attn.v_proj", # llama4 + ), + + # Attention output + MODEL_TENSOR.ATTN_OUT: ( + "gpt_neox.layers.{bid}.attention.dense", # gptneox + "transformer.h.{bid}.attn.c_proj", # gpt2 refact qwen jais + "transformer.blocks.{bid}.attn.out_proj", # mpt + "transformer.h.{bid}.self_attention.dense", # falcon + "h.{bid}.self_attention.dense", # bloom + "model.layers.{bid}.self_attn.o_proj", # llama-hf nemotron olmoe olmo2 phimoe + "model.layers.{bid}.self_attn.linear_attn", # deci + "layers.{bid}.attention.wo", # llama-pth + "encoder.layer.{bid}.attention.output.dense", # bert + "transformer.layer.{bid}.attention.out_lin", # distillbert + "transformer.h.{bid}.attn.out_proj", # gpt-j + "language_model.encoder.layers.{bid}.self_attention.dense", # persimmon + "model.layers.{bid}.self_attn.dense", # persimmon + "h.{bid}.attn.c_proj", # gpt2 + "transformer.h.{bid}.mixer.out_proj", # phi2 + "model.layers.layers.{bid}.self_attn.o_proj", # plamo + "model.layers.{bid}.attention.wo", # internlm2 + "encoder.layers.{bid}.attn.out_proj", # nomic-bert + "encoder.layers.{bid}.mixer.out_proj", # jina + "transformer.decoder_layer.{bid}.multi_head_attention.linear", # Grok + "transformer.blocks.{bid}.norm_attn_norm.attn.out_proj", # dbrx + "encoder.layers.{bid}.self_attention.dense", # chatglm + "transformer.layers.{bid}.attn.out_proj", # openelm + "transformer.h.{bid}.attn.attention.out_proj", # exaone + "model.layers.{bid}.self_attn.o_proj", # llama4 + "transformer_encoder.{bid}.wo", # neobert + ), + + # Attention output norm + MODEL_TENSOR.ATTN_OUT_NORM: ( + "encoder.layer.{bid}.attention.output.LayerNorm", # bert + "transformer.layer.{bid}.sa_layer_norm", # distillbert + "encoder.layers.{bid}.norm1", # nomic-bert + "transformer.decoder_layer.{bid}.rms_norm_1", # Grok + "transformer.blocks.{bid}.norm_attn_norm.norm_2", # dbrx + ), + + MODEL_TENSOR.ATTN_POST_NORM: ( + "model.layers.{bid}.post_attention_layernorm", # gemma2 olmo2 # ge + "model.layers.{bid}.post_self_attn_layernorm", # glm-4-0414 + ), + + # Rotary embeddings + MODEL_TENSOR.ATTN_ROT_EMBD: ( + "model.layers.{bid}.self_attn.rotary_emb.inv_freq", # llama-hf + "layers.{bid}.attention.inner_attention.rope.freqs", # llama-pth + "model.layers.layers.{bid}.self_attn.rotary_emb.inv_freq", # plamo + "transformer.h.{bid}.attn.rotary_emb.inv_freq", # codeshell + ), + + # Feed-forward norm + MODEL_TENSOR.FFN_NORM: ( + "gpt_neox.layers.{bid}.post_attention_layernorm", # gptneox + "transformer.h.{bid}.ln_2", # gpt2 refact qwen jais exaone + "h.{bid}.post_attention_layernorm", # bloom + "transformer.blocks.{bid}.norm_2", # mpt + "model.layers.{bid}.post_attention_layernorm", # llama-hf nemotron olmoe phimoe + "layers.{bid}.ffn_norm", # llama-pth + "language_model.encoder.layers.{bid}.post_attention_layernorm", # persimmon + "model.layers.{bid}.ln2", # yi + "h.{bid}.ln_2", # gpt2 + "model.layers.{bid}.ffn_norm", # internlm2 + "transformer.decoder_layer.{bid}.rms_norm_2", # Grok + "encoder.layers.{bid}.post_attention_layernorm", # chatglm + "transformer.layers.{bid}.ffn_norm", # openelm + "model.layers.{bid}.post_attention_layernorm", # llama4 + "transformer_encoder.{bid}.ffn_norm", # neobert + ), + + # Post feed-forward norm + MODEL_TENSOR.FFN_PRE_NORM: ( + "model.layers.{bid}.pre_feedforward_layernorm", # gemma2 + ), + + # Post feed-forward norm + MODEL_TENSOR.FFN_POST_NORM: ( + "model.layers.{bid}.post_feedforward_layernorm", # gemma2 olmo2 + "model.layers.{bid}.post_mlp_layernorm", # glm-4-0414 + ), + + MODEL_TENSOR.FFN_GATE_INP: ( + "layers.{bid}.feed_forward.gate", # mixtral + "model.layers.{bid}.block_sparse_moe.gate", # mixtral phimoe + "model.layers.{bid}.mlp.gate", # qwen2moe olmoe + "transformer.decoder_layer.{bid}.router", # Grok + "transformer.blocks.{bid}.ffn.router.layer", # dbrx + "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe + "model.layers.{bid}.feed_forward.router", # llama4 + "encoder.layers.{bid}.mlp.router.layer", # nomic-bert-moe + ), + + MODEL_TENSOR.FFN_GATE_INP_SHEXP: ( + "model.layers.{bid}.mlp.shared_expert_gate", # qwen2moe + ), + + MODEL_TENSOR.FFN_EXP_PROBS_B: ( + "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3 dots1 + ), + + # Feed-forward up + MODEL_TENSOR.FFN_UP: ( + "gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox + "transformer.h.{bid}.mlp.c_fc", # gpt2 jais + "transformer.blocks.{bid}.ffn.up_proj", # mpt + "transformer.h.{bid}.mlp.dense_h_to_4h", # falcon + "h.{bid}.mlp.dense_h_to_4h", # bloom + "model.layers.{bid}.mlp.up_proj", # llama-hf refact nemotron olmo2 + "layers.{bid}.feed_forward.w3", # llama-pth + "encoder.layer.{bid}.intermediate.dense", # bert + "transformer.layer.{bid}.ffn.lin1", # distillbert + "transformer.h.{bid}.mlp.fc_in", # gpt-j + "transformer.h.{bid}.mlp.linear_3", # refact + "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon + "model.layers.{bid}.mlp.dense_h_to_4h", # persimmon + "transformer.h.{bid}.mlp.w1", # qwen + "h.{bid}.mlp.c_fc", # gpt2 + "transformer.h.{bid}.mlp.fc1", # phi2 + "model.layers.{bid}.mlp.fc1", # phi2 + "model.layers.{bid}.mlp.gate_up_proj", # phi3 glm-4-0414 + "model.layers.layers.{bid}.mlp.up_proj", # plamo + "model.layers.{bid}.feed_forward.w3", # internlm2 + "encoder.layers.{bid}.mlp.fc11", # nomic-bert + "encoder.layers.{bid}.mlp.fc1", # nomic-bert-moe + "model.layers.{bid}.mlp.c_fc", # starcoder2 + "encoder.layer.{bid}.mlp.gated_layers_v", # jina-bert-v2 (split up/gate, no longer used) + "encoder.layer.{bid}.mlp.gated_layers", # jina-bert-v2 (GEGLU) + "encoder.layer.{bid}.mlp.up_gated_layer", # jina-v2-code (GEGLU) + "model.layers.{bid}.residual_mlp.w3", # arctic + "encoder.layers.{bid}.mlp.dense_h_to_4h", # chatglm + "transformer.h.{bid}.mlp.c_fc_1", # exaone + "model.layers.{bid}.feed_forward.up_proj", # llama4 + "transformer_encoder.{bid}.ffn.w12", # neobert + ), + + MODEL_TENSOR.FFN_UP_EXP: ( + "layers.{bid}.feed_forward.experts.w3", # mixtral (merged) + "transformer.decoder_layer.{bid}.moe.linear_v", # Grok (merged) + "transformer.blocks.{bid}.ffn.experts.mlp.v1", # dbrx + "model.layers.{bid}.mlp.experts.up_proj", # qwen2moe olmoe (merged) + "model.layers.{bid}.block_sparse_moe.experts.w3", # phimoe (merged) + "model.layers.{bid}.feed_forward.experts.up_proj", # llama4 + "encoder.layers.{bid}.mlp.experts.mlp.w1", # nomic-bert-moe + ), + + MODEL_TENSOR.FFN_UP_SHEXP: ( + "model.layers.{bid}.mlp.shared_expert.up_proj", # qwen2moe + "model.layers.{bid}.mlp.shared_experts.up_proj", # deepseek deepseek2 + "model.layers.{bid}.feed_forward.shared_expert.up_proj", # llama4 + ), + + # AWQ-activation gate + MODEL_TENSOR.FFN_ACT: ( + "transformer.blocks.{bid}.ffn.act", # mpt + ), + + # Feed-forward gate + MODEL_TENSOR.FFN_GATE: ( + "model.layers.{bid}.mlp.gate_proj", # llama-hf refact olmo2 + "layers.{bid}.feed_forward.w1", # llama-pth + "transformer.h.{bid}.mlp.w2", # qwen + "transformer.h.{bid}.mlp.c_fc2", # jais + "model.layers.layers.{bid}.mlp.gate_proj", # plamo + "model.layers.{bid}.feed_forward.w1", # internlm2 + "encoder.layers.{bid}.mlp.fc12", # nomic-bert + "encoder.layer.{bid}.mlp.gated_layers_w", # jina-bert-v2 (split up/gate, no longer used) + "transformer.h.{bid}.mlp.linear_1", # refact + "model.layers.{bid}.residual_mlp.w1", # arctic + "transformer.h.{bid}.mlp.c_fc_0", # exaone + "model.layers.{bid}.feed_forward.gate_proj", # llama4 + ), + + MODEL_TENSOR.FFN_GATE_EXP: ( + "layers.{bid}.feed_forward.experts.w1", # mixtral (merged) + "transformer.decoder_layer.{bid}.moe.linear", # Grok (merged) + "transformer.blocks.{bid}.ffn.experts.mlp.w1", # dbrx + "model.layers.{bid}.mlp.experts.gate_proj", # qwen2moe olmoe (merged) + "model.layers.{bid}.block_sparse_moe.experts.w1", # phimoe (merged) + "model.layers.{bid}.feed_forward.experts.gate_proj", # llama4 + ), + + MODEL_TENSOR.FFN_GATE_SHEXP: ( + "model.layers.{bid}.mlp.shared_expert.gate_proj", # qwen2moe + "model.layers.{bid}.mlp.shared_experts.gate_proj", # deepseek deepseek2 + "model.layers.{bid}.feed_forward.shared_expert.gate_proj", # llama4 + ), + + # Feed-forward down + MODEL_TENSOR.FFN_DOWN: ( + "gpt_neox.layers.{bid}.mlp.dense_4h_to_h", # gptneox + "transformer.h.{bid}.mlp.c_proj", # gpt2 refact qwen jais + "transformer.blocks.{bid}.ffn.down_proj", # mpt + "transformer.h.{bid}.mlp.dense_4h_to_h", # falcon + "h.{bid}.mlp.dense_4h_to_h", # bloom + "model.layers.{bid}.mlp.down_proj", # llama-hf nemotron olmo2 + "layers.{bid}.feed_forward.w2", # llama-pth + "encoder.layer.{bid}.output.dense", # bert + "transformer.layer.{bid}.ffn.lin2", # distillbert + "transformer.h.{bid}.mlp.fc_out", # gpt-j + "language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon + "model.layers.{bid}.mlp.dense_4h_to_h", # persimmon + "h.{bid}.mlp.c_proj", # gpt2 + "transformer.h.{bid}.mlp.fc2", # phi2 + "model.layers.{bid}.mlp.fc2", # phi2 + "model.layers.layers.{bid}.mlp.down_proj", # plamo + "model.layers.{bid}.feed_forward.w2", # internlm2 + "encoder.layers.{bid}.mlp.fc2", # nomic-bert + "model.layers.{bid}.mlp.c_proj", # starcoder2 + "encoder.layer.{bid}.mlp.wo", # jina-bert-v2 + "transformer.layers.{bid}.ffn.proj_2", # openelm + "model.layers.{bid}.residual_mlp.w2", # arctic + "encoder.layer.{bid}.mlp.down_layer", # jina-bert-v2 + "encoder.layers.{bid}.mlp.dense_4h_to_h", # chatglm + "model.layers.h.{bid}.mlp.c_proj", # exaone + "model.layers.{bid}.feed_forward.down_proj", # llama4 + "transformer_encoder.{bid}.ffn.w3", # neobert + ), + + MODEL_TENSOR.FFN_DOWN_EXP: ( + "layers.{bid}.feed_forward.experts.w2", # mixtral (merged) + "transformer.decoder_layer.{bid}.moe.linear_1", # Grok (merged) + "transformer.blocks.{bid}.ffn.experts.mlp.w2", # dbrx + "model.layers.{bid}.mlp.experts.down_proj", # qwen2moe olmoe (merged) + "model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe + "model.layers.{bid}.block_sparse_moe.experts.w2", # phimoe (merged) + "model.layers.{bid}.feed_forward.experts.down_proj", # llama4 + "encoder.layers.{bid}.mlp.experts.mlp.w2", # nomic-bert-moe + ), + + MODEL_TENSOR.FFN_DOWN_SHEXP: ( + "model.layers.{bid}.mlp.shared_expert.down_proj", # qwen2moe + "model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek deepseek2 + "model.layers.{bid}.feed_forward.shared_expert.down_proj", # llama4 + "model.layers.{bid}.shared_mlp.output_linear", # granitemoe + ), + + MODEL_TENSOR.ATTN_Q_NORM: ( + "language_model.encoder.layers.{bid}.self_attention.q_layernorm", + "model.layers.{bid}.self_attn.q_layernorm", # persimmon + "model.layers.{bid}.self_attn.q_norm", # cohere olmoe chameleon olmo2 + "transformer.blocks.{bid}.attn.q_ln", # sea-lion + "encoder.layer.{bid}.attention.self.layer_norm_q", # jina-bert-v2 + "transformer.layers.{bid}.attn.q_norm", # openelm + ), + + MODEL_TENSOR.ATTN_K_NORM: ( + "language_model.encoder.layers.{bid}.self_attention.k_layernorm", + "model.layers.{bid}.self_attn.k_layernorm", # persimmon + "model.layers.{bid}.self_attn.k_norm", # cohere olmoe chameleon olmo2 + "transformer.blocks.{bid}.attn.k_ln", # sea-lion + "encoder.layer.{bid}.attention.self.layer_norm_k", # jina-bert-v2 + "transformer.layers.{bid}.attn.k_norm", # openelm + ), + + MODEL_TENSOR.ROPE_FREQS: ( + "language_model.encoder.layers.{bid}.self_attention.rotary_emb.inv_freq", # persimmon + ), + + MODEL_TENSOR.LAYER_OUT_NORM: ( + "encoder.layer.{bid}.output.LayerNorm", # bert + "transformer.layer.{bid}.output_layer_norm", # distillbert + "encoder.layers.{bid}.norm2", # nomic-bert + "transformer.decoder_layer.{bid}.rms_norm_3", # Grok + "encoder.layer.{bid}.mlp.layernorm", # jina-bert-v2 + "encoder.layer.{bid}.layer_norm_2" # jina-v2-code + ), + + MODEL_TENSOR.SSM_IN: ( + "model.layers.{bid}.in_proj", + "backbone.layers.{bid}.mixer.in_proj", + ), + + MODEL_TENSOR.SSM_CONV1D: ( + "model.layers.{bid}.conv1d", + "backbone.layers.{bid}.mixer.conv1d", + ), + + MODEL_TENSOR.SSM_X: ( + "model.layers.{bid}.x_proj", + "backbone.layers.{bid}.mixer.x_proj", + ), + + MODEL_TENSOR.SSM_DT: ( + "model.layers.{bid}.dt_proj", + "backbone.layers.{bid}.mixer.dt_proj", + ), + + MODEL_TENSOR.SSM_A: ( + "model.layers.{bid}.A_log", + "backbone.layers.{bid}.mixer.A_log", + ), + + MODEL_TENSOR.SSM_D: ( + "model.layers.{bid}.D", + "backbone.layers.{bid}.mixer.D", + ), + + MODEL_TENSOR.SSM_OUT: ( + "model.layers.{bid}.out_proj", + "backbone.layers.{bid}.mixer.out_proj", + ), + + MODEL_TENSOR.TIME_MIX_W0: ( + "model.layers.{bid}.attention.w0", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_W1: ( + "rwkv.blocks.{bid}.attention.time_maa_w1", # rwkv6 + "model.layers.{bid}.self_attn.time_maa_w1", # rwkv6qwen2 + "model.layers.{bid}.attention.w1", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_W2: ( + "rwkv.blocks.{bid}.attention.time_maa_w2", # rwkv6 + "model.layers.{bid}.self_attn.time_maa_w2", # rwkv6qwen2 + "model.layers.{bid}.attention.w2", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_A0: ( + "model.layers.{bid}.attention.a0", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_A1: ( + "model.layers.{bid}.attention.a1", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_A2: ( + "model.layers.{bid}.attention.a2", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_V0: ( + "model.layers.{bid}.attention.v0", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_V1: ( + "model.layers.{bid}.attention.v1", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_V2: ( + "model.layers.{bid}.attention.v2", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_G1: ( + "model.layers.{bid}.attention.g1", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_G2: ( + "model.layers.{bid}.attention.g2", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_K_K: ( + "model.layers.{bid}.attention.k_k", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_K_A: ( + "model.layers.{bid}.attention.k_a", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_R_K: ( + "model.layers.{bid}.attention.r_k", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_LERP_X: ( + "rwkv.blocks.{bid}.attention.time_maa_x", # rwkv6 + "model.layers.{bid}.self_attn.time_maa_x", # rwkv6qwen2 + ), + + MODEL_TENSOR.TIME_MIX_LERP_K: ( + "rwkv.blocks.{bid}.attention.time_maa_k", # rwkv6 + "model.layers.{bid}.self_attn.time_maa_k", # rwkv6qwen2 + ), + + MODEL_TENSOR.TIME_MIX_LERP_V: ( + "rwkv.blocks.{bid}.attention.time_maa_v", # rwkv6 + "model.layers.{bid}.self_attn.time_maa_v", # rwkv6qwen2 + ), + + MODEL_TENSOR.TIME_MIX_LERP_R: ( + "rwkv.blocks.{bid}.attention.time_maa_r", # rwkv6 + "model.layers.{bid}.self_attn.time_maa_r", # rwkv6qwen2 + ), + + MODEL_TENSOR.TIME_MIX_LERP_G: ( + "rwkv.blocks.{bid}.attention.time_maa_g", # rwkv6 + "model.layers.{bid}.self_attn.time_maa_g", # rwkv6qwen2 + ), + + MODEL_TENSOR.TIME_MIX_LERP_W: ( + "rwkv.blocks.{bid}.attention.time_maa_w", # rwkv6 + "model.layers.{bid}.self_attn.time_maa_w", # rwkv6qwen2 + ), + + MODEL_TENSOR.TIME_MIX_FIRST: ( + "rwkv.blocks.{bid}.attention.time_faaaa", # rwkv6 + ), + + MODEL_TENSOR.TIME_MIX_DECAY: ( + "rwkv.blocks.{bid}.attention.time_decay", # rwkv6 + "model.layers.{bid}.self_attn.time_decay", # rwkv6qwen2 + ), + + MODEL_TENSOR.TIME_MIX_DECAY_W1: ( + "rwkv.blocks.{bid}.attention.time_decay_w1", # rwkv6 + "model.layers.{bid}.self_attn.time_decay_w1", # rwkv6qwen2 + ), + + MODEL_TENSOR.TIME_MIX_DECAY_W2: ( + "rwkv.blocks.{bid}.attention.time_decay_w2", # rwkv6 + "model.layers.{bid}.self_attn.time_decay_w2", # rwkv6qwen2 + ), + + MODEL_TENSOR.TIME_MIX_KEY: ( + "rwkv.blocks.{bid}.attention.key", # rwkv6 + "model.layers.{bid}.self_attn.k_proj", # rwkv6qwen2 + "model.layers.{bid}.attention.key", # rwkv7 + "model.layers.{bid}.attention.k_proj", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_VALUE: ( + "rwkv.blocks.{bid}.attention.value", # rwkv6 + "model.layers.{bid}.self_attn.v_proj", # rwkv6qwen2 + "model.layers.{bid}.attention.value", # rwkv7 + "model.layers.{bid}.attention.v_proj", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_RECEPTANCE: ( + "rwkv.blocks.{bid}.attention.receptance", # rwkv6 + "model.layers.{bid}.self_attn.q_proj", # rwkv6qwen2 + "model.layers.{bid}.attention.receptance", # rwkv7 + "model.layers.{bid}.attention.r_proj", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_GATE: ( + "rwkv.blocks.{bid}.attention.gate", # rwkv6 + "model.layers.{bid}.self_attn.gate", # rwkv6qwen2 + ), + + MODEL_TENSOR.TIME_MIX_LN: ( + "rwkv.blocks.{bid}.attention.ln_x", # rwkv6 + "model.layers.{bid}.attention.ln_x" # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_OUTPUT: ( + "rwkv.blocks.{bid}.attention.output", # rwkv6 + "model.layers.{bid}.self_attn.o_proj", # rwkv6qwen2 + "model.layers.{bid}.attention.output", # rwkv7 + "model.layers.{bid}.attention.o_proj", # rwkv7 + ), + + MODEL_TENSOR.CHANNEL_MIX_LERP_K: ( + "rwkv.blocks.{bid}.feed_forward.time_maa_k", # rwkv6 + "model.layers.{bid}.feed_forward.x_k", # rwkv7 + ), + + MODEL_TENSOR.CHANNEL_MIX_LERP_R: ( + "rwkv.blocks.{bid}.feed_forward.time_maa_r", # rwkv6 + ), + + MODEL_TENSOR.CHANNEL_MIX_KEY: ( + "rwkv.blocks.{bid}.feed_forward.key", # rwkv6 + "model.layers.{bid}.feed_forward.key", # rwkv7 + ), + + MODEL_TENSOR.CHANNEL_MIX_RECEPTANCE: ( + "rwkv.blocks.{bid}.feed_forward.receptance", # rwkv6 + ), + + MODEL_TENSOR.CHANNEL_MIX_VALUE: ( + "rwkv.blocks.{bid}.feed_forward.value", # rwkv6 + "model.layers.{bid}.feed_forward.value", # rwkv7 + ), + + MODEL_TENSOR.ATTN_Q_A: ( + "model.layers.{bid}.self_attn.q_a_proj", # deepseek2 + ), + + MODEL_TENSOR.ATTN_Q_B: ( + "model.layers.{bid}.self_attn.q_b_proj", # deepseek2 + ), + + MODEL_TENSOR.ATTN_KV_A_MQA: ( + "model.layers.{bid}.self_attn.kv_a_proj_with_mqa", # deepseek2 + ), + + MODEL_TENSOR.ATTN_KV_B: ( + "model.layers.{bid}.self_attn.kv_b_proj", # deepseek2 + ), + + MODEL_TENSOR.ATTN_K_B: ( + "model.layers.{bid}.self_attn.k_b_proj", # deepseek2 + ), + + MODEL_TENSOR.ATTN_V_B: ( + "model.layers.{bid}.self_attn.v_b_proj", # deepseek2 + ), + + MODEL_TENSOR.ATTN_Q_A_NORM: ( + "model.layers.{bid}.self_attn.q_a_layernorm", # deepseek2 + ), + + MODEL_TENSOR.ATTN_KV_A_NORM: ( + "model.layers.{bid}.self_attn.kv_a_layernorm", # deepseek2 + ), + + MODEL_TENSOR.ATTN_SUB_NORM: ( + "model.layers.{bid}.self_attn.inner_attn_ln", # bitnet + ), + + MODEL_TENSOR.FFN_SUB_NORM: ( + "model.layers.{bid}.mlp.ffn_layernorm", # bitnet + ), + + MODEL_TENSOR.DEC_ATTN_NORM: ( + "decoder.block.{bid}.layer.0.layer_norm", # t5 + ), + + MODEL_TENSOR.DEC_ATTN_Q: ( + "decoder.block.{bid}.layer.0.SelfAttention.q", # t5 + ), + + MODEL_TENSOR.DEC_ATTN_K: ( + "decoder.block.{bid}.layer.0.SelfAttention.k", # t5 + ), + + MODEL_TENSOR.DEC_ATTN_V: ( + "decoder.block.{bid}.layer.0.SelfAttention.v", # t5 + ), + + MODEL_TENSOR.DEC_ATTN_OUT: ( + "decoder.block.{bid}.layer.0.SelfAttention.o", # t5 + ), + + MODEL_TENSOR.DEC_ATTN_REL_B: ( + "decoder.block.{bid}.layer.0.SelfAttention.relative_attention_bias", # t5 + ), + + MODEL_TENSOR.DEC_CROSS_ATTN_NORM: ( + "decoder.block.{bid}.layer.1.layer_norm", # t5 + ), + + MODEL_TENSOR.DEC_CROSS_ATTN_Q: ( + "decoder.block.{bid}.layer.1.EncDecAttention.q", # t5 + ), + + MODEL_TENSOR.DEC_CROSS_ATTN_K: ( + "decoder.block.{bid}.layer.1.EncDecAttention.k", # t5 + ), + + MODEL_TENSOR.DEC_CROSS_ATTN_V: ( + "decoder.block.{bid}.layer.1.EncDecAttention.v", # t5 + ), + + MODEL_TENSOR.DEC_CROSS_ATTN_OUT: ( + "decoder.block.{bid}.layer.1.EncDecAttention.o", # t5 + ), + + MODEL_TENSOR.DEC_CROSS_ATTN_REL_B: ( + "decoder.block.{bid}.layer.1.EncDecAttention.relative_attention_bias", # t5 + ), + + MODEL_TENSOR.DEC_FFN_NORM: ( + "decoder.block.{bid}.layer.2.layer_norm", # t5 + ), + + MODEL_TENSOR.DEC_FFN_GATE: ( + "decoder.block.{bid}.layer.2.DenseReluDense.wi_0", # flan-t5 + ), + + MODEL_TENSOR.DEC_FFN_UP: ( + "decoder.block.{bid}.layer.2.DenseReluDense.wi", # t5 + "decoder.block.{bid}.layer.2.DenseReluDense.wi_1", # flan-t5 + ), + + MODEL_TENSOR.DEC_FFN_DOWN: ( + "decoder.block.{bid}.layer.2.DenseReluDense.wo", # t5 + ), + + MODEL_TENSOR.DEC_OUTPUT_NORM: ( + "decoder.final_layer_norm", # t5 + ), + + MODEL_TENSOR.ENC_ATTN_NORM: ( + "encoder.block.{bid}.layer.0.layer_norm", # t5 + ), + + MODEL_TENSOR.ENC_ATTN_Q: ( + "encoder.block.{bid}.layer.0.SelfAttention.q", # t5 + ), + + MODEL_TENSOR.ENC_ATTN_K: ( + "encoder.block.{bid}.layer.0.SelfAttention.k", # t5 + ), + + MODEL_TENSOR.ENC_ATTN_V: ( + "encoder.block.{bid}.layer.0.SelfAttention.v", # t5 + ), + + MODEL_TENSOR.ENC_ATTN_OUT: ( + "encoder.block.{bid}.layer.0.SelfAttention.o", # t5 + ), + + MODEL_TENSOR.ENC_ATTN_REL_B: ( + "encoder.block.{bid}.layer.0.SelfAttention.relative_attention_bias", # t5 + ), + + MODEL_TENSOR.ENC_FFN_NORM: ( + "encoder.block.{bid}.layer.1.layer_norm", # t5 + ), + + MODEL_TENSOR.ENC_FFN_GATE: ( + "encoder.block.{bid}.layer.1.DenseReluDense.wi_0", # flan-t5 + ), + + MODEL_TENSOR.ENC_FFN_UP: ( + "encoder.block.{bid}.layer.1.DenseReluDense.wi", # t5 + "encoder.block.{bid}.layer.1.DenseReluDense.wi_1", # flan-t5 + ), + + MODEL_TENSOR.ENC_FFN_DOWN: ( + "encoder.block.{bid}.layer.1.DenseReluDense.wo", # t5 + ), + + ############################################################################ + # TODO: these do not belong to block_mappings_cfg - move them to mappings_cfg + MODEL_TENSOR.ENC_OUTPUT_NORM: ( + "encoder.final_layer_norm", # t5 + "layer_norm", # neobert + ), + + MODEL_TENSOR.CLS: ( + "classifier", # jina + "classifier.dense", # roberta + "pre_classifier", # distillbert + "dense", # neobert + ), + + MODEL_TENSOR.CLS_OUT: ( + "classifier.out_proj", # roberta + ), + ############################################################################# + + MODEL_TENSOR.CONVNEXT_DW: ( + "backbone.convnext.{bid}.dwconv", # wavtokenizer + ), + + MODEL_TENSOR.CONVNEXT_NORM: ( + "backbone.convnext.{bid}.norm", # wavtokenizer + ), + + MODEL_TENSOR.CONVNEXT_PW1: ( + "backbone.convnext.{bid}.pwconv1", # wavtokenizer + ), + + MODEL_TENSOR.CONVNEXT_PW2: ( + "backbone.convnext.{bid}.pwconv2", # wavtokenizer + ), + + MODEL_TENSOR.CONVNEXT_GAMMA: ( + "backbone.convnext.{bid}.gamma", # wavtokenizer + ), + + MODEL_TENSOR.POSNET_CONV1: ( + "backbone.posnet.{bid}.conv1", # wavtokenizer + ), + + MODEL_TENSOR.POSNET_CONV2: ( + "backbone.posnet.{bid}.conv2", # wavtokenizer + ), + + MODEL_TENSOR.POSNET_NORM: ( + "backbone.posnet.{bid}.norm", # wavtokenizer + ), + + MODEL_TENSOR.POSNET_NORM1: ( + "backbone.posnet.{bid}.norm1", # wavtokenizer + ), + + MODEL_TENSOR.POSNET_NORM2: ( + "backbone.posnet.{bid}.norm2", # wavtokenizer + ), + + MODEL_TENSOR.POSNET_ATTN_NORM: ( + "backbone.posnet.{bid}.norm", # wavtokenizer + ), + + MODEL_TENSOR.POSNET_ATTN_Q: ( + "backbone.posnet.{bid}.q", # wavtokenizer + ), + + MODEL_TENSOR.POSNET_ATTN_K: ( + "backbone.posnet.{bid}.k", # wavtokenizer + ), + + MODEL_TENSOR.POSNET_ATTN_V: ( + "backbone.posnet.{bid}.v", # wavtokenizer + ), + + MODEL_TENSOR.POSNET_ATTN_OUT: ( + "backbone.posnet.{bid}.proj_out", # wavtokenizer + ), + + ############################################################################# + ## Vision encoder + + MODEL_TENSOR.V_MMPROJ: ( + "multi_modal_projector.linear_{bid}", + "visual.merger.mlp.{bid}", # qwen2vl + ), + + MODEL_TENSOR.V_MMPROJ_FC: ( + "model.connector.modality_projection.proj", # SmolVLM + ), + + MODEL_TENSOR.V_MMPROJ_MLP: ( + "model.mm_projector.mlp.mlp.{bid}", + "vision_model.vision_adapter.mlp.fc{bid}", # llama 4 + "mlp1.{bid}", # InternVL + ), + + MODEL_TENSOR.V_MMPROJ_PEG: ( + "model.mm_projector.peg.peg.{bid}", + ), + + MODEL_TENSOR.V_ENC_EMBD_CLS: ( + "vision_tower.vision_model.embeddings.class_embedding", + "vision_model.class_embedding", # llama 4 + ), + + MODEL_TENSOR.V_ENC_EMBD_PATCH: ( + "vision_tower.vision_model.embeddings.patch_embedding", + "vpm.embeddings.patch_embedding", + "model.vision_model.embeddings.patch_embedding", # SmolVLM + "vision_tower.patch_conv", # pixtral + "vision_model.patch_embedding.linear", # llama 4 + "visual.patch_embed.proj", # qwen2vl + ), + + MODEL_TENSOR.V_ENC_EMBD_POS: ( + "vision_tower.vision_model.embeddings.position_embedding", + "vpm.embeddings.position_embedding", + "model.vision_model.embeddings.position_embedding", # SmolVLM + "vision_model.positional_embedding_vlm", # llama 4 + ), + + MODEL_TENSOR.V_ENC_ATTN_Q: ( + "vision_tower.vision_model.encoder.layers.{bid}.self_attn.q_proj", + "vpm.encoder.layers.{bid}.self_attn.q_proj", + "model.vision_model.encoder.layers.{bid}.self_attn.q_proj", # SmolVLM + "vision_model.model.layers.{bid}.self_attn.q_proj", # llama4 + "vision_tower.transformer.layers.{bid}.attention.q_proj", # pixtral + "visual.blocks.{bid}.attn.q", # qwen2vl, generated + ), + + MODEL_TENSOR.V_ENC_ATTN_Q_NORM: ( + "vision_tower.vision_model.encoder.layers.{bid}.attn.q_norm", # InternVL + ), + + MODEL_TENSOR.V_ENC_ATTN_K: ( + "vision_tower.vision_model.encoder.layers.{bid}.self_attn.k_proj", + "vpm.encoder.layers.{bid}.self_attn.k_proj", + "model.vision_model.encoder.layers.{bid}.self_attn.k_proj", # SmolVLM + "vision_model.model.layers.{bid}.self_attn.k_proj", # llama4 + "vision_tower.transformer.layers.{bid}.attention.k_proj", # pixtral + "visual.blocks.{bid}.attn.k", # qwen2vl, generated + ), + + MODEL_TENSOR.V_ENC_ATTN_K_NORM: ( + "vision_tower.vision_model.encoder.layers.{bid}.attn.k_norm", # InternVL + ), + + MODEL_TENSOR.V_ENC_ATTN_V: ( + "vision_tower.vision_model.encoder.layers.{bid}.self_attn.v_proj", + "vpm.encoder.layers.{bid}.self_attn.v_proj", + "model.vision_model.encoder.layers.{bid}.self_attn.v_proj", # SmolVLM + "vision_model.model.layers.{bid}.self_attn.v_proj", # llama4 + "vision_tower.transformer.layers.{bid}.attention.v_proj", # pixtral + "visual.blocks.{bid}.attn.v", # qwen2vl, generated + ), + + MODEL_TENSOR.V_ENC_INPUT_NORM: ( + "vision_tower.vision_model.encoder.layers.{bid}.layer_norm1", + "vision_tower.vision_model.encoder.layers.{bid}.norm1", # InternVL + "vpm.encoder.layers.{bid}.layer_norm1", + "model.vision_model.encoder.layers.{bid}.layer_norm1", # SmolVLM + "vision_tower.transformer.layers.{bid}.attention_norm", # pixtral + "vision_model.model.layers.{bid}.input_layernorm", # llama4 + "visual.blocks.{bid}.norm1", # qwen2vl + ), + + MODEL_TENSOR.V_ENC_ATTN_O: ( + "vision_tower.vision_model.encoder.layers.{bid}.self_attn.out_proj", + "vision_tower.vision_model.encoder.layers.{bid}.attn.proj", # InternVL + "vpm.encoder.layers.{bid}.self_attn.out_proj", + "model.vision_model.encoder.layers.{bid}.self_attn.out_proj", # SmolVLM + "vision_model.model.layers.{bid}.self_attn.o_proj", # llama4 + "vision_tower.transformer.layers.{bid}.attention.o_proj", # pixtral + "visual.blocks.{bid}.attn.proj", # qwen2vl + ), + + MODEL_TENSOR.V_ENC_POST_ATTN_NORM: ( + "vision_tower.vision_model.encoder.layers.{bid}.layer_norm2", + "vision_tower.vision_model.encoder.layers.{bid}.norm2", # InternVL + "vpm.encoder.layers.{bid}.layer_norm2", + "model.vision_model.encoder.layers.{bid}.layer_norm2", # SmolVLM + "vision_model.model.layers.{bid}.post_attention_layernorm", # llama4 + "vision_tower.transformer.layers.{bid}.ffn_norm", # pixtral + "visual.blocks.{bid}.norm2", # qwen2vl + ), + + MODEL_TENSOR.V_ENC_FFN_UP: ( + "vision_tower.vision_model.encoder.layers.{bid}.mlp.fc1", + "vpm.encoder.layers.{bid}.mlp.fc1", + "model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3 + "vision_tower.transformer.layers.{bid}.feed_forward.up_proj", # pixtral + "vision_model.model.layers.{bid}.mlp.fc1", # llama4 + "visual.blocks.{bid}.mlp.fc1", # qwen2vl + "visual.blocks.{bid}.mlp.up_proj", # qwen2.5vl + ), + + MODEL_TENSOR.V_ENC_FFN_GATE: ( + "vision_tower.transformer.layers.{bid}.feed_forward.gate_proj", # pixtral + "visual.blocks.{bid}.mlp.gate_proj", # qwen2.5vl + ), + + MODEL_TENSOR.V_ENC_FFN_DOWN: ( + "vision_tower.vision_model.encoder.layers.{bid}.mlp.fc2", + "vpm.encoder.layers.{bid}.mlp.fc2", + "model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3 + "vision_tower.transformer.layers.{bid}.feed_forward.down_proj", # pixtral + "vision_model.model.layers.{bid}.mlp.fc2", # llama4 + "visual.blocks.{bid}.mlp.fc2", # qwen2vl + "visual.blocks.{bid}.mlp.down_proj", # qwen2.5vl + ), + + MODEL_TENSOR.V_LAYER_SCALE_1: ( + "vision_tower.vision_model.encoder.layers.{bid}.ls1", # InternVL + ), + + MODEL_TENSOR.V_LAYER_SCALE_2: ( + "vision_tower.vision_model.encoder.layers.{bid}.ls2", # InternVL + ), + + MODEL_TENSOR.V_PRE_NORM: ( + "vision_tower.vision_model.pre_layrnorm", + "vision_tower.ln_pre", # pixtral + "vision_model.layernorm_pre", # llama4 + ), + + MODEL_TENSOR.V_POST_NORM: ( + "vision_tower.vision_model.post_layernorm", + "model.vision_model.post_layernorm", # SmolVLM + "vision_model.layernorm_post", # llama4 + "visual.merger.ln_q", # qwen2vl + ), + + MODEL_TENSOR.V_MM_INP_PROJ: ( + "multi_modal_projector.mm_input_projection", + ), + + MODEL_TENSOR.V_MM_INP_NORM: ( + "multi_modal_projector.norm", + ), + + MODEL_TENSOR.V_MM_SOFT_EMB_NORM: ( + "multi_modal_projector.mm_soft_emb_norm", + ), + + MODEL_TENSOR.V_RESMPL_POS_EMBD_K: ( + "resampler.pos_embed_k", + ), + + MODEL_TENSOR.V_RESMPL_ATTN_Q: ( + "resampler.attn.in_proj_q", # tensor generated from resampler.attn.in_proj + ), + + MODEL_TENSOR.V_RESMPL_ATTN_K: ( + "resampler.attn.in_proj_k", # tensor generated from resampler.attn.in_proj + ), + + MODEL_TENSOR.V_RESMPL_ATTN_V: ( + "resampler.attn.in_proj_v", # tensor generated from resampler.attn.in_proj + ), + + MODEL_TENSOR.V_RESMPL_ATTN_OUT: ( + "resampler.attn.out_proj", + ), + + MODEL_TENSOR.V_RESMPL_KV: ( + "resampler.kv_proj", + ), + + MODEL_TENSOR.V_RESMPL_POST_NORM: ( + "resampler.ln_post", + ), + + MODEL_TENSOR.V_RESMPL_KV_NORM: ( + "resampler.ln_kv", + ), + + MODEL_TENSOR.V_RESMPL_Q_NORM: ( + "resampler.ln_q", + ), + + MODEL_TENSOR.V_RESMPL_PROJ: ( + "resampler.proj", + ), + + MODEL_TENSOR.V_RESMPL_QUERY: ( + "resampler.query", + ), + + MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK: ( + "v.token_embd.img_break", # for pixtral, this is a generated vector + ), + + MODEL_TENSOR.V_MM_PATCH_MERGER: ( + "multi_modal_projector.patch_merger.merging_layer", # mistral small 3.1 + ), + + # audio (mtmd) + + MODEL_TENSOR.A_ENC_EMBD_POS: ( + "audio_tower.embed_positions", # ultravox + ), + + MODEL_TENSOR.A_ENC_CONV1D: ( + "audio_tower.conv{bid}", # ultravox + ), + + MODEL_TENSOR.A_PRE_NORM: (), + + MODEL_TENSOR.A_POST_NORM: ( + "audio_tower.layer_norm", # ultravox + "audio_tower.ln_post", # qwen2omni + ), + + MODEL_TENSOR.A_ENC_ATTN_Q: ( + "audio_tower.layers.{bid}.self_attn.q_proj", # ultravox + ), + + MODEL_TENSOR.A_ENC_ATTN_K: ( + "audio_tower.layers.{bid}.self_attn.k_proj", # ultravox + ), + + MODEL_TENSOR.A_ENC_ATTN_V: ( + "audio_tower.layers.{bid}.self_attn.v_proj", # ultravox + ), + + MODEL_TENSOR.A_ENC_INPUT_NORM: ( + "audio_tower.layers.{bid}.self_attn_layer_norm", # ultravox + ), + + MODEL_TENSOR.A_ENC_OUTPUT: ( + "audio_tower.layers.{bid}.self_attn.out_proj", # ultravox + ), + + MODEL_TENSOR.A_ENC_OUTPUT_NORM: ( + "audio_tower.layers.{bid}.final_layer_norm", # ultravox + ), + + MODEL_TENSOR.A_ENC_FFN_UP: ( + "audio_tower.layers.{bid}.fc1", # ultravox + ), + + MODEL_TENSOR.A_ENC_FFN_GATE: (), + + MODEL_TENSOR.A_ENC_FFN_DOWN: ( + "audio_tower.layers.{bid}.fc2", # ultravox + ), + + # note: some tensors below has "audio." pseudo-prefix, to prevent conflicts with vision tensors + # this prefix is added in the conversion code in modify_tensors() + + MODEL_TENSOR.A_MMPROJ: ( + "audio.multi_modal_projector.linear_{bid}", # ultravox + ), + + MODEL_TENSOR.A_MMPROJ_FC: ( + "audio.multi_modal_projector.linear", # qwen2audio + "audio_tower.proj", # qwen2omni + ), + + MODEL_TENSOR.A_MM_NORM_PRE: ( + "audio.multi_modal_projector.ln_pre", # ultravox + ), + + MODEL_TENSOR.A_MM_NORM_MID: ( + "audio.multi_modal_projector.ln_mid", # ultravox + ), + } + + # architecture-specific block mappings + arch_block_mappings_cfg: dict[MODEL_ARCH, dict[MODEL_TENSOR, tuple[str, ...]]] = { + MODEL_ARCH.ARCTIC: { + MODEL_TENSOR.FFN_NORM: ( + "model.layers.{bid}.residual_layernorm", + ), + MODEL_TENSOR.FFN_NORM_EXP: ( + "model.layers.{bid}.post_attention_layernorm", + ), + }, + } + + mapping: dict[str, tuple[MODEL_TENSOR, str]] + + def __init__(self, arch: MODEL_ARCH, n_blocks: int): + self.mapping = {} + for tensor, keys in self.mappings_cfg.items(): + if tensor not in MODEL_TENSORS[arch]: + continue + tensor_name = TENSOR_NAMES[tensor] + self.mapping[tensor_name] = (tensor, tensor_name) + for key in keys: + self.mapping[key] = (tensor, tensor_name) + if arch in self.arch_block_mappings_cfg: + self.block_mappings_cfg.update(self.arch_block_mappings_cfg[arch]) + for bid in range(n_blocks): + for tensor, keys in self.block_mappings_cfg.items(): + if tensor not in MODEL_TENSORS[arch]: + continue + + tensor_name = TENSOR_NAMES[tensor].format(bid = bid) + self.mapping[tensor_name] = (tensor, tensor_name) + for key in keys: + key = key.format(bid = bid) + self.mapping[key] = (tensor, tensor_name) + + def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None: + result = self.mapping.get(key) + if result is not None: + return result + for suffix in try_suffixes: + if key.endswith(suffix): + result = self.mapping.get(key[:-len(suffix)]) + if result is not None: + return result[0], result[1] + suffix + return None + + def get_name(self, key: str, try_suffixes: Sequence[str] = ()) -> str | None: + result = self.get_type_and_name(key, try_suffixes = try_suffixes) + if result is None: + return None + return result[1] + + def get_type(self, key: str, try_suffixes: Sequence[str] = ()) -> MODEL_TENSOR | None: + result = self.get_type_and_name(key, try_suffixes = try_suffixes) + if result is None: + return None + return result[0] + + def __getitem__(self, key: str) -> str: + try: + return self.mapping[key][1] + except KeyError: + raise KeyError(key) + + def __contains__(self, key: str) -> bool: + return key in self.mapping + + def __repr__(self) -> str: + return repr(self.mapping) + + +def get_tensor_name_map(arch: MODEL_ARCH, n_blocks: int) -> TensorNameMap: + return TensorNameMap(arch, n_blocks) diff --git a/venv/lib/python3.10/site-packages/gguf/utility.py b/venv/lib/python3.10/site-packages/gguf/utility.py new file mode 100644 index 0000000000000000000000000000000000000000..00adcbc937398d1e7a9d4b7159a54052bb79ae64 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/utility.py @@ -0,0 +1,264 @@ +from __future__ import annotations + +from dataclasses import dataclass +from typing import Literal + +import os +import json + + +def fill_templated_filename(filename: str, output_type: str | None) -> str: + # Given a file name fill in any type templates e.g. 'some-model-name.{ftype}.gguf' + ftype_lowercase: str = output_type.lower() if output_type is not None else "" + ftype_uppercase: str = output_type.upper() if output_type is not None else "" + return filename.format(ftype_lowercase, + outtype=ftype_lowercase, ftype=ftype_lowercase, + OUTTYPE=ftype_uppercase, FTYPE=ftype_uppercase) + + +def model_weight_count_rounded_notation(model_params_count: int, min_digits: int = 2) -> str: + if model_params_count > 1e12 : + # Trillions Of Parameters + scaled_model_params = model_params_count * 1e-12 + scale_suffix = "T" + elif model_params_count > 1e9 : + # Billions Of Parameters + scaled_model_params = model_params_count * 1e-9 + scale_suffix = "B" + elif model_params_count > 1e6 : + # Millions Of Parameters + scaled_model_params = model_params_count * 1e-6 + scale_suffix = "M" + else: + # Thousands Of Parameters + scaled_model_params = model_params_count * 1e-3 + scale_suffix = "K" + + fix = max(min_digits - len(str(round(scaled_model_params)).lstrip('0')), 0) + + return f"{scaled_model_params:.{fix}f}{scale_suffix}" + + +def size_label(total_params: int, shared_params: int, expert_params: int, expert_count: int) -> str: + + if expert_count > 0: + pretty_size = model_weight_count_rounded_notation(abs(shared_params) + abs(expert_params), min_digits=2) + size_class = f"{expert_count}x{pretty_size}" + else: + size_class = model_weight_count_rounded_notation(abs(total_params), min_digits=2) + + return size_class + + +def naming_convention(model_name: str | None, base_name: str | None, finetune_string: str | None, version_string: str | None, size_label: str | None, output_type: str | None, model_type: Literal['vocab', 'LoRA'] | None = None) -> str: + # Reference: https://github.com/ggml-org/ggml/blob/master/docs/gguf.md#gguf-naming-convention + + if base_name is not None: + name = base_name.strip().replace(' ', '-').replace('/', '-') + elif model_name is not None: + name = model_name.strip().replace(' ', '-').replace('/', '-') + else: + name = "ggml-model" + + parameters = f"-{size_label}" if size_label is not None else "" + + finetune = f"-{finetune_string.strip().replace(' ', '-')}" if finetune_string is not None else "" + + version = f"-{version_string.strip().replace(' ', '-')}" if version_string is not None else "" + + encoding = f"-{output_type.strip().replace(' ', '-').upper()}" if output_type is not None else "" + + kind = f"-{model_type.strip().replace(' ', '-')}" if model_type is not None else "" + + return f"{name}{parameters}{finetune}{version}{encoding}{kind}" + + +@dataclass +class RemoteTensor: + dtype: str + shape: tuple[int, ...] + offset_start: int + size: int + url: str + + def data(self) -> bytearray: + # TODO: handle request errors (maybe with limited retries?) + # NOTE: using a bytearray, otherwise PyTorch complains the buffer is not writeable + data = bytearray(SafetensorRemote.get_data_by_range(url=self.url, start=self.offset_start, size=self.size)) + return data + + +class SafetensorRemote: + """ + Uility class to handle remote safetensor files. + This class is designed to work with Hugging Face model repositories. + + Example (one model has single safetensor file, the other has multiple): + for model_id in ["ngxson/TEST-Tiny-Llama4", "Qwen/Qwen2.5-7B-Instruct"]: + tensors = SafetensorRemote.get_list_tensors_hf_model(model_id) + print(tensors) + + Example reading tensor data: + tensors = SafetensorRemote.get_list_tensors_hf_model(model_id) + for name, meta in tensors.items(): + dtype, shape, offset_start, size, remote_safetensor_url = meta + # read the tensor data + data = SafetensorRemote.get_data_by_range(remote_safetensor_url, offset_start, size) + print(data) + """ + + BASE_DOMAIN = "https://huggingface.co" + ALIGNMENT = 8 # bytes + + @classmethod + def get_list_tensors_hf_model(cls, model_id: str) -> dict[str, RemoteTensor]: + """ + Get list of tensors from a Hugging Face model repository. + + Returns a dictionary of tensor names and their metadata. + Each tensor is represented as a tuple of (dtype, shape, offset_start, size, remote_safetensor_url) + """ + # case 1: model has only one single model.safetensor file + is_single_file = cls.check_file_exist(f"{cls.BASE_DOMAIN}/{model_id}/resolve/main/model.safetensors") + if is_single_file: + url = f"{cls.BASE_DOMAIN}/{model_id}/resolve/main/model.safetensors" + return cls.get_list_tensors(url) + + # case 2: model has multiple files + index_url = f"{cls.BASE_DOMAIN}/{model_id}/resolve/main/model.safetensors.index.json" + is_multiple_files = cls.check_file_exist(index_url) + if is_multiple_files: + # read the index file + index_data = cls.get_data_by_range(index_url, 0) + index_str = index_data.decode('utf-8') + index_json = json.loads(index_str) + assert index_json.get("weight_map") is not None, "weight_map not found in index file" + weight_map = index_json["weight_map"] + # get the list of files + all_files = list(set(weight_map.values())) + all_files.sort() # make sure we load shard files in order + # get the list of tensors + tensors: dict[str, RemoteTensor] = {} + for file in all_files: + url = f"{cls.BASE_DOMAIN}/{model_id}/resolve/main/{file}" + for key, val in cls.get_list_tensors(url).items(): + tensors[key] = val + return tensors + + raise ValueError(f"Model {model_id} does not have any safetensor files") + + @classmethod + def get_list_tensors(cls, url: str) -> dict[str, RemoteTensor]: + """ + Get list of tensors from a remote safetensor file. + + Returns a dictionary of tensor names and their metadata. + Each tensor is represented as a tuple of (dtype, shape, offset_start, size) + """ + metadata, data_start_offset = cls.get_metadata(url) + res: dict[str, RemoteTensor] = {} + + for name, meta in metadata.items(): + if name == "__metadata__": + continue + if not isinstance(meta, dict): + raise ValueError(f"Invalid metadata for tensor '{name}': {meta}") + try: + dtype = meta["dtype"] + shape = meta["shape"] + offset_start_relative, offset_end_relative = meta["data_offsets"] + size = offset_end_relative - offset_start_relative + offset_start = data_start_offset + offset_start_relative + res[name] = RemoteTensor(dtype=dtype, shape=tuple(shape), offset_start=offset_start, size=size, url=url) + except KeyError as e: + raise ValueError(f"Missing key in metadata for tensor '{name}': {e}, meta = {meta}") + + return res + + @classmethod + def get_metadata(cls, url: str) -> tuple[dict, int]: + """ + Get JSON metadata from a remote safetensor file. + + Returns tuple of (metadata, data_start_offset) + """ + # Request first 5MB of the file (hopefully enough for metadata) + read_size = 5 * 1024 * 1024 + raw_data = cls.get_data_by_range(url, 0, read_size) + + # Parse header + # First 8 bytes contain the metadata length as u64 little-endian + if len(raw_data) < 8: + raise ValueError("Not enough data to read metadata size") + metadata_length = int.from_bytes(raw_data[:8], byteorder='little') + + # Calculate the data start offset + data_start_offset = 8 + metadata_length + alignment = SafetensorRemote.ALIGNMENT + if data_start_offset % alignment != 0: + data_start_offset += alignment - (data_start_offset % alignment) + + # Check if we have enough data to read the metadata + if len(raw_data) < 8 + metadata_length: + raise ValueError(f"Could not read complete metadata. Need {8 + metadata_length} bytes, got {len(raw_data)}") + + # Extract metadata bytes and parse as JSON + metadata_bytes = raw_data[8:8 + metadata_length] + metadata_str = metadata_bytes.decode('utf-8') + try: + metadata = json.loads(metadata_str) + return metadata, data_start_offset + except json.JSONDecodeError as e: + raise ValueError(f"Failed to parse safetensor metadata as JSON: {e}") + + @classmethod + def get_data_by_range(cls, url: str, start: int, size: int = -1) -> bytes: + """ + Get raw byte data from a remote file by range. + If size is not specified, it will read the entire file. + """ + import requests + from urllib.parse import urlparse + + parsed_url = urlparse(url) + if not parsed_url.scheme or not parsed_url.netloc: + raise ValueError(f"Invalid URL: {url}") + + headers = cls._get_request_headers() + if size > -1: + headers["Range"] = f"bytes={start}-{start + size}" + response = requests.get(url, allow_redirects=True, headers=headers) + response.raise_for_status() + + # Get raw byte data + return response.content[slice(size if size > -1 else None)] + + @classmethod + def check_file_exist(cls, url: str) -> bool: + """ + Check if a file exists at the given URL. + Returns True if the file exists, False otherwise. + """ + import requests + from urllib.parse import urlparse + + parsed_url = urlparse(url) + if not parsed_url.scheme or not parsed_url.netloc: + raise ValueError(f"Invalid URL: {url}") + + try: + headers = cls._get_request_headers() + headers["Range"] = "bytes=0-0" + response = requests.head(url, allow_redirects=True, headers=headers) + # Success (2xx) or redirect (3xx) + return 200 <= response.status_code < 400 + except requests.RequestException: + return False + + @classmethod + def _get_request_headers(cls) -> dict[str, str]: + """Prepare common headers for requests.""" + headers = {"User-Agent": "convert_hf_to_gguf"} + if os.environ.get("HF_TOKEN"): + headers["Authorization"] = f"Bearer {os.environ['HF_TOKEN']}" + return headers diff --git a/venv/lib/python3.10/site-packages/gguf/vocab.py b/venv/lib/python3.10/site-packages/gguf/vocab.py new file mode 100644 index 0000000000000000000000000000000000000000..44d066ee75a7ece48dabb168c5054fb4053f5d48 --- /dev/null +++ b/venv/lib/python3.10/site-packages/gguf/vocab.py @@ -0,0 +1,498 @@ +from __future__ import annotations + +import re +import logging +import json +import os +from pathlib import Path +from typing import Any, Callable, Sequence, Mapping, Iterable, Protocol, ClassVar, runtime_checkable + +try: + from sentencepiece import SentencePieceProcessor +except ImportError: + SentencePieceProcessor = None + +import gguf + +from .gguf_writer import GGUFWriter + +logger = logging.getLogger(__name__) + + +class SpecialVocab: + merges: list[str] + add_special_token: dict[str, bool] + special_token_ids: dict[str, int] + chat_template: str | Sequence[Mapping[str, str]] | None + + def __init__( + self, path: str | os.PathLike[str], load_merges: bool = False, + special_token_types: Iterable[str] | None = None, + n_vocab: int | None = None, + ): + self.special_token_ids = {} + self.add_special_token = {} + self.n_vocab = n_vocab + self.load_merges = load_merges + self.merges = [] + self.chat_template = None + if special_token_types is not None: + self.special_token_types = special_token_types + else: + self.special_token_types = ('bos', 'eos', 'unk', 'sep', 'pad', 'cls', 'mask') + self._load(Path(path)) + + def __repr__(self) -> str: + return ''.format( + len(self.merges), self.special_token_ids or "unset", self.add_special_token or "unset", + ) + + def add_to_gguf(self, gw: GGUFWriter, quiet: bool = False) -> None: + if self.merges: + if not quiet: + logger.info(f'Adding {len(self.merges)} merge(s).') + gw.add_token_merges(self.merges) + elif self.load_merges: + logger.warning('Adding merges requested but no merges found, output may be non-functional.') + for typ, tokid in self.special_token_ids.items(): + id_handler: Callable[[int], None] | None = getattr(gw, f'add_{typ}_token_id', None) + if id_handler is None: + logger.warning(f'No handler for special token type {typ} with id {tokid} - skipping') + continue + if not quiet: + logger.info(f'Setting special token type {typ} to {tokid}') + id_handler(tokid) + for typ, value in self.add_special_token.items(): + add_handler: Callable[[bool], None] | None = getattr(gw, f'add_add_{typ}_token', None) + if add_handler is None: + logger.warning(f'No handler for add_{typ}_token with value {value} - skipping') + continue + if not quiet: + logger.info(f'Setting add_{typ}_token to {value}') + add_handler(value) + if self.chat_template is not None: + if not quiet: + logger.info(f'Setting chat_template to {self.chat_template}') + gw.add_chat_template(self.chat_template) + + def _load(self, path: Path) -> None: + self._try_load_from_tokenizer_json(path) + self._try_load_from_config_json(path) + if self.load_merges and not self.merges: + self._try_load_merges_txt(path) + + def _try_load_merges_txt(self, path: Path) -> bool: + merges_file = path / 'merges.txt' + if not merges_file.is_file(): + return False + with open(merges_file, 'r', encoding = 'utf-8') as fp: + first_line = next(fp, '').strip() + if not first_line.startswith('#'): + fp.seek(0) + line_num = 0 + else: + line_num = 1 + merges = [] + for line in fp: + line_num += 1 + line = line.strip() + if not line: + continue + parts = line.split(None, 3) + if len(parts) != 2: + logger.warning(f'{merges_file.name}: Line {line_num}: Entry malformed, ignoring') + continue + merges.append(f'{parts[0]} {parts[1]}') + self.merges = merges + return True + + def _set_special_token(self, typ: str, tid: Any) -> None: + if not isinstance(tid, int): + return + if tid < 0: + raise ValueError(f'invalid value for special token type {typ}: {tid}') + if self.n_vocab is None or tid < self.n_vocab: + if typ in self.special_token_ids: + return + self.special_token_ids[typ] = tid + return + logger.warning(f'Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping') + + def _try_load_from_tokenizer_json(self, path: Path) -> bool: + tokenizer_file = path / 'tokenizer.json' + if tokenizer_file.is_file(): + with open(tokenizer_file, encoding = 'utf-8') as f: + tokenizer = json.load(f) + if self.load_merges: + merges = tokenizer.get('model', {}).get('merges') + if isinstance(merges, list) and merges: + if isinstance(merges[0], str): + self.merges = merges + elif isinstance(merges[0], list) and len(merges[0]) == 2 and isinstance(merges[0][0], str): + # New format since transformers 4.45 to support spaces in merges + # ref: https://github.com/ggml-org/llama.cpp/issues/9692 + # TODO: internally store as the new format instead of converting to old + if any(' ' in s for pair in merges for s in pair): + logger.warning(f'Spaces in merges detected, encoding as {chr(ord(" ") + 256)!r}') + self.merges = [ + ' '.join( + [ + # ensure the spaces are properly encoded + ''.join( + chr(ord(c) + 256) if c == ' ' else c + for c in part + ) + for part in pair + ] + ) + for pair in merges + ] + else: + raise ValueError("Unknown tokenizer merges format") + added_tokens = tokenizer.get('added_tokens', {}) + else: + added_tokens = {} + tokenizer_config_file = path / 'tokenizer_config.json' + if not tokenizer_config_file.is_file(): + return True + with open(tokenizer_config_file, encoding = 'utf-8') as f: + tokenizer_config = json.load(f) + chat_template_alt = None + chat_template_file = path / 'chat_template.json' + if chat_template_file.is_file(): + with open(chat_template_file, encoding = 'utf-8') as f: + chat_template_alt = json.load(f).get('chat_template') + chat_template = tokenizer_config.get('chat_template', chat_template_alt) + if chat_template is None or isinstance(chat_template, (str, list)): + self.chat_template = chat_template + else: + logger.warning(f'Bad type for chat_template field in {tokenizer_config_file!r} - ignoring') + for typ in self.special_token_types: + add_entry = tokenizer_config.get(f'add_{typ}_token') + if isinstance(add_entry, bool): + self.add_special_token[typ] = add_entry + entry = tokenizer_config.get(f'{typ}_token') + if isinstance(entry, str): + tc_content = entry + elif isinstance(entry, dict): + entry_content = entry.get('content') + if not isinstance(entry_content, str): + continue + tc_content = entry_content + else: + continue + # We only need the first match here. + maybe_token_id = next( + (atok.get('id') for atok in added_tokens if atok.get('content') == tc_content), + None, + ) + self._set_special_token(typ, maybe_token_id) + return True + + def _try_load_from_config_json(self, path: Path) -> bool: + config_file = path / 'config.json' + if not config_file.is_file(): + return False + with open(config_file, encoding = 'utf-8') as f: + config = json.load(f) + for typ in self.special_token_types: + self._set_special_token(typ, config.get(f'{typ}_token_id')) + return True + + +@runtime_checkable +class BaseVocab(Protocol): + tokenizer_model: ClassVar[str] + name: ClassVar[str] + + +@runtime_checkable +class Vocab(BaseVocab, Protocol): + vocab_size: int + added_tokens_dict: dict[str, int] + added_tokens_list: list[str] + fname_tokenizer: Path + + def __init__(self, base_path: Path): ... + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: ... + + +class NoVocab(BaseVocab): + tokenizer_model = "no_vocab" + name = "no_vocab" + + def __repr__(self) -> str: + return "" + + +class BpeVocab(Vocab): + tokenizer_model = "gpt2" + name = "bpe" + + def __init__(self, base_path: Path): + added_tokens: dict[str, int] = {} + + if (fname_tokenizer := base_path / 'vocab.json').exists(): + # "slow" tokenizer + with open(fname_tokenizer, encoding="utf-8") as f: + self.vocab = json.load(f) + + try: + # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab. + with open(base_path / 'added_tokens.json', encoding="utf-8") as f: + added_tokens = json.load(f) + except FileNotFoundError: + pass + else: + # "fast" tokenizer + fname_tokenizer = base_path / 'tokenizer.json' + + # if this fails, FileNotFoundError propagates to caller + with open(fname_tokenizer, encoding="utf-8") as f: + tokenizer_json = json.load(f) + + tokenizer_model: dict[str, Any] = tokenizer_json['model'] + if ( + tokenizer_model['type'] != 'BPE' or tokenizer_model.get('byte_fallback', False) + or tokenizer_json['decoder']['type'] != 'ByteLevel' + ): + raise FileNotFoundError('Cannot find GPT-2 BPE tokenizer') + + self.vocab = tokenizer_model["vocab"] + + if (added := tokenizer_json.get('added_tokens')) is not None: + # Added tokens here can be duplicates of the main vocabulary. + added_tokens = {item['content']: item['id'] + for item in added + if item['content'] not in self.vocab} + + vocab_size = len(self.vocab) + expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) + actual_ids = sorted(added_tokens.values()) + if expected_ids != actual_ids: + expected_end_id = vocab_size + len(actual_ids) - 1 + raise ValueError(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range " + f"{vocab_size} - {expected_end_id}; got {actual_ids}") + + items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1]) + self.added_tokens_dict = added_tokens + self.added_tokens_list = [text for (text, idx) in items] + self.vocab_size_base = vocab_size + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + self.fname_tokenizer = fname_tokenizer + + def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()} + + for i, _ in enumerate(self.vocab): + yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL + + def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + for text in self.added_tokens_list: + score = -1000.0 + yield text.encode("utf-8"), score, gguf.TokenType.CONTROL + + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + yield from self.bpe_tokens() + yield from self.added_tokens() + + def __repr__(self) -> str: + return f"" + + +class SentencePieceVocab(Vocab): + tokenizer_model = "llama" + name = "spm" + + def __init__(self, base_path: Path): + if SentencePieceProcessor is None: + raise RuntimeError("sentencepiece is not installed") + + added_tokens: dict[str, int] = {} + if (fname_tokenizer := base_path / 'tokenizer.model').exists(): + # normal location + try: + with open(base_path / 'added_tokens.json', encoding="utf-8") as f: + added_tokens = json.load(f) + except FileNotFoundError: + pass + elif not (fname_tokenizer := base_path.parent / 'tokenizer.model').exists(): + # not found in alternate location either + raise FileNotFoundError('Cannot find tokenizer.model') + + self.sentencepiece_tokenizer = SentencePieceProcessor() + self.sentencepiece_tokenizer.LoadFromFile(str(fname_tokenizer)) + vocab_size = self.sentencepiece_tokenizer.vocab_size() + + new_tokens = {id: piece for piece, id in added_tokens.items() if id >= vocab_size} + expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens))) + actual_new_ids = sorted(new_tokens.keys()) + + if expected_new_ids != actual_new_ids: + raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}") + + # Token pieces that were added to the base vocabulary. + self.added_tokens_dict = added_tokens + self.added_tokens_list = [new_tokens[id] for id in actual_new_ids] + self.vocab_size_base = vocab_size + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + self.fname_tokenizer = fname_tokenizer + + def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + tokenizer = self.sentencepiece_tokenizer + for i in range(tokenizer.vocab_size()): + piece = tokenizer.IdToPiece(i) + text = piece.encode("utf-8") + score: float = tokenizer.GetScore(i) + + toktype = gguf.TokenType.NORMAL + if tokenizer.IsUnknown(i): + toktype = gguf.TokenType.UNKNOWN + if tokenizer.IsControl(i): + toktype = gguf.TokenType.CONTROL + + # NOTE: I think added_tokens are user defined. + # ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto + # if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED + + if tokenizer.IsUnused(i): + toktype = gguf.TokenType.UNUSED + if tokenizer.IsByte(i): + toktype = gguf.TokenType.BYTE + + yield text, score, toktype + + def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + for text in self.added_tokens_list: + score = -1000.0 + yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED + + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + yield from self.sentencepiece_tokens() + yield from self.added_tokens() + + def __repr__(self) -> str: + return f"" + + +class LlamaHfVocab(Vocab): + tokenizer_model = "llama" + name = "hfft" + + def __init__(self, base_path: Path): + fname_tokenizer = base_path / 'tokenizer.json' + # if this fails, FileNotFoundError propagates to caller + with open(fname_tokenizer, encoding='utf-8') as f: + tokenizer_json = json.load(f) + + # pre-check so we know if we need transformers + tokenizer_model: dict[str, Any] = tokenizer_json['model'] + is_llama3 = ( + tokenizer_model['type'] == 'BPE' and tokenizer_model.get('ignore_merges', False) + and not tokenizer_model.get('byte_fallback', True) + ) + if is_llama3: + raise TypeError('Llama 3 must be converted with BpeVocab') + + if not is_llama3 and ( + tokenizer_model['type'] != 'BPE' or not tokenizer_model.get('byte_fallback', False) + or tokenizer_json['decoder']['type'] != 'Sequence' + ): + raise FileNotFoundError('Cannot find Llama BPE tokenizer') + + try: + from transformers import AutoTokenizer + except ImportError as e: + raise ImportError( + "To use LlamaHfVocab, please install the `transformers` package. " + "You can install it with `pip install transformers`." + ) from e + + # Allow the tokenizer to default to slow or fast versions. + # Explicitly set tokenizer to use local paths. + self.tokenizer = AutoTokenizer.from_pretrained( + base_path, + cache_dir=base_path, + local_files_only=True, + ) + assert self.tokenizer.is_fast # assume tokenizer.json is used + + # Initialize lists and dictionaries for added tokens + self.added_tokens_list = [] + self.added_tokens_dict = dict() + self.added_tokens_ids = set() + + # Process added tokens + for tok, tokidx in sorted( + self.tokenizer.get_added_vocab().items(), key=lambda x: x[1] + ): + # Only consider added tokens that are not in the base vocabulary + if tokidx >= self.tokenizer.vocab_size: + self.added_tokens_list.append(tok) + self.added_tokens_dict[tok] = tokidx + self.added_tokens_ids.add(tokidx) + + # Store special tokens and their IDs + self.specials = { + tok: self.tokenizer.get_vocab()[tok] + for tok in self.tokenizer.all_special_tokens + } + self.special_ids = set(self.tokenizer.all_special_ids) + + # Set vocabulary sizes + self.vocab_size_base = self.tokenizer.vocab_size + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + + self.fname_tokenizer = fname_tokenizer + + def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + reverse_vocab = { + id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items() + } + + for token_id in range(self.vocab_size_base): + # Skip processing added tokens here + if token_id in self.added_tokens_ids: + continue + + # Convert token text to bytes + token_text = reverse_vocab[token_id].encode("utf-8") + + # Yield token text, score, and type + yield token_text, self.get_token_score(token_id), self.get_token_type( + token_id, token_text, self.special_ids # Reuse already stored special IDs + ) + + def get_token_type(self, token_id: int, token_text: bytes, special_ids: set[int]) -> gguf.TokenType: + # Special case for byte tokens + if re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text): + return gguf.TokenType.BYTE + + # Determine token type based on whether it's a special token + return gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL + + def get_token_score(self, token_id: int) -> float: + # Placeholder for actual logic to determine the token's score + # This needs to be implemented based on specific requirements + return -1000.0 # Default score + + def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + for text in self.added_tokens_list: + if text in self.specials: + toktype = self.get_token_type(self.specials[text], b'', self.special_ids) + score = self.get_token_score(self.specials[text]) + else: + toktype = gguf.TokenType.USER_DEFINED + score = -1000.0 + + yield text.encode("utf-8"), score, toktype + + def has_newline_token(self): + return "<0x0A>" in self.tokenizer.vocab or "\n" in self.tokenizer.vocab + + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + yield from self.hf_tokens() + yield from self.added_tokens() + + def __repr__(self) -> str: + return f"" diff --git a/venv/lib/python3.10/site-packages/jwt/__init__.py b/venv/lib/python3.10/site-packages/jwt/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..457a4e3580ccc21cb69317008478d20d600abcbc --- /dev/null +++ b/venv/lib/python3.10/site-packages/jwt/__init__.py @@ -0,0 +1,75 @@ +from .api_jwk import PyJWK, PyJWKSet +from .api_jws import ( + PyJWS, + get_algorithm_by_name, + get_unverified_header, + register_algorithm, + unregister_algorithm, +) +from .api_jwt import PyJWT, decode, decode_complete, encode +from .exceptions import ( + DecodeError, + ExpiredSignatureError, + ImmatureSignatureError, + InvalidAlgorithmError, + InvalidAudienceError, + InvalidIssuedAtError, + InvalidIssuerError, + InvalidKeyError, + InvalidSignatureError, + InvalidTokenError, + MissingRequiredClaimError, + PyJWKClientConnectionError, + PyJWKClientError, + PyJWKError, + PyJWKSetError, + PyJWTError, +) +from .jwks_client import PyJWKClient + +__version__ = "2.10.1" + +__title__ = "PyJWT" +__description__ = "JSON Web Token implementation in Python" +__url__ = "https://pyjwt.readthedocs.io" +__uri__ = __url__ +__doc__ = f"{__description__} <{__uri__}>" + +__author__ = "José Padilla" +__email__ = "hello@jpadilla.com" + +__license__ = "MIT" +__copyright__ = "Copyright 2015-2022 José Padilla" + + +__all__ = [ + "PyJWS", + "PyJWT", + "PyJWKClient", + "PyJWK", + "PyJWKSet", + "decode", + "decode_complete", + "encode", + "get_unverified_header", + "register_algorithm", + "unregister_algorithm", + "get_algorithm_by_name", + # Exceptions + "DecodeError", + "ExpiredSignatureError", + "ImmatureSignatureError", + "InvalidAlgorithmError", + "InvalidAudienceError", + "InvalidIssuedAtError", + "InvalidIssuerError", + "InvalidKeyError", + "InvalidSignatureError", + "InvalidTokenError", + "MissingRequiredClaimError", + "PyJWKClientConnectionError", + "PyJWKClientError", + "PyJWKError", + "PyJWKSetError", + "PyJWTError", +] diff --git a/venv/lib/python3.10/site-packages/jwt/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/jwt/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ff2faa524b84e554a1a5d6949d1308589c273c53 Binary files /dev/null and b/venv/lib/python3.10/site-packages/jwt/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/jwt/__pycache__/algorithms.cpython-310.pyc b/venv/lib/python3.10/site-packages/jwt/__pycache__/algorithms.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ca3cc5fd4129c9848a75e497e71689e8996d9d83 Binary files /dev/null and b/venv/lib/python3.10/site-packages/jwt/__pycache__/algorithms.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/jwt/__pycache__/api_jwk.cpython-310.pyc b/venv/lib/python3.10/site-packages/jwt/__pycache__/api_jwk.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..84ce5a517088949aec751778ac629f81849312ce Binary files /dev/null and b/venv/lib/python3.10/site-packages/jwt/__pycache__/api_jwk.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/jwt/__pycache__/api_jws.cpython-310.pyc b/venv/lib/python3.10/site-packages/jwt/__pycache__/api_jws.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..477b90d1eefaa92af19e27631f4e084b5f060823 Binary files /dev/null and b/venv/lib/python3.10/site-packages/jwt/__pycache__/api_jws.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/jwt/__pycache__/api_jwt.cpython-310.pyc b/venv/lib/python3.10/site-packages/jwt/__pycache__/api_jwt.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..28165f968d45fd8d3f952ad9e6f46a31c7c01b37 Binary files /dev/null and b/venv/lib/python3.10/site-packages/jwt/__pycache__/api_jwt.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/jwt/__pycache__/exceptions.cpython-310.pyc b/venv/lib/python3.10/site-packages/jwt/__pycache__/exceptions.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f8e891dad8e2c3e1de3c5cb1a7f4d29571d44490 Binary files /dev/null and b/venv/lib/python3.10/site-packages/jwt/__pycache__/exceptions.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/jwt/__pycache__/help.cpython-310.pyc b/venv/lib/python3.10/site-packages/jwt/__pycache__/help.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cfe7f920398d00fca6df57578df683000fd62119 Binary files /dev/null and b/venv/lib/python3.10/site-packages/jwt/__pycache__/help.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/jwt/__pycache__/jwk_set_cache.cpython-310.pyc b/venv/lib/python3.10/site-packages/jwt/__pycache__/jwk_set_cache.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..05294f158d536d76e0e724684e3401991f1f5115 Binary files /dev/null and b/venv/lib/python3.10/site-packages/jwt/__pycache__/jwk_set_cache.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/jwt/__pycache__/jwks_client.cpython-310.pyc b/venv/lib/python3.10/site-packages/jwt/__pycache__/jwks_client.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0a876cf7c0915eab2aaba4680a1676c28f585f8d Binary files /dev/null and b/venv/lib/python3.10/site-packages/jwt/__pycache__/jwks_client.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/jwt/__pycache__/types.cpython-310.pyc b/venv/lib/python3.10/site-packages/jwt/__pycache__/types.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e95b105648b0420746041ba89d05b53df797a437 Binary files /dev/null and b/venv/lib/python3.10/site-packages/jwt/__pycache__/types.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/jwt/__pycache__/utils.cpython-310.pyc b/venv/lib/python3.10/site-packages/jwt/__pycache__/utils.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9a74e05d6737a82568165de4e3d7711aa2b73f99 Binary files /dev/null and b/venv/lib/python3.10/site-packages/jwt/__pycache__/utils.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/jwt/__pycache__/warnings.cpython-310.pyc b/venv/lib/python3.10/site-packages/jwt/__pycache__/warnings.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..44e82dad77244e4890b196e4939b3aac8486f00f Binary files /dev/null and b/venv/lib/python3.10/site-packages/jwt/__pycache__/warnings.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/jwt/algorithms.py b/venv/lib/python3.10/site-packages/jwt/algorithms.py new file mode 100644 index 0000000000000000000000000000000000000000..ccb1500fcf827b8a2c1eed0e0acf8c1699c0dbf6 --- /dev/null +++ b/venv/lib/python3.10/site-packages/jwt/algorithms.py @@ -0,0 +1,875 @@ +from __future__ import annotations + +import hashlib +import hmac +import json +from abc import ABC, abstractmethod +from typing import TYPE_CHECKING, Any, ClassVar, Literal, NoReturn, cast, overload + +from .exceptions import InvalidKeyError +from .types import HashlibHash, JWKDict +from .utils import ( + base64url_decode, + base64url_encode, + der_to_raw_signature, + force_bytes, + from_base64url_uint, + is_pem_format, + is_ssh_key, + raw_to_der_signature, + to_base64url_uint, +) + +try: + from cryptography.exceptions import InvalidSignature, UnsupportedAlgorithm + from cryptography.hazmat.backends import default_backend + from cryptography.hazmat.primitives import hashes + from cryptography.hazmat.primitives.asymmetric import padding + from cryptography.hazmat.primitives.asymmetric.ec import ( + ECDSA, + SECP256K1, + SECP256R1, + SECP384R1, + SECP521R1, + EllipticCurve, + EllipticCurvePrivateKey, + EllipticCurvePrivateNumbers, + EllipticCurvePublicKey, + EllipticCurvePublicNumbers, + ) + from cryptography.hazmat.primitives.asymmetric.ed448 import ( + Ed448PrivateKey, + Ed448PublicKey, + ) + from cryptography.hazmat.primitives.asymmetric.ed25519 import ( + Ed25519PrivateKey, + Ed25519PublicKey, + ) + from cryptography.hazmat.primitives.asymmetric.rsa import ( + RSAPrivateKey, + RSAPrivateNumbers, + RSAPublicKey, + RSAPublicNumbers, + rsa_crt_dmp1, + rsa_crt_dmq1, + rsa_crt_iqmp, + rsa_recover_prime_factors, + ) + from cryptography.hazmat.primitives.serialization import ( + Encoding, + NoEncryption, + PrivateFormat, + PublicFormat, + load_pem_private_key, + load_pem_public_key, + load_ssh_public_key, + ) + + has_crypto = True +except ModuleNotFoundError: + has_crypto = False + + +if TYPE_CHECKING: + # Type aliases for convenience in algorithms method signatures + AllowedRSAKeys = RSAPrivateKey | RSAPublicKey + AllowedECKeys = EllipticCurvePrivateKey | EllipticCurvePublicKey + AllowedOKPKeys = ( + Ed25519PrivateKey | Ed25519PublicKey | Ed448PrivateKey | Ed448PublicKey + ) + AllowedKeys = AllowedRSAKeys | AllowedECKeys | AllowedOKPKeys + AllowedPrivateKeys = ( + RSAPrivateKey | EllipticCurvePrivateKey | Ed25519PrivateKey | Ed448PrivateKey + ) + AllowedPublicKeys = ( + RSAPublicKey | EllipticCurvePublicKey | Ed25519PublicKey | Ed448PublicKey + ) + + +requires_cryptography = { + "RS256", + "RS384", + "RS512", + "ES256", + "ES256K", + "ES384", + "ES521", + "ES512", + "PS256", + "PS384", + "PS512", + "EdDSA", +} + + +def get_default_algorithms() -> dict[str, Algorithm]: + """ + Returns the algorithms that are implemented by the library. + """ + default_algorithms = { + "none": NoneAlgorithm(), + "HS256": HMACAlgorithm(HMACAlgorithm.SHA256), + "HS384": HMACAlgorithm(HMACAlgorithm.SHA384), + "HS512": HMACAlgorithm(HMACAlgorithm.SHA512), + } + + if has_crypto: + default_algorithms.update( + { + "RS256": RSAAlgorithm(RSAAlgorithm.SHA256), + "RS384": RSAAlgorithm(RSAAlgorithm.SHA384), + "RS512": RSAAlgorithm(RSAAlgorithm.SHA512), + "ES256": ECAlgorithm(ECAlgorithm.SHA256), + "ES256K": ECAlgorithm(ECAlgorithm.SHA256), + "ES384": ECAlgorithm(ECAlgorithm.SHA384), + "ES521": ECAlgorithm(ECAlgorithm.SHA512), + "ES512": ECAlgorithm( + ECAlgorithm.SHA512 + ), # Backward compat for #219 fix + "PS256": RSAPSSAlgorithm(RSAPSSAlgorithm.SHA256), + "PS384": RSAPSSAlgorithm(RSAPSSAlgorithm.SHA384), + "PS512": RSAPSSAlgorithm(RSAPSSAlgorithm.SHA512), + "EdDSA": OKPAlgorithm(), + } + ) + + return default_algorithms + + +class Algorithm(ABC): + """ + The interface for an algorithm used to sign and verify tokens. + """ + + def compute_hash_digest(self, bytestr: bytes) -> bytes: + """ + Compute a hash digest using the specified algorithm's hash algorithm. + + If there is no hash algorithm, raises a NotImplementedError. + """ + # lookup self.hash_alg if defined in a way that mypy can understand + hash_alg = getattr(self, "hash_alg", None) + if hash_alg is None: + raise NotImplementedError + + if ( + has_crypto + and isinstance(hash_alg, type) + and issubclass(hash_alg, hashes.HashAlgorithm) + ): + digest = hashes.Hash(hash_alg(), backend=default_backend()) + digest.update(bytestr) + return bytes(digest.finalize()) + else: + return bytes(hash_alg(bytestr).digest()) + + @abstractmethod + def prepare_key(self, key: Any) -> Any: + """ + Performs necessary validation and conversions on the key and returns + the key value in the proper format for sign() and verify(). + """ + + @abstractmethod + def sign(self, msg: bytes, key: Any) -> bytes: + """ + Returns a digital signature for the specified message + using the specified key value. + """ + + @abstractmethod + def verify(self, msg: bytes, key: Any, sig: bytes) -> bool: + """ + Verifies that the specified digital signature is valid + for the specified message and key values. + """ + + @overload + @staticmethod + @abstractmethod + def to_jwk(key_obj, as_dict: Literal[True]) -> JWKDict: ... # pragma: no cover + + @overload + @staticmethod + @abstractmethod + def to_jwk(key_obj, as_dict: Literal[False] = False) -> str: ... # pragma: no cover + + @staticmethod + @abstractmethod + def to_jwk(key_obj, as_dict: bool = False) -> JWKDict | str: + """ + Serializes a given key into a JWK + """ + + @staticmethod + @abstractmethod + def from_jwk(jwk: str | JWKDict) -> Any: + """ + Deserializes a given key from JWK back into a key object + """ + + +class NoneAlgorithm(Algorithm): + """ + Placeholder for use when no signing or verification + operations are required. + """ + + def prepare_key(self, key: str | None) -> None: + if key == "": + key = None + + if key is not None: + raise InvalidKeyError('When alg = "none", key value must be None.') + + return key + + def sign(self, msg: bytes, key: None) -> bytes: + return b"" + + def verify(self, msg: bytes, key: None, sig: bytes) -> bool: + return False + + @staticmethod + def to_jwk(key_obj: Any, as_dict: bool = False) -> NoReturn: + raise NotImplementedError() + + @staticmethod + def from_jwk(jwk: str | JWKDict) -> NoReturn: + raise NotImplementedError() + + +class HMACAlgorithm(Algorithm): + """ + Performs signing and verification operations using HMAC + and the specified hash function. + """ + + SHA256: ClassVar[HashlibHash] = hashlib.sha256 + SHA384: ClassVar[HashlibHash] = hashlib.sha384 + SHA512: ClassVar[HashlibHash] = hashlib.sha512 + + def __init__(self, hash_alg: HashlibHash) -> None: + self.hash_alg = hash_alg + + def prepare_key(self, key: str | bytes) -> bytes: + key_bytes = force_bytes(key) + + if is_pem_format(key_bytes) or is_ssh_key(key_bytes): + raise InvalidKeyError( + "The specified key is an asymmetric key or x509 certificate and" + " should not be used as an HMAC secret." + ) + + return key_bytes + + @overload + @staticmethod + def to_jwk( + key_obj: str | bytes, as_dict: Literal[True] + ) -> JWKDict: ... # pragma: no cover + + @overload + @staticmethod + def to_jwk( + key_obj: str | bytes, as_dict: Literal[False] = False + ) -> str: ... # pragma: no cover + + @staticmethod + def to_jwk(key_obj: str | bytes, as_dict: bool = False) -> JWKDict | str: + jwk = { + "k": base64url_encode(force_bytes(key_obj)).decode(), + "kty": "oct", + } + + if as_dict: + return jwk + else: + return json.dumps(jwk) + + @staticmethod + def from_jwk(jwk: str | JWKDict) -> bytes: + try: + if isinstance(jwk, str): + obj: JWKDict = json.loads(jwk) + elif isinstance(jwk, dict): + obj = jwk + else: + raise ValueError + except ValueError: + raise InvalidKeyError("Key is not valid JSON") from None + + if obj.get("kty") != "oct": + raise InvalidKeyError("Not an HMAC key") + + return base64url_decode(obj["k"]) + + def sign(self, msg: bytes, key: bytes) -> bytes: + return hmac.new(key, msg, self.hash_alg).digest() + + def verify(self, msg: bytes, key: bytes, sig: bytes) -> bool: + return hmac.compare_digest(sig, self.sign(msg, key)) + + +if has_crypto: + + class RSAAlgorithm(Algorithm): + """ + Performs signing and verification operations using + RSASSA-PKCS-v1_5 and the specified hash function. + """ + + SHA256: ClassVar[type[hashes.HashAlgorithm]] = hashes.SHA256 + SHA384: ClassVar[type[hashes.HashAlgorithm]] = hashes.SHA384 + SHA512: ClassVar[type[hashes.HashAlgorithm]] = hashes.SHA512 + + def __init__(self, hash_alg: type[hashes.HashAlgorithm]) -> None: + self.hash_alg = hash_alg + + def prepare_key(self, key: AllowedRSAKeys | str | bytes) -> AllowedRSAKeys: + if isinstance(key, (RSAPrivateKey, RSAPublicKey)): + return key + + if not isinstance(key, (bytes, str)): + raise TypeError("Expecting a PEM-formatted key.") + + key_bytes = force_bytes(key) + + try: + if key_bytes.startswith(b"ssh-rsa"): + return cast(RSAPublicKey, load_ssh_public_key(key_bytes)) + else: + return cast( + RSAPrivateKey, load_pem_private_key(key_bytes, password=None) + ) + except ValueError: + try: + return cast(RSAPublicKey, load_pem_public_key(key_bytes)) + except (ValueError, UnsupportedAlgorithm): + raise InvalidKeyError( + "Could not parse the provided public key." + ) from None + + @overload + @staticmethod + def to_jwk( + key_obj: AllowedRSAKeys, as_dict: Literal[True] + ) -> JWKDict: ... # pragma: no cover + + @overload + @staticmethod + def to_jwk( + key_obj: AllowedRSAKeys, as_dict: Literal[False] = False + ) -> str: ... # pragma: no cover + + @staticmethod + def to_jwk(key_obj: AllowedRSAKeys, as_dict: bool = False) -> JWKDict | str: + obj: dict[str, Any] | None = None + + if hasattr(key_obj, "private_numbers"): + # Private key + numbers = key_obj.private_numbers() + + obj = { + "kty": "RSA", + "key_ops": ["sign"], + "n": to_base64url_uint(numbers.public_numbers.n).decode(), + "e": to_base64url_uint(numbers.public_numbers.e).decode(), + "d": to_base64url_uint(numbers.d).decode(), + "p": to_base64url_uint(numbers.p).decode(), + "q": to_base64url_uint(numbers.q).decode(), + "dp": to_base64url_uint(numbers.dmp1).decode(), + "dq": to_base64url_uint(numbers.dmq1).decode(), + "qi": to_base64url_uint(numbers.iqmp).decode(), + } + + elif hasattr(key_obj, "verify"): + # Public key + numbers = key_obj.public_numbers() + + obj = { + "kty": "RSA", + "key_ops": ["verify"], + "n": to_base64url_uint(numbers.n).decode(), + "e": to_base64url_uint(numbers.e).decode(), + } + else: + raise InvalidKeyError("Not a public or private key") + + if as_dict: + return obj + else: + return json.dumps(obj) + + @staticmethod + def from_jwk(jwk: str | JWKDict) -> AllowedRSAKeys: + try: + if isinstance(jwk, str): + obj = json.loads(jwk) + elif isinstance(jwk, dict): + obj = jwk + else: + raise ValueError + except ValueError: + raise InvalidKeyError("Key is not valid JSON") from None + + if obj.get("kty") != "RSA": + raise InvalidKeyError("Not an RSA key") from None + + if "d" in obj and "e" in obj and "n" in obj: + # Private key + if "oth" in obj: + raise InvalidKeyError( + "Unsupported RSA private key: > 2 primes not supported" + ) + + other_props = ["p", "q", "dp", "dq", "qi"] + props_found = [prop in obj for prop in other_props] + any_props_found = any(props_found) + + if any_props_found and not all(props_found): + raise InvalidKeyError( + "RSA key must include all parameters if any are present besides d" + ) from None + + public_numbers = RSAPublicNumbers( + from_base64url_uint(obj["e"]), + from_base64url_uint(obj["n"]), + ) + + if any_props_found: + numbers = RSAPrivateNumbers( + d=from_base64url_uint(obj["d"]), + p=from_base64url_uint(obj["p"]), + q=from_base64url_uint(obj["q"]), + dmp1=from_base64url_uint(obj["dp"]), + dmq1=from_base64url_uint(obj["dq"]), + iqmp=from_base64url_uint(obj["qi"]), + public_numbers=public_numbers, + ) + else: + d = from_base64url_uint(obj["d"]) + p, q = rsa_recover_prime_factors( + public_numbers.n, d, public_numbers.e + ) + + numbers = RSAPrivateNumbers( + d=d, + p=p, + q=q, + dmp1=rsa_crt_dmp1(d, p), + dmq1=rsa_crt_dmq1(d, q), + iqmp=rsa_crt_iqmp(p, q), + public_numbers=public_numbers, + ) + + return numbers.private_key() + elif "n" in obj and "e" in obj: + # Public key + return RSAPublicNumbers( + from_base64url_uint(obj["e"]), + from_base64url_uint(obj["n"]), + ).public_key() + else: + raise InvalidKeyError("Not a public or private key") + + def sign(self, msg: bytes, key: RSAPrivateKey) -> bytes: + return key.sign(msg, padding.PKCS1v15(), self.hash_alg()) + + def verify(self, msg: bytes, key: RSAPublicKey, sig: bytes) -> bool: + try: + key.verify(sig, msg, padding.PKCS1v15(), self.hash_alg()) + return True + except InvalidSignature: + return False + + class ECAlgorithm(Algorithm): + """ + Performs signing and verification operations using + ECDSA and the specified hash function + """ + + SHA256: ClassVar[type[hashes.HashAlgorithm]] = hashes.SHA256 + SHA384: ClassVar[type[hashes.HashAlgorithm]] = hashes.SHA384 + SHA512: ClassVar[type[hashes.HashAlgorithm]] = hashes.SHA512 + + def __init__(self, hash_alg: type[hashes.HashAlgorithm]) -> None: + self.hash_alg = hash_alg + + def prepare_key(self, key: AllowedECKeys | str | bytes) -> AllowedECKeys: + if isinstance(key, (EllipticCurvePrivateKey, EllipticCurvePublicKey)): + return key + + if not isinstance(key, (bytes, str)): + raise TypeError("Expecting a PEM-formatted key.") + + key_bytes = force_bytes(key) + + # Attempt to load key. We don't know if it's + # a Signing Key or a Verifying Key, so we try + # the Verifying Key first. + try: + if key_bytes.startswith(b"ecdsa-sha2-"): + crypto_key = load_ssh_public_key(key_bytes) + else: + crypto_key = load_pem_public_key(key_bytes) # type: ignore[assignment] + except ValueError: + crypto_key = load_pem_private_key(key_bytes, password=None) # type: ignore[assignment] + + # Explicit check the key to prevent confusing errors from cryptography + if not isinstance( + crypto_key, (EllipticCurvePrivateKey, EllipticCurvePublicKey) + ): + raise InvalidKeyError( + "Expecting a EllipticCurvePrivateKey/EllipticCurvePublicKey. Wrong key provided for ECDSA algorithms" + ) from None + + return crypto_key + + def sign(self, msg: bytes, key: EllipticCurvePrivateKey) -> bytes: + der_sig = key.sign(msg, ECDSA(self.hash_alg())) + + return der_to_raw_signature(der_sig, key.curve) + + def verify(self, msg: bytes, key: AllowedECKeys, sig: bytes) -> bool: + try: + der_sig = raw_to_der_signature(sig, key.curve) + except ValueError: + return False + + try: + public_key = ( + key.public_key() + if isinstance(key, EllipticCurvePrivateKey) + else key + ) + public_key.verify(der_sig, msg, ECDSA(self.hash_alg())) + return True + except InvalidSignature: + return False + + @overload + @staticmethod + def to_jwk( + key_obj: AllowedECKeys, as_dict: Literal[True] + ) -> JWKDict: ... # pragma: no cover + + @overload + @staticmethod + def to_jwk( + key_obj: AllowedECKeys, as_dict: Literal[False] = False + ) -> str: ... # pragma: no cover + + @staticmethod + def to_jwk(key_obj: AllowedECKeys, as_dict: bool = False) -> JWKDict | str: + if isinstance(key_obj, EllipticCurvePrivateKey): + public_numbers = key_obj.public_key().public_numbers() + elif isinstance(key_obj, EllipticCurvePublicKey): + public_numbers = key_obj.public_numbers() + else: + raise InvalidKeyError("Not a public or private key") + + if isinstance(key_obj.curve, SECP256R1): + crv = "P-256" + elif isinstance(key_obj.curve, SECP384R1): + crv = "P-384" + elif isinstance(key_obj.curve, SECP521R1): + crv = "P-521" + elif isinstance(key_obj.curve, SECP256K1): + crv = "secp256k1" + else: + raise InvalidKeyError(f"Invalid curve: {key_obj.curve}") + + obj: dict[str, Any] = { + "kty": "EC", + "crv": crv, + "x": to_base64url_uint( + public_numbers.x, + bit_length=key_obj.curve.key_size, + ).decode(), + "y": to_base64url_uint( + public_numbers.y, + bit_length=key_obj.curve.key_size, + ).decode(), + } + + if isinstance(key_obj, EllipticCurvePrivateKey): + obj["d"] = to_base64url_uint( + key_obj.private_numbers().private_value, + bit_length=key_obj.curve.key_size, + ).decode() + + if as_dict: + return obj + else: + return json.dumps(obj) + + @staticmethod + def from_jwk(jwk: str | JWKDict) -> AllowedECKeys: + try: + if isinstance(jwk, str): + obj = json.loads(jwk) + elif isinstance(jwk, dict): + obj = jwk + else: + raise ValueError + except ValueError: + raise InvalidKeyError("Key is not valid JSON") from None + + if obj.get("kty") != "EC": + raise InvalidKeyError("Not an Elliptic curve key") from None + + if "x" not in obj or "y" not in obj: + raise InvalidKeyError("Not an Elliptic curve key") from None + + x = base64url_decode(obj.get("x")) + y = base64url_decode(obj.get("y")) + + curve = obj.get("crv") + curve_obj: EllipticCurve + + if curve == "P-256": + if len(x) == len(y) == 32: + curve_obj = SECP256R1() + else: + raise InvalidKeyError( + "Coords should be 32 bytes for curve P-256" + ) from None + elif curve == "P-384": + if len(x) == len(y) == 48: + curve_obj = SECP384R1() + else: + raise InvalidKeyError( + "Coords should be 48 bytes for curve P-384" + ) from None + elif curve == "P-521": + if len(x) == len(y) == 66: + curve_obj = SECP521R1() + else: + raise InvalidKeyError( + "Coords should be 66 bytes for curve P-521" + ) from None + elif curve == "secp256k1": + if len(x) == len(y) == 32: + curve_obj = SECP256K1() + else: + raise InvalidKeyError( + "Coords should be 32 bytes for curve secp256k1" + ) + else: + raise InvalidKeyError(f"Invalid curve: {curve}") + + public_numbers = EllipticCurvePublicNumbers( + x=int.from_bytes(x, byteorder="big"), + y=int.from_bytes(y, byteorder="big"), + curve=curve_obj, + ) + + if "d" not in obj: + return public_numbers.public_key() + + d = base64url_decode(obj.get("d")) + if len(d) != len(x): + raise InvalidKeyError( + "D should be {} bytes for curve {}", len(x), curve + ) + + return EllipticCurvePrivateNumbers( + int.from_bytes(d, byteorder="big"), public_numbers + ).private_key() + + class RSAPSSAlgorithm(RSAAlgorithm): + """ + Performs a signature using RSASSA-PSS with MGF1 + """ + + def sign(self, msg: bytes, key: RSAPrivateKey) -> bytes: + return key.sign( + msg, + padding.PSS( + mgf=padding.MGF1(self.hash_alg()), + salt_length=self.hash_alg().digest_size, + ), + self.hash_alg(), + ) + + def verify(self, msg: bytes, key: RSAPublicKey, sig: bytes) -> bool: + try: + key.verify( + sig, + msg, + padding.PSS( + mgf=padding.MGF1(self.hash_alg()), + salt_length=self.hash_alg().digest_size, + ), + self.hash_alg(), + ) + return True + except InvalidSignature: + return False + + class OKPAlgorithm(Algorithm): + """ + Performs signing and verification operations using EdDSA + + This class requires ``cryptography>=2.6`` to be installed. + """ + + def __init__(self, **kwargs: Any) -> None: + pass + + def prepare_key(self, key: AllowedOKPKeys | str | bytes) -> AllowedOKPKeys: + if isinstance(key, (bytes, str)): + key_str = key.decode("utf-8") if isinstance(key, bytes) else key + key_bytes = key.encode("utf-8") if isinstance(key, str) else key + + if "-----BEGIN PUBLIC" in key_str: + key = load_pem_public_key(key_bytes) # type: ignore[assignment] + elif "-----BEGIN PRIVATE" in key_str: + key = load_pem_private_key(key_bytes, password=None) # type: ignore[assignment] + elif key_str[0:4] == "ssh-": + key = load_ssh_public_key(key_bytes) # type: ignore[assignment] + + # Explicit check the key to prevent confusing errors from cryptography + if not isinstance( + key, + (Ed25519PrivateKey, Ed25519PublicKey, Ed448PrivateKey, Ed448PublicKey), + ): + raise InvalidKeyError( + "Expecting a EllipticCurvePrivateKey/EllipticCurvePublicKey. Wrong key provided for EdDSA algorithms" + ) + + return key + + def sign( + self, msg: str | bytes, key: Ed25519PrivateKey | Ed448PrivateKey + ) -> bytes: + """ + Sign a message ``msg`` using the EdDSA private key ``key`` + :param str|bytes msg: Message to sign + :param Ed25519PrivateKey}Ed448PrivateKey key: A :class:`.Ed25519PrivateKey` + or :class:`.Ed448PrivateKey` isinstance + :return bytes signature: The signature, as bytes + """ + msg_bytes = msg.encode("utf-8") if isinstance(msg, str) else msg + return key.sign(msg_bytes) + + def verify( + self, msg: str | bytes, key: AllowedOKPKeys, sig: str | bytes + ) -> bool: + """ + Verify a given ``msg`` against a signature ``sig`` using the EdDSA key ``key`` + + :param str|bytes sig: EdDSA signature to check ``msg`` against + :param str|bytes msg: Message to sign + :param Ed25519PrivateKey|Ed25519PublicKey|Ed448PrivateKey|Ed448PublicKey key: + A private or public EdDSA key instance + :return bool verified: True if signature is valid, False if not. + """ + try: + msg_bytes = msg.encode("utf-8") if isinstance(msg, str) else msg + sig_bytes = sig.encode("utf-8") if isinstance(sig, str) else sig + + public_key = ( + key.public_key() + if isinstance(key, (Ed25519PrivateKey, Ed448PrivateKey)) + else key + ) + public_key.verify(sig_bytes, msg_bytes) + return True # If no exception was raised, the signature is valid. + except InvalidSignature: + return False + + @overload + @staticmethod + def to_jwk( + key: AllowedOKPKeys, as_dict: Literal[True] + ) -> JWKDict: ... # pragma: no cover + + @overload + @staticmethod + def to_jwk( + key: AllowedOKPKeys, as_dict: Literal[False] = False + ) -> str: ... # pragma: no cover + + @staticmethod + def to_jwk(key: AllowedOKPKeys, as_dict: bool = False) -> JWKDict | str: + if isinstance(key, (Ed25519PublicKey, Ed448PublicKey)): + x = key.public_bytes( + encoding=Encoding.Raw, + format=PublicFormat.Raw, + ) + crv = "Ed25519" if isinstance(key, Ed25519PublicKey) else "Ed448" + + obj = { + "x": base64url_encode(force_bytes(x)).decode(), + "kty": "OKP", + "crv": crv, + } + + if as_dict: + return obj + else: + return json.dumps(obj) + + if isinstance(key, (Ed25519PrivateKey, Ed448PrivateKey)): + d = key.private_bytes( + encoding=Encoding.Raw, + format=PrivateFormat.Raw, + encryption_algorithm=NoEncryption(), + ) + + x = key.public_key().public_bytes( + encoding=Encoding.Raw, + format=PublicFormat.Raw, + ) + + crv = "Ed25519" if isinstance(key, Ed25519PrivateKey) else "Ed448" + obj = { + "x": base64url_encode(force_bytes(x)).decode(), + "d": base64url_encode(force_bytes(d)).decode(), + "kty": "OKP", + "crv": crv, + } + + if as_dict: + return obj + else: + return json.dumps(obj) + + raise InvalidKeyError("Not a public or private key") + + @staticmethod + def from_jwk(jwk: str | JWKDict) -> AllowedOKPKeys: + try: + if isinstance(jwk, str): + obj = json.loads(jwk) + elif isinstance(jwk, dict): + obj = jwk + else: + raise ValueError + except ValueError: + raise InvalidKeyError("Key is not valid JSON") from None + + if obj.get("kty") != "OKP": + raise InvalidKeyError("Not an Octet Key Pair") + + curve = obj.get("crv") + if curve != "Ed25519" and curve != "Ed448": + raise InvalidKeyError(f"Invalid curve: {curve}") + + if "x" not in obj: + raise InvalidKeyError('OKP should have "x" parameter') + x = base64url_decode(obj.get("x")) + + try: + if "d" not in obj: + if curve == "Ed25519": + return Ed25519PublicKey.from_public_bytes(x) + return Ed448PublicKey.from_public_bytes(x) + d = base64url_decode(obj.get("d")) + if curve == "Ed25519": + return Ed25519PrivateKey.from_private_bytes(d) + return Ed448PrivateKey.from_private_bytes(d) + except ValueError as err: + raise InvalidKeyError("Invalid key parameter") from err diff --git a/venv/lib/python3.10/site-packages/jwt/api_jwk.py b/venv/lib/python3.10/site-packages/jwt/api_jwk.py new file mode 100644 index 0000000000000000000000000000000000000000..02f4679cbb55a854419b18af322d93db1b295312 --- /dev/null +++ b/venv/lib/python3.10/site-packages/jwt/api_jwk.py @@ -0,0 +1,144 @@ +from __future__ import annotations + +import json +import time +from typing import Any + +from .algorithms import get_default_algorithms, has_crypto, requires_cryptography +from .exceptions import ( + InvalidKeyError, + MissingCryptographyError, + PyJWKError, + PyJWKSetError, + PyJWTError, +) +from .types import JWKDict + + +class PyJWK: + def __init__(self, jwk_data: JWKDict, algorithm: str | None = None) -> None: + self._algorithms = get_default_algorithms() + self._jwk_data = jwk_data + + kty = self._jwk_data.get("kty", None) + if not kty: + raise InvalidKeyError(f"kty is not found: {self._jwk_data}") + + if not algorithm and isinstance(self._jwk_data, dict): + algorithm = self._jwk_data.get("alg", None) + + if not algorithm: + # Determine alg with kty (and crv). + crv = self._jwk_data.get("crv", None) + if kty == "EC": + if crv == "P-256" or not crv: + algorithm = "ES256" + elif crv == "P-384": + algorithm = "ES384" + elif crv == "P-521": + algorithm = "ES512" + elif crv == "secp256k1": + algorithm = "ES256K" + else: + raise InvalidKeyError(f"Unsupported crv: {crv}") + elif kty == "RSA": + algorithm = "RS256" + elif kty == "oct": + algorithm = "HS256" + elif kty == "OKP": + if not crv: + raise InvalidKeyError(f"crv is not found: {self._jwk_data}") + if crv == "Ed25519": + algorithm = "EdDSA" + else: + raise InvalidKeyError(f"Unsupported crv: {crv}") + else: + raise InvalidKeyError(f"Unsupported kty: {kty}") + + if not has_crypto and algorithm in requires_cryptography: + raise MissingCryptographyError( + f"{algorithm} requires 'cryptography' to be installed." + ) + + self.algorithm_name = algorithm + + if algorithm in self._algorithms: + self.Algorithm = self._algorithms[algorithm] + else: + raise PyJWKError(f"Unable to find an algorithm for key: {self._jwk_data}") + + self.key = self.Algorithm.from_jwk(self._jwk_data) + + @staticmethod + def from_dict(obj: JWKDict, algorithm: str | None = None) -> PyJWK: + return PyJWK(obj, algorithm) + + @staticmethod + def from_json(data: str, algorithm: None = None) -> PyJWK: + obj = json.loads(data) + return PyJWK.from_dict(obj, algorithm) + + @property + def key_type(self) -> str | None: + return self._jwk_data.get("kty", None) + + @property + def key_id(self) -> str | None: + return self._jwk_data.get("kid", None) + + @property + def public_key_use(self) -> str | None: + return self._jwk_data.get("use", None) + + +class PyJWKSet: + def __init__(self, keys: list[JWKDict]) -> None: + self.keys = [] + + if not keys: + raise PyJWKSetError("The JWK Set did not contain any keys") + + if not isinstance(keys, list): + raise PyJWKSetError("Invalid JWK Set value") + + for key in keys: + try: + self.keys.append(PyJWK(key)) + except PyJWTError as error: + if isinstance(error, MissingCryptographyError): + raise error + # skip unusable keys + continue + + if len(self.keys) == 0: + raise PyJWKSetError( + "The JWK Set did not contain any usable keys. Perhaps 'cryptography' is not installed?" + ) + + @staticmethod + def from_dict(obj: dict[str, Any]) -> PyJWKSet: + keys = obj.get("keys", []) + return PyJWKSet(keys) + + @staticmethod + def from_json(data: str) -> PyJWKSet: + obj = json.loads(data) + return PyJWKSet.from_dict(obj) + + def __getitem__(self, kid: str) -> PyJWK: + for key in self.keys: + if key.key_id == kid: + return key + raise KeyError(f"keyset has no key for kid: {kid}") + + +class PyJWTSetWithTimestamp: + def __init__(self, jwk_set: PyJWKSet): + self.jwk_set = jwk_set + self.timestamp = time.monotonic() + + def get_jwk_set(self) -> PyJWKSet: + return self.jwk_set + + def get_timestamp(self) -> float: + return self.timestamp diff --git a/venv/lib/python3.10/site-packages/jwt/api_jws.py b/venv/lib/python3.10/site-packages/jwt/api_jws.py new file mode 100644 index 0000000000000000000000000000000000000000..654ee0b744b35a81bb01fe9c4f1bccd041e2cab8 --- /dev/null +++ b/venv/lib/python3.10/site-packages/jwt/api_jws.py @@ -0,0 +1,346 @@ +from __future__ import annotations + +import binascii +import json +import warnings +from collections.abc import Sequence +from typing import TYPE_CHECKING, Any + +from .algorithms import ( + Algorithm, + get_default_algorithms, + has_crypto, + requires_cryptography, +) +from .api_jwk import PyJWK +from .exceptions import ( + DecodeError, + InvalidAlgorithmError, + InvalidSignatureError, + InvalidTokenError, +) +from .utils import base64url_decode, base64url_encode +from .warnings import RemovedInPyjwt3Warning + +if TYPE_CHECKING: + from .algorithms import AllowedPrivateKeys, AllowedPublicKeys + + +class PyJWS: + header_typ = "JWT" + + def __init__( + self, + algorithms: Sequence[str] | None = None, + options: dict[str, Any] | None = None, + ) -> None: + self._algorithms = get_default_algorithms() + self._valid_algs = ( + set(algorithms) if algorithms is not None else set(self._algorithms) + ) + + # Remove algorithms that aren't on the whitelist + for key in list(self._algorithms.keys()): + if key not in self._valid_algs: + del self._algorithms[key] + + if options is None: + options = {} + self.options = {**self._get_default_options(), **options} + + @staticmethod + def _get_default_options() -> dict[str, bool]: + return {"verify_signature": True} + + def register_algorithm(self, alg_id: str, alg_obj: Algorithm) -> None: + """ + Registers a new Algorithm for use when creating and verifying tokens. + """ + if alg_id in self._algorithms: + raise ValueError("Algorithm already has a handler.") + + if not isinstance(alg_obj, Algorithm): + raise TypeError("Object is not of type `Algorithm`") + + self._algorithms[alg_id] = alg_obj + self._valid_algs.add(alg_id) + + def unregister_algorithm(self, alg_id: str) -> None: + """ + Unregisters an Algorithm for use when creating and verifying tokens + Throws KeyError if algorithm is not registered. + """ + if alg_id not in self._algorithms: + raise KeyError( + "The specified algorithm could not be removed" + " because it is not registered." + ) + + del self._algorithms[alg_id] + self._valid_algs.remove(alg_id) + + def get_algorithms(self) -> list[str]: + """ + Returns a list of supported values for the 'alg' parameter. + """ + return list(self._valid_algs) + + def get_algorithm_by_name(self, alg_name: str) -> Algorithm: + """ + For a given string name, return the matching Algorithm object. + + Example usage: + + >>> jws_obj.get_algorithm_by_name("RS256") + """ + try: + return self._algorithms[alg_name] + except KeyError as e: + if not has_crypto and alg_name in requires_cryptography: + raise NotImplementedError( + f"Algorithm '{alg_name}' could not be found. Do you have cryptography installed?" + ) from e + raise NotImplementedError("Algorithm not supported") from e + + def encode( + self, + payload: bytes, + key: AllowedPrivateKeys | PyJWK | str | bytes, + algorithm: str | None = None, + headers: dict[str, Any] | None = None, + json_encoder: type[json.JSONEncoder] | None = None, + is_payload_detached: bool = False, + sort_headers: bool = True, + ) -> str: + segments = [] + + # declare a new var to narrow the type for type checkers + if algorithm is None: + if isinstance(key, PyJWK): + algorithm_ = key.algorithm_name + else: + algorithm_ = "HS256" + else: + algorithm_ = algorithm + + # Prefer headers values if present to function parameters. + if headers: + headers_alg = headers.get("alg") + if headers_alg: + algorithm_ = headers["alg"] + + headers_b64 = headers.get("b64") + if headers_b64 is False: + is_payload_detached = True + + # Header + header: dict[str, Any] = {"typ": self.header_typ, "alg": algorithm_} + + if headers: + self._validate_headers(headers) + header.update(headers) + + if not header["typ"]: + del header["typ"] + + if is_payload_detached: + header["b64"] = False + elif "b64" in header: + # True is the standard value for b64, so no need for it + del header["b64"] + + json_header = json.dumps( + header, separators=(",", ":"), cls=json_encoder, sort_keys=sort_headers + ).encode() + + segments.append(base64url_encode(json_header)) + + if is_payload_detached: + msg_payload = payload + else: + msg_payload = base64url_encode(payload) + segments.append(msg_payload) + + # Segments + signing_input = b".".join(segments) + + alg_obj = self.get_algorithm_by_name(algorithm_) + if isinstance(key, PyJWK): + key = key.key + key = alg_obj.prepare_key(key) + signature = alg_obj.sign(signing_input, key) + + segments.append(base64url_encode(signature)) + + # Don't put the payload content inside the encoded token when detached + if is_payload_detached: + segments[1] = b"" + encoded_string = b".".join(segments) + + return encoded_string.decode("utf-8") + + def decode_complete( + self, + jwt: str | bytes, + key: AllowedPublicKeys | PyJWK | str | bytes = "", + algorithms: Sequence[str] | None = None, + options: dict[str, Any] | None = None, + detached_payload: bytes | None = None, + **kwargs, + ) -> dict[str, Any]: + if kwargs: + warnings.warn( + "passing additional kwargs to decode_complete() is deprecated " + "and will be removed in pyjwt version 3. " + f"Unsupported kwargs: {tuple(kwargs.keys())}", + RemovedInPyjwt3Warning, + stacklevel=2, + ) + if options is None: + options = {} + merged_options = {**self.options, **options} + verify_signature = merged_options["verify_signature"] + + if verify_signature and not algorithms and not isinstance(key, PyJWK): + raise DecodeError( + 'It is required that you pass in a value for the "algorithms" argument when calling decode().' + ) + + payload, signing_input, header, signature = self._load(jwt) + + if header.get("b64", True) is False: + if detached_payload is None: + raise DecodeError( + 'It is required that you pass in a value for the "detached_payload" argument to decode a message having the b64 header set to false.' + ) + payload = detached_payload + signing_input = b".".join([signing_input.rsplit(b".", 1)[0], payload]) + + if verify_signature: + self._verify_signature(signing_input, header, signature, key, algorithms) + + return { + "payload": payload, + "header": header, + "signature": signature, + } + + def decode( + self, + jwt: str | bytes, + key: AllowedPublicKeys | PyJWK | str | bytes = "", + algorithms: Sequence[str] | None = None, + options: dict[str, Any] | None = None, + detached_payload: bytes | None = None, + **kwargs, + ) -> Any: + if kwargs: + warnings.warn( + "passing additional kwargs to decode() is deprecated " + "and will be removed in pyjwt version 3. " + f"Unsupported kwargs: {tuple(kwargs.keys())}", + RemovedInPyjwt3Warning, + stacklevel=2, + ) + decoded = self.decode_complete( + jwt, key, algorithms, options, detached_payload=detached_payload + ) + return decoded["payload"] + + def get_unverified_header(self, jwt: str | bytes) -> dict[str, Any]: + """Returns back the JWT header parameters as a dict() + + Note: The signature is not verified so the header parameters + should not be fully trusted until signature verification is complete + """ + headers = self._load(jwt)[2] + self._validate_headers(headers) + + return headers + + def _load(self, jwt: str | bytes) -> tuple[bytes, bytes, dict[str, Any], bytes]: + if isinstance(jwt, str): + jwt = jwt.encode("utf-8") + + if not isinstance(jwt, bytes): + raise DecodeError(f"Invalid token type. Token must be a {bytes}") + + try: + signing_input, crypto_segment = jwt.rsplit(b".", 1) + header_segment, payload_segment = signing_input.split(b".", 1) + except ValueError as err: + raise DecodeError("Not enough segments") from err + + try: + header_data = base64url_decode(header_segment) + except (TypeError, binascii.Error) as err: + raise DecodeError("Invalid header padding") from err + + try: + header = json.loads(header_data) + except ValueError as e: + raise DecodeError(f"Invalid header string: {e}") from e + + if not isinstance(header, dict): + raise DecodeError("Invalid header string: must be a json object") + + try: + payload = base64url_decode(payload_segment) + except (TypeError, binascii.Error) as err: + raise DecodeError("Invalid payload padding") from err + + try: + signature = base64url_decode(crypto_segment) + except (TypeError, binascii.Error) as err: + raise DecodeError("Invalid crypto padding") from err + + return (payload, signing_input, header, signature) + + def _verify_signature( + self, + signing_input: bytes, + header: dict[str, Any], + signature: bytes, + key: AllowedPublicKeys | PyJWK | str | bytes = "", + algorithms: Sequence[str] | None = None, + ) -> None: + if algorithms is None and isinstance(key, PyJWK): + algorithms = [key.algorithm_name] + try: + alg = header["alg"] + except KeyError: + raise InvalidAlgorithmError("Algorithm not specified") from None + + if not alg or (algorithms is not None and alg not in algorithms): + raise InvalidAlgorithmError("The specified alg value is not allowed") + + if isinstance(key, PyJWK): + alg_obj = key.Algorithm + prepared_key = key.key + else: + try: + alg_obj = self.get_algorithm_by_name(alg) + except NotImplementedError as e: + raise InvalidAlgorithmError("Algorithm not supported") from e + prepared_key = alg_obj.prepare_key(key) + + if not alg_obj.verify(signing_input, prepared_key, signature): + raise InvalidSignatureError("Signature verification failed") + + def _validate_headers(self, headers: dict[str, Any]) -> None: + if "kid" in headers: + self._validate_kid(headers["kid"]) + + def _validate_kid(self, kid: Any) -> None: + if not isinstance(kid, str): + raise InvalidTokenError("Key ID header parameter must be a string") + + +_jws_global_obj = PyJWS() +encode = _jws_global_obj.encode +decode_complete = _jws_global_obj.decode_complete +decode = _jws_global_obj.decode +register_algorithm = _jws_global_obj.register_algorithm +unregister_algorithm = _jws_global_obj.unregister_algorithm +get_algorithm_by_name = _jws_global_obj.get_algorithm_by_name +get_unverified_header = _jws_global_obj.get_unverified_header diff --git a/venv/lib/python3.10/site-packages/jwt/api_jwt.py b/venv/lib/python3.10/site-packages/jwt/api_jwt.py new file mode 100644 index 0000000000000000000000000000000000000000..3a201436c5b5064c32f117f6a1590a8af2af3b84 --- /dev/null +++ b/venv/lib/python3.10/site-packages/jwt/api_jwt.py @@ -0,0 +1,433 @@ +from __future__ import annotations + +import json +import warnings +from calendar import timegm +from collections.abc import Iterable, Sequence +from datetime import datetime, timedelta, timezone +from typing import TYPE_CHECKING, Any + +from . import api_jws +from .exceptions import ( + DecodeError, + ExpiredSignatureError, + ImmatureSignatureError, + InvalidAudienceError, + InvalidIssuedAtError, + InvalidIssuerError, + InvalidJTIError, + InvalidSubjectError, + MissingRequiredClaimError, +) +from .warnings import RemovedInPyjwt3Warning + +if TYPE_CHECKING: + from .algorithms import AllowedPrivateKeys, AllowedPublicKeys + from .api_jwk import PyJWK + + +class PyJWT: + def __init__(self, options: dict[str, Any] | None = None) -> None: + if options is None: + options = {} + self.options: dict[str, Any] = {**self._get_default_options(), **options} + + @staticmethod + def _get_default_options() -> dict[str, bool | list[str]]: + return { + "verify_signature": True, + "verify_exp": True, + "verify_nbf": True, + "verify_iat": True, + "verify_aud": True, + "verify_iss": True, + "verify_sub": True, + "verify_jti": True, + "require": [], + } + + def encode( + self, + payload: dict[str, Any], + key: AllowedPrivateKeys | PyJWK | str | bytes, + algorithm: str | None = None, + headers: dict[str, Any] | None = None, + json_encoder: type[json.JSONEncoder] | None = None, + sort_headers: bool = True, + ) -> str: + # Check that we get a dict + if not isinstance(payload, dict): + raise TypeError( + "Expecting a dict object, as JWT only supports " + "JSON objects as payloads." + ) + + # Payload + payload = payload.copy() + for time_claim in ["exp", "iat", "nbf"]: + # Convert datetime to a intDate value in known time-format claims + if isinstance(payload.get(time_claim), datetime): + payload[time_claim] = timegm(payload[time_claim].utctimetuple()) + + json_payload = self._encode_payload( + payload, + headers=headers, + json_encoder=json_encoder, + ) + + return api_jws.encode( + json_payload, + key, + algorithm, + headers, + json_encoder, + sort_headers=sort_headers, + ) + + def _encode_payload( + self, + payload: dict[str, Any], + headers: dict[str, Any] | None = None, + json_encoder: type[json.JSONEncoder] | None = None, + ) -> bytes: + """ + Encode a given payload to the bytes to be signed. + + This method is intended to be overridden by subclasses that need to + encode the payload in a different way, e.g. compress the payload. + """ + return json.dumps( + payload, + separators=(",", ":"), + cls=json_encoder, + ).encode("utf-8") + + def decode_complete( + self, + jwt: str | bytes, + key: AllowedPublicKeys | PyJWK | str | bytes = "", + algorithms: Sequence[str] | None = None, + options: dict[str, Any] | None = None, + # deprecated arg, remove in pyjwt3 + verify: bool | None = None, + # could be used as passthrough to api_jws, consider removal in pyjwt3 + detached_payload: bytes | None = None, + # passthrough arguments to _validate_claims + # consider putting in options + audience: str | Iterable[str] | None = None, + issuer: str | Sequence[str] | None = None, + subject: str | None = None, + leeway: float | timedelta = 0, + # kwargs + **kwargs: Any, + ) -> dict[str, Any]: + if kwargs: + warnings.warn( + "passing additional kwargs to decode_complete() is deprecated " + "and will be removed in pyjwt version 3. " + f"Unsupported kwargs: {tuple(kwargs.keys())}", + RemovedInPyjwt3Warning, + stacklevel=2, + ) + options = dict(options or {}) # shallow-copy or initialize an empty dict + options.setdefault("verify_signature", True) + + # If the user has set the legacy `verify` argument, and it doesn't match + # what the relevant `options` entry for the argument is, inform the user + # that they're likely making a mistake. + if verify is not None and verify != options["verify_signature"]: + warnings.warn( + "The `verify` argument to `decode` does nothing in PyJWT 2.0 and newer. " + "The equivalent is setting `verify_signature` to False in the `options` dictionary. " + "This invocation has a mismatch between the kwarg and the option entry.", + category=DeprecationWarning, + stacklevel=2, + ) + + if not options["verify_signature"]: + options.setdefault("verify_exp", False) + options.setdefault("verify_nbf", False) + options.setdefault("verify_iat", False) + options.setdefault("verify_aud", False) + options.setdefault("verify_iss", False) + options.setdefault("verify_sub", False) + options.setdefault("verify_jti", False) + + decoded = api_jws.decode_complete( + jwt, + key=key, + algorithms=algorithms, + options=options, + detached_payload=detached_payload, + ) + + payload = self._decode_payload(decoded) + + merged_options = {**self.options, **options} + self._validate_claims( + payload, + merged_options, + audience=audience, + issuer=issuer, + leeway=leeway, + subject=subject, + ) + + decoded["payload"] = payload + return decoded + + def _decode_payload(self, decoded: dict[str, Any]) -> Any: + """ + Decode the payload from a JWS dictionary (payload, signature, header). + + This method is intended to be overridden by subclasses that need to + decode the payload in a different way, e.g. decompress compressed + payloads. + """ + try: + payload = json.loads(decoded["payload"]) + except ValueError as e: + raise DecodeError(f"Invalid payload string: {e}") from e + if not isinstance(payload, dict): + raise DecodeError("Invalid payload string: must be a json object") + return payload + + def decode( + self, + jwt: str | bytes, + key: AllowedPublicKeys | PyJWK | str | bytes = "", + algorithms: Sequence[str] | None = None, + options: dict[str, Any] | None = None, + # deprecated arg, remove in pyjwt3 + verify: bool | None = None, + # could be used as passthrough to api_jws, consider removal in pyjwt3 + detached_payload: bytes | None = None, + # passthrough arguments to _validate_claims + # consider putting in options + audience: str | Iterable[str] | None = None, + subject: str | None = None, + issuer: str | Sequence[str] | None = None, + leeway: float | timedelta = 0, + # kwargs + **kwargs: Any, + ) -> Any: + if kwargs: + warnings.warn( + "passing additional kwargs to decode() is deprecated " + "and will be removed in pyjwt version 3. " + f"Unsupported kwargs: {tuple(kwargs.keys())}", + RemovedInPyjwt3Warning, + stacklevel=2, + ) + decoded = self.decode_complete( + jwt, + key, + algorithms, + options, + verify=verify, + detached_payload=detached_payload, + audience=audience, + subject=subject, + issuer=issuer, + leeway=leeway, + ) + return decoded["payload"] + + def _validate_claims( + self, + payload: dict[str, Any], + options: dict[str, Any], + audience=None, + issuer=None, + subject: str | None = None, + leeway: float | timedelta = 0, + ) -> None: + if isinstance(leeway, timedelta): + leeway = leeway.total_seconds() + + if audience is not None and not isinstance(audience, (str, Iterable)): + raise TypeError("audience must be a string, iterable or None") + + self._validate_required_claims(payload, options) + + now = datetime.now(tz=timezone.utc).timestamp() + + if "iat" in payload and options["verify_iat"]: + self._validate_iat(payload, now, leeway) + + if "nbf" in payload and options["verify_nbf"]: + self._validate_nbf(payload, now, leeway) + + if "exp" in payload and options["verify_exp"]: + self._validate_exp(payload, now, leeway) + + if options["verify_iss"]: + self._validate_iss(payload, issuer) + + if options["verify_aud"]: + self._validate_aud( + payload, audience, strict=options.get("strict_aud", False) + ) + + if options["verify_sub"]: + self._validate_sub(payload, subject) + + if options["verify_jti"]: + self._validate_jti(payload) + + def _validate_required_claims( + self, + payload: dict[str, Any], + options: dict[str, Any], + ) -> None: + for claim in options["require"]: + if payload.get(claim) is None: + raise MissingRequiredClaimError(claim) + + def _validate_sub(self, payload: dict[str, Any], subject=None) -> None: + """ + Checks whether "sub" if in the payload is valid ot not. + This is an Optional claim + + :param payload(dict): The payload which needs to be validated + :param subject(str): The subject of the token + """ + + if "sub" not in payload: + return + + if not isinstance(payload["sub"], str): + raise InvalidSubjectError("Subject must be a string") + + if subject is not None: + if payload.get("sub") != subject: + raise InvalidSubjectError("Invalid subject") + + def _validate_jti(self, payload: dict[str, Any]) -> None: + """ + Checks whether "jti" if in the payload is valid ot not + This is an Optional claim + + :param payload(dict): The payload which needs to be validated + """ + + if "jti" not in payload: + return + + if not isinstance(payload.get("jti"), str): + raise InvalidJTIError("JWT ID must be a string") + + def _validate_iat( + self, + payload: dict[str, Any], + now: float, + leeway: float, + ) -> None: + try: + iat = int(payload["iat"]) + except ValueError: + raise InvalidIssuedAtError( + "Issued At claim (iat) must be an integer." + ) from None + if iat > (now + leeway): + raise ImmatureSignatureError("The token is not yet valid (iat)") + + def _validate_nbf( + self, + payload: dict[str, Any], + now: float, + leeway: float, + ) -> None: + try: + nbf = int(payload["nbf"]) + except ValueError: + raise DecodeError("Not Before claim (nbf) must be an integer.") from None + + if nbf > (now + leeway): + raise ImmatureSignatureError("The token is not yet valid (nbf)") + + def _validate_exp( + self, + payload: dict[str, Any], + now: float, + leeway: float, + ) -> None: + try: + exp = int(payload["exp"]) + except ValueError: + raise DecodeError( + "Expiration Time claim (exp) must be an integer." + ) from None + + if exp <= (now - leeway): + raise ExpiredSignatureError("Signature has expired") + + def _validate_aud( + self, + payload: dict[str, Any], + audience: str | Iterable[str] | None, + *, + strict: bool = False, + ) -> None: + if audience is None: + if "aud" not in payload or not payload["aud"]: + return + # Application did not specify an audience, but + # the token has the 'aud' claim + raise InvalidAudienceError("Invalid audience") + + if "aud" not in payload or not payload["aud"]: + # Application specified an audience, but it could not be + # verified since the token does not contain a claim. + raise MissingRequiredClaimError("aud") + + audience_claims = payload["aud"] + + # In strict mode, we forbid list matching: the supplied audience + # must be a string, and it must exactly match the audience claim. + if strict: + # Only a single audience is allowed in strict mode. + if not isinstance(audience, str): + raise InvalidAudienceError("Invalid audience (strict)") + + # Only a single audience claim is allowed in strict mode. + if not isinstance(audience_claims, str): + raise InvalidAudienceError("Invalid claim format in token (strict)") + + if audience != audience_claims: + raise InvalidAudienceError("Audience doesn't match (strict)") + + return + + if isinstance(audience_claims, str): + audience_claims = [audience_claims] + if not isinstance(audience_claims, list): + raise InvalidAudienceError("Invalid claim format in token") + if any(not isinstance(c, str) for c in audience_claims): + raise InvalidAudienceError("Invalid claim format in token") + + if isinstance(audience, str): + audience = [audience] + + if all(aud not in audience_claims for aud in audience): + raise InvalidAudienceError("Audience doesn't match") + + def _validate_iss(self, payload: dict[str, Any], issuer: Any) -> None: + if issuer is None: + return + + if "iss" not in payload: + raise MissingRequiredClaimError("iss") + + if isinstance(issuer, str): + if payload["iss"] != issuer: + raise InvalidIssuerError("Invalid issuer") + else: + if payload["iss"] not in issuer: + raise InvalidIssuerError("Invalid issuer") + + +_jwt_global_obj = PyJWT() +encode = _jwt_global_obj.encode +decode_complete = _jwt_global_obj.decode_complete +decode = _jwt_global_obj.decode diff --git a/venv/lib/python3.10/site-packages/jwt/exceptions.py b/venv/lib/python3.10/site-packages/jwt/exceptions.py new file mode 100644 index 0000000000000000000000000000000000000000..9b45ae48c4e0f2c7627c152bf0059e57f17b73ca --- /dev/null +++ b/venv/lib/python3.10/site-packages/jwt/exceptions.py @@ -0,0 +1,82 @@ +class PyJWTError(Exception): + """ + Base class for all exceptions + """ + + pass + + +class InvalidTokenError(PyJWTError): + pass + + +class DecodeError(InvalidTokenError): + pass + + +class InvalidSignatureError(DecodeError): + pass + + +class ExpiredSignatureError(InvalidTokenError): + pass + + +class InvalidAudienceError(InvalidTokenError): + pass + + +class InvalidIssuerError(InvalidTokenError): + pass + + +class InvalidIssuedAtError(InvalidTokenError): + pass + + +class ImmatureSignatureError(InvalidTokenError): + pass + + +class InvalidKeyError(PyJWTError): + pass + + +class InvalidAlgorithmError(InvalidTokenError): + pass + + +class MissingRequiredClaimError(InvalidTokenError): + def __init__(self, claim: str) -> None: + self.claim = claim + + def __str__(self) -> str: + return f'Token is missing the "{self.claim}" claim' + + +class PyJWKError(PyJWTError): + pass + + +class MissingCryptographyError(PyJWKError): + pass + + +class PyJWKSetError(PyJWTError): + pass + + +class PyJWKClientError(PyJWTError): + pass + + +class PyJWKClientConnectionError(PyJWKClientError): + pass + + +class InvalidSubjectError(InvalidTokenError): + pass + + +class InvalidJTIError(InvalidTokenError): + pass diff --git a/venv/lib/python3.10/site-packages/jwt/help.py b/venv/lib/python3.10/site-packages/jwt/help.py new file mode 100644 index 0000000000000000000000000000000000000000..8e1c2286e7ea936292f525c44bcc1b0ee30c6560 --- /dev/null +++ b/venv/lib/python3.10/site-packages/jwt/help.py @@ -0,0 +1,67 @@ +import json +import platform +import sys +from typing import Dict + +from . import __version__ as pyjwt_version + +try: + import cryptography + + cryptography_version = cryptography.__version__ +except ModuleNotFoundError: + cryptography_version = "" + + +def info() -> Dict[str, Dict[str, str]]: + """ + Generate information for a bug report. + Based on the requests package help utility module. + """ + try: + platform_info = { + "system": platform.system(), + "release": platform.release(), + } + except OSError: + platform_info = {"system": "Unknown", "release": "Unknown"} + + implementation = platform.python_implementation() + + if implementation == "CPython": + implementation_version = platform.python_version() + elif implementation == "PyPy": + pypy_version_info = sys.pypy_version_info # type: ignore[attr-defined] + implementation_version = ( + f"{pypy_version_info.major}." + f"{pypy_version_info.minor}." + f"{pypy_version_info.micro}" + ) + if pypy_version_info.releaselevel != "final": + implementation_version = "".join( + [ + implementation_version, + pypy_version_info.releaselevel, + ] + ) + else: + implementation_version = "Unknown" + + return { + "platform": platform_info, + "implementation": { + "name": implementation, + "version": implementation_version, + }, + "cryptography": {"version": cryptography_version}, + "pyjwt": {"version": pyjwt_version}, + } + + +def main() -> None: + """Pretty-print the bug information as JSON.""" + print(json.dumps(info(), sort_keys=True, indent=2)) + + +if __name__ == "__main__": + main() diff --git a/venv/lib/python3.10/site-packages/jwt/jwk_set_cache.py b/venv/lib/python3.10/site-packages/jwt/jwk_set_cache.py new file mode 100644 index 0000000000000000000000000000000000000000..2432563056a460c59d6e615968df456da3ec49cd --- /dev/null +++ b/venv/lib/python3.10/site-packages/jwt/jwk_set_cache.py @@ -0,0 +1,31 @@ +import time +from typing import Optional + +from .api_jwk import PyJWKSet, PyJWTSetWithTimestamp + + +class JWKSetCache: + def __init__(self, lifespan: int) -> None: + self.jwk_set_with_timestamp: Optional[PyJWTSetWithTimestamp] = None + self.lifespan = lifespan + + def put(self, jwk_set: PyJWKSet) -> None: + if jwk_set is not None: + self.jwk_set_with_timestamp = PyJWTSetWithTimestamp(jwk_set) + else: + # clear cache + self.jwk_set_with_timestamp = None + + def get(self) -> Optional[PyJWKSet]: + if self.jwk_set_with_timestamp is None or self.is_expired(): + return None + + return self.jwk_set_with_timestamp.get_jwk_set() + + def is_expired(self) -> bool: + return ( + self.jwk_set_with_timestamp is not None + and self.lifespan > -1 + and time.monotonic() + > self.jwk_set_with_timestamp.get_timestamp() + self.lifespan + ) diff --git a/venv/lib/python3.10/site-packages/jwt/jwks_client.py b/venv/lib/python3.10/site-packages/jwt/jwks_client.py new file mode 100644 index 0000000000000000000000000000000000000000..9a8992ca6b9d2be153e203b11465a8e6e7cdc570 --- /dev/null +++ b/venv/lib/python3.10/site-packages/jwt/jwks_client.py @@ -0,0 +1,126 @@ +import json +import urllib.request +from functools import lru_cache +from ssl import SSLContext +from typing import Any, Dict, List, Optional +from urllib.error import URLError + +from .api_jwk import PyJWK, PyJWKSet +from .api_jwt import decode_complete as decode_token +from .exceptions import PyJWKClientConnectionError, PyJWKClientError +from .jwk_set_cache import JWKSetCache + + +class PyJWKClient: + def __init__( + self, + uri: str, + cache_keys: bool = False, + max_cached_keys: int = 16, + cache_jwk_set: bool = True, + lifespan: int = 300, + headers: Optional[Dict[str, Any]] = None, + timeout: int = 30, + ssl_context: Optional[SSLContext] = None, + ): + if headers is None: + headers = {} + self.uri = uri + self.jwk_set_cache: Optional[JWKSetCache] = None + self.headers = headers + self.timeout = timeout + self.ssl_context = ssl_context + + if cache_jwk_set: + # Init jwt set cache with default or given lifespan. + # Default lifespan is 300 seconds (5 minutes). + if lifespan <= 0: + raise PyJWKClientError( + f'Lifespan must be greater than 0, the input is "{lifespan}"' + ) + self.jwk_set_cache = JWKSetCache(lifespan) + else: + self.jwk_set_cache = None + + if cache_keys: + # Cache signing keys + # Ignore mypy (https://github.com/python/mypy/issues/2427) + self.get_signing_key = lru_cache(maxsize=max_cached_keys)( + self.get_signing_key + ) # type: ignore + + def fetch_data(self) -> Any: + jwk_set: Any = None + try: + r = urllib.request.Request(url=self.uri, headers=self.headers) + with urllib.request.urlopen( + r, timeout=self.timeout, context=self.ssl_context + ) as response: + jwk_set = json.load(response) + except (URLError, TimeoutError) as e: + raise PyJWKClientConnectionError( + f'Fail to fetch data from the url, err: "{e}"' + ) from e + else: + return jwk_set + finally: + if self.jwk_set_cache is not None: + self.jwk_set_cache.put(jwk_set) + + def get_jwk_set(self, refresh: bool = False) -> PyJWKSet: + data = None + if self.jwk_set_cache is not None and not refresh: + data = self.jwk_set_cache.get() + + if data is None: + data = self.fetch_data() + + if not isinstance(data, dict): + raise PyJWKClientError("The JWKS endpoint did not return a JSON object") + + return PyJWKSet.from_dict(data) + + def get_signing_keys(self, refresh: bool = False) -> List[PyJWK]: + jwk_set = self.get_jwk_set(refresh) + signing_keys = [ + jwk_set_key + for jwk_set_key in jwk_set.keys + if jwk_set_key.public_key_use in ["sig", None] and jwk_set_key.key_id + ] + + if not signing_keys: + raise PyJWKClientError("The JWKS endpoint did not contain any signing keys") + + return signing_keys + + def get_signing_key(self, kid: str) -> PyJWK: + signing_keys = self.get_signing_keys() + signing_key = self.match_kid(signing_keys, kid) + + if not signing_key: + # If no matching signing key from the jwk set, refresh the jwk set and try again. + signing_keys = self.get_signing_keys(refresh=True) + signing_key = self.match_kid(signing_keys, kid) + + if not signing_key: + raise PyJWKClientError( + f'Unable to find a signing key that matches: "{kid}"' + ) + + return signing_key + + def get_signing_key_from_jwt(self, token: str) -> PyJWK: + unverified = decode_token(token, options={"verify_signature": False}) + header = unverified["header"] + return self.get_signing_key(header.get("kid")) + + @staticmethod + def match_kid(signing_keys: List[PyJWK], kid: str) -> Optional[PyJWK]: + signing_key = None + + for key in signing_keys: + if key.key_id == kid: + signing_key = key + break + + return signing_key diff --git a/venv/lib/python3.10/site-packages/jwt/py.typed b/venv/lib/python3.10/site-packages/jwt/py.typed new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/venv/lib/python3.10/site-packages/jwt/types.py b/venv/lib/python3.10/site-packages/jwt/types.py new file mode 100644 index 0000000000000000000000000000000000000000..7d9935205688a3cb45902017884633a490d59055 --- /dev/null +++ b/venv/lib/python3.10/site-packages/jwt/types.py @@ -0,0 +1,5 @@ +from typing import Any, Callable, Dict + +JWKDict = Dict[str, Any] + +HashlibHash = Callable[..., Any] diff --git a/venv/lib/python3.10/site-packages/jwt/utils.py b/venv/lib/python3.10/site-packages/jwt/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..56e89bb7782f4b5f61c9b2197258dcbf45335d33 --- /dev/null +++ b/venv/lib/python3.10/site-packages/jwt/utils.py @@ -0,0 +1,142 @@ +import base64 +import binascii +import re +from typing import Optional, Union + +try: + from cryptography.hazmat.primitives.asymmetric.ec import EllipticCurve + from cryptography.hazmat.primitives.asymmetric.utils import ( + decode_dss_signature, + encode_dss_signature, + ) +except ModuleNotFoundError: + pass + + +def force_bytes(value: Union[bytes, str]) -> bytes: + if isinstance(value, str): + return value.encode("utf-8") + elif isinstance(value, bytes): + return value + else: + raise TypeError("Expected a string value") + + +def base64url_decode(input: Union[bytes, str]) -> bytes: + input_bytes = force_bytes(input) + + rem = len(input_bytes) % 4 + + if rem > 0: + input_bytes += b"=" * (4 - rem) + + return base64.urlsafe_b64decode(input_bytes) + + +def base64url_encode(input: bytes) -> bytes: + return base64.urlsafe_b64encode(input).replace(b"=", b"") + + +def to_base64url_uint(val: int, *, bit_length: Optional[int] = None) -> bytes: + if val < 0: + raise ValueError("Must be a positive integer") + + int_bytes = bytes_from_int(val, bit_length=bit_length) + + if len(int_bytes) == 0: + int_bytes = b"\x00" + + return base64url_encode(int_bytes) + + +def from_base64url_uint(val: Union[bytes, str]) -> int: + data = base64url_decode(force_bytes(val)) + return int.from_bytes(data, byteorder="big") + + +def number_to_bytes(num: int, num_bytes: int) -> bytes: + padded_hex = "%0*x" % (2 * num_bytes, num) + return binascii.a2b_hex(padded_hex.encode("ascii")) + + +def bytes_to_number(string: bytes) -> int: + return int(binascii.b2a_hex(string), 16) + + +def bytes_from_int(val: int, *, bit_length: Optional[int] = None) -> bytes: + if bit_length is None: + bit_length = val.bit_length() + byte_length = (bit_length + 7) // 8 + + return val.to_bytes(byte_length, "big", signed=False) + + +def der_to_raw_signature(der_sig: bytes, curve: "EllipticCurve") -> bytes: + num_bits = curve.key_size + num_bytes = (num_bits + 7) // 8 + + r, s = decode_dss_signature(der_sig) + + return number_to_bytes(r, num_bytes) + number_to_bytes(s, num_bytes) + + +def raw_to_der_signature(raw_sig: bytes, curve: "EllipticCurve") -> bytes: + num_bits = curve.key_size + num_bytes = (num_bits + 7) // 8 + + if len(raw_sig) != 2 * num_bytes: + raise ValueError("Invalid signature") + + r = bytes_to_number(raw_sig[:num_bytes]) + s = bytes_to_number(raw_sig[num_bytes:]) + + return bytes(encode_dss_signature(r, s)) + + +# Based on https://github.com/hynek/pem/blob/7ad94db26b0bc21d10953f5dbad3acfdfacf57aa/src/pem/_core.py#L224-L252 +_PEMS = { + b"CERTIFICATE", + b"TRUSTED CERTIFICATE", + b"PRIVATE KEY", + b"PUBLIC KEY", + b"ENCRYPTED PRIVATE KEY", + b"OPENSSH PRIVATE KEY", + b"DSA PRIVATE KEY", + b"RSA PRIVATE KEY", + b"RSA PUBLIC KEY", + b"EC PRIVATE KEY", + b"DH PARAMETERS", + b"NEW CERTIFICATE REQUEST", + b"CERTIFICATE REQUEST", + b"SSH2 PUBLIC KEY", + b"SSH2 ENCRYPTED PRIVATE KEY", + b"X509 CRL", +} + +_PEM_RE = re.compile( + b"----[- ]BEGIN (" + + b"|".join(_PEMS) + + b""")[- ]----\r? +.+?\r? +----[- ]END \\1[- ]----\r?\n?""", + re.DOTALL, +) + + +def is_pem_format(key: bytes) -> bool: + return bool(_PEM_RE.search(key)) + + +# Based on https://github.com/pyca/cryptography/blob/bcb70852d577b3f490f015378c75cba74986297b/src/cryptography/hazmat/primitives/serialization/ssh.py#L40-L46 +_SSH_KEY_FORMATS = ( + b"ssh-ed25519", + b"ssh-rsa", + b"ssh-dss", + b"ecdsa-sha2-nistp256", + b"ecdsa-sha2-nistp384", + b"ecdsa-sha2-nistp521", +) + + +def is_ssh_key(key: bytes) -> bool: + return key.startswith(_SSH_KEY_FORMATS) diff --git a/venv/lib/python3.10/site-packages/jwt/warnings.py b/venv/lib/python3.10/site-packages/jwt/warnings.py new file mode 100644 index 0000000000000000000000000000000000000000..8762a8cbb3408b5b30901d4dd39fc53097eb31d8 --- /dev/null +++ b/venv/lib/python3.10/site-packages/jwt/warnings.py @@ -0,0 +1,2 @@ +class RemovedInPyjwt3Warning(DeprecationWarning): + pass diff --git a/venv/lib/python3.10/site-packages/pygments/filters/__init__.py b/venv/lib/python3.10/site-packages/pygments/filters/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..2fed761a01ee5e9e25918b9a0619fcb47c8811b1 --- /dev/null +++ b/venv/lib/python3.10/site-packages/pygments/filters/__init__.py @@ -0,0 +1,940 @@ +""" + pygments.filters + ~~~~~~~~~~~~~~~~ + + Module containing filter lookup functions and default + filters. + + :copyright: Copyright 2006-2025 by the Pygments team, see AUTHORS. + :license: BSD, see LICENSE for details. +""" + +import re + +from pygments.token import String, Comment, Keyword, Name, Error, Whitespace, \ + string_to_tokentype +from pygments.filter import Filter +from pygments.util import get_list_opt, get_int_opt, get_bool_opt, \ + get_choice_opt, ClassNotFound, OptionError +from pygments.plugin import find_plugin_filters + + +def find_filter_class(filtername): + """Lookup a filter by name. Return None if not found.""" + if filtername in FILTERS: + return FILTERS[filtername] + for name, cls in find_plugin_filters(): + if name == filtername: + return cls + return None + + +def get_filter_by_name(filtername, **options): + """Return an instantiated filter. + + Options are passed to the filter initializer if wanted. + Raise a ClassNotFound if not found. + """ + cls = find_filter_class(filtername) + if cls: + return cls(**options) + else: + raise ClassNotFound(f'filter {filtername!r} not found') + + +def get_all_filters(): + """Return a generator of all filter names.""" + yield from FILTERS + for name, _ in find_plugin_filters(): + yield name + + +def _replace_special(ttype, value, regex, specialttype, + replacefunc=lambda x: x): + last = 0 + for match in regex.finditer(value): + start, end = match.start(), match.end() + if start != last: + yield ttype, value[last:start] + yield specialttype, replacefunc(value[start:end]) + last = end + if last != len(value): + yield ttype, value[last:] + + +class CodeTagFilter(Filter): + """Highlight special code tags in comments and docstrings. + + Options accepted: + + `codetags` : list of strings + A list of strings that are flagged as code tags. The default is to + highlight ``XXX``, ``TODO``, ``FIXME``, ``BUG`` and ``NOTE``. + + .. versionchanged:: 2.13 + Now recognizes ``FIXME`` by default. + """ + + def __init__(self, **options): + Filter.__init__(self, **options) + tags = get_list_opt(options, 'codetags', + ['XXX', 'TODO', 'FIXME', 'BUG', 'NOTE']) + self.tag_re = re.compile(r'\b({})\b'.format('|'.join([ + re.escape(tag) for tag in tags if tag + ]))) + + def filter(self, lexer, stream): + regex = self.tag_re + for ttype, value in stream: + if ttype in String.Doc or \ + ttype in Comment and \ + ttype not in Comment.Preproc: + yield from _replace_special(ttype, value, regex, Comment.Special) + else: + yield ttype, value + + +class SymbolFilter(Filter): + """Convert mathematical symbols such as \\ in Isabelle + or \\longrightarrow in LaTeX into Unicode characters. + + This is mostly useful for HTML or console output when you want to + approximate the source rendering you'd see in an IDE. + + Options accepted: + + `lang` : string + The symbol language. Must be one of ``'isabelle'`` or + ``'latex'``. The default is ``'isabelle'``. + """ + + latex_symbols = { + '\\alpha' : '\U000003b1', + '\\beta' : '\U000003b2', + '\\gamma' : '\U000003b3', + '\\delta' : '\U000003b4', + '\\varepsilon' : '\U000003b5', + '\\zeta' : '\U000003b6', + '\\eta' : '\U000003b7', + '\\vartheta' : '\U000003b8', + '\\iota' : '\U000003b9', + '\\kappa' : '\U000003ba', + '\\lambda' : '\U000003bb', + '\\mu' : '\U000003bc', + '\\nu' : '\U000003bd', + '\\xi' : '\U000003be', + '\\pi' : '\U000003c0', + '\\varrho' : '\U000003c1', + '\\sigma' : '\U000003c3', + '\\tau' : '\U000003c4', + '\\upsilon' : '\U000003c5', + '\\varphi' : '\U000003c6', + '\\chi' : '\U000003c7', + '\\psi' : '\U000003c8', + '\\omega' : '\U000003c9', + '\\Gamma' : '\U00000393', + '\\Delta' : '\U00000394', + '\\Theta' : '\U00000398', + '\\Lambda' : '\U0000039b', + '\\Xi' : '\U0000039e', + '\\Pi' : '\U000003a0', + '\\Sigma' : '\U000003a3', + '\\Upsilon' : '\U000003a5', + '\\Phi' : '\U000003a6', + '\\Psi' : '\U000003a8', + '\\Omega' : '\U000003a9', + '\\leftarrow' : '\U00002190', + '\\longleftarrow' : '\U000027f5', + '\\rightarrow' : '\U00002192', + '\\longrightarrow' : '\U000027f6', + '\\Leftarrow' : '\U000021d0', + '\\Longleftarrow' : '\U000027f8', + '\\Rightarrow' : '\U000021d2', + '\\Longrightarrow' : '\U000027f9', + '\\leftrightarrow' : '\U00002194', + '\\longleftrightarrow' : '\U000027f7', + '\\Leftrightarrow' : '\U000021d4', + '\\Longleftrightarrow' : '\U000027fa', + '\\mapsto' : '\U000021a6', + '\\longmapsto' : '\U000027fc', + '\\relbar' : '\U00002500', + '\\Relbar' : '\U00002550', + '\\hookleftarrow' : '\U000021a9', + '\\hookrightarrow' : '\U000021aa', + '\\leftharpoondown' : '\U000021bd', + '\\rightharpoondown' : '\U000021c1', + '\\leftharpoonup' : '\U000021bc', + '\\rightharpoonup' : '\U000021c0', + '\\rightleftharpoons' : '\U000021cc', + '\\leadsto' : '\U0000219d', + '\\downharpoonleft' : '\U000021c3', + '\\downharpoonright' : '\U000021c2', + '\\upharpoonleft' : '\U000021bf', + '\\upharpoonright' : '\U000021be', + '\\restriction' : '\U000021be', + '\\uparrow' : '\U00002191', + '\\Uparrow' : '\U000021d1', + '\\downarrow' : '\U00002193', + '\\Downarrow' : '\U000021d3', + '\\updownarrow' : '\U00002195', + '\\Updownarrow' : '\U000021d5', + '\\langle' : '\U000027e8', + '\\rangle' : '\U000027e9', + '\\lceil' : '\U00002308', + '\\rceil' : '\U00002309', + '\\lfloor' : '\U0000230a', + '\\rfloor' : '\U0000230b', + '\\flqq' : '\U000000ab', + '\\frqq' : '\U000000bb', + '\\bot' : '\U000022a5', + '\\top' : '\U000022a4', + '\\wedge' : '\U00002227', + '\\bigwedge' : '\U000022c0', + '\\vee' : '\U00002228', + '\\bigvee' : '\U000022c1', + '\\forall' : '\U00002200', + '\\exists' : '\U00002203', + '\\nexists' : '\U00002204', + '\\neg' : '\U000000ac', + '\\Box' : '\U000025a1', + '\\Diamond' : '\U000025c7', + '\\vdash' : '\U000022a2', + '\\models' : '\U000022a8', + '\\dashv' : '\U000022a3', + '\\surd' : '\U0000221a', + '\\le' : '\U00002264', + '\\ge' : '\U00002265', + '\\ll' : '\U0000226a', + '\\gg' : '\U0000226b', + '\\lesssim' : '\U00002272', + '\\gtrsim' : '\U00002273', + '\\lessapprox' : '\U00002a85', + '\\gtrapprox' : '\U00002a86', + '\\in' : '\U00002208', + '\\notin' : '\U00002209', + '\\subset' : '\U00002282', + '\\supset' : '\U00002283', + '\\subseteq' : '\U00002286', + '\\supseteq' : '\U00002287', + '\\sqsubset' : '\U0000228f', + '\\sqsupset' : '\U00002290', + '\\sqsubseteq' : '\U00002291', + '\\sqsupseteq' : '\U00002292', + '\\cap' : '\U00002229', + '\\bigcap' : '\U000022c2', + '\\cup' : '\U0000222a', + '\\bigcup' : '\U000022c3', + '\\sqcup' : '\U00002294', + '\\bigsqcup' : '\U00002a06', + '\\sqcap' : '\U00002293', + '\\Bigsqcap' : '\U00002a05', + '\\setminus' : '\U00002216', + '\\propto' : '\U0000221d', + '\\uplus' : '\U0000228e', + '\\bigplus' : '\U00002a04', + '\\sim' : '\U0000223c', + '\\doteq' : '\U00002250', + '\\simeq' : '\U00002243', + '\\approx' : '\U00002248', + '\\asymp' : '\U0000224d', + '\\cong' : '\U00002245', + '\\equiv' : '\U00002261', + '\\Join' : '\U000022c8', + '\\bowtie' : '\U00002a1d', + '\\prec' : '\U0000227a', + '\\succ' : '\U0000227b', + '\\preceq' : '\U0000227c', + '\\succeq' : '\U0000227d', + '\\parallel' : '\U00002225', + '\\mid' : '\U000000a6', + '\\pm' : '\U000000b1', + '\\mp' : '\U00002213', + '\\times' : '\U000000d7', + '\\div' : '\U000000f7', + '\\cdot' : '\U000022c5', + '\\star' : '\U000022c6', + '\\circ' : '\U00002218', + '\\dagger' : '\U00002020', + '\\ddagger' : '\U00002021', + '\\lhd' : '\U000022b2', + '\\rhd' : '\U000022b3', + '\\unlhd' : '\U000022b4', + '\\unrhd' : '\U000022b5', + '\\triangleleft' : '\U000025c3', + '\\triangleright' : '\U000025b9', + '\\triangle' : '\U000025b3', + '\\triangleq' : '\U0000225c', + '\\oplus' : '\U00002295', + '\\bigoplus' : '\U00002a01', + '\\otimes' : '\U00002297', + '\\bigotimes' : '\U00002a02', + '\\odot' : '\U00002299', + '\\bigodot' : '\U00002a00', + '\\ominus' : '\U00002296', + '\\oslash' : '\U00002298', + '\\dots' : '\U00002026', + '\\cdots' : '\U000022ef', + '\\sum' : '\U00002211', + '\\prod' : '\U0000220f', + '\\coprod' : '\U00002210', + '\\infty' : '\U0000221e', + '\\int' : '\U0000222b', + '\\oint' : '\U0000222e', + '\\clubsuit' : '\U00002663', + '\\diamondsuit' : '\U00002662', + '\\heartsuit' : '\U00002661', + '\\spadesuit' : '\U00002660', + '\\aleph' : '\U00002135', + '\\emptyset' : '\U00002205', + '\\nabla' : '\U00002207', + '\\partial' : '\U00002202', + '\\flat' : '\U0000266d', + '\\natural' : '\U0000266e', + '\\sharp' : '\U0000266f', + '\\angle' : '\U00002220', + '\\copyright' : '\U000000a9', + '\\textregistered' : '\U000000ae', + '\\textonequarter' : '\U000000bc', + '\\textonehalf' : '\U000000bd', + '\\textthreequarters' : '\U000000be', + '\\textordfeminine' : '\U000000aa', + '\\textordmasculine' : '\U000000ba', + '\\euro' : '\U000020ac', + '\\pounds' : '\U000000a3', + '\\yen' : '\U000000a5', + '\\textcent' : '\U000000a2', + '\\textcurrency' : '\U000000a4', + '\\textdegree' : '\U000000b0', + } + + isabelle_symbols = { + '\\' : '\U0001d7ec', + '\\' : '\U0001d7ed', + '\\' : '\U0001d7ee', + '\\' : '\U0001d7ef', + '\\' : '\U0001d7f0', + '\\' : '\U0001d7f1', + '\\' : '\U0001d7f2', + '\\' : '\U0001d7f3', + '\\' : '\U0001d7f4', + '\\' : '\U0001d7f5', + '\\' : '\U0001d49c', + '\\' : '\U0000212c', + '\\' : '\U0001d49e', + '\\' : '\U0001d49f', + '\\' : '\U00002130', + '\\' : '\U00002131', + '\\' : '\U0001d4a2', + '\\' : '\U0000210b', + '\\' : '\U00002110', + '\\' : '\U0001d4a5', + '\\' : '\U0001d4a6', + '\\' : '\U00002112', + '\\' : '\U00002133', + '\\' : '\U0001d4a9', + '\\' : '\U0001d4aa', + '\\

' : '\U0001d5c9', + '\\' : '\U0001d5ca', + '\\' : '\U0001d5cb', + '\\' : '\U0001d5cc', + '\\' : '\U0001d5cd', + '\\' : '\U0001d5ce', + '\\' : '\U0001d5cf', + '\\' : '\U0001d5d0', + '\\' : '\U0001d5d1', + '\\' : '\U0001d5d2', + '\\' : '\U0001d5d3', + '\\' : '\U0001d504', + '\\' : '\U0001d505', + '\\' : '\U0000212d', + '\\

' : '\U0001d507', + '\\' : '\U0001d508', + '\\' : '\U0001d509', + '\\' : '\U0001d50a', + '\\' : '\U0000210c', + '\\' : '\U00002111', + '\\' : '\U0001d50d', + '\\' : '\U0001d50e', + '\\' : '\U0001d50f', + '\\' : '\U0001d510', + '\\' : '\U0001d511', + '\\' : '\U0001d512', + '\\' : '\U0001d513', + '\\' : '\U0001d514', + '\\' : '\U0000211c', + '\\' : '\U0001d516', + '\\' : '\U0001d517', + '\\' : '\U0001d518', + '\\' : '\U0001d519', + '\\' : '\U0001d51a', + '\\' : '\U0001d51b', + '\\' : '\U0001d51c', + '\\' : '\U00002128', + '\\' : '\U0001d51e', + '\\' : '\U0001d51f', + '\\' : '\U0001d520', + '\\
' : '\U0001d521', + '\\' : '\U0001d522', + '\\' : '\U0001d523', + '\\' : '\U0001d524', + '\\' : '\U0001d525', + '\\' : '\U0001d526', + '\\' : '\U0001d527', + '\\' : '\U0001d528', + '\\' : '\U0001d529', + '\\' : '\U0001d52a', + '\\' : '\U0001d52b', + '\\' : '\U0001d52c', + '\\' : '\U0001d52d', + '\\' : '\U0001d52e', + '\\' : '\U0001d52f', + '\\' : '\U0001d530', + '\\' : '\U0001d531', + '\\' : '\U0001d532', + '\\' : '\U0001d533', + '\\' : '\U0001d534', + '\\' : '\U0001d535', + '\\' : '\U0001d536', + '\\' : '\U0001d537', + '\\' : '\U000003b1', + '\\' : '\U000003b2', + '\\' : '\U000003b3', + '\\' : '\U000003b4', + '\\' : '\U000003b5', + '\\' : '\U000003b6', + '\\' : '\U000003b7', + '\\' : '\U000003b8', + '\\' : '\U000003b9', + '\\' : '\U000003ba', + '\\' : '\U000003bb', + '\\' : '\U000003bc', + '\\' : '\U000003bd', + '\\' : '\U000003be', + '\\' : '\U000003c0', + '\\' : '\U000003c1', + '\\' : '\U000003c3', + '\\' : '\U000003c4', + '\\' : '\U000003c5', + '\\' : '\U000003c6', + '\\' : '\U000003c7', + '\\' : '\U000003c8', + '\\' : '\U000003c9', + '\\' : '\U00000393', + '\\' : '\U00000394', + '\\' : '\U00000398', + '\\' : '\U0000039b', + '\\' : '\U0000039e', + '\\' : '\U000003a0', + '\\' : '\U000003a3', + '\\' : '\U000003a5', + '\\' : '\U000003a6', + '\\' : '\U000003a8', + '\\' : '\U000003a9', + '\\' : '\U0001d539', + '\\' : '\U00002102', + '\\' : '\U00002115', + '\\' : '\U0000211a', + '\\' : '\U0000211d', + '\\' : '\U00002124', + '\\' : '\U00002190', + '\\' : '\U000027f5', + '\\' : '\U00002192', + '\\' : '\U000027f6', + '\\' : '\U000021d0', + '\\' : '\U000027f8', + '\\' : '\U000021d2', + '\\' : '\U000027f9', + '\\' : '\U00002194', + '\\' : '\U000027f7', + '\\' : '\U000021d4', + '\\' : '\U000027fa', + '\\' : '\U000021a6', + '\\' : '\U000027fc', + '\\' : '\U00002500', + '\\' : '\U00002550', + '\\' : '\U000021a9', + '\\' : '\U000021aa', + '\\' : '\U000021bd', + '\\' : '\U000021c1', + '\\' : '\U000021bc', + '\\' : '\U000021c0', + '\\' : '\U000021cc', + '\\' : '\U0000219d', + '\\' : '\U000021c3', + '\\' : '\U000021c2', + '\\' : '\U000021bf', + '\\' : '\U000021be', + '\\' : '\U000021be', + '\\' : '\U00002237', + '\\' : '\U00002191', + '\\' : '\U000021d1', + '\\' : '\U00002193', + '\\' : '\U000021d3', + '\\' : '\U00002195', + '\\' : '\U000021d5', + '\\' : '\U000027e8', + '\\' : '\U000027e9', + '\\' : '\U00002308', + '\\' : '\U00002309', + '\\' : '\U0000230a', + '\\' : '\U0000230b', + '\\' : '\U00002987', + '\\' : '\U00002988', + '\\' : '\U000027e6', + '\\' : '\U000027e7', + '\\' : '\U00002983', + '\\' : '\U00002984', + '\\' : '\U000000ab', + '\\' : '\U000000bb', + '\\' : '\U000022a5', + '\\' : '\U000022a4', + '\\' : '\U00002227', + '\\' : '\U000022c0', + '\\' : '\U00002228', + '\\' : '\U000022c1', + '\\' : '\U00002200', + '\\' : '\U00002203', + '\\' : '\U00002204', + '\\' : '\U000000ac', + '\\' : '\U000025a1', + '\\' : '\U000025c7', + '\\' : '\U000022a2', + '\\' : '\U000022a8', + '\\' : '\U000022a9', + '\\' : '\U000022ab', + '\\' : '\U000022a3', + '\\' : '\U0000221a', + '\\' : '\U00002264', + '\\' : '\U00002265', + '\\' : '\U0000226a', + '\\' : '\U0000226b', + '\\' : '\U00002272', + '\\' : '\U00002273', + '\\' : '\U00002a85', + '\\' : '\U00002a86', + '\\' : '\U00002208', + '\\' : '\U00002209', + '\\' : '\U00002282', + '\\' : '\U00002283', + '\\' : '\U00002286', + '\\' : '\U00002287', + '\\' : '\U0000228f', + '\\' : '\U00002290', + '\\' : '\U00002291', + '\\' : '\U00002292', + '\\' : '\U00002229', + '\\' : '\U000022c2', + '\\' : '\U0000222a', + '\\' : '\U000022c3', + '\\' : '\U00002294', + '\\' : '\U00002a06', + '\\' : '\U00002293', + '\\' : '\U00002a05', + '\\' : '\U00002216', + '\\' : '\U0000221d', + '\\' : '\U0000228e', + '\\' : '\U00002a04', + '\\' : '\U00002260', + '\\' : '\U0000223c', + '\\' : '\U00002250', + '\\' : '\U00002243', + '\\' : '\U00002248', + '\\' : '\U0000224d', + '\\' : '\U00002245', + '\\' : '\U00002323', + '\\' : '\U00002261', + '\\' : '\U00002322', + '\\' : '\U000022c8', + '\\' : '\U00002a1d', + '\\' : '\U0000227a', + '\\' : '\U0000227b', + '\\' : '\U0000227c', + '\\' : '\U0000227d', + '\\' : '\U00002225', + '\\' : '\U000000a6', + '\\' : '\U000000b1', + '\\' : '\U00002213', + '\\' : '\U000000d7', + '\\
' : '\U000000f7', + '\\' : '\U000022c5', + '\\' : '\U000022c6', + '\\' : '\U00002219', + '\\' : '\U00002218', + '\\' : '\U00002020', + '\\' : '\U00002021', + '\\' : '\U000022b2', + '\\' : '\U000022b3', + '\\' : '\U000022b4', + '\\' : '\U000022b5', + '\\' : '\U000025c3', + '\\' : '\U000025b9', + '\\' : '\U000025b3', + '\\' : '\U0000225c', + '\\' : '\U00002295', + '\\' : '\U00002a01', + '\\' : '\U00002297', + '\\' : '\U00002a02', + '\\' : '\U00002299', + '\\' : '\U00002a00', + '\\' : '\U00002296', + '\\' : '\U00002298', + '\\' : '\U00002026', + '\\' : '\U000022ef', + '\\' : '\U00002211', + '\\' : '\U0000220f', + '\\' : '\U00002210', + '\\' : '\U0000221e', + '\\' : '\U0000222b', + '\\' : '\U0000222e', + '\\' : '\U00002663', + '\\' : '\U00002662', + '\\' : '\U00002661', + '\\' : '\U00002660', + '\\' : '\U00002135', + '\\' : '\U00002205', + '\\' : '\U00002207', + '\\' : '\U00002202', + '\\' : '\U0000266d', + '\\' : '\U0000266e', + '\\' : '\U0000266f', + '\\' : '\U00002220', + '\\' : '\U000000a9', + '\\' : '\U000000ae', + '\\' : '\U000000ad', + '\\' : '\U000000af', + '\\' : '\U000000bc', + '\\' : '\U000000bd', + '\\' : '\U000000be', + '\\' : '\U000000aa', + '\\' : '\U000000ba', + '\\
' : '\U000000a7', + '\\' : '\U000000b6', + '\\' : '\U000000a1', + '\\' : '\U000000bf', + '\\' : '\U000020ac', + '\\' : '\U000000a3', + '\\' : '\U000000a5', + '\\' : '\U000000a2', + '\\' : '\U000000a4', + '\\' : '\U000000b0', + '\\' : '\U00002a3f', + '\\' : '\U00002127', + '\\' : '\U000025ca', + '\\' : '\U00002118', + '\\' : '\U00002240', + '\\' : '\U000022c4', + '\\' : '\U000000b4', + '\\' : '\U00000131', + '\\' : '\U000000a8', + '\\' : '\U000000b8', + '\\' : '\U000002dd', + '\\' : '\U000003f5', + '\\' : '\U000023ce', + '\\' : '\U00002039', + '\\' : '\U0000203a', + '\\' : '\U00002302', + '\\<^sub>' : '\U000021e9', + '\\<^sup>' : '\U000021e7', + '\\<^bold>' : '\U00002759', + '\\<^bsub>' : '\U000021d8', + '\\<^esub>' : '\U000021d9', + '\\<^bsup>' : '\U000021d7', + '\\<^esup>' : '\U000021d6', + } + + lang_map = {'isabelle' : isabelle_symbols, 'latex' : latex_symbols} + + def __init__(self, **options): + Filter.__init__(self, **options) + lang = get_choice_opt(options, 'lang', + ['isabelle', 'latex'], 'isabelle') + self.symbols = self.lang_map[lang] + + def filter(self, lexer, stream): + for ttype, value in stream: + if value in self.symbols: + yield ttype, self.symbols[value] + else: + yield ttype, value + + +class KeywordCaseFilter(Filter): + """Convert keywords to lowercase or uppercase or capitalize them, which + means first letter uppercase, rest lowercase. + + This can be useful e.g. if you highlight Pascal code and want to adapt the + code to your styleguide. + + Options accepted: + + `case` : string + The casing to convert keywords to. Must be one of ``'lower'``, + ``'upper'`` or ``'capitalize'``. The default is ``'lower'``. + """ + + def __init__(self, **options): + Filter.__init__(self, **options) + case = get_choice_opt(options, 'case', + ['lower', 'upper', 'capitalize'], 'lower') + self.convert = getattr(str, case) + + def filter(self, lexer, stream): + for ttype, value in stream: + if ttype in Keyword: + yield ttype, self.convert(value) + else: + yield ttype, value + + +class NameHighlightFilter(Filter): + """Highlight a normal Name (and Name.*) token with a different token type. + + Example:: + + filter = NameHighlightFilter( + names=['foo', 'bar', 'baz'], + tokentype=Name.Function, + ) + + This would highlight the names "foo", "bar" and "baz" + as functions. `Name.Function` is the default token type. + + Options accepted: + + `names` : list of strings + A list of names that should be given the different token type. + There is no default. + `tokentype` : TokenType or string + A token type or a string containing a token type name that is + used for highlighting the strings in `names`. The default is + `Name.Function`. + """ + + def __init__(self, **options): + Filter.__init__(self, **options) + self.names = set(get_list_opt(options, 'names', [])) + tokentype = options.get('tokentype') + if tokentype: + self.tokentype = string_to_tokentype(tokentype) + else: + self.tokentype = Name.Function + + def filter(self, lexer, stream): + for ttype, value in stream: + if ttype in Name and value in self.names: + yield self.tokentype, value + else: + yield ttype, value + + +class ErrorToken(Exception): + pass + + +class RaiseOnErrorTokenFilter(Filter): + """Raise an exception when the lexer generates an error token. + + Options accepted: + + `excclass` : Exception class + The exception class to raise. + The default is `pygments.filters.ErrorToken`. + + .. versionadded:: 0.8 + """ + + def __init__(self, **options): + Filter.__init__(self, **options) + self.exception = options.get('excclass', ErrorToken) + try: + # issubclass() will raise TypeError if first argument is not a class + if not issubclass(self.exception, Exception): + raise TypeError + except TypeError: + raise OptionError('excclass option is not an exception class') + + def filter(self, lexer, stream): + for ttype, value in stream: + if ttype is Error: + raise self.exception(value) + yield ttype, value + + +class VisibleWhitespaceFilter(Filter): + """Convert tabs, newlines and/or spaces to visible characters. + + Options accepted: + + `spaces` : string or bool + If this is a one-character string, spaces will be replaces by this string. + If it is another true value, spaces will be replaced by ``·`` (unicode + MIDDLE DOT). If it is a false value, spaces will not be replaced. The + default is ``False``. + `tabs` : string or bool + The same as for `spaces`, but the default replacement character is ``»`` + (unicode RIGHT-POINTING DOUBLE ANGLE QUOTATION MARK). The default value + is ``False``. Note: this will not work if the `tabsize` option for the + lexer is nonzero, as tabs will already have been expanded then. + `tabsize` : int + If tabs are to be replaced by this filter (see the `tabs` option), this + is the total number of characters that a tab should be expanded to. + The default is ``8``. + `newlines` : string or bool + The same as for `spaces`, but the default replacement character is ``¶`` + (unicode PILCROW SIGN). The default value is ``False``. + `wstokentype` : bool + If true, give whitespace the special `Whitespace` token type. This allows + styling the visible whitespace differently (e.g. greyed out), but it can + disrupt background colors. The default is ``True``. + + .. versionadded:: 0.8 + """ + + def __init__(self, **options): + Filter.__init__(self, **options) + for name, default in [('spaces', '·'), + ('tabs', '»'), + ('newlines', '¶')]: + opt = options.get(name, False) + if isinstance(opt, str) and len(opt) == 1: + setattr(self, name, opt) + else: + setattr(self, name, (opt and default or '')) + tabsize = get_int_opt(options, 'tabsize', 8) + if self.tabs: + self.tabs += ' ' * (tabsize - 1) + if self.newlines: + self.newlines += '\n' + self.wstt = get_bool_opt(options, 'wstokentype', True) + + def filter(self, lexer, stream): + if self.wstt: + spaces = self.spaces or ' ' + tabs = self.tabs or '\t' + newlines = self.newlines or '\n' + regex = re.compile(r'\s') + + def replacefunc(wschar): + if wschar == ' ': + return spaces + elif wschar == '\t': + return tabs + elif wschar == '\n': + return newlines + return wschar + + for ttype, value in stream: + yield from _replace_special(ttype, value, regex, Whitespace, + replacefunc) + else: + spaces, tabs, newlines = self.spaces, self.tabs, self.newlines + # simpler processing + for ttype, value in stream: + if spaces: + value = value.replace(' ', spaces) + if tabs: + value = value.replace('\t', tabs) + if newlines: + value = value.replace('\n', newlines) + yield ttype, value + + +class GobbleFilter(Filter): + """Gobbles source code lines (eats initial characters). + + This filter drops the first ``n`` characters off every line of code. This + may be useful when the source code fed to the lexer is indented by a fixed + amount of space that isn't desired in the output. + + Options accepted: + + `n` : int + The number of characters to gobble. + + .. versionadded:: 1.2 + """ + def __init__(self, **options): + Filter.__init__(self, **options) + self.n = get_int_opt(options, 'n', 0) + + def gobble(self, value, left): + if left < len(value): + return value[left:], 0 + else: + return '', left - len(value) + + def filter(self, lexer, stream): + n = self.n + left = n # How many characters left to gobble. + for ttype, value in stream: + # Remove ``left`` tokens from first line, ``n`` from all others. + parts = value.split('\n') + (parts[0], left) = self.gobble(parts[0], left) + for i in range(1, len(parts)): + (parts[i], left) = self.gobble(parts[i], n) + value = '\n'.join(parts) + + if value != '': + yield ttype, value + + +class TokenMergeFilter(Filter): + """Merges consecutive tokens with the same token type in the output + stream of a lexer. + + .. versionadded:: 1.2 + """ + def __init__(self, **options): + Filter.__init__(self, **options) + + def filter(self, lexer, stream): + current_type = None + current_value = None + for ttype, value in stream: + if ttype is current_type: + current_value += value + else: + if current_type is not None: + yield current_type, current_value + current_type = ttype + current_value = value + if current_type is not None: + yield current_type, current_value + + +FILTERS = { + 'codetagify': CodeTagFilter, + 'keywordcase': KeywordCaseFilter, + 'highlight': NameHighlightFilter, + 'raiseonerror': RaiseOnErrorTokenFilter, + 'whitespace': VisibleWhitespaceFilter, + 'gobble': GobbleFilter, + 'tokenmerge': TokenMergeFilter, + 'symbols': SymbolFilter, +} diff --git a/venv/lib/python3.10/site-packages/pygments/filters/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/filters/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..04be0ab53d64f3d726f221d4406614ce97643bca Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/filters/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__init__.py b/venv/lib/python3.10/site-packages/pygments/formatters/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b24931cd191efe2a83010b46422face50f050f37 --- /dev/null +++ b/venv/lib/python3.10/site-packages/pygments/formatters/__init__.py @@ -0,0 +1,157 @@ +""" + pygments.formatters + ~~~~~~~~~~~~~~~~~~~ + + Pygments formatters. + + :copyright: Copyright 2006-2025 by the Pygments team, see AUTHORS. + :license: BSD, see LICENSE for details. +""" + +import re +import sys +import types +import fnmatch +from os.path import basename + +from pygments.formatters._mapping import FORMATTERS +from pygments.plugin import find_plugin_formatters +from pygments.util import ClassNotFound + +__all__ = ['get_formatter_by_name', 'get_formatter_for_filename', + 'get_all_formatters', 'load_formatter_from_file'] + list(FORMATTERS) + +_formatter_cache = {} # classes by name +_pattern_cache = {} + + +def _fn_matches(fn, glob): + """Return whether the supplied file name fn matches pattern filename.""" + if glob not in _pattern_cache: + pattern = _pattern_cache[glob] = re.compile(fnmatch.translate(glob)) + return pattern.match(fn) + return _pattern_cache[glob].match(fn) + + +def _load_formatters(module_name): + """Load a formatter (and all others in the module too).""" + mod = __import__(module_name, None, None, ['__all__']) + for formatter_name in mod.__all__: + cls = getattr(mod, formatter_name) + _formatter_cache[cls.name] = cls + + +def get_all_formatters(): + """Return a generator for all formatter classes.""" + # NB: this returns formatter classes, not info like get_all_lexers(). + for info in FORMATTERS.values(): + if info[1] not in _formatter_cache: + _load_formatters(info[0]) + yield _formatter_cache[info[1]] + for _, formatter in find_plugin_formatters(): + yield formatter + + +def find_formatter_class(alias): + """Lookup a formatter by alias. + + Returns None if not found. + """ + for module_name, name, aliases, _, _ in FORMATTERS.values(): + if alias in aliases: + if name not in _formatter_cache: + _load_formatters(module_name) + return _formatter_cache[name] + for _, cls in find_plugin_formatters(): + if alias in cls.aliases: + return cls + + +def get_formatter_by_name(_alias, **options): + """ + Return an instance of a :class:`.Formatter` subclass that has `alias` in its + aliases list. The formatter is given the `options` at its instantiation. + + Will raise :exc:`pygments.util.ClassNotFound` if no formatter with that + alias is found. + """ + cls = find_formatter_class(_alias) + if cls is None: + raise ClassNotFound(f"no formatter found for name {_alias!r}") + return cls(**options) + + +def load_formatter_from_file(filename, formattername="CustomFormatter", **options): + """ + Return a `Formatter` subclass instance loaded from the provided file, relative + to the current directory. + + The file is expected to contain a Formatter class named ``formattername`` + (by default, CustomFormatter). Users should be very careful with the input, because + this method is equivalent to running ``eval()`` on the input file. The formatter is + given the `options` at its instantiation. + + :exc:`pygments.util.ClassNotFound` is raised if there are any errors loading + the formatter. + + .. versionadded:: 2.2 + """ + try: + # This empty dict will contain the namespace for the exec'd file + custom_namespace = {} + with open(filename, 'rb') as f: + exec(f.read(), custom_namespace) + # Retrieve the class `formattername` from that namespace + if formattername not in custom_namespace: + raise ClassNotFound(f'no valid {formattername} class found in {filename}') + formatter_class = custom_namespace[formattername] + # And finally instantiate it with the options + return formatter_class(**options) + except OSError as err: + raise ClassNotFound(f'cannot read {filename}: {err}') + except ClassNotFound: + raise + except Exception as err: + raise ClassNotFound(f'error when loading custom formatter: {err}') + + +def get_formatter_for_filename(fn, **options): + """ + Return a :class:`.Formatter` subclass instance that has a filename pattern + matching `fn`. The formatter is given the `options` at its instantiation. + + Will raise :exc:`pygments.util.ClassNotFound` if no formatter for that filename + is found. + """ + fn = basename(fn) + for modname, name, _, filenames, _ in FORMATTERS.values(): + for filename in filenames: + if _fn_matches(fn, filename): + if name not in _formatter_cache: + _load_formatters(modname) + return _formatter_cache[name](**options) + for _name, cls in find_plugin_formatters(): + for filename in cls.filenames: + if _fn_matches(fn, filename): + return cls(**options) + raise ClassNotFound(f"no formatter found for file name {fn!r}") + + +class _automodule(types.ModuleType): + """Automatically import formatters.""" + + def __getattr__(self, name): + info = FORMATTERS.get(name) + if info: + _load_formatters(info[0]) + cls = _formatter_cache[info[1]] + setattr(self, name, cls) + return cls + raise AttributeError(name) + + +oldmod = sys.modules[__name__] +newmod = _automodule(__name__) +newmod.__dict__.update(oldmod.__dict__) +sys.modules[__name__] = newmod +del newmod.newmod, newmod.oldmod, newmod.sys, newmod.types diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/__init__.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/__init__.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c8f370da95efe50efaec8e92193097503730c2dc Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/__init__.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/_mapping.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/_mapping.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..02ee2cb1bf77b75a63c790184dc67b94d4b521f6 Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/_mapping.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/bbcode.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/bbcode.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9e5a0b5f71fd87a89d059e30b61280dc16251f01 Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/bbcode.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/groff.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/groff.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8147011192035e09d1d1c1b3ed694adf196d3b19 Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/groff.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/html.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/html.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..267a9c848cef35c8a0d64913cc6e37fdf4bd568b Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/html.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/img.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/img.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ef3a33a5ab128027c7e091c3fe8d7945cc6b2857 Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/img.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/irc.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/irc.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..450732a5a2eceafc18efac6246a74c1b1a5955c5 Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/irc.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/latex.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/latex.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..dd33e1fb950b2b5e66895afdf3b0948d672970d0 Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/latex.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/other.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/other.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5cb2a032cb56523f20b58fd9e2488ed166afb2b4 Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/other.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/pangomarkup.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/pangomarkup.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1d670a067dd2559dd78c7b5e62c68e380eaa87e9 Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/pangomarkup.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/rtf.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/rtf.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..79a55dec052b01d86445b1fececd38fc974d7351 Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/rtf.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/svg.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/svg.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c324cf10cd627793b27b454f757f944e71b233af Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/svg.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/terminal.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/terminal.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5625316193e1fb69b18fe4b546a3e17453eaba22 Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/terminal.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/terminal256.cpython-310.pyc b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/terminal256.cpython-310.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9c7f908745cf701a4eaf3b45c11cf054c81d745a Binary files /dev/null and b/venv/lib/python3.10/site-packages/pygments/formatters/__pycache__/terminal256.cpython-310.pyc differ diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/_mapping.py b/venv/lib/python3.10/site-packages/pygments/formatters/_mapping.py new file mode 100644 index 0000000000000000000000000000000000000000..72ca84040b626183e3328679db600c13472021be --- /dev/null +++ b/venv/lib/python3.10/site-packages/pygments/formatters/_mapping.py @@ -0,0 +1,23 @@ +# Automatically generated by scripts/gen_mapfiles.py. +# DO NOT EDIT BY HAND; run `tox -e mapfiles` instead. + +FORMATTERS = { + 'BBCodeFormatter': ('pygments.formatters.bbcode', 'BBCode', ('bbcode', 'bb'), (), 'Format tokens with BBcodes. These formatting codes are used by many bulletin boards, so you can highlight your sourcecode with pygments before posting it there.'), + 'BmpImageFormatter': ('pygments.formatters.img', 'img_bmp', ('bmp', 'bitmap'), ('*.bmp',), 'Create a bitmap image from source code. This uses the Python Imaging Library to generate a pixmap from the source code.'), + 'GifImageFormatter': ('pygments.formatters.img', 'img_gif', ('gif',), ('*.gif',), 'Create a GIF image from source code. This uses the Python Imaging Library to generate a pixmap from the source code.'), + 'GroffFormatter': ('pygments.formatters.groff', 'groff', ('groff', 'troff', 'roff'), (), 'Format tokens with groff escapes to change their color and font style.'), + 'HtmlFormatter': ('pygments.formatters.html', 'HTML', ('html',), ('*.html', '*.htm'), "Format tokens as HTML 4 ```` tags. By default, the content is enclosed in a ``
`` tag, itself wrapped in a ``
`` tag (but see the `nowrap` option). The ``
``'s CSS class can be set by the `cssclass` option."),
+    'IRCFormatter': ('pygments.formatters.irc', 'IRC', ('irc', 'IRC'), (), 'Format tokens with IRC color sequences'),
+    'ImageFormatter': ('pygments.formatters.img', 'img', ('img', 'IMG', 'png'), ('*.png',), 'Create a PNG image from source code. This uses the Python Imaging Library to generate a pixmap from the source code.'),
+    'JpgImageFormatter': ('pygments.formatters.img', 'img_jpg', ('jpg', 'jpeg'), ('*.jpg',), 'Create a JPEG image from source code. This uses the Python Imaging Library to generate a pixmap from the source code.'),
+    'LatexFormatter': ('pygments.formatters.latex', 'LaTeX', ('latex', 'tex'), ('*.tex',), 'Format tokens as LaTeX code. This needs the `fancyvrb` and `color` standard packages.'),
+    'NullFormatter': ('pygments.formatters.other', 'Text only', ('text', 'null'), ('*.txt',), 'Output the text unchanged without any formatting.'),
+    'PangoMarkupFormatter': ('pygments.formatters.pangomarkup', 'Pango Markup', ('pango', 'pangomarkup'), (), 'Format tokens as Pango Markup code. It can then be rendered to an SVG.'),
+    'RawTokenFormatter': ('pygments.formatters.other', 'Raw tokens', ('raw', 'tokens'), ('*.raw',), 'Format tokens as a raw representation for storing token streams.'),
+    'RtfFormatter': ('pygments.formatters.rtf', 'RTF', ('rtf',), ('*.rtf',), 'Format tokens as RTF markup. This formatter automatically outputs full RTF documents with color information and other useful stuff. Perfect for Copy and Paste into Microsoft(R) Word(R) documents.'),
+    'SvgFormatter': ('pygments.formatters.svg', 'SVG', ('svg',), ('*.svg',), 'Format tokens as an SVG graphics file.  This formatter is still experimental. Each line of code is a ```` element with explicit ``x`` and ``y`` coordinates containing ```` elements with the individual token styles.'),
+    'Terminal256Formatter': ('pygments.formatters.terminal256', 'Terminal256', ('terminal256', 'console256', '256'), (), 'Format tokens with ANSI color sequences, for output in a 256-color terminal or console.  Like in `TerminalFormatter` color sequences are terminated at newlines, so that paging the output works correctly.'),
+    'TerminalFormatter': ('pygments.formatters.terminal', 'Terminal', ('terminal', 'console'), (), 'Format tokens with ANSI color sequences, for output in a text console. Color sequences are terminated at newlines, so that paging the output works correctly.'),
+    'TerminalTrueColorFormatter': ('pygments.formatters.terminal256', 'TerminalTrueColor', ('terminal16m', 'console16m', '16m'), (), 'Format tokens with ANSI color sequences, for output in a true-color terminal or console.  Like in `TerminalFormatter` color sequences are terminated at newlines, so that paging the output works correctly.'),
+    'TestcaseFormatter': ('pygments.formatters.other', 'Testcase', ('testcase',), (), 'Format tokens as appropriate for a new testcase.'),
+}
diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/bbcode.py b/venv/lib/python3.10/site-packages/pygments/formatters/bbcode.py
new file mode 100644
index 0000000000000000000000000000000000000000..339edf9dc9a87829457de5cf27f2e7081569cc75
--- /dev/null
+++ b/venv/lib/python3.10/site-packages/pygments/formatters/bbcode.py
@@ -0,0 +1,108 @@
+"""
+    pygments.formatters.bbcode
+    ~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    BBcode formatter.
+
+    :copyright: Copyright 2006-2025 by the Pygments team, see AUTHORS.
+    :license: BSD, see LICENSE for details.
+"""
+
+
+from pygments.formatter import Formatter
+from pygments.util import get_bool_opt
+
+__all__ = ['BBCodeFormatter']
+
+
+class BBCodeFormatter(Formatter):
+    """
+    Format tokens with BBcodes. These formatting codes are used by many
+    bulletin boards, so you can highlight your sourcecode with pygments before
+    posting it there.
+
+    This formatter has no support for background colors and borders, as there
+    are no common BBcode tags for that.
+
+    Some board systems (e.g. phpBB) don't support colors in their [code] tag,
+    so you can't use the highlighting together with that tag.
+    Text in a [code] tag usually is shown with a monospace font (which this
+    formatter can do with the ``monofont`` option) and no spaces (which you
+    need for indentation) are removed.
+
+    Additional options accepted:
+
+    `style`
+        The style to use, can be a string or a Style subclass (default:
+        ``'default'``).
+
+    `codetag`
+        If set to true, put the output into ``[code]`` tags (default:
+        ``false``)
+
+    `monofont`
+        If set to true, add a tag to show the code with a monospace font
+        (default: ``false``).
+    """
+    name = 'BBCode'
+    aliases = ['bbcode', 'bb']
+    filenames = []
+
+    def __init__(self, **options):
+        Formatter.__init__(self, **options)
+        self._code = get_bool_opt(options, 'codetag', False)
+        self._mono = get_bool_opt(options, 'monofont', False)
+
+        self.styles = {}
+        self._make_styles()
+
+    def _make_styles(self):
+        for ttype, ndef in self.style:
+            start = end = ''
+            if ndef['color']:
+                start += '[color=#{}]'.format(ndef['color'])
+                end = '[/color]' + end
+            if ndef['bold']:
+                start += '[b]'
+                end = '[/b]' + end
+            if ndef['italic']:
+                start += '[i]'
+                end = '[/i]' + end
+            if ndef['underline']:
+                start += '[u]'
+                end = '[/u]' + end
+            # there are no common BBcodes for background-color and border
+
+            self.styles[ttype] = start, end
+
+    def format_unencoded(self, tokensource, outfile):
+        if self._code:
+            outfile.write('[code]')
+        if self._mono:
+            outfile.write('[font=monospace]')
+
+        lastval = ''
+        lasttype = None
+
+        for ttype, value in tokensource:
+            while ttype not in self.styles:
+                ttype = ttype.parent
+            if ttype == lasttype:
+                lastval += value
+            else:
+                if lastval:
+                    start, end = self.styles[lasttype]
+                    outfile.write(''.join((start, lastval, end)))
+                lastval = value
+                lasttype = ttype
+
+        if lastval:
+            start, end = self.styles[lasttype]
+            outfile.write(''.join((start, lastval, end)))
+
+        if self._mono:
+            outfile.write('[/font]')
+        if self._code:
+            outfile.write('[/code]')
+        if self._code or self._mono:
+            outfile.write('\n')
diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/groff.py b/venv/lib/python3.10/site-packages/pygments/formatters/groff.py
new file mode 100644
index 0000000000000000000000000000000000000000..028fec4ea49c5ba2eb78fb032fb98c886f4de90c
--- /dev/null
+++ b/venv/lib/python3.10/site-packages/pygments/formatters/groff.py
@@ -0,0 +1,170 @@
+"""
+    pygments.formatters.groff
+    ~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    Formatter for groff output.
+
+    :copyright: Copyright 2006-2025 by the Pygments team, see AUTHORS.
+    :license: BSD, see LICENSE for details.
+"""
+
+import math
+from pygments.formatter import Formatter
+from pygments.util import get_bool_opt, get_int_opt
+
+__all__ = ['GroffFormatter']
+
+
+class GroffFormatter(Formatter):
+    """
+    Format tokens with groff escapes to change their color and font style.
+
+    .. versionadded:: 2.11
+
+    Additional options accepted:
+
+    `style`
+        The style to use, can be a string or a Style subclass (default:
+        ``'default'``).
+
+    `monospaced`
+        If set to true, monospace font will be used (default: ``true``).
+
+    `linenos`
+        If set to true, print the line numbers (default: ``false``).
+
+    `wrap`
+        Wrap lines to the specified number of characters. Disabled if set to 0
+        (default: ``0``).
+    """
+
+    name = 'groff'
+    aliases = ['groff','troff','roff']
+    filenames = []
+
+    def __init__(self, **options):
+        Formatter.__init__(self, **options)
+
+        self.monospaced = get_bool_opt(options, 'monospaced', True)
+        self.linenos = get_bool_opt(options, 'linenos', False)
+        self._lineno = 0
+        self.wrap = get_int_opt(options, 'wrap', 0)
+        self._linelen = 0
+
+        self.styles = {}
+        self._make_styles()
+
+
+    def _make_styles(self):
+        regular = '\\f[CR]' if self.monospaced else '\\f[R]'
+        bold = '\\f[CB]' if self.monospaced else '\\f[B]'
+        italic = '\\f[CI]' if self.monospaced else '\\f[I]'
+
+        for ttype, ndef in self.style:
+            start = end = ''
+            if ndef['color']:
+                start += '\\m[{}]'.format(ndef['color'])
+                end = '\\m[]' + end
+            if ndef['bold']:
+                start += bold
+                end = regular + end
+            if ndef['italic']:
+                start += italic
+                end = regular + end
+            if ndef['bgcolor']:
+                start += '\\M[{}]'.format(ndef['bgcolor'])
+                end = '\\M[]' + end
+
+            self.styles[ttype] = start, end
+
+
+    def _define_colors(self, outfile):
+        colors = set()
+        for _, ndef in self.style:
+            if ndef['color'] is not None:
+                colors.add(ndef['color'])
+
+        for color in sorted(colors):
+            outfile.write('.defcolor ' + color + ' rgb #' + color + '\n')
+
+
+    def _write_lineno(self, outfile):
+        self._lineno += 1
+        outfile.write("%s% 4d " % (self._lineno != 1 and '\n' or '', self._lineno))
+
+
+    def _wrap_line(self, line):
+        length = len(line.rstrip('\n'))
+        space = '     ' if self.linenos else ''
+        newline = ''
+
+        if length > self.wrap:
+            for i in range(0, math.floor(length / self.wrap)):
+                chunk = line[i*self.wrap:i*self.wrap+self.wrap]
+                newline += (chunk + '\n' + space)
+            remainder = length % self.wrap
+            if remainder > 0:
+                newline += line[-remainder-1:]
+                self._linelen = remainder
+        elif self._linelen + length > self.wrap:
+            newline = ('\n' + space) + line
+            self._linelen = length
+        else:
+            newline = line
+            self._linelen += length
+
+        return newline
+
+
+    def _escape_chars(self, text):
+        text = text.replace('\\', '\\[u005C]'). \
+                    replace('.', '\\[char46]'). \
+                    replace('\'', '\\[u0027]'). \
+                    replace('`', '\\[u0060]'). \
+                    replace('~', '\\[u007E]')
+        copy = text
+
+        for char in copy:
+            if len(char) != len(char.encode()):
+                uni = char.encode('unicode_escape') \
+                    .decode()[1:] \
+                    .replace('x', 'u00') \
+                    .upper()
+                text = text.replace(char, '\\[u' + uni[1:] + ']')
+
+        return text
+
+
+    def format_unencoded(self, tokensource, outfile):
+        self._define_colors(outfile)
+
+        outfile.write('.nf\n\\f[CR]\n')
+
+        if self.linenos:
+            self._write_lineno(outfile)
+
+        for ttype, value in tokensource:
+            while ttype not in self.styles:
+                ttype = ttype.parent
+            start, end = self.styles[ttype]
+
+            for line in value.splitlines(True):
+                if self.wrap > 0:
+                    line = self._wrap_line(line)
+
+                if start and end:
+                    text = self._escape_chars(line.rstrip('\n'))
+                    if text != '':
+                        outfile.write(''.join((start, text, end)))
+                else:
+                    outfile.write(self._escape_chars(line.rstrip('\n')))
+
+                if line.endswith('\n'):
+                    if self.linenos:
+                        self._write_lineno(outfile)
+                        self._linelen = 0
+                    else:
+                        outfile.write('\n')
+                        self._linelen = 0
+
+        outfile.write('\n.fi')
diff --git a/venv/lib/python3.10/site-packages/pygments/formatters/html.py b/venv/lib/python3.10/site-packages/pygments/formatters/html.py
new file mode 100644
index 0000000000000000000000000000000000000000..4ef183681fd1e325831caecffc531dcb305516f3
--- /dev/null
+++ b/venv/lib/python3.10/site-packages/pygments/formatters/html.py
@@ -0,0 +1,995 @@
+"""
+    pygments.formatters.html
+    ~~~~~~~~~~~~~~~~~~~~~~~~
+
+    Formatter for HTML output.
+
+    :copyright: Copyright 2006-2025 by the Pygments team, see AUTHORS.
+    :license: BSD, see LICENSE for details.
+"""
+
+import functools
+import os
+import sys
+import os.path
+from io import StringIO
+
+from pygments.formatter import Formatter
+from pygments.token import Token, Text, STANDARD_TYPES
+from pygments.util import get_bool_opt, get_int_opt, get_list_opt
+
+try:
+    import ctags
+except ImportError:
+    ctags = None
+
+__all__ = ['HtmlFormatter']
+
+
+_escape_html_table = {
+    ord('&'): '&',
+    ord('<'): '<',
+    ord('>'): '>',
+    ord('"'): '"',
+    ord("'"): ''',
+}
+
+
+def escape_html(text, table=_escape_html_table):
+    """Escape &, <, > as well as single and double quotes for HTML."""
+    return text.translate(table)
+
+
+def webify(color):
+    if color.startswith('calc') or color.startswith('var'):
+        return color
+    else:
+        # Check if the color can be shortened from 6 to 3 characters
+        color = color.upper()
+        if (len(color) == 6 and
+            (    color[0] == color[1]
+             and color[2] == color[3]
+             and color[4] == color[5])):
+            return f'#{color[0]}{color[2]}{color[4]}'
+        else:
+            return f'#{color}'
+
+
+def _get_ttype_class(ttype):
+    fname = STANDARD_TYPES.get(ttype)
+    if fname:
+        return fname
+    aname = ''
+    while fname is None:
+        aname = '-' + ttype[-1] + aname
+        ttype = ttype.parent
+        fname = STANDARD_TYPES.get(ttype)
+    return fname + aname
+
+
+CSSFILE_TEMPLATE = '''\
+/*
+generated by Pygments 
+Copyright 2006-2025 by the Pygments team.
+Licensed under the BSD license, see LICENSE for details.
+*/
+%(styledefs)s
+'''
+
+DOC_HEADER = '''\
+
+
+
+
+  %(title)s
+  
+  
+
+
+
%(title)s

+
+'''
+
+DOC_HEADER_EXTERNALCSS = '''\
+
+
+
+
+  %(title)s
+  
+  
+
+
+
%(title)s

+
+'''
+
+DOC_FOOTER = '''\
+
+
+'''
+
+
+class HtmlFormatter(Formatter):
+    r"""
+    Format tokens as HTML 4 ```` tags. By default, the content is enclosed
+    in a ``
`` tag, itself wrapped in a ``
`` tag (but see the `nowrap` option).
+    The ``
``'s CSS class can be set by the `cssclass` option.
+
+    If the `linenos` option is set to ``"table"``, the ``
`` is
+    additionally wrapped inside a ```` which has one row and two
+    cells: one containing the line numbers and one containing the code.
+    Example:
+
+    .. sourcecode:: html
+
+        

+        

+          
+          
+        

+            
1
+            2

+          		
+            
def foo(bar):
+              pass
+            

+          

+
+    (whitespace added to improve clarity).
+
+    A list of lines can be specified using the `hl_lines` option to make these
+    lines highlighted (as of Pygments 0.11).
+
+    With the `full` option, a complete HTML 4 document is output, including
+    the style definitions inside a ``

' : '\U0001d4ab', + '\\' : '\U0001d4ac', + '\\' : '\U0000211b', + '\\' : '\U0001d4ae', + '\\' : '\U0001d4af', + '\\' : '\U0001d4b0', + '\\' : '\U0001d4b1', + '\\' : '\U0001d4b2', + '\\' : '\U0001d4b3', + '\\' : '\U0001d4b4', + '\\' : '\U0001d4b5', + '\\' : '\U0001d5ba', + '\\' : '\U0001d5bb', + '\\' : '\U0001d5bc', + '\\' : '\U0001d5bd', + '\\' : '\U0001d5be', + '\\' : '\U0001d5bf', + '\\' : '\U0001d5c0', + '\\' : '\U0001d5c1', + '\\' : '\U0001d5c2', + '\\' : '\U0001d5c3', + '\\' : '\U0001d5c4', + '\\' : '\U0001d5c5', + '\\' : '\U0001d5c6', + '\\' : '\U0001d5c7', + '\\' : '\U0001d5c8', + '\\