mirror of
https://github.com/huggingface/diffusers.git
synced 2026-01-29 07:22:12 +03:00
Merge branch 'main' into animatediff/freenoise-improvements
This commit is contained in:
Aryan
177
.github/workflows/nightly_tests.yml
vendored
177
.github/workflows/nightly_tests.yml
vendored
@@ -116,6 +116,7 @@ jobs:
|
||||
run:
|
||||
shell: bash
|
||||
strategy:
|
||||
fail-fast: false
|
||||
max-parallel: 2
|
||||
matrix:
|
||||
module: [models, schedulers, lora, others, single_file, examples]
|
||||
@@ -290,64 +291,118 @@ jobs:
|
||||
pip install slack_sdk tabulate
|
||||
python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
|
||||
|
||||
run_nightly_tests_apple_m1:
|
||||
name: Nightly PyTorch MPS tests on MacOS
|
||||
runs-on: [ self-hosted, apple-m1 ]
|
||||
if: github.event_name == 'schedule'
|
||||
|
||||
steps:
|
||||
- name: Checkout diffusers
|
||||
uses: actions/checkout@v3
|
||||
with:
|
||||
fetch-depth: 2
|
||||
|
||||
- name: Clean checkout
|
||||
shell: arch -arch arm64 bash {0}
|
||||
run: |
|
||||
git clean -fxd
|
||||
|
||||
- name: Setup miniconda
|
||||
uses: ./.github/actions/setup-miniconda
|
||||
with:
|
||||
python-version: 3.9
|
||||
|
||||
- name: Install dependencies
|
||||
shell: arch -arch arm64 bash {0}
|
||||
run: |
|
||||
${CONDA_RUN} python -m pip install --upgrade pip uv
|
||||
${CONDA_RUN} python -m uv pip install -e [quality,test]
|
||||
${CONDA_RUN} python -m uv pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu
|
||||
${CONDA_RUN} python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate
|
||||
${CONDA_RUN} python -m uv pip install pytest-reportlog
|
||||
|
||||
- name: Environment
|
||||
shell: arch -arch arm64 bash {0}
|
||||
run: |
|
||||
${CONDA_RUN} python utils/print_env.py
|
||||
|
||||
- name: Run nightly PyTorch tests on M1 (MPS)
|
||||
shell: arch -arch arm64 bash {0}
|
||||
env:
|
||||
HF_HOME: /System/Volumes/Data/mnt/cache
|
||||
HF_TOKEN: ${{ secrets.HF_TOKEN }}
|
||||
run: |
|
||||
${CONDA_RUN} python -m pytest -n 1 -s -v --make-reports=tests_torch_mps \
|
||||
--report-log=tests_torch_mps.log \
|
||||
tests/
|
||||
|
||||
- name: Failure short reports
|
||||
if: ${{ failure() }}
|
||||
run: cat reports/tests_torch_mps_failures_short.txt
|
||||
|
||||
- name: Test suite reports artifacts
|
||||
if: ${{ always() }}
|
||||
uses: actions/upload-artifact@v2
|
||||
with:
|
||||
name: torch_mps_test_reports
|
||||
path: reports
|
||||
|
||||
- name: Generate Report and Notify Channel
|
||||
if: always()
|
||||
run: |
|
||||
pip install slack_sdk tabulate
|
||||
python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
|
||||
# M1 runner currently not well supported
|
||||
# TODO: (Dhruv) add these back when we setup better testing for Apple Silicon
|
||||
# run_nightly_tests_apple_m1:
|
||||
# name: Nightly PyTorch MPS tests on MacOS
|
||||
# runs-on: [ self-hosted, apple-m1 ]
|
||||
# if: github.event_name == 'schedule'
|
||||
#
|
||||
# steps:
|
||||
# - name: Checkout diffusers
|
||||
# uses: actions/checkout@v3
|
||||
# with:
|
||||
# fetch-depth: 2
|
||||
#
|
||||
# - name: Clean checkout
|
||||
# shell: arch -arch arm64 bash {0}
|
||||
# run: |
|
||||
# git clean -fxd
|
||||
# - name: Setup miniconda
|
||||
# uses: ./.github/actions/setup-miniconda
|
||||
# with:
|
||||
# python-version: 3.9
|
||||
#
|
||||
# - name: Install dependencies
|
||||
# shell: arch -arch arm64 bash {0}
|
||||
# run: |
|
||||
# ${CONDA_RUN} python -m pip install --upgrade pip uv
|
||||
# ${CONDA_RUN} python -m uv pip install -e [quality,test]
|
||||
# ${CONDA_RUN} python -m uv pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu
|
||||
# ${CONDA_RUN} python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate
|
||||
# ${CONDA_RUN} python -m uv pip install pytest-reportlog
|
||||
# - name: Environment
|
||||
# shell: arch -arch arm64 bash {0}
|
||||
# run: |
|
||||
# ${CONDA_RUN} python utils/print_env.py
|
||||
# - name: Run nightly PyTorch tests on M1 (MPS)
|
||||
# shell: arch -arch arm64 bash {0}
|
||||
# env:
|
||||
# HF_HOME: /System/Volumes/Data/mnt/cache
|
||||
# HF_TOKEN: ${{ secrets.HF_TOKEN }}
|
||||
# run: |
|
||||
# ${CONDA_RUN} python -m pytest -n 1 -s -v --make-reports=tests_torch_mps \
|
||||
# --report-log=tests_torch_mps.log \
|
||||
# tests/
|
||||
# - name: Failure short reports
|
||||
# if: ${{ failure() }}
|
||||
# run: cat reports/tests_torch_mps_failures_short.txt
|
||||
#
|
||||
# - name: Test suite reports artifacts
|
||||
# if: ${{ always() }}
|
||||
# uses: actions/upload-artifact@v2
|
||||
# with:
|
||||
# name: torch_mps_test_reports
|
||||
# path: reports
|
||||
#
|
||||
# - name: Generate Report and Notify Channel
|
||||
# if: always()
|
||||
# run: |
|
||||
# pip install slack_sdk tabulate
|
||||
# python utils/log_reports.py >> $GITHUB_STEP_SUMMARY run_nightly_tests_apple_m1:
|
||||
# name: Nightly PyTorch MPS tests on MacOS
|
||||
# runs-on: [ self-hosted, apple-m1 ]
|
||||
# if: github.event_name == 'schedule'
|
||||
#
|
||||
# steps:
|
||||
# - name: Checkout diffusers
|
||||
# uses: actions/checkout@v3
|
||||
# with:
|
||||
# fetch-depth: 2
|
||||
#
|
||||
# - name: Clean checkout
|
||||
# shell: arch -arch arm64 bash {0}
|
||||
# run: |
|
||||
# git clean -fxd
|
||||
# - name: Setup miniconda
|
||||
# uses: ./.github/actions/setup-miniconda
|
||||
# with:
|
||||
# python-version: 3.9
|
||||
#
|
||||
# - name: Install dependencies
|
||||
# shell: arch -arch arm64 bash {0}
|
||||
# run: |
|
||||
# ${CONDA_RUN} python -m pip install --upgrade pip uv
|
||||
# ${CONDA_RUN} python -m uv pip install -e [quality,test]
|
||||
# ${CONDA_RUN} python -m uv pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu
|
||||
# ${CONDA_RUN} python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate
|
||||
# ${CONDA_RUN} python -m uv pip install pytest-reportlog
|
||||
# - name: Environment
|
||||
# shell: arch -arch arm64 bash {0}
|
||||
# run: |
|
||||
# ${CONDA_RUN} python utils/print_env.py
|
||||
# - name: Run nightly PyTorch tests on M1 (MPS)
|
||||
# shell: arch -arch arm64 bash {0}
|
||||
# env:
|
||||
# HF_HOME: /System/Volumes/Data/mnt/cache
|
||||
# HF_TOKEN: ${{ secrets.HF_TOKEN }}
|
||||
# run: |
|
||||
# ${CONDA_RUN} python -m pytest -n 1 -s -v --make-reports=tests_torch_mps \
|
||||
# --report-log=tests_torch_mps.log \
|
||||
# tests/
|
||||
# - name: Failure short reports
|
||||
# if: ${{ failure() }}
|
||||
# run: cat reports/tests_torch_mps_failures_short.txt
|
||||
#
|
||||
# - name: Test suite reports artifacts
|
||||
# if: ${{ always() }}
|
||||
# uses: actions/upload-artifact@v2
|
||||
# with:
|
||||
# name: torch_mps_test_reports
|
||||
# path: reports
|
||||
#
|
||||
# - name: Generate Report and Notify Channel
|
||||
# if: always()
|
||||
# run: |
|
||||
# pip install slack_sdk tabulate
|
||||
# python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
|
||||
6
.github/workflows/push_tests.yml
vendored
6
.github/workflows/push_tests.yml
vendored
@@ -1,9 +1,11 @@
|
||||
name: Slow Tests on main
|
||||
name: Fast GPU Tests on main
|
||||
|
||||
on:
|
||||
push:
|
||||
branches:
|
||||
- main
|
||||
- "v*.*.*-release"
|
||||
- "v*.*.*-patch"
|
||||
paths:
|
||||
- "src/diffusers/**.py"
|
||||
- "examples/**.py"
|
||||
@@ -112,6 +114,8 @@ jobs:
|
||||
run:
|
||||
shell: bash
|
||||
strategy:
|
||||
fail-fast: false
|
||||
max-parallel: 2
|
||||
matrix:
|
||||
module: [models, schedulers, lora, others, single_file]
|
||||
steps:
|
||||
|
||||
@@ -57,7 +57,7 @@ Any question or comment related to the Diffusers library can be asked on the [di
|
||||
- ...
|
||||
|
||||
Every question that is asked on the forum or on Discord actively encourages the community to publicly
|
||||
share knowledge and might very well help a beginner in the future that has the same question you're
|
||||
share knowledge and might very well help a beginner in the future who has the same question you're
|
||||
having. Please do pose any questions you might have.
|
||||
In the same spirit, you are of immense help to the community by answering such questions because this way you are publicly documenting knowledge for everybody to learn from.
|
||||
|
||||
@@ -503,4 +503,4 @@ $ git push --set-upstream origin your-branch-for-syncing
|
||||
|
||||
### Style guide
|
||||
|
||||
For documentation strings, 🧨 Diffusers follows the [Google style](https://google.github.io/styleguide/pyguide.html).
|
||||
For documentation strings, 🧨 Diffusers follows the [Google style](https://google.github.io/styleguide/pyguide.html).
|
||||
|
||||
@@ -15,7 +15,7 @@ specific language governing permissions and limitations under the License.
|
||||
🧨 Diffusers provides **state-of-the-art** pretrained diffusion models across multiple modalities.
|
||||
Its purpose is to serve as a **modular toolbox** for both inference and training.
|
||||
|
||||
We aim at building a library that stands the test of time and therefore take API design very seriously.
|
||||
We aim to build a library that stands the test of time and therefore take API design very seriously.
|
||||
|
||||
In a nutshell, Diffusers is built to be a natural extension of PyTorch. Therefore, most of our design choices are based on [PyTorch's Design Principles](https://pytorch.org/docs/stable/community/design.html#pytorch-design-philosophy). Let's go over the most important ones:
|
||||
|
||||
@@ -107,4 +107,4 @@ The following design principles are followed:
|
||||
- Every scheduler exposes the timesteps to be "looped over" via a `timesteps` attribute, which is an array of timesteps the model will be called upon.
|
||||
- The `step(...)` function takes a predicted model output and the "current" sample (x_t) and returns the "previous", slightly more denoised sample (x_t-1).
|
||||
- Given the complexity of diffusion schedulers, the `step` function does not expose all the complexity and can be a bit of a "black box".
|
||||
- In almost all cases, novel schedulers shall be implemented in a new scheduling file.
|
||||
- In almost all cases, novel schedulers shall be implemented in a new scheduling file.
|
||||
|
||||
@@ -29,6 +29,10 @@ Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.m
|
||||
|
||||
This pipeline was contributed by [zRzRzRzRzRzRzR](https://github.com/zRzRzRzRzRzRzR). The original codebase can be found [here](https://huggingface.co/THUDM). The original weights can be found under [hf.co/THUDM](https://huggingface.co/THUDM).
|
||||
|
||||
There are two models available that can be used with the CogVideoX pipeline:
|
||||
- [`THUDM/CogVideoX-2b`](https://huggingface.co/THUDM/CogVideoX-2b)
|
||||
- [`THUDM/CogVideoX-5b`](https://huggingface.co/THUDM/CogVideoX-5b)
|
||||
|
||||
## Inference
|
||||
|
||||
Use [`torch.compile`](https://huggingface.co/docs/diffusers/main/en/tutorials/fast_diffusion#torchcompile) to reduce the inference latency.
|
||||
@@ -68,7 +72,7 @@ With torch.compile(): Average inference time: 76.27 seconds.
|
||||
|
||||
### Memory optimization
|
||||
|
||||
CogVideoX requires about 19 GB of GPU memory to decode 49 frames (6 seconds of video at 8 FPS) with output resolution 720x480 (W x H), which makes it not possible to run on consumer GPUs or free-tier T4 Colab. The following memory optimizations could be used to reduce the memory footprint. For replication, you can refer to [this](https://gist.github.com/a-r-r-o-w/3959a03f15be5c9bd1fe545b09dfcc93) script.
|
||||
CogVideoX-2b requires about 19 GB of GPU memory to decode 49 frames (6 seconds of video at 8 FPS) with output resolution 720x480 (W x H), which makes it not possible to run on consumer GPUs or free-tier T4 Colab. The following memory optimizations could be used to reduce the memory footprint. For replication, you can refer to [this](https://gist.github.com/a-r-r-o-w/3959a03f15be5c9bd1fe545b09dfcc93) script.
|
||||
|
||||
- `pipe.enable_model_cpu_offload()`:
|
||||
- Without enabling cpu offloading, memory usage is `33 GB`
|
||||
|
||||
@@ -78,6 +78,10 @@ Since RegEx is supported as a way for matching layer identifiers, it is crucial
|
||||
- all
|
||||
- __call__
|
||||
|
||||
## StableDiffusionXLControlNetPAGImg2ImgPipeline
|
||||
[[autodoc]] StableDiffusionXLControlNetPAGImg2ImgPipeline
|
||||
- all
|
||||
- __call__
|
||||
|
||||
## StableDiffusion3PAGPipeline
|
||||
[[autodoc]] StableDiffusion3PAGPipeline
|
||||
|
||||
@@ -238,7 +238,7 @@ Pretty impressive! Let's tweak the second image - corresponding to the `Generato
|
||||
```python
|
||||
prompts = [
|
||||
"portrait photo of the oldest warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3 --beta --upbeta",
|
||||
"portrait photo of a old warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3 --beta --upbeta",
|
||||
"portrait photo of an old warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3 --beta --upbeta",
|
||||
"portrait photo of a warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3 --beta --upbeta",
|
||||
"portrait photo of a young warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3 --beta --upbeta",
|
||||
]
|
||||
|
||||
@@ -10,7 +10,7 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
|
||||
specific language governing permissions and limitations under the License.
|
||||
-->
|
||||
|
||||
# Diffusers에 기여하는 방법 🧨
|
||||
# Diffusers에 기여하는 방법 🧨 [[how-to-contribute-to-diffusers-]]
|
||||
|
||||
오픈 소스 커뮤니티에서의 기여를 환영합니다! 누구나 참여할 수 있으며, 코드뿐만 아니라 질문에 답변하거나 문서를 개선하는 등 모든 유형의 참여가 가치 있고 감사히 여겨집니다. 질문에 답변하고 다른 사람들을 도와주며 소통하고 문서를 개선하는 것은 모두 커뮤니티에게 큰 도움이 됩니다. 따라서 관심이 있다면 두려워하지 말고 참여해보세요!
|
||||
|
||||
@@ -18,9 +18,9 @@ specific language governing permissions and limitations under the License.
|
||||
|
||||
어떤 방식으로든 기여하려는 경우, 우리는 개방적이고 환영하며 친근한 커뮤니티의 일부가 되기 위해 노력하고 있습니다. 우리의 [행동 강령](https://github.com/huggingface/diffusers/blob/main/CODE_OF_CONDUCT.md)을 읽고 상호 작용 중에 이를 존중하도록 주의해주시기 바랍니다. 또한 프로젝트를 안내하는 [윤리 지침](https://huggingface.co/docs/diffusers/conceptual/ethical_guidelines)에 익숙해지고 동일한 투명성과 책임성의 원칙을 준수해주시기를 부탁드립니다.
|
||||
|
||||
우리는 커뮤니티로부터의 피드백을 매우 중요하게 생각하므로, 라이브러리를 개선하는 데 도움이 될 가치 있는 피드백이 있다고 생각되면 망설이지 말고 의견을 제시해주세요 - 모든 메시지, 댓글, 이슈, 풀 리퀘스트(PR)는 읽히고 고려됩니다.
|
||||
우리는 커뮤니티로부터의 피드백을 매우 중요하게 생각하므로, 라이브러리를 개선하는 데 도움이 될 가치 있는 피드백이 있다고 생각되면 망설이지 말고 의견을 제시해주세요 - 모든 메시지, 댓글, 이슈, Pull Request(PR)는 읽히고 고려됩니다.
|
||||
|
||||
## 개요
|
||||
## 개요 [[overview]]
|
||||
|
||||
이슈에 있는 질문에 답변하는 것에서부터 코어 라이브러리에 새로운 diffusion 모델을 추가하는 것까지 다양한 방법으로 기여를 할 수 있습니다.
|
||||
|
||||
@@ -38,9 +38,9 @@ specific language governing permissions and limitations under the License.
|
||||
|
||||
앞서 말한 대로, **모든 기여는 커뮤니티에게 가치가 있습니다**. 이어지는 부분에서 각 기여에 대해 조금 더 자세히 설명하겠습니다.
|
||||
|
||||
4부터 9까지의 모든 기여에는 PR을 열어야 합니다. [PR을 열기](#how-to-open-a-pr)에서 자세히 설명되어 있습니다.
|
||||
4부터 9까지의 모든 기여에는 Pull Request을 열어야 합니다. [Pull Request 열기](#how-to-open-a-pr)에서 자세히 설명되어 있습니다.
|
||||
|
||||
### 1. Diffusers 토론 포럼이나 Diffusers Discord에서 질문하고 답변하기
|
||||
### 1. Diffusers 토론 포럼이나 Diffusers Discord에서 질문하고 답변하기 [[1-asking-and-answering-questions-on-the-diffusers-discussion-forum-or-on-the-diffusers-discord]]
|
||||
|
||||
Diffusers 라이브러리와 관련된 모든 질문이나 의견은 [토론 포럼](https://discuss.huggingface.co/c/discussion-related-to-httpsgithubcomhuggingfacediffusers/63)이나 [Discord](https://discord.gg/G7tWnz98XR)에서 할 수 있습니다. 이러한 질문과 의견에는 다음과 같은 내용이 포함됩니다(하지만 이에 국한되지는 않습니다):
|
||||
- 지식을 공유하기 위해서 훈련 또는 추론 실험에 대한 결과 보고
|
||||
@@ -54,7 +54,7 @@ Diffusers 라이브러리와 관련된 모든 질문이나 의견은 [토론 포
|
||||
- Diffusion 모델에 대한 윤리적 질문
|
||||
- ...
|
||||
|
||||
포럼이나 Discord에서 질문을 하면 커뮤니티가 지식을 공개적으로 공유하도록 장려되며, 미래에 동일한 질문을 가진 초보자에게도 도움이 될 수 있습니다. 따라서 궁금한 질문은 언제든지 하시기 바랍니다.
|
||||
포럼이나 Discord에서 질문을 하면 커뮤니티가 지식을 공개적으로 공유하도록 장려되며, 향후 동일한 질문을 가진 초보자에게도 도움이 될 수 있습니다. 따라서 궁금한 질문은 언제든지 하시기 바랍니다.
|
||||
또한, 이러한 질문에 답변하는 것은 커뮤니티에게 매우 큰 도움이 됩니다. 왜냐하면 이렇게 하면 모두가 학습할 수 있는 공개적인 지식을 문서화하기 때문입니다.
|
||||
|
||||
**주의**하십시오. 질문이나 답변에 투자하는 노력이 많을수록 공개적으로 문서화된 지식의 품질이 높아집니다. 마찬가지로, 잘 정의되고 잘 답변된 질문은 모두에게 접근 가능한 고품질 지식 데이터베이스를 만들어줍니다. 반면에 잘못된 질문이나 답변은 공개 지식 데이터베이스의 전반적인 품질을 낮출 수 있습니다.
|
||||
@@ -64,9 +64,9 @@ Diffusers 라이브러리와 관련된 모든 질문이나 의견은 [토론 포
|
||||
[*포럼*](https://discuss.huggingface.co/c/discussion-related-to-httpsgithubcomhuggingfacediffusers/63)은 구글과 같은 검색 엔진에서 더 잘 색인화됩니다. 게시물은 인기에 따라 순위가 매겨지며, 시간순으로 정렬되지 않습니다. 따라서 이전에 게시한 질문과 답변을 쉽게 찾을 수 있습니다.
|
||||
또한, 포럼에 게시된 질문과 답변은 쉽게 링크할 수 있습니다.
|
||||
반면 *Discord*는 채팅 형식으로 되어 있어 빠른 대화를 유도합니다.
|
||||
질문에 대한 답변을 빠르게 받을 수는 있겠지만, 시간이 지나면 질문이 더 이상 보이지 않습니다. 또한, Discord에서 이전에 게시된 정보를 찾는 것은 훨씬 어렵습니다. 따라서 포럼을 사용하여 고품질의 질문과 답변을 하여 커뮤니티를 위한 오래 지속되는 지식을 만들기를 권장합니다. Discord에서의 토론이 매우 흥미로운 답변과 결론을 이끌어내는 경우, 해당 정보를 포럼에 게시하여 미래 독자들에게 더 쉽게 액세스할 수 있도록 권장합니다.
|
||||
질문에 대한 답변을 빠르게 받을 수는 있겠지만, 시간이 지나면 질문이 더 이상 보이지 않습니다. 또한, Discord에서 이전에 게시된 정보를 찾는 것은 훨씬 어렵습니다. 따라서 포럼을 사용하여 고품질의 질문과 답변을 하여 커뮤니티를 위한 오래 지속되는 지식을 만들기를 권장합니다. Discord에서의 토론이 매우 흥미로운 답변과 결론을 이끌어내는 경우, 해당 정보를 포럼에 게시하여 향후 독자들에게 더 쉽게 액세스할 수 있도록 권장합니다.
|
||||
|
||||
### 2. GitHub 이슈 탭에서 새로운 이슈 열기
|
||||
### 2. GitHub 이슈 탭에서 새로운 이슈 열기 [[2-opening-new-issues-on-the-github-issues-tab]]
|
||||
|
||||
🧨 Diffusers 라이브러리는 사용자들이 마주치는 문제를 알려주는 덕분에 견고하고 신뢰할 수 있습니다. 따라서 이슈를 보고해주셔서 감사합니다.
|
||||
|
||||
@@ -81,53 +81,52 @@ Diffusers 라이브러리와 관련된 모든 질문이나 의견은 [토론 포
|
||||
- 이슈가 최신 Diffusers 버전으로 업데이트하면 해결될 수 있는지 확인해주세요. 이슈를 게시하기 전에 `python -c "import diffusers; print(diffusers.__version__)"` 명령을 실행하여 현재 사용 중인 Diffusers 버전이 최신 버전과 일치하거나 더 높은지 확인해주세요.
|
||||
- 새로운 이슈를 열 때 투자하는 노력이 많을수록 답변의 품질이 높아지고 Diffusers 이슈 전체의 품질도 향상됩니다.
|
||||
|
||||
#### 2.1 재현가능하고 최소한인 버그 리포트
|
||||
#### 2.1 재현 가능한 최소한의 버그 리포트 [[21-reproducible-minimal-bug-reports]]
|
||||
|
||||
새로운 이슈는 일반적으로 다음과 같은 내용을 포함합니다.
|
||||
|
||||
버그 보고서는 항상 재현 가능한 코드 조각을 포함하고 가능한 한 최소한이어야 하며 간결해야 합니다.
|
||||
버그 리포트는 항상 재현 가능한 코드 조각을 포함하고 가능한 한 최소한이어야 하며 간결해야 합니다.
|
||||
자세히 말하면:
|
||||
- 버그를 가능한 한 좁혀야 합니다. **전체 코드 파일을 그냥 던지지 마세요**.
|
||||
- 코드의 서식을 지정해야 합니다.
|
||||
- Diffusers가 의존하는 외부 라이브러리를 제외한 다른 외부 라이브러리는 포함하지 마십시오.
|
||||
- **반드시** 환경에 대한 모든 필요한 정보를 제공해야 합니다. 이를 위해 쉘에서 `diffusers-cli env`를 실행하고 표시된 정보를 이슈에 복사하여 붙여넣을 수 있습니다.
|
||||
- 이슈를 설명해야 합니다. 독자가 문제가 무엇이며 왜 문제인지 모르면 해결할 수 없습니다.
|
||||
- **항상** 독자가 가능한 한 적은 노력으로 문제를 재현할 수 있도록 해야 합니다. 코드 조각이 라이브러리가 없거나 정의되지 않은 변수 때문에 실행되지 않는 경우 독자가 도움을 줄 수 없습니다. 재현 가능한 코드 조각이 가능한 한 최소화되고 간단한 Python 셸에 복사하여 붙여넣을 수 있도록 해야 합니다.
|
||||
- **항상** 사용자 환경에 대한 모든 필요한 정보를 제공하세요. 이를 위해 쉘에서 `diffusers-cli env`를 실행하고 표시된 정보를 이슈에 복사하여 붙여넣을 수 있습니다.
|
||||
- 이슈를 설명해야 합니다. 독자가 문제가 무엇인지, 왜 문제가 되는지 모른다면 이슈를 해결할 수 없습니다.
|
||||
- **항상** 독자가 가능한 한 적은 노력으로 문제를 재현할 수 있어야 합니다. 코드 조각이 라이브러리가 없거나 정의되지 않은 변수 때문에 실행되지 않는 경우 독자가 도움을 줄 수 없습니다. 재현 가능한 코드 조각이 가능한 한 최소화되고 간단한 Python 셸에 복사하여 붙여넣을 수 있도록 해야 합니다.
|
||||
- 문제를 재현하기 위해 모델과/또는 데이터셋이 필요한 경우 독자가 해당 모델이나 데이터셋에 접근할 수 있도록 해야 합니다. 모델이나 데이터셋을 [Hub](https://huggingface.co)에 업로드하여 쉽게 다운로드할 수 있도록 할 수 있습니다. 문제 재현을 가능한 한 쉽게하기 위해 모델과 데이터셋을 가능한 한 작게 유지하려고 노력하세요.
|
||||
|
||||
자세한 내용은 [좋은 이슈 작성 방법](#how-to-write-a-good-issue) 섹션을 참조하세요.
|
||||
|
||||
버그 보고서를 열려면 [여기](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=bug&projects=&template=bug-report.yml)를 클릭하세요.
|
||||
버그 리포트를 열려면 [여기](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=bug&projects=&template=bug-report.yml)를 클릭하세요.
|
||||
|
||||
|
||||
#### 2.2. 기능 요청
|
||||
#### 2.2. 기능 요청 [[22-feature-requests]]
|
||||
|
||||
세계적인 기능 요청은 다음 사항을 다룹니다:
|
||||
|
||||
1. 먼저 동기부여:
|
||||
* 라이브러리와 관련된 문제/불만이 있는가요? 그렇다면 왜 그런지 설명해주세요. 문제를 보여주는 코드 조각을 제공하는 것이 가장 좋습니다.
|
||||
* 라이브러리와 관련된 문제/불만이 있나요? 그렇다면 왜 그런지 설명해주세요. 문제를 보여주는 코드 조각을 제공하는 것이 가장 좋습니다.
|
||||
* 프로젝트에 필요한 기능인가요? 우리는 그에 대해 듣고 싶습니다!
|
||||
* 커뮤니티에 도움이 될 수 있는 것을 작업했고 그것에 대해 생각하고 있는가요? 멋지네요! 어떤 문제를 해결했는지 알려주세요.
|
||||
2. 기능을 *상세히 설명하는* 문단을 작성해주세요;
|
||||
3. 미래 사용을 보여주는 **코드 조각**을 제공해주세요;
|
||||
4. 이것이 논문과 관련된 경우 링크를 첨부해주세요;
|
||||
5. 도움이 될 수 있는 추가 정보(그림, 스크린샷 등)를 첨부해주세요.
|
||||
3. 향후 사용을 보여주는 **코드 조각**을 제공해주세요;
|
||||
4. 논문과 관련된 내용인 경우 링크를 첨부해주세요;
|
||||
5. 도움이 될 수 있다고 생각되는 추가 정보(그림, 스크린샷 등)를 첨부해주세요.
|
||||
|
||||
기능 요청은 [여기](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=&template=feature_request.md&title=)에서 열 수 있습니다.
|
||||
|
||||
#### 2.3 피드백
|
||||
#### 2.3 피드백 [[23-feedback]]
|
||||
|
||||
라이브러리 디자인과 그것이 왜 좋은지 또는 나쁜지에 대한 이유에 대한 피드백은 핵심 메인테이너가 사용자 친화적인 라이브러리를 만드는 데 엄청난 도움이 됩니다. 현재 디자인 철학을 이해하려면 [여기](https://huggingface.co/docs/diffusers/conceptual/philosophy)를 참조해 주세요. 특정 디자인 선택이 현재 디자인 철학과 맞지 않는다고 생각되면, 그 이유와 어떻게 변경되어야 하는지 설명해 주세요. 반대로 특정 디자인 선택이 디자인 철학을 너무 따르기 때문에 사용 사례를 제한한다고 생각되면, 그 이유와 어떻게 변경되어야 하는지 설명해 주세요. 특정 디자인 선택이 매우 유용하다고 생각되면, 미래의 디자인 결정에 큰 도움이 되므로 이에 대한 의견을 남겨 주세요.
|
||||
라이브러리 디자인과 그것이 왜 좋은지 또는 나쁜지에 대한 이유에 대한 피드백은 핵심 메인테이너가 사용자 친화적인 라이브러리를 만드는 데 엄청난 도움이 됩니다. 현재 디자인 철학을 이해하려면 [여기](https://huggingface.co/docs/diffusers/conceptual/philosophy)를 참조해 주세요. 특정 디자인 선택이 현재 디자인 철학과 맞지 않는다고 생각되면, 그 이유와 어떻게 변경되어야 하는지 설명해 주세요. 반대로 특정 디자인 선택이 디자인 철학을 너무 따르기 때문에 사용 사례를 제한한다고 생각되면, 그 이유와 어떻게 변경되어야 하는지 설명해 주세요. 특정 디자인 선택이 매우 유용하다고 생각되면, 향후 디자인 결정에 큰 도움이 되므로 이에 대한 의견을 남겨 주세요.
|
||||
|
||||
피드백에 관한 이슈는 [여기](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=&template=feedback.md&title=)에서 열 수 있습니다.
|
||||
|
||||
#### 2.4 기술적인 질문
|
||||
#### 2.4 기술적인 질문 [[24-technical-questions]]
|
||||
|
||||
기술적인 질문은 주로 라이브러리의 특정 코드가 왜 특정 방식으로 작성되었는지 또는 코드의 특정 부분이 무엇을 하는지에 대한 질문입니다. 질문하신 코드 부분에 대한 링크를 제공하고 해당 코드 부분이 이해하기 어려운 이유에 대한 자세한 설명을 해주시기 바랍니다.
|
||||
|
||||
기술적인 질문에 관한 이슈를 [여기](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=bug&template=bug-report.yml)에서 열 수 있습니다.
|
||||
|
||||
#### 2.5 새로운 모델, 스케줄러 또는 파이프라인 추가 제안
|
||||
#### 2.5 새로운 모델, 스케줄러 또는 파이프라인 추가 제안 [[25-proposal-to-add-a-new-model-scheduler-or-pipeline]]
|
||||
|
||||
만약 diffusion 모델 커뮤니티에서 Diffusers 라이브러리에 추가하고 싶은 새로운 모델, 파이프라인 또는 스케줄러가 있다면, 다음 정보를 제공해주세요:
|
||||
|
||||
@@ -135,34 +134,34 @@ Diffusers 라이브러리와 관련된 모든 질문이나 의견은 [토론 포
|
||||
* 해당 모델의 오픈 소스 구현에 대한 링크
|
||||
* 모델 가중치가 있는 경우, 가중치의 링크
|
||||
|
||||
모델에 직접 기여하고자 하는 경우, 최선의 안내를 위해 우리에게 알려주세요. 또한, 가능하다면 구성 요소(모델, 스케줄러, 파이프라인 등)의 원래 저자를 GitHub 핸들로 태그하는 것을 잊지 마세요.
|
||||
직접 모델에 기여하고 싶다면, 가장 잘 안내해드릴 수 있습니다. 또한, 가능하다면 구성 요소(모델, 스케줄러, 파이프라인 등)의 원저자를 GitHub 핸들로 태그하는 것을 잊지 마세요.
|
||||
|
||||
모델/파이프라인/스케줄러에 대한 요청을 [여기](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=New+model%2Fpipeline%2Fscheduler&template=new-model-addition.yml)에서 열 수 있습니다.
|
||||
|
||||
### 3. GitHub 이슈 탭에서 문제에 대한 답변하기
|
||||
### 3. GitHub 이슈 탭에서 문제에 대한 답변하기 [[3-answering-issues-on-the-github-issues-tab]]
|
||||
|
||||
GitHub에서 이슈에 대한 답변을 하기 위해서는 Diffusers에 대한 기술적인 지식이 필요할 수 있지만, 정확한 답변이 아니더라도 모두가 시도해기를 권장합니다. 이슈에 대한 고품질 답변을 제공하기 위한 몇 가지 팁:
|
||||
- 가능한 한 간결하고 최소한으로 유지합니다.
|
||||
- 주제에 집중합니다. 이슈에 대한 답변은 해당 이슈에 관련된 내용에만 집중해야 합니다.
|
||||
- 코드, 논문 또는 다른 소스를 제공하여 답변을 증명하거나 지지합니다.
|
||||
- 자신의 주장을 증명하거나 장려하는 코드, 논문 또는 기타 출처는 링크를 제공하세요.
|
||||
- 코드로 답변합니다. 간단한 코드 조각이 이슈에 대한 답변이거나 이슈를 해결하는 방법을 보여준다면, 완전히 재현 가능한 코드 조각을 제공해주세요.
|
||||
|
||||
또한, 많은 이슈들은 단순히 주제와 무관하거나 다른 이슈의 중복이거나 관련이 없는 경우가 많습니다. 이러한 이슈들에 대한 답변을 제공하고, 이슈 작성자에게 더 정확한 정보를 제공하거나, 중복된 이슈에 대한 링크를 제공하거나, [포럼](https://discuss.huggingface.co/c/discussion-related-to-httpsgithubcomhuggingfacediffusers/63) 이나 [Discord](https://discord.gg/G7tWnz98XR)로 리디렉션하는 것은 메인테이너에게 큰 도움이 됩니다.
|
||||
|
||||
이슈가 올바른 버그 보고서이고 소스 코드에서 수정이 필요하다고 확인한 경우, 다음 섹션을 살펴보세요.
|
||||
|
||||
다음 모든 기여에 대해서는 PR을 열여야 합니다. [PR 열기](#how-to-open-a-pr) 섹션에서 자세히 설명되어 있습니다.
|
||||
다음 모든 기여에 대해서는 PR을 열여야 합니다. [Pull Request 열기](#how-to-open-a-pr) 섹션에서 자세히 설명되어 있습니다.
|
||||
|
||||
### 4. "Good first issue" 고치기
|
||||
### 4. "Good first issue" 고치기 [[4-fixing-a-good-first-issue]]
|
||||
|
||||
*Good first issues*는 [Good first issue](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22good+first+issue%22) 라벨로 표시됩니다. 일반적으로, 이슈는 이미 잠재적인 해결책이 어떻게 보이는지 설명하고 있어서 수정하기 쉽습니다.
|
||||
만약 이슈가 아직 닫히지 않았고 이 문제를 해결해보고 싶다면, "이 이슈를 해결해보고 싶습니다."라는 메시지를 남기면 됩니다. 일반적으로 세 가지 시나리오가 있습니다:
|
||||
- a.) 이슈 설명이 이미 해결책을 제안합니다. 이 경우, 해결책이 이해되고 합리적으로 보인다면, PR 또는 드래프트 PR을 열어서 수정할 수 있습니다.
|
||||
- b.) 이슈 설명이 해결책을 제안하지 않습니다. 이 경우, 어떤 해결책이 가능할지 물어볼 수 있고, Diffusers 팀의 누군가가 곧 답변해줄 것입니다. 만약 어떻게 수정할지 좋은 아이디어가 있다면, 직접 PR을 열어도 됩니다.
|
||||
- a.) 이슈 설명에 이미 수정 사항을 제안하는 경우, 해결책이 이해되고 합리적으로 보인다면, PR 또는 드래프트 PR을 열어서 수정할 수 있습니다.
|
||||
- b.) 이슈 설명에 수정 사항이 제안되어 있지 않은 경우, 제안한 수정 사항이 가능할지 물어볼 수 있고, Diffusers 팀의 누군가가 곧 답변해줄 것입니다. 만약 어떻게 수정할지 좋은 아이디어가 있다면, 직접 PR을 열어도 됩니다.
|
||||
- c.) 이미 이 문제를 해결하기 위해 열린 PR이 있지만, 이슈가 아직 닫히지 않았습니다. PR이 더 이상 진행되지 않았다면, 새로운 PR을 열고 이전 PR에 링크를 걸면 됩니다. PR은 종종 원래 기여자가 갑자기 시간을 내지 못해 더 이상 진행하지 못하는 경우에 더 이상 진행되지 않게 됩니다. 이는 오픈 소스에서 자주 발생하는 일이며 매우 정상적인 상황입니다. 이 경우, 커뮤니티는 새로 시도하고 기존 PR의 지식을 활용해주면 매우 기쁠 것입니다. 이미 PR이 있고 활성화되어 있다면, 제안을 해주거나 PR을 검토하거나 PR에 기여할 수 있는지 물어보는 등 작성자를 도와줄 수 있습니다.
|
||||
|
||||
|
||||
### 5. 문서에 기여하기
|
||||
### 5. 문서에 기여하기 [[5-contribute-to-the-documentation]]
|
||||
|
||||
좋은 라이브러리는 항상 좋은 문서를 갖고 있습니다! 공식 문서는 라이브러리를 처음 사용하는 사용자들에게 첫 번째 접점 중 하나이며, 따라서 문서에 기여하는 것은 매우 가치 있는 기여입니다.
|
||||
|
||||
@@ -180,7 +179,7 @@ GitHub에서 이슈에 대한 답변을 하기 위해서는 Diffusers에 대한
|
||||
문서에 대한 변경 사항을 로컬에서 확인하는 방법은 [이 페이지](https://github.com/huggingface/diffusers/tree/main/docs)를 참조해주세요.
|
||||
|
||||
|
||||
### 6. 커뮤니티 파이프라인에 기여하기
|
||||
### 6. 커뮤니티 파이프라인에 기여하기 [[6-contribute-a-community-pipeline]]
|
||||
|
||||
> [!TIP]
|
||||
> 커뮤니티 파이프라인에 대해 자세히 알아보려면 [커뮤니티 파이프라인](../using-diffusers/custom_pipeline_overview#community-pipelines) 가이드를 읽어보세요. 커뮤니티 파이프라인이 왜 필요한지 궁금하다면 GitHub 이슈 [#841](https://github.com/huggingface/diffusers/issues/841)를 확인해보세요 (기본적으로, 우리는 diffusion 모델이 추론에 사용될 수 있는 모든 방법을 유지할 수 없지만 커뮤니티가 이를 구축하는 것을 방해하고 싶지 않습니다).
|
||||
@@ -246,7 +245,7 @@ output = pipeline()
|
||||
<hfoptions id="pipeline type">
|
||||
<hfoption id="GitHub pipeline">
|
||||
|
||||
GitHub 파이프라인을 공유하려면 Diffusers [저장소](https://github.com/huggingface/diffusers)에서 PR을 열고 one_step_unet.py 파일을 [examples/community](https://github.com/huggingface/diffusers/tree/main/examples/community) 하위 폴더에 추가하세요.
|
||||
GitHub 파이프라인을 공유하려면 Diffusers [저장소](https://github.com/huggingface/diffusers)에서 Pull Request를 열고 one_step_unet.py 파일을 [examples/community](https://github.com/huggingface/diffusers/tree/main/examples/community) 하위 폴더에 추가하세요.
|
||||
|
||||
</hfoption>
|
||||
<hfoption id="Hub pipeline">
|
||||
@@ -256,7 +255,7 @@ Hub 파이프라인을 공유하려면, 허브에 모델 저장소를 생성하
|
||||
</hfoption>
|
||||
</hfoptions>
|
||||
|
||||
### 7. 훈련 예제에 기여하기
|
||||
### 7. 훈련 예제에 기여하기 [[7-contribute-to-training-examples]]
|
||||
|
||||
Diffusers 예제는 [examples](https://github.com/huggingface/diffusers/tree/main/examples) 폴더에 있는 훈련 스크립트의 모음입니다.
|
||||
|
||||
@@ -268,7 +267,7 @@ Diffusers 예제는 [examples](https://github.com/huggingface/diffusers/tree/mai
|
||||
연구용 훈련 예제는 [examples/research_projects](https://github.com/huggingface/diffusers/tree/main/examples/research_projects)에 위치하며, 공식 훈련 예제는 `research_projects` 및 `community` 폴더를 제외한 [examples](https://github.com/huggingface/diffusers/tree/main/examples)의 모든 폴더를 포함합니다.
|
||||
공식 훈련 예제는 Diffusers의 핵심 메인테이너가 유지 관리하며, 연구용 훈련 예제는 커뮤니티가 유지 관리합니다.
|
||||
이는 공식 파이프라인 vs 커뮤니티 파이프라인에 대한 [6. 커뮤니티 파이프라인 기여하기](#6-contribute-a-community-pipeline)에서 제시한 이유와 동일합니다: 핵심 메인테이너가 diffusion 모델의 모든 가능한 훈련 방법을 유지 관리하는 것은 현실적으로 불가능합니다.
|
||||
Diffusers 핵심 메인테잉너와 커뮤니티가 특정 훈련 패러다임을 너무 실험적이거나 충분히 인기 없는 것으로 간주하는 경우, 해당 훈련 코드는 `research_projects` 폴더에 넣고 작성자가 유지 관리해야 합니다.
|
||||
Diffusers 핵심 메인테이너와 커뮤니티가 특정 훈련 패러다임을 너무 실험적이거나 충분히 대중적이지 않다고 판단한다면, 해당 훈련 코드는 `research_projects` 폴더에 넣고 작성자에 의해 관리되어야 합니다.
|
||||
|
||||
공식 훈련 및 연구 예제는 하나 이상의 훈련 스크립트, requirements.txt 파일 및 README.md 파일을 포함하는 디렉토리로 구성됩니다. 사용자가 훈련 예제를 사용하려면 리포지토리를 복제해야 합니다:
|
||||
|
||||
@@ -298,14 +297,14 @@ Diffusers와 긴밀하게 통합되어 있기 때문에, 기여자들이 [Accele
|
||||
|
||||
만약 공식 훈련 예제에 기여하는 경우, [examples/test_examples.py](https://github.com/huggingface/diffusers/blob/main/examples/test_examples.py)에 테스트를 추가하는 것도 확인해주세요. 비공식 훈련 예제에는 이 작업이 필요하지 않습니다.
|
||||
|
||||
### 8. "Good second issue" 고치기
|
||||
### 8. "Good second issue" 고치기 [[8-fixing-a-good-second-issue]]
|
||||
|
||||
"Good second issue"는 [Good second issue](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22Good+second+issue%22) 라벨로 표시됩니다. Good second issue는 [Good first issues](https://github.com/huggingface/diffusers/issues?q=is%3Aopen+is%3Aissue+label%3A%22good+first+issue%22)보다 해결하기가 더 복잡합니다.
|
||||
이슈 설명은 일반적으로 이슈를 해결하는 방법에 대해 덜 구체적이며, 관심 있는 기여자는 라이브러리에 대한 꽤 깊은 이해가 필요합니다.
|
||||
Good second issue를 해결하고자 하는 경우, 해당 이슈를 해결하기 위해 PR을 열고 PR을 이슈에 링크하세요. 이미 해당 이슈에 대한 PR이 열려있지만 병합되지 않은 경우, 왜 병합되지 않았는지 이해하기 위해 살펴보고 개선된 PR을 열어보세요.
|
||||
Good second issue는 일반적으로 Good first issue 이슈보다 병합하기가 더 어려우므로, 핵심 메인테이너에게 도움을 요청하는 것이 좋습니다. PR이 거의 완료된 경우, 핵심 메인테이너는 PR에 참여하여 커밋하고 병합을 진행할 수 있습니다.
|
||||
|
||||
### 9. 파이프라인, 모델, 스케줄러 추가하기
|
||||
### 9. 파이프라인, 모델, 스케줄러 추가하기 [[9-adding-pipelines-models-schedulers]]
|
||||
|
||||
파이프라인, 모델, 스케줄러는 Diffusers 라이브러리에서 가장 중요한 부분입니다.
|
||||
이들은 최첨단 diffusion 기술에 쉽게 접근하도록 하며, 따라서 커뮤니티가 강력한 생성형 AI 애플리케이션을 만들 수 있도록 합니다.
|
||||
@@ -323,9 +322,9 @@ PR에 원본 코드베이스/논문 링크를 추가하고, 가능하면 PR에
|
||||
|
||||
PR에서 막힌 경우나 도움이 필요한 경우, 첫 번째 리뷰나 도움을 요청하는 메시지를 남기는 것을 주저하지 마세요.
|
||||
|
||||
#### Copied from mechanism
|
||||
#### Copied from mechanism [[copied-from-mechanism]]
|
||||
|
||||
`# Copied from mechanism` 은 파이프라인, 모델 또는 스케줄러 코드를 추가할 때 이해해야 할 독특하고 중요한 기능입니다. Diffusers 코드베이스 전체에서 이를 자주 볼 수 있는데, 이를 사용하는 이유는 코드베이스를 이해하기 쉽고 유지 관리하기 쉽게 유지하기 위함입니다. `# Copied from mechanism` 으로 표시된 코드는 복사한 코드와 정확히 동일하도록 강제됩니다. 이를 통해 `make fix-copies`를 실행할 때 많은 파일에 걸쳐 변경 사항을 쉽게 업데이트하고 전파할 수 있습니다.
|
||||
`# Copied from mechanism` 은 파이프라인, 모델 또는 스케줄러 코드를 추가할 때 이해해야 할 독특하고 중요한 기능입니다. 이것은 Diffusers 코드베이스 전반에서 볼 수 있으며, 이를 사용하는 이유는 코드베이스를 이해하고 유지 관리하기 쉽게 만들기 위해서입니다. `# Copied from mechanism` 으로 표시된 코드는 복사한 코드와 정확히 동일하도록 강제됩니다. 이렇게 하면 `make fix-copies`를 실행할 때마다 여러 파일에 걸쳐 변경 사항을 쉽게 업데이트하고 전파할 수 있습니다.
|
||||
|
||||
예를 들어, 아래 코드 예제에서 [`~diffusers.pipelines.stable_diffusion.StableDiffusionPipelineOutput`]은 원래 코드이며, `AltDiffusionPipelineOutput`은 `# Copied from mechanism`을 사용하여 복사합니다. 유일한 차이점은 클래스 접두사를 `Stable`에서 `Alt`로 변경한 것입니다.
|
||||
|
||||
@@ -347,7 +346,7 @@ class AltDiffusionPipelineOutput(BaseOutput):
|
||||
|
||||
더 자세히 알고 싶다면 [~Don't~ Repeat Yourself*](https://huggingface.co/blog/transformers-design-philosophy#4-machine-learning-models-are-static) 블로그 포스트의 이 섹션을 읽어보세요.
|
||||
|
||||
## 좋은 이슈 작성 방법
|
||||
## 좋은 이슈 작성 방법 [[how-to-write-a-good-issue]]
|
||||
|
||||
**이슈를 잘 작성할수록 빠르게 해결될 가능성이 높아집니다.**
|
||||
|
||||
@@ -356,16 +355,16 @@ class AltDiffusionPipelineOutput(BaseOutput):
|
||||
3. **재현 가능성**: 재현 가능한 코드 조각이 없으면 해결할 수 없습니다. 버그를 발견한 경우, 유지 관리자는 그 버그를 재현할 수 있어야 합니다. 이슈에 재현 가능한 코드 조각을 포함해야 합니다. 코드 조각은 Python 인터프리터에 복사하여 붙여넣을 수 있는 형태여야 합니다. 코드 조각이 작동해야 합니다. 즉, 누락된 import나 이미지에 대한 링크가 없어야 합니다. 이슈에는 오류 메시지와 정확히 동일한 오류 메시지를 재현하기 위해 수정하지 않고 복사하여 붙여넣을 수 있는 코드 조각이 포함되어야 합니다. 이슈에 사용자의 로컬 모델 가중치나 로컬 데이터를 사용하는 경우, 독자가 액세스할 수 없는 경우 이슈를 해결할 수 없습니다. 데이터나 모델을 공유할 수 없는 경우, 더미 모델이나 더미 데이터를 만들어 사용해보세요.
|
||||
4. **간결성**: 가능한 한 간결하게 유지하여 독자가 문제를 빠르게 이해할 수 있도록 도와주세요. 문제와 관련이 없는 코드나 정보는 모두 제거해주세요. 버그를 발견한 경우, 문제를 설명하는 가장 간단한 코드 예제를 만들어보세요. 버그를 발견한 후에는 작업 흐름 전체를 문제에 던지는 것이 아니라, 에러가 발생하는 훈련 코드의 어느 부분이 문제인지 먼저 이해하고 몇 줄로 재현해보세요. 전체 데이터셋 대신 더미 데이터를 사용해보세요.
|
||||
5. 링크 추가하기. 특정한 이름, 메서드, 또는 모델을 참조하는 경우, 독자가 더 잘 이해할 수 있도록 링크를 제공해주세요. 특정 PR이나 이슈를 참조하는 경우, 해당 이슈에 링크를 걸어주세요. 독자가 무엇을 말하는지 알고 있다고 가정하지 마세요. 이슈에 링크를 추가할수록 좋습니다.
|
||||
6. 포맷팅. 파이썬 코드 구문으로 코드를 포맷팅하고, 일반 코드 구문으로 에러 메시지를 포맷팅해주세요. 자세한 내용은 [공식 GitHub 포맷팅 문서](https://docs.github.com/en/get-started/writing-on-github/getting-started-with-writing-and-formatting-on-github/basic-writing-and-formatting-syntax)를 참조하세요.
|
||||
7. 이슈를 해결해야 하는 티켓이 아니라, 잘 작성된 백과사전 항목으로 생각해보세요. 추가된 이슈는 공개적으로 사용 가능한 지식에 기여하는 것입니다. 잘 작성된 이슈를 추가함으로써 메인테이너가 문제를 해결하는 데 도움을 주는 것뿐만 아니라, 전체 커뮤니티가 라이브러리의 특정 측면을 더 잘 이해할 수 있도록 도움을 주는 것입니다.
|
||||
6. 포맷팅. 코드를 파이썬 코드 구문으로, 에러 메시지를 일반 코드 구문으로 형식화하여 이슈를 깔끔하게 작성하세요. 자세한 내용은 [GitHub 공식 포맷팅 문서](https://docs.github.com/en/get-started/writing-on-github/getting-started-with-writing-and-formatting-on-github/basic-writing-and-formatting-syntax)를 참조하세요.
|
||||
7. 여러분의 이슈를 단순히 해결해야 할 티켓으로 생각하지 말고, 잘 작성된 백과사전 항목으로 생각해보세요. 추가된 모든 이슈는 공개적으로 이용 가능한 지식에 대한 기여입니다. 잘 작성된 이슈를 추가함으로써 메인테이너가 여러분의 이슈를 더 쉽게 해결할 수 있게 할 뿐만 아니라, 전체 커뮤니티가 라이브러리의 특정 측면을 더 잘 이해할 수 있도록 도움을 주게 됩니다.
|
||||
|
||||
## 좋은 PR 작성 방법
|
||||
## 좋은 PR 작성 방법 [[how-to-write-a-good-pr]]
|
||||
|
||||
1. 카멜레온이 되세요. 기존의 디자인 패턴과 구문을 이해하고, 코드 추가가 기존 코드베이스에 매끄럽게 흐르도록 해야 합니다. 기존 디자인 패턴이나 사용자 인터페이스와 크게 다른 PR은 병합되지 않습니다.
|
||||
2. 초점을 맞추세요. 하나의 문제만 해결하는 PR을 작성해야 합니다. "추가하면서 다른 문제도 해결하기"에 빠지지 않도록 주의하세요. 여러 개의 관련 없는 문제를 해결하는 PR을 작성하는 것은 리뷰하기가 훨씬 어렵습니다.
|
||||
1. 카멜레온이 되세요. 기존의 디자인 패턴과 구문을 이해하고, 여러분이 추가하는 코드가 기존 코드베이스와 자연스럽게 어우러지도록 해야 합니다. 기존 디자인 패턴이나 사용자 인터페이스와 크게 다른 Pull Request들은 병합되지 않습니다.
|
||||
2. 레이저처럼 집중하세요. Pull Request는 하나의 문제, 오직 하나의 문제만 해결해야 합니다. "이왕 추가하는 김에 다른 문제도 고치자"는 함정에 빠지지 않도록 주의하세요. 여러 개의 관련 없는 문제를 해결하는 한 번에 해결하는 Pull Request들은 검토하기가 훨씬 더 어렵습니다.
|
||||
3. 도움이 되는 경우, 추가한 내용이 어떻게 사용되는지 예제 코드 조각을 추가해보세요.
|
||||
4. PR의 제목은 기여 내용을 요약해야 합니다.
|
||||
5. PR이 이슈를 해결하는 경우, PR 설명에 이슈 번호를 언급하여 연결되도록 해주세요 (이슈를 참조하는 사람들이 작업 중임을 알 수 있도록).
|
||||
4. Pull Request의 제목은 기여 내용을 요약해야 합니다.
|
||||
5. Pull Request가 이슈를 해결하는 경우, Pull Request의 설명에 이슈 번호를 언급하여 연결되도록 해주세요 (이슈를 참조하는 사람들이 작업 중임을 알 수 있도록).
|
||||
6. 진행 중인 작업을 나타내려면 제목에 `[WIP]`를 접두사로 붙여주세요. 이는 중복 작업을 피하고, 병합 준비가 된 PR과 구분할 수 있도록 도움이 됩니다.
|
||||
7. [좋은 이슈를 작성하는 방법](#how-to-write-a-good-issue)에 설명된 대로 텍스트를 구성하고 형식을 지정해보세요.
|
||||
8. 기존 테스트가 통과하는지 확인하세요
|
||||
@@ -374,10 +373,10 @@ class AltDiffusionPipelineOutput(BaseOutput):
|
||||
`RUN_SLOW=1 python -m pytest tests/test_my_new_model.py`.
|
||||
CircleCI는 느린 테스트를 실행하지 않지만, GitHub Actions는 매일 실행합니다!
|
||||
10. 모든 공개 메서드는 마크다운과 잘 작동하는 정보성 docstring을 가져야 합니다. 예시로 [`pipeline_latent_diffusion.py`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion.py)를 참조하세요.
|
||||
11. 레포지토리가 빠르게 성장하고 있기 때문에, 레포지토리에 큰 부담을 주는 파일이 추가되지 않도록 주의해야 합니다. 이미지, 비디오 및 기타 텍스트가 아닌 파일을 포함합니다. 이러한 파일을 배치하기 위해 hf.co 호스팅 `dataset`인 [`hf-internal-testing`](https://huggingface.co/hf-internal-testing) 또는 [huggingface/documentation-images](https://huggingface.co/datasets/huggingface/documentation-images)를 활용하는 것이 우선입니다.
|
||||
11. 리포지토리가 빠르게 성장하고 있기 때문에, 리포지토리에 큰 부담을 주는 파일이 추가되지 않도록 주의해야 합니다. 이미지, 비디오 및 기타 텍스트가 아닌 파일을 포함합니다. 이러한 파일을 배치하기 위해 hf.co 호스팅 `dataset`인 [`hf-internal-testing`](https://huggingface.co/hf-internal-testing) 또는 [huggingface/documentation-images](https://huggingface.co/datasets/huggingface/documentation-images)를 활용하는 것이 우선입니다.
|
||||
외부 기여인 경우, 이미지를 PR에 추가하고 Hugging Face 구성원에게 이미지를 이 데이터셋으로 이동하도록 요청하세요.
|
||||
|
||||
## PR을 열기 위한 방법
|
||||
## PR을 열기 위한 방법 [[how-to-open-a-pr]]
|
||||
|
||||
코드를 작성하기 전에, 이미 누군가가 같은 작업을 하고 있는지 확인하기 위해 기존의 PR이나 이슈를 검색하는 것이 좋습니다. 확실하지 않은 경우, 피드백을 받기 위해 이슈를 열어보는 것이 항상 좋은 아이디어입니다.
|
||||
|
||||
@@ -403,7 +402,7 @@ CircleCI는 느린 테스트를 실행하지 않지만, GitHub Actions는 매일
|
||||
|
||||
`main` 브랜치 위에서 **절대** 작업하지 마세요.
|
||||
|
||||
1. 가상 환경에서 다음 명령을 실행하여 개발 환경을 설정하세요:
|
||||
4. 가상 환경에서 다음 명령을 실행하여 개발 환경을 설정하세요:
|
||||
|
||||
```bash
|
||||
$ pip install -e ".[dev]"
|
||||
@@ -467,7 +466,7 @@ CircleCI는 느린 테스트를 실행하지 않지만, GitHub Actions는 매일
|
||||
|
||||
7. 메인테이너가 변경 사항을 요청하는 것은 괜찮습니다. 핵심 기여자들에게도 일어나는 일입니다! 따라서 변경 사항을 Pull request에서 볼 수 있도록 로컬 브랜치에서 작업하고 변경 사항을 포크에 푸시하면 자동으로 Pull request에 나타납니다.
|
||||
|
||||
### 테스트
|
||||
### 테스트 [[tests]]
|
||||
|
||||
라이브러리 동작과 여러 예제를 테스트하기 위해 포괄적인 테스트 묶음이 포함되어 있습니다. 라이브러리 테스트는 [tests 폴더](https://github.com/huggingface/diffusers/tree/main/tests)에서 찾을 수 있습니다.
|
||||
|
||||
@@ -494,7 +493,7 @@ $ python -m unittest discover -s tests -t . -v
|
||||
$ python -m unittest discover -s examples -t examples -v
|
||||
```
|
||||
|
||||
### upstream(main)과 forked main 동기화하기
|
||||
### upstream(HuggingFace) main과 forked main 동기화하기 [[syncing-forked-main-with-upstream-huggingface-main]]
|
||||
|
||||
upstream 저장소에 불필요한 참조 노트를 추가하고 관련 개발자에게 알림을 보내는 것을 피하기 위해,
|
||||
forked 저장소의 main 브랜치를 동기화할 때 다음 단계를 따르세요:
|
||||
@@ -507,6 +506,6 @@ $ git commit -m '<your message without GitHub references>'
|
||||
$ git push --set-upstream origin your-branch-for-syncing
|
||||
```
|
||||
|
||||
### 스타일 가이드
|
||||
### 스타일 가이드 [[style-guide]]
|
||||
|
||||
Documentation string에 대해서는, 🧨 Diffusers는 [Google 스타일](https://google.github.io/styleguide/pyguide.html)을 따릅니다.
|
||||
|
||||
@@ -1084,7 +1084,7 @@ def main(args):
|
||||
|
||||
# Add noise to the model input according to the noise magnitude at each timestep
|
||||
# (this is the forward diffusion process)
|
||||
noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
|
||||
noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps).to(dtype=weight_dtype)
|
||||
|
||||
# time ids
|
||||
def compute_time_ids(original_size, crops_coords_top_left):
|
||||
@@ -1101,7 +1101,7 @@ def main(args):
|
||||
|
||||
# Predict the noise residual
|
||||
unet_added_conditions = {"time_ids": add_time_ids}
|
||||
prompt_embeds = batch["prompt_embeds"].to(accelerator.device)
|
||||
prompt_embeds = batch["prompt_embeds"].to(accelerator.device, dtype=weight_dtype)
|
||||
pooled_prompt_embeds = batch["pooled_prompt_embeds"].to(accelerator.device)
|
||||
unet_added_conditions.update({"text_embeds": pooled_prompt_embeds})
|
||||
model_pred = unet(
|
||||
|
||||
@@ -109,6 +109,9 @@ import torch
|
||||
model_id = "path-to-your-trained-model"
|
||||
pipe = StableDiffusionPipeline.from_pretrained(model_id,torch_dtype=torch.float16).to("cuda")
|
||||
|
||||
repo_id_embeds = "path-to-your-learned-embeds"
|
||||
pipe.load_textual_inversion(repo_id_embeds)
|
||||
|
||||
prompt = "A <cat-toy> backpack"
|
||||
|
||||
image = pipe(prompt, num_inference_steps=50, guidance_scale=7.5).images[0]
|
||||
|
||||
@@ -86,6 +86,9 @@ TRANSFORMER_SPECIAL_KEYS_REMAP = {
|
||||
"key_layernorm_list": reassign_query_key_layernorm_inplace,
|
||||
"adaln_layer.adaLN_modulations": reassign_adaln_norm_inplace,
|
||||
"embed_tokens": remove_keys_inplace,
|
||||
"freqs_sin": remove_keys_inplace,
|
||||
"freqs_cos": remove_keys_inplace,
|
||||
"position_embedding": remove_keys_inplace,
|
||||
}
|
||||
|
||||
VAE_KEYS_RENAME_DICT = {
|
||||
@@ -123,11 +126,21 @@ def update_state_dict_inplace(state_dict: Dict[str, Any], old_key: str, new_key:
|
||||
state_dict[new_key] = state_dict.pop(old_key)
|
||||
|
||||
|
||||
def convert_transformer(ckpt_path: str):
|
||||
def convert_transformer(
|
||||
ckpt_path: str,
|
||||
num_layers: int,
|
||||
num_attention_heads: int,
|
||||
use_rotary_positional_embeddings: bool,
|
||||
dtype: torch.dtype,
|
||||
):
|
||||
PREFIX_KEY = "model.diffusion_model."
|
||||
|
||||
original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
|
||||
transformer = CogVideoXTransformer3DModel()
|
||||
transformer = CogVideoXTransformer3DModel(
|
||||
num_layers=num_layers,
|
||||
num_attention_heads=num_attention_heads,
|
||||
use_rotary_positional_embeddings=use_rotary_positional_embeddings,
|
||||
).to(dtype=dtype)
|
||||
|
||||
for key in list(original_state_dict.keys()):
|
||||
new_key = key[len(PREFIX_KEY) :]
|
||||
@@ -145,9 +158,9 @@ def convert_transformer(ckpt_path: str):
|
||||
return transformer
|
||||
|
||||
|
||||
def convert_vae(ckpt_path: str):
|
||||
def convert_vae(ckpt_path: str, scaling_factor: float, dtype: torch.dtype):
|
||||
original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
|
||||
vae = AutoencoderKLCogVideoX()
|
||||
vae = AutoencoderKLCogVideoX(scaling_factor=scaling_factor).to(dtype=dtype)
|
||||
|
||||
for key in list(original_state_dict.keys()):
|
||||
new_key = key[:]
|
||||
@@ -172,13 +185,26 @@ def get_args():
|
||||
)
|
||||
parser.add_argument("--vae_ckpt_path", type=str, default=None, help="Path to original vae checkpoint")
|
||||
parser.add_argument("--output_path", type=str, required=True, help="Path where converted model should be saved")
|
||||
parser.add_argument("--fp16", action="store_true", default=True, help="Whether to save the model weights in fp16")
|
||||
parser.add_argument("--fp16", action="store_true", default=False, help="Whether to save the model weights in fp16")
|
||||
parser.add_argument("--bf16", action="store_true", default=False, help="Whether to save the model weights in bf16")
|
||||
parser.add_argument(
|
||||
"--push_to_hub", action="store_true", default=False, help="Whether to push to HF Hub after saving"
|
||||
)
|
||||
parser.add_argument(
|
||||
"--text_encoder_cache_dir", type=str, default=None, help="Path to text encoder cache directory"
|
||||
)
|
||||
# For CogVideoX-2B, num_layers is 30. For 5B, it is 42
|
||||
parser.add_argument("--num_layers", type=int, default=30, help="Number of transformer blocks")
|
||||
# For CogVideoX-2B, num_attention_heads is 30. For 5B, it is 48
|
||||
parser.add_argument("--num_attention_heads", type=int, default=30, help="Number of attention heads")
|
||||
# For CogVideoX-2B, use_rotary_positional_embeddings is False. For 5B, it is True
|
||||
parser.add_argument(
|
||||
"--use_rotary_positional_embeddings", action="store_true", default=False, help="Whether to use RoPE or not"
|
||||
)
|
||||
# For CogVideoX-2B, scaling_factor is 1.15258426. For 5B, it is 0.7
|
||||
parser.add_argument("--scaling_factor", type=float, default=1.15258426, help="Scaling factor in the VAE")
|
||||
# For CogVideoX-2B, snr_shift_scale is 3.0. For 5B, it is 1.0
|
||||
parser.add_argument("--snr_shift_scale", type=float, default=3.0, help="Scaling factor in the VAE")
|
||||
return parser.parse_args()
|
||||
|
||||
|
||||
@@ -188,18 +214,33 @@ if __name__ == "__main__":
|
||||
transformer = None
|
||||
vae = None
|
||||
|
||||
if args.fp16 and args.bf16:
|
||||
raise ValueError("You cannot pass both --fp16 and --bf16 at the same time.")
|
||||
|
||||
dtype = torch.float16 if args.fp16 else torch.bfloat16 if args.bf16 else torch.float32
|
||||
|
||||
if args.transformer_ckpt_path is not None:
|
||||
transformer = convert_transformer(args.transformer_ckpt_path)
|
||||
transformer = convert_transformer(
|
||||
args.transformer_ckpt_path,
|
||||
args.num_layers,
|
||||
args.num_attention_heads,
|
||||
args.use_rotary_positional_embeddings,
|
||||
dtype,
|
||||
)
|
||||
if args.vae_ckpt_path is not None:
|
||||
vae = convert_vae(args.vae_ckpt_path)
|
||||
vae = convert_vae(args.vae_ckpt_path, args.scaling_factor, dtype)
|
||||
|
||||
text_encoder_id = "google/t5-v1_1-xxl"
|
||||
tokenizer = T5Tokenizer.from_pretrained(text_encoder_id, model_max_length=TOKENIZER_MAX_LENGTH)
|
||||
text_encoder = T5EncoderModel.from_pretrained(text_encoder_id, cache_dir=args.text_encoder_cache_dir)
|
||||
|
||||
# Apparently, the conversion does not work any more without this :shrug:
|
||||
for param in text_encoder.parameters():
|
||||
param.data = param.data.contiguous()
|
||||
|
||||
scheduler = CogVideoXDDIMScheduler.from_config(
|
||||
{
|
||||
"snr_shift_scale": 3.0,
|
||||
"snr_shift_scale": args.snr_shift_scale,
|
||||
"beta_end": 0.012,
|
||||
"beta_schedule": "scaled_linear",
|
||||
"beta_start": 0.00085,
|
||||
@@ -208,7 +249,7 @@ if __name__ == "__main__":
|
||||
"prediction_type": "v_prediction",
|
||||
"rescale_betas_zero_snr": True,
|
||||
"set_alpha_to_one": True,
|
||||
"timestep_spacing": "linspace",
|
||||
"timestep_spacing": "trailing",
|
||||
}
|
||||
)
|
||||
|
||||
@@ -218,5 +259,10 @@ if __name__ == "__main__":
|
||||
|
||||
if args.fp16:
|
||||
pipe = pipe.to(dtype=torch.float16)
|
||||
if args.bf16:
|
||||
pipe = pipe.to(dtype=torch.bfloat16)
|
||||
|
||||
# We don't use variant here because the model must be run in fp16 (2B) or bf16 (5B). It would be weird
|
||||
# for users to specify variant when the default is not fp32 and they want to run with the correct default (which
|
||||
# is either fp16/bf16 here).
|
||||
pipe.save_pretrained(args.output_path, safe_serialization=True, push_to_hub=args.push_to_hub)
|
||||
|
||||
@@ -346,6 +346,7 @@ else:
|
||||
"StableDiffusionXLAdapterPipeline",
|
||||
"StableDiffusionXLControlNetImg2ImgPipeline",
|
||||
"StableDiffusionXLControlNetInpaintPipeline",
|
||||
"StableDiffusionXLControlNetPAGImg2ImgPipeline",
|
||||
"StableDiffusionXLControlNetPAGPipeline",
|
||||
"StableDiffusionXLControlNetPipeline",
|
||||
"StableDiffusionXLControlNetXSPipeline",
|
||||
@@ -787,6 +788,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
|
||||
StableDiffusionXLAdapterPipeline,
|
||||
StableDiffusionXLControlNetImg2ImgPipeline,
|
||||
StableDiffusionXLControlNetInpaintPipeline,
|
||||
StableDiffusionXLControlNetPAGImg2ImgPipeline,
|
||||
StableDiffusionXLControlNetPAGPipeline,
|
||||
StableDiffusionXLControlNetPipeline,
|
||||
StableDiffusionXLControlNetXSPipeline,
|
||||
|
||||
@@ -280,7 +280,9 @@ class StableDiffusionLoraLoaderMixin(LoraBaseMixin):
|
||||
A standard state dict containing the lora layer parameters. The key should be prefixed with an
|
||||
additional `text_encoder` to distinguish between unet lora layers.
|
||||
network_alphas (`Dict[str, float]`):
|
||||
See `LoRALinearLayer` for more details.
|
||||
The value of the network alpha used for stable learning and preventing underflow. This value has the
|
||||
same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
|
||||
link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
|
||||
text_encoder (`CLIPTextModel`):
|
||||
The text encoder model to load the LoRA layers into.
|
||||
prefix (`str`):
|
||||
@@ -753,7 +755,9 @@ class StableDiffusionXLLoraLoaderMixin(LoraBaseMixin):
|
||||
A standard state dict containing the lora layer parameters. The key should be prefixed with an
|
||||
additional `text_encoder` to distinguish between unet lora layers.
|
||||
network_alphas (`Dict[str, float]`):
|
||||
See `LoRALinearLayer` for more details.
|
||||
The value of the network alpha used for stable learning and preventing underflow. This value has the
|
||||
same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
|
||||
link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
|
||||
text_encoder (`CLIPTextModel`):
|
||||
The text encoder model to load the LoRA layers into.
|
||||
prefix (`str`):
|
||||
@@ -1249,7 +1253,9 @@ class SD3LoraLoaderMixin(LoraBaseMixin):
|
||||
A standard state dict containing the lora layer parameters. The key should be prefixed with an
|
||||
additional `text_encoder` to distinguish between unet lora layers.
|
||||
network_alphas (`Dict[str, float]`):
|
||||
See `LoRALinearLayer` for more details.
|
||||
The value of the network alpha used for stable learning and preventing underflow. This value has the
|
||||
same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
|
||||
link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
|
||||
text_encoder (`CLIPTextModel`):
|
||||
The text encoder model to load the LoRA layers into.
|
||||
prefix (`str`):
|
||||
@@ -1489,10 +1495,10 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
|
||||
|
||||
@classmethod
|
||||
@validate_hf_hub_args
|
||||
# Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.lora_state_dict
|
||||
def lora_state_dict(
|
||||
cls,
|
||||
pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
|
||||
return_alphas: bool = False,
|
||||
**kwargs,
|
||||
):
|
||||
r"""
|
||||
@@ -1577,7 +1583,26 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
|
||||
allow_pickle=allow_pickle,
|
||||
)
|
||||
|
||||
return state_dict
|
||||
# For state dicts like
|
||||
# https://huggingface.co/TheLastBen/Jon_Snow_Flux_LoRA
|
||||
keys = list(state_dict.keys())
|
||||
network_alphas = {}
|
||||
for k in keys:
|
||||
if "alpha" in k:
|
||||
alpha_value = state_dict.get(k)
|
||||
if (torch.is_tensor(alpha_value) and torch.is_floating_point(alpha_value)) or isinstance(
|
||||
alpha_value, float
|
||||
):
|
||||
network_alphas[k] = state_dict.pop(k)
|
||||
else:
|
||||
raise ValueError(
|
||||
f"The alpha key ({k}) seems to be incorrect. If you think this error is unexpected, please open as issue."
|
||||
)
|
||||
|
||||
if return_alphas:
|
||||
return state_dict, network_alphas
|
||||
else:
|
||||
return state_dict
|
||||
|
||||
def load_lora_weights(
|
||||
self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], adapter_name=None, **kwargs
|
||||
@@ -1611,7 +1636,9 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
|
||||
pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()
|
||||
|
||||
# First, ensure that the checkpoint is a compatible one and can be successfully loaded.
|
||||
state_dict = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
|
||||
state_dict, network_alphas = self.lora_state_dict(
|
||||
pretrained_model_name_or_path_or_dict, return_alphas=True, **kwargs
|
||||
)
|
||||
|
||||
is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
|
||||
if not is_correct_format:
|
||||
@@ -1619,6 +1646,7 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
|
||||
|
||||
self.load_lora_into_transformer(
|
||||
state_dict,
|
||||
network_alphas=network_alphas,
|
||||
transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
|
||||
adapter_name=adapter_name,
|
||||
_pipeline=self,
|
||||
@@ -1628,7 +1656,7 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
|
||||
if len(text_encoder_state_dict) > 0:
|
||||
self.load_lora_into_text_encoder(
|
||||
text_encoder_state_dict,
|
||||
network_alphas=None,
|
||||
network_alphas=network_alphas,
|
||||
text_encoder=self.text_encoder,
|
||||
prefix="text_encoder",
|
||||
lora_scale=self.lora_scale,
|
||||
@@ -1637,8 +1665,7 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
|
||||
)
|
||||
|
||||
@classmethod
|
||||
# Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer
|
||||
def load_lora_into_transformer(cls, state_dict, transformer, adapter_name=None, _pipeline=None):
|
||||
def load_lora_into_transformer(cls, state_dict, network_alphas, transformer, adapter_name=None, _pipeline=None):
|
||||
"""
|
||||
This will load the LoRA layers specified in `state_dict` into `transformer`.
|
||||
|
||||
@@ -1647,6 +1674,10 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
|
||||
A standard state dict containing the lora layer parameters. The keys can either be indexed directly
|
||||
into the unet or prefixed with an additional `unet` which can be used to distinguish between text
|
||||
encoder lora layers.
|
||||
network_alphas (`Dict[str, float]`):
|
||||
The value of the network alpha used for stable learning and preventing underflow. This value has the
|
||||
same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
|
||||
link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
|
||||
transformer (`SD3Transformer2DModel`):
|
||||
The Transformer model to load the LoRA layers into.
|
||||
adapter_name (`str`, *optional*):
|
||||
@@ -1678,7 +1709,12 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
|
||||
if "lora_B" in key:
|
||||
rank[key] = val.shape[1]
|
||||
|
||||
lora_config_kwargs = get_peft_kwargs(rank, network_alpha_dict=None, peft_state_dict=state_dict)
|
||||
if network_alphas is not None and len(network_alphas) >= 1:
|
||||
prefix = cls.transformer_name
|
||||
alpha_keys = [k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix]
|
||||
network_alphas = {k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys}
|
||||
|
||||
lora_config_kwargs = get_peft_kwargs(rank, network_alpha_dict=network_alphas, peft_state_dict=state_dict)
|
||||
if "use_dora" in lora_config_kwargs:
|
||||
if lora_config_kwargs["use_dora"] and is_peft_version("<", "0.9.0"):
|
||||
raise ValueError(
|
||||
@@ -1735,7 +1771,9 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
|
||||
A standard state dict containing the lora layer parameters. The key should be prefixed with an
|
||||
additional `text_encoder` to distinguish between unet lora layers.
|
||||
network_alphas (`Dict[str, float]`):
|
||||
See `LoRALinearLayer` for more details.
|
||||
The value of the network alpha used for stable learning and preventing underflow. This value has the
|
||||
same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
|
||||
link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
|
||||
text_encoder (`CLIPTextModel`):
|
||||
The text encoder model to load the LoRA layers into.
|
||||
prefix (`str`):
|
||||
@@ -1968,7 +2006,9 @@ class AmusedLoraLoaderMixin(StableDiffusionLoraLoaderMixin):
|
||||
into the unet or prefixed with an additional `unet` which can be used to distinguish between text
|
||||
encoder lora layers.
|
||||
network_alphas (`Dict[str, float]`):
|
||||
See `LoRALinearLayer` for more details.
|
||||
The value of the network alpha used for stable learning and preventing underflow. This value has the
|
||||
same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
|
||||
link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
|
||||
unet (`UNet2DConditionModel`):
|
||||
The UNet model to load the LoRA layers into.
|
||||
adapter_name (`str`, *optional*):
|
||||
@@ -2061,7 +2101,9 @@ class AmusedLoraLoaderMixin(StableDiffusionLoraLoaderMixin):
|
||||
A standard state dict containing the lora layer parameters. The key should be prefixed with an
|
||||
additional `text_encoder` to distinguish between unet lora layers.
|
||||
network_alphas (`Dict[str, float]`):
|
||||
See `LoRALinearLayer` for more details.
|
||||
The value of the network alpha used for stable learning and preventing underflow. This value has the
|
||||
same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
|
||||
link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
|
||||
text_encoder (`CLIPTextModel`):
|
||||
The text encoder model to load the LoRA layers into.
|
||||
prefix (`str`):
|
||||
|
||||
@@ -23,6 +23,7 @@ from packaging import version
|
||||
from ..utils import deprecate, is_transformers_available, logging
|
||||
from .single_file_utils import (
|
||||
SingleFileComponentError,
|
||||
_is_legacy_scheduler_kwargs,
|
||||
_is_model_weights_in_cached_folder,
|
||||
_legacy_load_clip_tokenizer,
|
||||
_legacy_load_safety_checker,
|
||||
@@ -42,7 +43,6 @@ logger = logging.get_logger(__name__)
|
||||
# Legacy behaviour. `from_single_file` does not load the safety checker unless explicitly provided
|
||||
SINGLE_FILE_OPTIONAL_COMPONENTS = ["safety_checker"]
|
||||
|
||||
|
||||
if is_transformers_available():
|
||||
import transformers
|
||||
from transformers import PreTrainedModel, PreTrainedTokenizer
|
||||
@@ -135,7 +135,7 @@ def load_single_file_sub_model(
|
||||
class_obj, checkpoint=checkpoint, config=cached_model_config_path, local_files_only=local_files_only
|
||||
)
|
||||
|
||||
elif is_diffusers_scheduler and is_legacy_loading:
|
||||
elif is_diffusers_scheduler and (is_legacy_loading or _is_legacy_scheduler_kwargs(kwargs)):
|
||||
loaded_sub_model = _legacy_load_scheduler(
|
||||
class_obj, checkpoint=checkpoint, component_name=name, original_config=original_config, **kwargs
|
||||
)
|
||||
|
||||
@@ -79,7 +79,10 @@ CHECKPOINT_KEY_NAMES = {
|
||||
"animatediff_sdxl_beta": "up_blocks.2.motion_modules.0.temporal_transformer.norm.weight",
|
||||
"animatediff_scribble": "controlnet_cond_embedding.conv_in.weight",
|
||||
"animatediff_rgb": "controlnet_cond_embedding.weight",
|
||||
"flux": "double_blocks.0.img_attn.norm.key_norm.scale",
|
||||
"flux": [
|
||||
"double_blocks.0.img_attn.norm.key_norm.scale",
|
||||
"model.diffusion_model.double_blocks.0.img_attn.norm.key_norm.scale",
|
||||
],
|
||||
}
|
||||
|
||||
DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
|
||||
@@ -258,7 +261,7 @@ SCHEDULER_DEFAULT_CONFIG = {
|
||||
"timestep_spacing": "leading",
|
||||
}
|
||||
|
||||
LDM_VAE_KEY = "first_stage_model."
|
||||
LDM_VAE_KEYS = ["first_stage_model.", "vae."]
|
||||
LDM_VAE_DEFAULT_SCALING_FACTOR = 0.18215
|
||||
PLAYGROUND_VAE_SCALING_FACTOR = 0.5
|
||||
LDM_UNET_KEY = "model.diffusion_model."
|
||||
@@ -267,8 +270,8 @@ LDM_CLIP_PREFIX_TO_REMOVE = [
|
||||
"cond_stage_model.transformer.",
|
||||
"conditioner.embedders.0.transformer.",
|
||||
]
|
||||
OPEN_CLIP_PREFIX = "conditioner.embedders.0.model."
|
||||
LDM_OPEN_CLIP_TEXT_PROJECTION_DIM = 1024
|
||||
SCHEDULER_LEGACY_KWARGS = ["prediction_type", "scheduler_type"]
|
||||
|
||||
VALID_URL_PREFIXES = ["https://huggingface.co/", "huggingface.co/", "hf.co/", "https://hf.co/"]
|
||||
|
||||
@@ -318,6 +321,10 @@ def _is_model_weights_in_cached_folder(cached_folder, name):
|
||||
return weights_exist
|
||||
|
||||
|
||||
def _is_legacy_scheduler_kwargs(kwargs):
|
||||
return any(k in SCHEDULER_LEGACY_KWARGS for k in kwargs.keys())
|
||||
|
||||
|
||||
def load_single_file_checkpoint(
|
||||
pretrained_model_link_or_path,
|
||||
force_download=False,
|
||||
@@ -449,6 +456,8 @@ def infer_diffusers_model_type(checkpoint):
|
||||
):
|
||||
if CHECKPOINT_KEY_NAMES["v2"] in checkpoint and checkpoint[CHECKPOINT_KEY_NAMES["v2"]].shape[-1] == 1024:
|
||||
model_type = "inpainting_v2"
|
||||
elif CHECKPOINT_KEY_NAMES["xl_base"] in checkpoint:
|
||||
model_type = "xl_inpaint"
|
||||
else:
|
||||
model_type = "inpainting"
|
||||
|
||||
@@ -516,8 +525,10 @@ def infer_diffusers_model_type(checkpoint):
|
||||
else:
|
||||
model_type = "animatediff_v3"
|
||||
|
||||
elif CHECKPOINT_KEY_NAMES["flux"] in checkpoint:
|
||||
if "guidance_in.in_layer.bias" in checkpoint:
|
||||
elif any(key in checkpoint for key in CHECKPOINT_KEY_NAMES["flux"]):
|
||||
if any(
|
||||
g in checkpoint for g in ["guidance_in.in_layer.bias", "model.diffusion_model.guidance_in.in_layer.bias"]
|
||||
):
|
||||
model_type = "flux-dev"
|
||||
else:
|
||||
model_type = "flux-schnell"
|
||||
@@ -1176,7 +1187,11 @@ def convert_ldm_vae_checkpoint(checkpoint, config):
|
||||
# remove the LDM_VAE_KEY prefix from the ldm checkpoint keys so that it is easier to map them to diffusers keys
|
||||
vae_state_dict = {}
|
||||
keys = list(checkpoint.keys())
|
||||
vae_key = LDM_VAE_KEY if any(k.startswith(LDM_VAE_KEY) for k in keys) else ""
|
||||
vae_key = ""
|
||||
for ldm_vae_key in LDM_VAE_KEYS:
|
||||
if any(k.startswith(ldm_vae_key) for k in keys):
|
||||
vae_key = ldm_vae_key
|
||||
|
||||
for key in keys:
|
||||
if key.startswith(vae_key):
|
||||
vae_state_dict[key.replace(vae_key, "")] = checkpoint.get(key)
|
||||
@@ -1477,14 +1492,22 @@ def _legacy_load_scheduler(
|
||||
|
||||
if scheduler_type is not None:
|
||||
deprecation_message = (
|
||||
"Please pass an instance of a Scheduler object directly to the `scheduler` argument in `from_single_file`."
|
||||
"Please pass an instance of a Scheduler object directly to the `scheduler` argument in `from_single_file`\n\n"
|
||||
"Example:\n\n"
|
||||
"from diffusers import StableDiffusionPipeline, DDIMScheduler\n\n"
|
||||
"scheduler = DDIMScheduler()\n"
|
||||
"pipe = StableDiffusionPipeline.from_single_file(<checkpoint path>, scheduler=scheduler)\n"
|
||||
)
|
||||
deprecate("scheduler_type", "1.0.0", deprecation_message)
|
||||
|
||||
if prediction_type is not None:
|
||||
deprecation_message = (
|
||||
"Please configure an instance of a Scheduler with the appropriate `prediction_type` "
|
||||
"and pass the object directly to the `scheduler` argument in `from_single_file`."
|
||||
"Please configure an instance of a Scheduler with the appropriate `prediction_type` and "
|
||||
"pass the object directly to the `scheduler` argument in `from_single_file`.\n\n"
|
||||
"Example:\n\n"
|
||||
"from diffusers import StableDiffusionPipeline, DDIMScheduler\n\n"
|
||||
'scheduler = DDIMScheduler(prediction_type="v_prediction")\n'
|
||||
"pipe = StableDiffusionPipeline.from_single_file(<checkpoint path>, scheduler=scheduler)\n"
|
||||
)
|
||||
deprecate("prediction_type", "1.0.0", deprecation_message)
|
||||
|
||||
@@ -1881,6 +1904,10 @@ def convert_animatediff_checkpoint_to_diffusers(checkpoint, **kwargs):
|
||||
|
||||
def convert_flux_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
|
||||
converted_state_dict = {}
|
||||
keys = list(checkpoint.keys())
|
||||
for k in keys:
|
||||
if "model.diffusion_model." in k:
|
||||
checkpoint[k.replace("model.diffusion_model.", "")] = checkpoint.pop(k)
|
||||
|
||||
num_layers = list(set(int(k.split(".", 2)[1]) for k in checkpoint if "double_blocks." in k))[-1] + 1 # noqa: C401
|
||||
num_single_layers = list(set(int(k.split(".", 2)[1]) for k in checkpoint if "single_blocks." in k))[-1] + 1 # noqa: C401
|
||||
|
||||
@@ -1695,81 +1695,6 @@ class FusedAuraFlowAttnProcessor2_0:
|
||||
return hidden_states
|
||||
|
||||
|
||||
# YiYi to-do: refactor rope related functions/classes
|
||||
def apply_rope(xq, xk, freqs_cis):
|
||||
xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2)
|
||||
xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2)
|
||||
xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
|
||||
xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
|
||||
return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk)
|
||||
|
||||
|
||||
class FluxSingleAttnProcessor2_0:
|
||||
r"""
|
||||
Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
|
||||
"""
|
||||
|
||||
def __init__(self):
|
||||
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.")
|
||||
|
||||
def __call__(
|
||||
self,
|
||||
attn: Attention,
|
||||
hidden_states: torch.Tensor,
|
||||
encoder_hidden_states: Optional[torch.Tensor] = None,
|
||||
attention_mask: Optional[torch.FloatTensor] = None,
|
||||
image_rotary_emb: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
input_ndim = hidden_states.ndim
|
||||
|
||||
if input_ndim == 4:
|
||||
batch_size, channel, height, width = hidden_states.shape
|
||||
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
|
||||
|
||||
batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
||||
|
||||
query = attn.to_q(hidden_states)
|
||||
if encoder_hidden_states is None:
|
||||
encoder_hidden_states = hidden_states
|
||||
|
||||
key = attn.to_k(encoder_hidden_states)
|
||||
value = attn.to_v(encoder_hidden_states)
|
||||
|
||||
inner_dim = key.shape[-1]
|
||||
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)
|
||||
|
||||
if attn.norm_q is not None:
|
||||
query = attn.norm_q(query)
|
||||
if attn.norm_k is not None:
|
||||
key = attn.norm_k(key)
|
||||
|
||||
# Apply RoPE if needed
|
||||
if image_rotary_emb is not None:
|
||||
# YiYi to-do: update uising apply_rotary_emb
|
||||
# from ..embeddings import apply_rotary_emb
|
||||
# query = apply_rotary_emb(query, image_rotary_emb)
|
||||
# key = apply_rotary_emb(key, image_rotary_emb)
|
||||
query, key = apply_rope(query, key, image_rotary_emb)
|
||||
|
||||
# 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, 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)
|
||||
|
||||
if input_ndim == 4:
|
||||
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
|
||||
|
||||
return hidden_states
|
||||
|
||||
|
||||
class FluxAttnProcessor2_0:
|
||||
"""Attention processor used typically in processing the SD3-like self-attention projections."""
|
||||
|
||||
@@ -1785,16 +1710,7 @@ class FluxAttnProcessor2_0:
|
||||
attention_mask: Optional[torch.FloatTensor] = None,
|
||||
image_rotary_emb: Optional[torch.Tensor] = None,
|
||||
) -> torch.FloatTensor:
|
||||
input_ndim = hidden_states.ndim
|
||||
if input_ndim == 4:
|
||||
batch_size, channel, height, width = hidden_states.shape
|
||||
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
|
||||
context_input_ndim = encoder_hidden_states.ndim
|
||||
if context_input_ndim == 4:
|
||||
batch_size, channel, height, width = encoder_hidden_states.shape
|
||||
encoder_hidden_states = encoder_hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
|
||||
|
||||
batch_size = encoder_hidden_states.shape[0]
|
||||
batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
||||
|
||||
# `sample` projections.
|
||||
query = attn.to_q(hidden_states)
|
||||
@@ -1813,58 +1729,293 @@ class FluxAttnProcessor2_0:
|
||||
if attn.norm_k is not None:
|
||||
key = attn.norm_k(key)
|
||||
|
||||
# `context` projections.
|
||||
encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
|
||||
encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
|
||||
encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
|
||||
# the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
|
||||
if encoder_hidden_states is not None:
|
||||
# `context` projections.
|
||||
encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
|
||||
encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
|
||||
encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
|
||||
|
||||
encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
|
||||
batch_size, -1, attn.heads, head_dim
|
||||
).transpose(1, 2)
|
||||
encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
|
||||
batch_size, -1, attn.heads, head_dim
|
||||
).transpose(1, 2)
|
||||
encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
|
||||
batch_size, -1, attn.heads, head_dim
|
||||
).transpose(1, 2)
|
||||
encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
|
||||
batch_size, -1, attn.heads, head_dim
|
||||
).transpose(1, 2)
|
||||
encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
|
||||
batch_size, -1, attn.heads, head_dim
|
||||
).transpose(1, 2)
|
||||
encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
|
||||
batch_size, -1, attn.heads, head_dim
|
||||
).transpose(1, 2)
|
||||
|
||||
if attn.norm_added_q is not None:
|
||||
encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
|
||||
if attn.norm_added_k is not None:
|
||||
encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
|
||||
if attn.norm_added_q is not None:
|
||||
encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
|
||||
if attn.norm_added_k is not None:
|
||||
encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
|
||||
|
||||
# attention
|
||||
query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
|
||||
key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
|
||||
value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
|
||||
# attention
|
||||
query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
|
||||
key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
|
||||
value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
|
||||
|
||||
if image_rotary_emb is not None:
|
||||
# YiYi to-do: update uising apply_rotary_emb
|
||||
# from ..embeddings import apply_rotary_emb
|
||||
# query = apply_rotary_emb(query, image_rotary_emb)
|
||||
# key = apply_rotary_emb(key, image_rotary_emb)
|
||||
query, key = apply_rope(query, key, image_rotary_emb)
|
||||
from .embeddings import apply_rotary_emb
|
||||
|
||||
query = apply_rotary_emb(query, image_rotary_emb)
|
||||
key = apply_rotary_emb(key, image_rotary_emb)
|
||||
|
||||
hidden_states = F.scaled_dot_product_attention(query, key, value, 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)
|
||||
|
||||
encoder_hidden_states, hidden_states = (
|
||||
hidden_states[:, : encoder_hidden_states.shape[1]],
|
||||
hidden_states[:, encoder_hidden_states.shape[1] :],
|
||||
if encoder_hidden_states is not None:
|
||||
encoder_hidden_states, hidden_states = (
|
||||
hidden_states[:, : encoder_hidden_states.shape[1]],
|
||||
hidden_states[:, encoder_hidden_states.shape[1] :],
|
||||
)
|
||||
|
||||
# linear proj
|
||||
hidden_states = attn.to_out[0](hidden_states)
|
||||
# dropout
|
||||
hidden_states = attn.to_out[1](hidden_states)
|
||||
encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
|
||||
|
||||
return hidden_states, encoder_hidden_states
|
||||
else:
|
||||
return hidden_states
|
||||
|
||||
|
||||
class FusedFluxAttnProcessor2_0:
|
||||
"""Attention processor used typically in processing the SD3-like self-attention projections."""
|
||||
|
||||
def __init__(self):
|
||||
if not hasattr(F, "scaled_dot_product_attention"):
|
||||
raise ImportError(
|
||||
"FusedFluxAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
|
||||
)
|
||||
|
||||
def __call__(
|
||||
self,
|
||||
attn: Attention,
|
||||
hidden_states: torch.FloatTensor,
|
||||
encoder_hidden_states: torch.FloatTensor = None,
|
||||
attention_mask: Optional[torch.FloatTensor] = None,
|
||||
image_rotary_emb: Optional[torch.Tensor] = None,
|
||||
) -> torch.FloatTensor:
|
||||
batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
||||
|
||||
# `sample` projections.
|
||||
qkv = attn.to_qkv(hidden_states)
|
||||
split_size = qkv.shape[-1] // 3
|
||||
query, key, value = torch.split(qkv, split_size, dim=-1)
|
||||
|
||||
inner_dim = key.shape[-1]
|
||||
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)
|
||||
|
||||
if attn.norm_q is not None:
|
||||
query = attn.norm_q(query)
|
||||
if attn.norm_k is not None:
|
||||
key = attn.norm_k(key)
|
||||
|
||||
# the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
|
||||
# `context` projections.
|
||||
if encoder_hidden_states is not None:
|
||||
encoder_qkv = attn.to_added_qkv(encoder_hidden_states)
|
||||
split_size = encoder_qkv.shape[-1] // 3
|
||||
(
|
||||
encoder_hidden_states_query_proj,
|
||||
encoder_hidden_states_key_proj,
|
||||
encoder_hidden_states_value_proj,
|
||||
) = torch.split(encoder_qkv, split_size, dim=-1)
|
||||
|
||||
encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
|
||||
batch_size, -1, attn.heads, head_dim
|
||||
).transpose(1, 2)
|
||||
encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
|
||||
batch_size, -1, attn.heads, head_dim
|
||||
).transpose(1, 2)
|
||||
encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
|
||||
batch_size, -1, attn.heads, head_dim
|
||||
).transpose(1, 2)
|
||||
|
||||
if attn.norm_added_q is not None:
|
||||
encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
|
||||
if attn.norm_added_k is not None:
|
||||
encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
|
||||
|
||||
# attention
|
||||
query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
|
||||
key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
|
||||
value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
|
||||
|
||||
if image_rotary_emb is not None:
|
||||
from .embeddings import apply_rotary_emb
|
||||
|
||||
query = apply_rotary_emb(query, image_rotary_emb)
|
||||
key = apply_rotary_emb(key, image_rotary_emb)
|
||||
|
||||
hidden_states = F.scaled_dot_product_attention(query, key, value, 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)
|
||||
|
||||
if encoder_hidden_states is not None:
|
||||
encoder_hidden_states, hidden_states = (
|
||||
hidden_states[:, : encoder_hidden_states.shape[1]],
|
||||
hidden_states[:, encoder_hidden_states.shape[1] :],
|
||||
)
|
||||
|
||||
# linear proj
|
||||
hidden_states = attn.to_out[0](hidden_states)
|
||||
# dropout
|
||||
hidden_states = attn.to_out[1](hidden_states)
|
||||
encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
|
||||
|
||||
return hidden_states, encoder_hidden_states
|
||||
else:
|
||||
return hidden_states
|
||||
|
||||
|
||||
class CogVideoXAttnProcessor2_0:
|
||||
r"""
|
||||
Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on
|
||||
query and key vectors, but does not include spatial normalization.
|
||||
"""
|
||||
|
||||
def __init__(self):
|
||||
if not hasattr(F, "scaled_dot_product_attention"):
|
||||
raise ImportError("CogVideoXAttnProcessor requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
|
||||
|
||||
def __call__(
|
||||
self,
|
||||
attn: Attention,
|
||||
hidden_states: torch.Tensor,
|
||||
encoder_hidden_states: torch.Tensor,
|
||||
attention_mask: Optional[torch.Tensor] = None,
|
||||
image_rotary_emb: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
text_seq_length = encoder_hidden_states.size(1)
|
||||
|
||||
hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
|
||||
|
||||
batch_size, sequence_length, _ = (
|
||||
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
||||
)
|
||||
|
||||
if attention_mask is not None:
|
||||
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
||||
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
|
||||
|
||||
query = attn.to_q(hidden_states)
|
||||
key = attn.to_k(hidden_states)
|
||||
value = attn.to_v(hidden_states)
|
||||
|
||||
inner_dim = key.shape[-1]
|
||||
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)
|
||||
|
||||
if attn.norm_q is not None:
|
||||
query = attn.norm_q(query)
|
||||
if attn.norm_k is not None:
|
||||
key = attn.norm_k(key)
|
||||
|
||||
# Apply RoPE if needed
|
||||
if image_rotary_emb is not None:
|
||||
from .embeddings import apply_rotary_emb
|
||||
|
||||
query[:, :, text_seq_length:] = apply_rotary_emb(query[:, :, text_seq_length:], image_rotary_emb)
|
||||
if not attn.is_cross_attention:
|
||||
key[:, :, text_seq_length:] = apply_rotary_emb(key[:, :, text_seq_length:], image_rotary_emb)
|
||||
|
||||
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)
|
||||
|
||||
# linear proj
|
||||
hidden_states = attn.to_out[0](hidden_states)
|
||||
# dropout
|
||||
hidden_states = attn.to_out[1](hidden_states)
|
||||
encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
|
||||
|
||||
if input_ndim == 4:
|
||||
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
|
||||
if context_input_ndim == 4:
|
||||
encoder_hidden_states = encoder_hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
|
||||
encoder_hidden_states, hidden_states = hidden_states.split(
|
||||
[text_seq_length, hidden_states.size(1) - text_seq_length], dim=1
|
||||
)
|
||||
return hidden_states, encoder_hidden_states
|
||||
|
||||
|
||||
class FusedCogVideoXAttnProcessor2_0:
|
||||
r"""
|
||||
Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on
|
||||
query and key vectors, but does not include spatial normalization.
|
||||
"""
|
||||
|
||||
def __init__(self):
|
||||
if not hasattr(F, "scaled_dot_product_attention"):
|
||||
raise ImportError("CogVideoXAttnProcessor requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
|
||||
|
||||
def __call__(
|
||||
self,
|
||||
attn: Attention,
|
||||
hidden_states: torch.Tensor,
|
||||
encoder_hidden_states: torch.Tensor,
|
||||
attention_mask: Optional[torch.Tensor] = None,
|
||||
image_rotary_emb: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
text_seq_length = encoder_hidden_states.size(1)
|
||||
|
||||
hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
|
||||
|
||||
batch_size, sequence_length, _ = (
|
||||
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
||||
)
|
||||
|
||||
if attention_mask is not None:
|
||||
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
||||
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
|
||||
|
||||
qkv = attn.to_qkv(hidden_states)
|
||||
split_size = qkv.shape[-1] // 3
|
||||
query, key, value = torch.split(qkv, split_size, dim=-1)
|
||||
|
||||
inner_dim = key.shape[-1]
|
||||
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)
|
||||
|
||||
if attn.norm_q is not None:
|
||||
query = attn.norm_q(query)
|
||||
if attn.norm_k is not None:
|
||||
key = attn.norm_k(key)
|
||||
|
||||
# Apply RoPE if needed
|
||||
if image_rotary_emb is not None:
|
||||
from .embeddings import apply_rotary_emb
|
||||
|
||||
query[:, :, text_seq_length:] = apply_rotary_emb(query[:, :, text_seq_length:], image_rotary_emb)
|
||||
if not attn.is_cross_attention:
|
||||
key[:, :, text_seq_length:] = apply_rotary_emb(key[:, :, text_seq_length:], image_rotary_emb)
|
||||
|
||||
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)
|
||||
|
||||
# linear proj
|
||||
hidden_states = attn.to_out[0](hidden_states)
|
||||
# dropout
|
||||
hidden_states = attn.to_out[1](hidden_states)
|
||||
|
||||
encoder_hidden_states, hidden_states = hidden_states.split(
|
||||
[text_seq_length, hidden_states.size(1) - text_seq_length], dim=1
|
||||
)
|
||||
return hidden_states, encoder_hidden_states
|
||||
|
||||
|
||||
@@ -4107,6 +4258,17 @@ class LoRAAttnAddedKVProcessor:
|
||||
pass
|
||||
|
||||
|
||||
class FluxSingleAttnProcessor2_0(FluxAttnProcessor2_0):
|
||||
r"""
|
||||
Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
|
||||
"""
|
||||
|
||||
def __init__(self):
|
||||
deprecation_message = "`FluxSingleAttnProcessor2_0` is deprecated and will be removed in a future version. Please use `FluxAttnProcessor2_0` instead."
|
||||
deprecate("FluxSingleAttnProcessor2_0", "0.32.0", deprecation_message)
|
||||
super().__init__()
|
||||
|
||||
|
||||
ADDED_KV_ATTENTION_PROCESSORS = (
|
||||
AttnAddedKVProcessor,
|
||||
SlicedAttnAddedKVProcessor,
|
||||
|
||||
@@ -902,7 +902,7 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
|
||||
Tuple of block output channels.
|
||||
act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
|
||||
sample_size (`int`, *optional*, defaults to `32`): Sample input size.
|
||||
scaling_factor (`float`, *optional*, defaults to 0.18215):
|
||||
scaling_factor (`float`, *optional*, defaults to `1.15258426`):
|
||||
The component-wise standard deviation of the trained latent space computed using the first batch of the
|
||||
training set. This is used to scale the latent space to have unit variance when training the diffusion
|
||||
model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
|
||||
|
||||
@@ -24,9 +24,9 @@ from ..models.attention_processor import AttentionProcessor
|
||||
from ..models.modeling_utils import ModelMixin
|
||||
from ..utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
|
||||
from .controlnet import BaseOutput, zero_module
|
||||
from .embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings
|
||||
from .embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
|
||||
from .modeling_outputs import Transformer2DModelOutput
|
||||
from .transformers.transformer_flux import EmbedND, FluxSingleTransformerBlock, FluxTransformerBlock
|
||||
from .transformers.transformer_flux import FluxSingleTransformerBlock, FluxTransformerBlock
|
||||
|
||||
|
||||
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
||||
@@ -59,7 +59,7 @@ class FluxControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
|
||||
self.out_channels = in_channels
|
||||
self.inner_dim = num_attention_heads * attention_head_dim
|
||||
|
||||
self.pos_embed = EmbedND(dim=self.inner_dim, theta=10000, axes_dim=axes_dims_rope)
|
||||
self.pos_embed = FluxPosEmbed(theta=10000, axes_dim=axes_dims_rope)
|
||||
text_time_guidance_cls = (
|
||||
CombinedTimestepGuidanceTextProjEmbeddings if guidance_embeds else CombinedTimestepTextProjEmbeddings
|
||||
)
|
||||
@@ -272,8 +272,20 @@ class FluxControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
|
||||
)
|
||||
encoder_hidden_states = self.context_embedder(encoder_hidden_states)
|
||||
|
||||
txt_ids = txt_ids.expand(img_ids.size(0), -1, -1)
|
||||
ids = torch.cat((txt_ids, img_ids), dim=1)
|
||||
if txt_ids.ndim == 3:
|
||||
logger.warning(
|
||||
"Passing `txt_ids` 3d torch.Tensor is deprecated."
|
||||
"Please remove the batch dimension and pass it as a 2d torch Tensor"
|
||||
)
|
||||
txt_ids = txt_ids[0]
|
||||
if img_ids.ndim == 3:
|
||||
logger.warning(
|
||||
"Passing `img_ids` 3d torch.Tensor is deprecated."
|
||||
"Please remove the batch dimension and pass it as a 2d torch Tensor"
|
||||
)
|
||||
img_ids = img_ids[0]
|
||||
|
||||
ids = torch.cat((txt_ids, img_ids), dim=0)
|
||||
image_rotary_emb = self.pos_embed(ids)
|
||||
|
||||
block_samples = ()
|
||||
|
||||
@@ -374,6 +374,90 @@ class CogVideoXPatchEmbed(nn.Module):
|
||||
return embeds
|
||||
|
||||
|
||||
def get_3d_rotary_pos_embed(
|
||||
embed_dim, crops_coords, grid_size, temporal_size, theta: int = 10000, use_real: bool = True
|
||||
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
||||
"""
|
||||
RoPE for video tokens with 3D structure.
|
||||
|
||||
Args:
|
||||
embed_dim: (`int`):
|
||||
The embedding dimension size, corresponding to hidden_size_head.
|
||||
crops_coords (`Tuple[int]`):
|
||||
The top-left and bottom-right coordinates of the crop.
|
||||
grid_size (`Tuple[int]`):
|
||||
The grid size of the spatial positional embedding (height, width).
|
||||
temporal_size (`int`):
|
||||
The size of the temporal dimension.
|
||||
theta (`float`):
|
||||
Scaling factor for frequency computation.
|
||||
use_real (`bool`):
|
||||
If True, return real part and imaginary part separately. Otherwise, return complex numbers.
|
||||
|
||||
Returns:
|
||||
`torch.Tensor`: positional embedding with shape `(temporal_size * grid_size[0] * grid_size[1], embed_dim/2)`.
|
||||
"""
|
||||
start, stop = crops_coords
|
||||
grid_h = np.linspace(start[0], stop[0], grid_size[0], endpoint=False, dtype=np.float32)
|
||||
grid_w = np.linspace(start[1], stop[1], grid_size[1], endpoint=False, dtype=np.float32)
|
||||
grid_t = np.linspace(0, temporal_size, temporal_size, endpoint=False, dtype=np.float32)
|
||||
|
||||
# Compute dimensions for each axis
|
||||
dim_t = embed_dim // 4
|
||||
dim_h = embed_dim // 8 * 3
|
||||
dim_w = embed_dim // 8 * 3
|
||||
|
||||
# Temporal frequencies
|
||||
freqs_t = 1.0 / (theta ** (torch.arange(0, dim_t, 2).float() / dim_t))
|
||||
grid_t = torch.from_numpy(grid_t).float()
|
||||
freqs_t = torch.einsum("n , f -> n f", grid_t, freqs_t)
|
||||
freqs_t = freqs_t.repeat_interleave(2, dim=-1)
|
||||
|
||||
# Spatial frequencies for height and width
|
||||
freqs_h = 1.0 / (theta ** (torch.arange(0, dim_h, 2).float() / dim_h))
|
||||
freqs_w = 1.0 / (theta ** (torch.arange(0, dim_w, 2).float() / dim_w))
|
||||
grid_h = torch.from_numpy(grid_h).float()
|
||||
grid_w = torch.from_numpy(grid_w).float()
|
||||
freqs_h = torch.einsum("n , f -> n f", grid_h, freqs_h)
|
||||
freqs_w = torch.einsum("n , f -> n f", grid_w, freqs_w)
|
||||
freqs_h = freqs_h.repeat_interleave(2, dim=-1)
|
||||
freqs_w = freqs_w.repeat_interleave(2, dim=-1)
|
||||
|
||||
# Broadcast and concatenate tensors along specified dimension
|
||||
def broadcast(tensors, dim=-1):
|
||||
num_tensors = len(tensors)
|
||||
shape_lens = {len(t.shape) for t in tensors}
|
||||
assert len(shape_lens) == 1, "tensors must all have the same number of dimensions"
|
||||
shape_len = list(shape_lens)[0]
|
||||
dim = (dim + shape_len) if dim < 0 else dim
|
||||
dims = list(zip(*(list(t.shape) for t in tensors)))
|
||||
expandable_dims = [(i, val) for i, val in enumerate(dims) if i != dim]
|
||||
assert all(
|
||||
[*(len(set(t[1])) <= 2 for t in expandable_dims)]
|
||||
), "invalid dimensions for broadcastable concatenation"
|
||||
max_dims = [(t[0], max(t[1])) for t in expandable_dims]
|
||||
expanded_dims = [(t[0], (t[1],) * num_tensors) for t in max_dims]
|
||||
expanded_dims.insert(dim, (dim, dims[dim]))
|
||||
expandable_shapes = list(zip(*(t[1] for t in expanded_dims)))
|
||||
tensors = [t[0].expand(*t[1]) for t in zip(tensors, expandable_shapes)]
|
||||
return torch.cat(tensors, dim=dim)
|
||||
|
||||
freqs = broadcast((freqs_t[:, None, None, :], freqs_h[None, :, None, :], freqs_w[None, None, :, :]), dim=-1)
|
||||
|
||||
t, h, w, d = freqs.shape
|
||||
freqs = freqs.view(t * h * w, d)
|
||||
|
||||
# Generate sine and cosine components
|
||||
sin = freqs.sin()
|
||||
cos = freqs.cos()
|
||||
|
||||
if use_real:
|
||||
return cos, sin
|
||||
else:
|
||||
freqs_cis = torch.polar(torch.ones_like(freqs), freqs)
|
||||
return freqs_cis
|
||||
|
||||
|
||||
def get_2d_rotary_pos_embed(embed_dim, crops_coords, grid_size, use_real=True):
|
||||
"""
|
||||
RoPE for image tokens with 2d structure.
|
||||
@@ -446,6 +530,7 @@ def get_1d_rotary_pos_embed(
|
||||
linear_factor=1.0,
|
||||
ntk_factor=1.0,
|
||||
repeat_interleave_real=True,
|
||||
freqs_dtype=torch.float32, # torch.float32 (hunyuan, stable audio), torch.float64 (flux)
|
||||
):
|
||||
"""
|
||||
Precompute the frequency tensor for complex exponentials (cis) with given dimensions.
|
||||
@@ -468,6 +553,8 @@ def get_1d_rotary_pos_embed(
|
||||
repeat_interleave_real (`bool`, *optional*, defaults to `True`):
|
||||
If `True` and `use_real`, real part and imaginary part are each interleaved with themselves to reach `dim`.
|
||||
Otherwise, they are concateanted with themselves.
|
||||
freqs_dtype (`torch.float32` or `torch.float64`, *optional*, defaults to `torch.float32`):
|
||||
the dtype of the frequency tensor.
|
||||
Returns:
|
||||
`torch.Tensor`: Precomputed frequency tensor with complex exponentials. [S, D/2]
|
||||
"""
|
||||
@@ -476,19 +563,19 @@ def get_1d_rotary_pos_embed(
|
||||
if isinstance(pos, int):
|
||||
pos = np.arange(pos)
|
||||
theta = theta * ntk_factor
|
||||
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim)) / linear_factor # [D/2]
|
||||
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=freqs_dtype)[: (dim // 2)] / dim)) / linear_factor # [D/2]
|
||||
t = torch.from_numpy(pos).to(freqs.device) # type: ignore # [S]
|
||||
freqs = torch.outer(t, freqs).float() # type: ignore # [S, D/2]
|
||||
freqs = torch.outer(t, freqs) # type: ignore # [S, D/2]
|
||||
if use_real and repeat_interleave_real:
|
||||
freqs_cos = freqs.cos().repeat_interleave(2, dim=1) # [S, D]
|
||||
freqs_sin = freqs.sin().repeat_interleave(2, dim=1) # [S, D]
|
||||
freqs_cos = freqs.cos().repeat_interleave(2, dim=1).float() # [S, D]
|
||||
freqs_sin = freqs.sin().repeat_interleave(2, dim=1).float() # [S, D]
|
||||
return freqs_cos, freqs_sin
|
||||
elif use_real:
|
||||
freqs_cos = torch.cat([freqs.cos(), freqs.cos()], dim=-1) # [S, D]
|
||||
freqs_sin = torch.cat([freqs.sin(), freqs.sin()], dim=-1) # [S, D]
|
||||
freqs_cos = torch.cat([freqs.cos(), freqs.cos()], dim=-1).float() # [S, D]
|
||||
freqs_sin = torch.cat([freqs.sin(), freqs.sin()], dim=-1).float() # [S, D]
|
||||
return freqs_cos, freqs_sin
|
||||
else:
|
||||
freqs_cis = torch.polar(torch.ones_like(freqs), freqs) # complex64 # [S, D/2]
|
||||
freqs_cis = torch.polar(torch.ones_like(freqs), freqs).float() # complex64 # [S, D/2]
|
||||
return freqs_cis
|
||||
|
||||
|
||||
@@ -540,6 +627,31 @@ def apply_rotary_emb(
|
||||
return x_out.type_as(x)
|
||||
|
||||
|
||||
class FluxPosEmbed(nn.Module):
|
||||
# modified from https://github.com/black-forest-labs/flux/blob/c00d7c60b085fce8058b9df845e036090873f2ce/src/flux/modules/layers.py#L11
|
||||
def __init__(self, theta: int, axes_dim: List[int]):
|
||||
super().__init__()
|
||||
self.theta = theta
|
||||
self.axes_dim = axes_dim
|
||||
|
||||
def forward(self, ids: torch.Tensor) -> torch.Tensor:
|
||||
n_axes = ids.shape[-1]
|
||||
cos_out = []
|
||||
sin_out = []
|
||||
pos = ids.squeeze().float().cpu().numpy()
|
||||
is_mps = ids.device.type == "mps"
|
||||
freqs_dtype = torch.float32 if is_mps else torch.float64
|
||||
for i in range(n_axes):
|
||||
cos, sin = get_1d_rotary_pos_embed(
|
||||
self.axes_dim[i], pos[:, i], repeat_interleave_real=True, use_real=True, freqs_dtype=freqs_dtype
|
||||
)
|
||||
cos_out.append(cos)
|
||||
sin_out.append(sin)
|
||||
freqs_cos = torch.cat(cos_out, dim=-1).to(ids.device)
|
||||
freqs_sin = torch.cat(sin_out, dim=-1).to(ids.device)
|
||||
return freqs_cos, freqs_sin
|
||||
|
||||
|
||||
class TimestepEmbedding(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
|
||||
@@ -13,7 +13,7 @@
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
from typing import Any, Dict, Optional, Union
|
||||
from typing import Any, Dict, Optional, Tuple, Union
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
@@ -22,6 +22,7 @@ from ...configuration_utils import ConfigMixin, register_to_config
|
||||
from ...utils import is_torch_version, logging
|
||||
from ...utils.torch_utils import maybe_allow_in_graph
|
||||
from ..attention import Attention, FeedForward
|
||||
from ..attention_processor import AttentionProcessor, CogVideoXAttnProcessor2_0, FusedCogVideoXAttnProcessor2_0
|
||||
from ..embeddings import CogVideoXPatchEmbed, TimestepEmbedding, Timesteps, get_3d_sincos_pos_embed
|
||||
from ..modeling_outputs import Transformer2DModelOutput
|
||||
from ..modeling_utils import ModelMixin
|
||||
@@ -97,6 +98,7 @@ class CogVideoXBlock(nn.Module):
|
||||
eps=1e-6,
|
||||
bias=attention_bias,
|
||||
out_bias=attention_out_bias,
|
||||
processor=CogVideoXAttnProcessor2_0(),
|
||||
)
|
||||
|
||||
# 2. Feed Forward
|
||||
@@ -116,24 +118,24 @@ class CogVideoXBlock(nn.Module):
|
||||
hidden_states: torch.Tensor,
|
||||
encoder_hidden_states: torch.Tensor,
|
||||
temb: torch.Tensor,
|
||||
image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
||||
) -> torch.Tensor:
|
||||
text_seq_length = encoder_hidden_states.size(1)
|
||||
|
||||
# norm & modulate
|
||||
norm_hidden_states, norm_encoder_hidden_states, gate_msa, enc_gate_msa = self.norm1(
|
||||
hidden_states, encoder_hidden_states, temb
|
||||
)
|
||||
|
||||
# attention
|
||||
text_length = norm_encoder_hidden_states.size(1)
|
||||
|
||||
# CogVideoX uses concatenated text + video embeddings with self-attention instead of using
|
||||
# them in cross-attention individually
|
||||
norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
|
||||
attn_output = self.attn1(
|
||||
attn_hidden_states, attn_encoder_hidden_states = self.attn1(
|
||||
hidden_states=norm_hidden_states,
|
||||
encoder_hidden_states=None,
|
||||
encoder_hidden_states=norm_encoder_hidden_states,
|
||||
image_rotary_emb=image_rotary_emb,
|
||||
)
|
||||
|
||||
hidden_states = hidden_states + gate_msa * attn_output[:, text_length:]
|
||||
encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_output[:, :text_length]
|
||||
hidden_states = hidden_states + gate_msa * attn_hidden_states
|
||||
encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_encoder_hidden_states
|
||||
|
||||
# norm & modulate
|
||||
norm_hidden_states, norm_encoder_hidden_states, gate_ff, enc_gate_ff = self.norm2(
|
||||
@@ -144,8 +146,9 @@ class CogVideoXBlock(nn.Module):
|
||||
norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
|
||||
ff_output = self.ff(norm_hidden_states)
|
||||
|
||||
hidden_states = hidden_states + gate_ff * ff_output[:, text_length:]
|
||||
encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_length]
|
||||
hidden_states = hidden_states + gate_ff * ff_output[:, text_seq_length:]
|
||||
encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_seq_length]
|
||||
|
||||
return hidden_states, encoder_hidden_states
|
||||
|
||||
|
||||
@@ -231,6 +234,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
|
||||
norm_eps: float = 1e-5,
|
||||
spatial_interpolation_scale: float = 1.875,
|
||||
temporal_interpolation_scale: float = 1.0,
|
||||
use_rotary_positional_embeddings: bool = False,
|
||||
):
|
||||
super().__init__()
|
||||
inner_dim = num_attention_heads * attention_head_dim
|
||||
@@ -295,12 +299,113 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
|
||||
def _set_gradient_checkpointing(self, module, value=False):
|
||||
self.gradient_checkpointing = value
|
||||
|
||||
@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.fuse_qkv_projections with FusedAttnProcessor2_0->FusedCogVideoXAttnProcessor2_0
|
||||
def fuse_qkv_projections(self):
|
||||
"""
|
||||
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.
|
||||
|
||||
<Tip warning={true}>
|
||||
|
||||
This API is 🧪 experimental.
|
||||
|
||||
</Tip>
|
||||
"""
|
||||
self.original_attn_processors = None
|
||||
|
||||
for _, attn_processor in self.attn_processors.items():
|
||||
if "Added" in str(attn_processor.__class__.__name__):
|
||||
raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
|
||||
|
||||
self.original_attn_processors = self.attn_processors
|
||||
|
||||
for module in self.modules():
|
||||
if isinstance(module, Attention):
|
||||
module.fuse_projections(fuse=True)
|
||||
|
||||
self.set_attn_processor(FusedCogVideoXAttnProcessor2_0())
|
||||
|
||||
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
|
||||
def unfuse_qkv_projections(self):
|
||||
"""Disables the fused QKV projection if enabled.
|
||||
|
||||
<Tip warning={true}>
|
||||
|
||||
This API is 🧪 experimental.
|
||||
|
||||
</Tip>
|
||||
|
||||
"""
|
||||
if self.original_attn_processors is not None:
|
||||
self.set_attn_processor(self.original_attn_processors)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
encoder_hidden_states: torch.Tensor,
|
||||
timestep: Union[int, float, torch.LongTensor],
|
||||
timestep_cond: Optional[torch.Tensor] = None,
|
||||
image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
||||
return_dict: bool = True,
|
||||
):
|
||||
batch_size, num_frames, channels, height, width = hidden_states.shape
|
||||
@@ -319,14 +424,16 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
|
||||
hidden_states = self.patch_embed(encoder_hidden_states, hidden_states)
|
||||
|
||||
# 3. Position embedding
|
||||
seq_length = height * width * num_frames // (self.config.patch_size**2)
|
||||
text_seq_length = encoder_hidden_states.shape[1]
|
||||
if not self.config.use_rotary_positional_embeddings:
|
||||
seq_length = height * width * num_frames // (self.config.patch_size**2)
|
||||
|
||||
pos_embeds = self.pos_embedding[:, : self.config.max_text_seq_length + seq_length]
|
||||
hidden_states = hidden_states + pos_embeds
|
||||
hidden_states = self.embedding_dropout(hidden_states)
|
||||
pos_embeds = self.pos_embedding[:, : text_seq_length + seq_length]
|
||||
hidden_states = hidden_states + pos_embeds
|
||||
hidden_states = self.embedding_dropout(hidden_states)
|
||||
|
||||
encoder_hidden_states = hidden_states[:, : self.config.max_text_seq_length]
|
||||
hidden_states = hidden_states[:, self.config.max_text_seq_length :]
|
||||
encoder_hidden_states = hidden_states[:, :text_seq_length]
|
||||
hidden_states = hidden_states[:, text_seq_length:]
|
||||
|
||||
# 4. Transformer blocks
|
||||
for i, block in enumerate(self.transformer_blocks):
|
||||
@@ -344,6 +451,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
|
||||
hidden_states,
|
||||
encoder_hidden_states,
|
||||
emb,
|
||||
image_rotary_emb,
|
||||
**ckpt_kwargs,
|
||||
)
|
||||
else:
|
||||
@@ -351,9 +459,17 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
|
||||
hidden_states=hidden_states,
|
||||
encoder_hidden_states=encoder_hidden_states,
|
||||
temb=emb,
|
||||
image_rotary_emb=image_rotary_emb,
|
||||
)
|
||||
|
||||
hidden_states = self.norm_final(hidden_states)
|
||||
if not self.config.use_rotary_positional_embeddings:
|
||||
# CogVideoX-2B
|
||||
hidden_states = self.norm_final(hidden_states)
|
||||
else:
|
||||
# CogVideoX-5B
|
||||
hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
|
||||
hidden_states = self.norm_final(hidden_states)
|
||||
hidden_states = hidden_states[:, text_seq_length:]
|
||||
|
||||
# 5. Final block
|
||||
hidden_states = self.norm_out(hidden_states, temb=emb)
|
||||
|
||||
@@ -13,7 +13,7 @@
|
||||
# limitations under the License.
|
||||
|
||||
|
||||
from typing import Any, Dict, List, Optional, Union
|
||||
from typing import Any, Dict, Optional, Tuple, Union
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
@@ -23,52 +23,23 @@ import torch.nn.functional as F
|
||||
from ...configuration_utils import ConfigMixin, register_to_config
|
||||
from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
|
||||
from ...models.attention import FeedForward
|
||||
from ...models.attention_processor import Attention, FluxAttnProcessor2_0, FluxSingleAttnProcessor2_0
|
||||
from ...models.attention_processor import (
|
||||
Attention,
|
||||
AttentionProcessor,
|
||||
FluxAttnProcessor2_0,
|
||||
FusedFluxAttnProcessor2_0,
|
||||
)
|
||||
from ...models.modeling_utils import ModelMixin
|
||||
from ...models.normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle
|
||||
from ...utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
|
||||
from ...utils.torch_utils import maybe_allow_in_graph
|
||||
from ..embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings
|
||||
from ..embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
|
||||
from ..modeling_outputs import Transformer2DModelOutput
|
||||
|
||||
|
||||
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
||||
|
||||
|
||||
# YiYi to-do: refactor rope related functions/classes
|
||||
def rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor:
|
||||
assert dim % 2 == 0, "The dimension must be even."
|
||||
|
||||
scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
|
||||
omega = 1.0 / (theta**scale)
|
||||
|
||||
batch_size, seq_length = pos.shape
|
||||
out = torch.einsum("...n,d->...nd", pos, omega)
|
||||
cos_out = torch.cos(out)
|
||||
sin_out = torch.sin(out)
|
||||
|
||||
stacked_out = torch.stack([cos_out, -sin_out, sin_out, cos_out], dim=-1)
|
||||
out = stacked_out.view(batch_size, -1, dim // 2, 2, 2)
|
||||
return out.float()
|
||||
|
||||
|
||||
# YiYi to-do: refactor rope related functions/classes
|
||||
class EmbedND(nn.Module):
|
||||
def __init__(self, dim: int, theta: int, axes_dim: List[int]):
|
||||
super().__init__()
|
||||
self.dim = dim
|
||||
self.theta = theta
|
||||
self.axes_dim = axes_dim
|
||||
|
||||
def forward(self, ids: torch.Tensor) -> torch.Tensor:
|
||||
n_axes = ids.shape[-1]
|
||||
emb = torch.cat(
|
||||
[rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
|
||||
dim=-3,
|
||||
)
|
||||
return emb.unsqueeze(1)
|
||||
|
||||
|
||||
@maybe_allow_in_graph
|
||||
class FluxSingleTransformerBlock(nn.Module):
|
||||
r"""
|
||||
@@ -93,7 +64,7 @@ class FluxSingleTransformerBlock(nn.Module):
|
||||
self.act_mlp = nn.GELU(approximate="tanh")
|
||||
self.proj_out = nn.Linear(dim + self.mlp_hidden_dim, dim)
|
||||
|
||||
processor = FluxSingleAttnProcessor2_0()
|
||||
processor = FluxAttnProcessor2_0()
|
||||
self.attn = Attention(
|
||||
query_dim=dim,
|
||||
cross_attention_dim=None,
|
||||
@@ -265,13 +236,14 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
|
||||
joint_attention_dim: int = 4096,
|
||||
pooled_projection_dim: int = 768,
|
||||
guidance_embeds: bool = False,
|
||||
axes_dims_rope: List[int] = [16, 56, 56],
|
||||
axes_dims_rope: Tuple[int] = (16, 56, 56),
|
||||
):
|
||||
super().__init__()
|
||||
self.out_channels = in_channels
|
||||
self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
|
||||
|
||||
self.pos_embed = EmbedND(dim=self.inner_dim, theta=10000, axes_dim=axes_dims_rope)
|
||||
self.pos_embed = FluxPosEmbed(theta=10000, axes_dim=axes_dims_rope)
|
||||
|
||||
text_time_guidance_cls = (
|
||||
CombinedTimestepGuidanceTextProjEmbeddings if guidance_embeds else CombinedTimestepTextProjEmbeddings
|
||||
)
|
||||
@@ -309,6 +281,106 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
|
||||
|
||||
self.gradient_checkpointing = False
|
||||
|
||||
@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.fuse_qkv_projections with FusedAttnProcessor2_0->FusedFluxAttnProcessor2_0
|
||||
def fuse_qkv_projections(self):
|
||||
"""
|
||||
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.
|
||||
|
||||
<Tip warning={true}>
|
||||
|
||||
This API is 🧪 experimental.
|
||||
|
||||
</Tip>
|
||||
"""
|
||||
self.original_attn_processors = None
|
||||
|
||||
for _, attn_processor in self.attn_processors.items():
|
||||
if "Added" in str(attn_processor.__class__.__name__):
|
||||
raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
|
||||
|
||||
self.original_attn_processors = self.attn_processors
|
||||
|
||||
for module in self.modules():
|
||||
if isinstance(module, Attention):
|
||||
module.fuse_projections(fuse=True)
|
||||
|
||||
self.set_attn_processor(FusedFluxAttnProcessor2_0())
|
||||
|
||||
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
|
||||
def unfuse_qkv_projections(self):
|
||||
"""Disables the fused QKV projection if enabled.
|
||||
|
||||
<Tip warning={true}>
|
||||
|
||||
This API is 🧪 experimental.
|
||||
|
||||
</Tip>
|
||||
|
||||
"""
|
||||
if self.original_attn_processors is not None:
|
||||
self.set_attn_processor(self.original_attn_processors)
|
||||
|
||||
def _set_gradient_checkpointing(self, module, value=False):
|
||||
if hasattr(module, "gradient_checkpointing"):
|
||||
module.gradient_checkpointing = value
|
||||
@@ -381,8 +453,19 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
|
||||
)
|
||||
encoder_hidden_states = self.context_embedder(encoder_hidden_states)
|
||||
|
||||
txt_ids = txt_ids.expand(img_ids.size(0), -1, -1)
|
||||
ids = torch.cat((txt_ids, img_ids), dim=1)
|
||||
if txt_ids.ndim == 3:
|
||||
logger.warning(
|
||||
"Passing `txt_ids` 3d torch.Tensor is deprecated."
|
||||
"Please remove the batch dimension and pass it as a 2d torch Tensor"
|
||||
)
|
||||
txt_ids = txt_ids[0]
|
||||
if img_ids.ndim == 3:
|
||||
logger.warning(
|
||||
"Passing `img_ids` 3d torch.Tensor is deprecated."
|
||||
"Please remove the batch dimension and pass it as a 2d torch Tensor"
|
||||
)
|
||||
img_ids = img_ids[0]
|
||||
ids = torch.cat((txt_ids, img_ids), dim=0)
|
||||
image_rotary_emb = self.pos_embed(ids)
|
||||
|
||||
for index_block, block in enumerate(self.transformer_blocks):
|
||||
|
||||
@@ -154,6 +154,7 @@ else:
|
||||
"StableDiffusionControlNetPAGPipeline",
|
||||
"StableDiffusionXLPAGPipeline",
|
||||
"StableDiffusionXLPAGInpaintPipeline",
|
||||
"StableDiffusionXLControlNetPAGImg2ImgPipeline",
|
||||
"StableDiffusionXLControlNetPAGPipeline",
|
||||
"StableDiffusionXLPAGImg2ImgPipeline",
|
||||
"PixArtSigmaPAGPipeline",
|
||||
@@ -547,6 +548,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
|
||||
StableDiffusion3PAGPipeline,
|
||||
StableDiffusionControlNetPAGPipeline,
|
||||
StableDiffusionPAGPipeline,
|
||||
StableDiffusionXLControlNetPAGImg2ImgPipeline,
|
||||
StableDiffusionXLControlNetPAGPipeline,
|
||||
StableDiffusionXLPAGImg2ImgPipeline,
|
||||
StableDiffusionXLPAGInpaintPipeline,
|
||||
|
||||
@@ -56,6 +56,7 @@ from .pag import (
|
||||
StableDiffusion3PAGPipeline,
|
||||
StableDiffusionControlNetPAGPipeline,
|
||||
StableDiffusionPAGPipeline,
|
||||
StableDiffusionXLControlNetPAGImg2ImgPipeline,
|
||||
StableDiffusionXLControlNetPAGPipeline,
|
||||
StableDiffusionXLPAGImg2ImgPipeline,
|
||||
StableDiffusionXLPAGInpaintPipeline,
|
||||
@@ -123,6 +124,7 @@ AUTO_IMAGE2IMAGE_PIPELINES_MAPPING = OrderedDict(
|
||||
("stable-diffusion-controlnet", StableDiffusionControlNetImg2ImgPipeline),
|
||||
("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetImg2ImgPipeline),
|
||||
("stable-diffusion-xl-pag", StableDiffusionXLPAGImg2ImgPipeline),
|
||||
("stable-diffusion-xl-controlnet-pag", StableDiffusionXLControlNetPAGImg2ImgPipeline),
|
||||
("lcm", LatentConsistencyModelImg2ImgPipeline),
|
||||
]
|
||||
)
|
||||
@@ -163,12 +165,12 @@ _AUTO_INPAINT_DECODER_PIPELINES_MAPPING = OrderedDict(
|
||||
)
|
||||
|
||||
if is_sentencepiece_available():
|
||||
from .kolors import KolorsPipeline
|
||||
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"] = KolorsPipeline
|
||||
AUTO_IMAGE2IMAGE_PIPELINES_MAPPING["kolors"] = KolorsImg2ImgPipeline
|
||||
|
||||
SUPPORTED_TASKS_MAPPINGS = [
|
||||
AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
|
||||
@@ -955,7 +957,8 @@ class AutoPipelineForInpainting(ConfigMixin):
|
||||
if "enable_pag" in kwargs:
|
||||
enable_pag = kwargs.pop("enable_pag")
|
||||
if enable_pag:
|
||||
orig_class_name = config["_class_name"].replace("Pipeline", "PAGPipeline")
|
||||
to_replace = "InpaintPipeline" if "Inpaint" in config["_class_name"] else "Pipeline"
|
||||
orig_class_name = config["_class_name"].replace(to_replace, "PAG" + to_replace)
|
||||
|
||||
inpainting_cls = _get_task_class(AUTO_INPAINT_PIPELINES_MAPPING, orig_class_name)
|
||||
|
||||
|
||||
@@ -23,6 +23,7 @@ from transformers import T5EncoderModel, T5Tokenizer
|
||||
|
||||
from ...callbacks import MultiPipelineCallbacks, PipelineCallback
|
||||
from ...models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
|
||||
from ...models.embeddings import get_3d_rotary_pos_embed
|
||||
from ...pipelines.pipeline_utils import DiffusionPipeline
|
||||
from ...schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
|
||||
from ...utils import BaseOutput, logging, replace_example_docstring
|
||||
@@ -40,6 +41,7 @@ EXAMPLE_DOC_STRING = """
|
||||
>>> from diffusers import CogVideoXPipeline
|
||||
>>> from diffusers.utils import export_to_video
|
||||
|
||||
>>> # Models: "THUDM/CogVideoX-2b" or "THUDM/CogVideoX-5b"
|
||||
>>> pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b", torch_dtype=torch.float16).to("cuda")
|
||||
>>> prompt = (
|
||||
... "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. "
|
||||
@@ -55,6 +57,25 @@ EXAMPLE_DOC_STRING = """
|
||||
"""
|
||||
|
||||
|
||||
# Similar to diffusers.pipelines.hunyuandit.pipeline_hunyuandit.get_resize_crop_region_for_grid
|
||||
def get_resize_crop_region_for_grid(src, tgt_width, tgt_height):
|
||||
tw = tgt_width
|
||||
th = tgt_height
|
||||
h, w = src
|
||||
r = h / w
|
||||
if r > (th / tw):
|
||||
resize_height = th
|
||||
resize_width = int(round(th / h * w))
|
||||
else:
|
||||
resize_width = tw
|
||||
resize_height = int(round(tw / w * h))
|
||||
|
||||
crop_top = int(round((th - resize_height) / 2.0))
|
||||
crop_left = int(round((tw - resize_width) / 2.0))
|
||||
|
||||
return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width)
|
||||
|
||||
|
||||
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
|
||||
def retrieve_timesteps(
|
||||
scheduler,
|
||||
@@ -409,6 +430,46 @@ class CogVideoXPipeline(DiffusionPipeline):
|
||||
f" {negative_prompt_embeds.shape}."
|
||||
)
|
||||
|
||||
def fuse_qkv_projections(self) -> None:
|
||||
r"""Enables fused QKV projections."""
|
||||
self.fusing_transformer = True
|
||||
self.transformer.fuse_qkv_projections()
|
||||
|
||||
def unfuse_qkv_projections(self) -> None:
|
||||
r"""Disable QKV projection fusion if enabled."""
|
||||
if not self.fusing_transformer:
|
||||
logger.warning("The Transformer was not initially fused for QKV projections. Doing nothing.")
|
||||
else:
|
||||
self.transformer.unfuse_qkv_projections()
|
||||
self.fusing_transformer = False
|
||||
|
||||
def _prepare_rotary_positional_embeddings(
|
||||
self,
|
||||
height: int,
|
||||
width: int,
|
||||
num_frames: int,
|
||||
device: torch.device,
|
||||
) -> Tuple[torch.Tensor, torch.Tensor]:
|
||||
grid_height = height // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
|
||||
grid_width = width // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
|
||||
base_size_width = 720 // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
|
||||
base_size_height = 480 // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
|
||||
|
||||
grid_crops_coords = get_resize_crop_region_for_grid(
|
||||
(grid_height, grid_width), base_size_width, base_size_height
|
||||
)
|
||||
freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
|
||||
embed_dim=self.transformer.config.attention_head_dim,
|
||||
crops_coords=grid_crops_coords,
|
||||
grid_size=(grid_height, grid_width),
|
||||
temporal_size=num_frames,
|
||||
use_real=True,
|
||||
)
|
||||
|
||||
freqs_cos = freqs_cos.to(device=device)
|
||||
freqs_sin = freqs_sin.to(device=device)
|
||||
return freqs_cos, freqs_sin
|
||||
|
||||
@property
|
||||
def guidance_scale(self):
|
||||
return self._guidance_scale
|
||||
@@ -599,7 +660,14 @@ class CogVideoXPipeline(DiffusionPipeline):
|
||||
# 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
|
||||
# 7. Create rotary embeds if required
|
||||
image_rotary_emb = (
|
||||
self._prepare_rotary_positional_embeddings(height, width, latents.size(1), device)
|
||||
if self.transformer.config.use_rotary_positional_embeddings
|
||||
else None
|
||||
)
|
||||
|
||||
# 8. Denoising loop
|
||||
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
||||
|
||||
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
||||
@@ -620,6 +688,7 @@ class CogVideoXPipeline(DiffusionPipeline):
|
||||
hidden_states=latent_model_input,
|
||||
encoder_hidden_states=prompt_embeds,
|
||||
timestep=timestep,
|
||||
image_rotary_emb=image_rotary_emb,
|
||||
return_dict=False,
|
||||
)[0]
|
||||
noise_pred = noise_pred.float()
|
||||
|
||||
@@ -1538,7 +1538,6 @@ class StableDiffusionXLControlNetImg2ImgPipeline(
|
||||
if isinstance(controlnet_cond_scale, list):
|
||||
controlnet_cond_scale = controlnet_cond_scale[0]
|
||||
cond_scale = controlnet_cond_scale * controlnet_keep[i]
|
||||
|
||||
down_block_res_samples, mid_block_res_sample = self.controlnet(
|
||||
control_model_input,
|
||||
t,
|
||||
|
||||
@@ -20,7 +20,7 @@ import torch
|
||||
from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast
|
||||
|
||||
from ...image_processor import VaeImageProcessor
|
||||
from ...loaders import FluxLoraLoaderMixin
|
||||
from ...loaders import FluxLoraLoaderMixin, FromSingleFileMixin
|
||||
from ...models.autoencoders import AutoencoderKL
|
||||
from ...models.transformers import FluxTransformer2DModel
|
||||
from ...schedulers import FlowMatchEulerDiscreteScheduler
|
||||
@@ -137,7 +137,7 @@ def retrieve_timesteps(
|
||||
return timesteps, num_inference_steps
|
||||
|
||||
|
||||
class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
|
||||
class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin):
|
||||
r"""
|
||||
The Flux pipeline for text-to-image generation.
|
||||
|
||||
@@ -331,10 +331,6 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
|
||||
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
|
||||
@@ -364,8 +360,7 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
|
||||
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(batch_size, prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
|
||||
text_ids = text_ids.repeat(num_images_per_prompt, 1, 1)
|
||||
text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
|
||||
|
||||
return prompt_embeds, pooled_prompt_embeds, text_ids
|
||||
|
||||
@@ -425,9 +420,8 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
|
||||
|
||||
latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape
|
||||
|
||||
latent_image_ids = latent_image_ids[None, :].repeat(batch_size, 1, 1, 1)
|
||||
latent_image_ids = latent_image_ids.reshape(
|
||||
batch_size, latent_image_id_height * latent_image_id_width, latent_image_id_channels
|
||||
latent_image_id_height * latent_image_id_width, latent_image_id_channels
|
||||
)
|
||||
|
||||
return latent_image_ids.to(device=device, dtype=dtype)
|
||||
@@ -454,6 +448,35 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
|
||||
|
||||
return 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,
|
||||
batch_size,
|
||||
@@ -695,7 +718,6 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
|
||||
|
||||
noise_pred = self.transformer(
|
||||
hidden_states=latents,
|
||||
# YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
|
||||
timestep=timestep / 1000,
|
||||
guidance=guidance,
|
||||
pooled_projections=pooled_prompt_embeds,
|
||||
|
||||
@@ -25,7 +25,7 @@ from transformers import (
|
||||
)
|
||||
|
||||
from ...image_processor import PipelineImageInput, VaeImageProcessor
|
||||
from ...loaders import FluxLoraLoaderMixin
|
||||
from ...loaders import FluxLoraLoaderMixin, FromSingleFileMixin
|
||||
from ...models.autoencoders import AutoencoderKL
|
||||
from ...models.controlnet_flux import FluxControlNetModel
|
||||
from ...models.transformers import FluxTransformer2DModel
|
||||
@@ -155,7 +155,7 @@ def retrieve_timesteps(
|
||||
return timesteps, num_inference_steps
|
||||
|
||||
|
||||
class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
|
||||
class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin):
|
||||
r"""
|
||||
The Flux pipeline for text-to-image generation.
|
||||
|
||||
@@ -354,10 +354,6 @@ class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
|
||||
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
|
||||
@@ -387,8 +383,7 @@ class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
|
||||
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(batch_size, prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
|
||||
text_ids = text_ids.repeat(num_images_per_prompt, 1, 1)
|
||||
text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
|
||||
|
||||
return prompt_embeds, pooled_prompt_embeds, text_ids
|
||||
|
||||
@@ -449,9 +444,8 @@ class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
|
||||
|
||||
latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape
|
||||
|
||||
latent_image_ids = latent_image_ids[None, :].repeat(batch_size, 1, 1, 1)
|
||||
latent_image_ids = latent_image_ids.reshape(
|
||||
batch_size, latent_image_id_height * latent_image_id_width, latent_image_id_channels
|
||||
latent_image_id_height * latent_image_id_width, latent_image_id_channels
|
||||
)
|
||||
|
||||
return latent_image_ids.to(device=device, dtype=dtype)
|
||||
@@ -804,7 +798,6 @@ class FluxControlNetPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
|
||||
|
||||
noise_pred = self.transformer(
|
||||
hidden_states=latents,
|
||||
# YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
|
||||
timestep=timestep / 1000,
|
||||
guidance=guidance,
|
||||
pooled_projections=pooled_prompt_embeds,
|
||||
|
||||
@@ -547,7 +547,7 @@ class KandinskyV22Img2ImgCombinedPipeline(DiffusionPipeline):
|
||||
negative_image_embeds = prior_outputs[1]
|
||||
|
||||
prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt
|
||||
image = [image] if isinstance(prompt, PIL.Image.Image) else image
|
||||
image = [image] if isinstance(image, PIL.Image.Image) else image
|
||||
|
||||
if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0:
|
||||
prompt = (image_embeds.shape[0] // len(prompt)) * prompt
|
||||
@@ -813,7 +813,7 @@ class KandinskyV22InpaintCombinedPipeline(DiffusionPipeline):
|
||||
negative_image_embeds = prior_outputs[1]
|
||||
|
||||
prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt
|
||||
image = [image] if isinstance(prompt, PIL.Image.Image) else image
|
||||
image = [image] if isinstance(image, PIL.Image.Image) else image
|
||||
mask_image = [mask_image] if isinstance(mask_image, PIL.Image.Image) else mask_image
|
||||
|
||||
if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0:
|
||||
|
||||
@@ -24,6 +24,7 @@ except OptionalDependencyNotAvailable:
|
||||
else:
|
||||
_import_structure["pipeline_pag_controlnet_sd"] = ["StableDiffusionControlNetPAGPipeline"]
|
||||
_import_structure["pipeline_pag_controlnet_sd_xl"] = ["StableDiffusionXLControlNetPAGPipeline"]
|
||||
_import_structure["pipeline_pag_controlnet_sd_xl_img2img"] = ["StableDiffusionXLControlNetPAGImg2ImgPipeline"]
|
||||
_import_structure["pipeline_pag_hunyuandit"] = ["HunyuanDiTPAGPipeline"]
|
||||
_import_structure["pipeline_pag_kolors"] = ["KolorsPAGPipeline"]
|
||||
_import_structure["pipeline_pag_pixart_sigma"] = ["PixArtSigmaPAGPipeline"]
|
||||
@@ -44,6 +45,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
|
||||
else:
|
||||
from .pipeline_pag_controlnet_sd import StableDiffusionControlNetPAGPipeline
|
||||
from .pipeline_pag_controlnet_sd_xl import StableDiffusionXLControlNetPAGPipeline
|
||||
from .pipeline_pag_controlnet_sd_xl_img2img import StableDiffusionXLControlNetPAGImg2ImgPipeline
|
||||
from .pipeline_pag_hunyuandit import HunyuanDiTPAGPipeline
|
||||
from .pipeline_pag_kolors import KolorsPAGPipeline
|
||||
from .pipeline_pag_pixart_sigma import PixArtSigmaPAGPipeline
|
||||
|
||||
1685
src/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl_img2img.py
Normal file
1685
src/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl_img2img.py
Normal file
File diff suppressed because it is too large
Load Diff
@@ -1471,6 +1471,14 @@ class StableDiffusionXLPAGInpaintPipeline(
|
||||
generator,
|
||||
self.do_classifier_free_guidance,
|
||||
)
|
||||
if self.do_perturbed_attention_guidance:
|
||||
if self.do_classifier_free_guidance:
|
||||
mask, _ = mask.chunk(2)
|
||||
masked_image_latents, _ = masked_image_latents.chunk(2)
|
||||
mask = self._prepare_perturbed_attention_guidance(mask, mask, self.do_classifier_free_guidance)
|
||||
masked_image_latents = self._prepare_perturbed_attention_guidance(
|
||||
masked_image_latents, masked_image_latents, self.do_classifier_free_guidance
|
||||
)
|
||||
|
||||
# 8. Check that sizes of mask, masked image and latents match
|
||||
if num_channels_unet == 9:
|
||||
@@ -1659,10 +1667,10 @@ class StableDiffusionXLPAGInpaintPipeline(
|
||||
|
||||
if num_channels_unet == 4:
|
||||
init_latents_proper = image_latents
|
||||
if self.do_classifier_free_guidance:
|
||||
init_mask, _ = mask.chunk(2)
|
||||
if self.do_perturbed_attention_guidance:
|
||||
init_mask, *_ = mask.chunk(3) if self.do_classifier_free_guidance else mask.chunk(2)
|
||||
else:
|
||||
init_mask = mask
|
||||
init_mask, *_ = mask.chunk(2) if self.do_classifier_free_guidance else mask
|
||||
|
||||
if i < len(timesteps) - 1:
|
||||
noise_timestep = timesteps[i + 1]
|
||||
|
||||
@@ -22,7 +22,7 @@ from pathlib import Path
|
||||
from typing import Any, Dict, List, Optional, Union
|
||||
|
||||
import torch
|
||||
from huggingface_hub import model_info
|
||||
from huggingface_hub import ModelCard, model_info
|
||||
from huggingface_hub.utils import validate_hf_hub_args
|
||||
from packaging import version
|
||||
|
||||
@@ -33,6 +33,7 @@ from ..utils import (
|
||||
ONNX_WEIGHTS_NAME,
|
||||
SAFETENSORS_WEIGHTS_NAME,
|
||||
WEIGHTS_NAME,
|
||||
deprecate,
|
||||
get_class_from_dynamic_module,
|
||||
is_accelerate_available,
|
||||
is_peft_available,
|
||||
@@ -89,49 +90,46 @@ for library in LOADABLE_CLASSES:
|
||||
ALL_IMPORTABLE_CLASSES.update(LOADABLE_CLASSES[library])
|
||||
|
||||
|
||||
def is_safetensors_compatible(filenames, variant=None, passed_components=None) -> bool:
|
||||
def is_safetensors_compatible(filenames, passed_components=None, folder_names=None) -> bool:
|
||||
"""
|
||||
Checking for safetensors compatibility:
|
||||
- By default, all models are saved with the default pytorch serialization, so we use the list of default pytorch
|
||||
files to know which safetensors files are needed.
|
||||
- The model is safetensors compatible only if there is a matching safetensors file for every default pytorch file.
|
||||
- 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"
|
||||
"""
|
||||
pt_filenames = []
|
||||
|
||||
sf_filenames = set()
|
||||
|
||||
passed_components = passed_components or []
|
||||
if folder_names is not None:
|
||||
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:
|
||||
_, extension = os.path.splitext(filename)
|
||||
|
||||
if len(filename.split("/")) == 2 and filename.split("/")[0] in passed_components:
|
||||
if not len(filename.split("/")) == 2:
|
||||
continue
|
||||
|
||||
if extension == ".bin":
|
||||
pt_filenames.append(os.path.normpath(filename))
|
||||
elif extension == ".safetensors":
|
||||
sf_filenames.add(os.path.normpath(filename))
|
||||
component, component_filename = filename.split("/")
|
||||
if component in passed_components:
|
||||
continue
|
||||
|
||||
for filename in pt_filenames:
|
||||
# filename = 'foo/bar/baz.bam' -> path = 'foo/bar', filename = 'baz', extension = '.bam'
|
||||
path, filename = os.path.split(filename)
|
||||
filename, extension = os.path.splitext(filename)
|
||||
components.setdefault(component, [])
|
||||
components[component].append(component_filename)
|
||||
|
||||
if filename.startswith("pytorch_model"):
|
||||
filename = filename.replace("pytorch_model", "model")
|
||||
else:
|
||||
filename = filename
|
||||
# iterate over all files of a component
|
||||
# check if safetensor files exist for that component
|
||||
# if variant is provided check if the variant of the safetensors exists
|
||||
for component, component_filenames in components.items():
|
||||
matches = []
|
||||
for component_filename in component_filenames:
|
||||
filename, extension = os.path.splitext(component_filename)
|
||||
|
||||
expected_sf_filename = os.path.normpath(os.path.join(path, filename))
|
||||
expected_sf_filename = f"{expected_sf_filename}.safetensors"
|
||||
if expected_sf_filename not in sf_filenames:
|
||||
logger.warning(f"{expected_sf_filename} not found")
|
||||
match_exists = extension == ".safetensors"
|
||||
matches.append(match_exists)
|
||||
|
||||
if not any(matches):
|
||||
return False
|
||||
|
||||
return True
|
||||
@@ -749,3 +747,92 @@ def _fetch_class_library_tuple(module):
|
||||
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/runwayml/stable-diffusion-inpainting instead or adapting your"
|
||||
f" checkpoint {pretrained_model_name_or_path} to the format of"
|
||||
" https://huggingface.co/runwayml/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 <prefix>_<component_name>, e.g. "prior_text_encoder"
|
||||
init_kwargs.update(
|
||||
{"_".join([prefix, name]): component for name, component in connected_pipe.components.items()}
|
||||
)
|
||||
|
||||
return init_kwargs
|
||||
|
||||
@@ -75,7 +75,11 @@ from .pipeline_loading_utils import (
|
||||
_get_custom_pipeline_class,
|
||||
_get_final_device_map,
|
||||
_get_pipeline_class,
|
||||
_identify_model_variants,
|
||||
_maybe_raise_warning_for_inpainting,
|
||||
_resolve_custom_pipeline_and_cls,
|
||||
_unwrap_model,
|
||||
_update_init_kwargs_with_connected_pipeline,
|
||||
is_safetensors_compatible,
|
||||
load_sub_model,
|
||||
maybe_raise_or_warn,
|
||||
@@ -622,6 +626,9 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
|
||||
>>> 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)
|
||||
@@ -722,33 +729,19 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
|
||||
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
|
||||
model_variants = {}
|
||||
if variant is not None:
|
||||
for folder in os.listdir(cached_folder):
|
||||
folder_path = os.path.join(cached_folder, folder)
|
||||
is_folder = os.path.isdir(folder_path) and folder in config_dict
|
||||
variant_exists = is_folder and any(
|
||||
p.split(".")[1].startswith(variant) for p in os.listdir(folder_path)
|
||||
)
|
||||
if variant_exists:
|
||||
model_variants[folder] = variant
|
||||
# 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)
|
||||
|
||||
# 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_class_name = None
|
||||
if os.path.isfile(os.path.join(cached_folder, f"{custom_pipeline}.py")):
|
||||
custom_pipeline = os.path.join(cached_folder, f"{custom_pipeline}.py")
|
||||
elif isinstance(config_dict["_class_name"], (list, tuple)) and os.path.isfile(
|
||||
os.path.join(cached_folder, f"{config_dict['_class_name'][0]}.py")
|
||||
):
|
||||
custom_pipeline = os.path.join(cached_folder, f"{config_dict['_class_name'][0]}.py")
|
||||
custom_class_name = config_dict["_class_name"][1]
|
||||
|
||||
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_dict,
|
||||
config=config_dict,
|
||||
load_connected_pipeline=load_connected_pipeline,
|
||||
custom_pipeline=custom_pipeline,
|
||||
class_name=custom_class_name,
|
||||
@@ -760,23 +753,13 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
|
||||
raise NotImplementedError("`device_map` is not yet supported for connected pipelines.")
|
||||
|
||||
# DEPRECATED: To be removed in 1.0.0
|
||||
if pipeline_class.__name__ == "StableDiffusionInpaintPipeline" and version.parse(
|
||||
version.parse(config_dict["_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/runwayml/stable-diffusion-inpainting instead or adapting your"
|
||||
f" checkpoint {pretrained_model_name_or_path} to the format of"
|
||||
" https://huggingface.co/runwayml/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)
|
||||
# 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`)
|
||||
@@ -787,7 +770,6 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
|
||||
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 and make sure that optional component modules are filtered out
|
||||
@@ -847,6 +829,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
|
||||
# 7. Load each module in the pipeline
|
||||
current_device_map = None
|
||||
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:
|
||||
@@ -854,15 +837,15 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
|
||||
else:
|
||||
current_device_map = None
|
||||
|
||||
# 7.1 - now that JAX/Flax is an official framework of the library, we might load from Flax names
|
||||
# 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
|
||||
|
||||
# 7.2 Define all importable classes
|
||||
# 7.3 Define all importable classes
|
||||
is_pipeline_module = hasattr(pipelines, library_name)
|
||||
importable_classes = ALL_IMPORTABLE_CLASSES
|
||||
loaded_sub_model = None
|
||||
|
||||
# 7.3 Use passed sub model or load class_name from library_name
|
||||
# 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
|
||||
@@ -900,56 +883,17 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
|
||||
|
||||
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")):
|
||||
modelcard = ModelCard.load(os.path.join(cached_folder, "README.md"))
|
||||
connected_pipes = {prefix: getattr(modelcard.data, prefix, [None])[0] for prefix in CONNECTED_PIPES_KEYS}
|
||||
load_kwargs = {
|
||||
"cache_dir": cache_dir,
|
||||
"force_download": force_download,
|
||||
"proxies": proxies,
|
||||
"local_files_only": local_files_only,
|
||||
"token": token,
|
||||
"revision": revision,
|
||||
"torch_dtype": torch_dtype,
|
||||
"custom_pipeline": custom_pipeline,
|
||||
"custom_revision": custom_revision,
|
||||
"provider": provider,
|
||||
"sess_options": sess_options,
|
||||
"device_map": device_map,
|
||||
"max_memory": max_memory,
|
||||
"offload_folder": offload_folder,
|
||||
"offload_state_dict": offload_state_dict,
|
||||
"low_cpu_mem_usage": low_cpu_mem_usage,
|
||||
"variant": variant,
|
||||
"use_safetensors": use_safetensors,
|
||||
}
|
||||
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,
|
||||
)
|
||||
|
||||
def get_connected_passed_kwargs(prefix):
|
||||
connected_passed_class_obj = {
|
||||
k.replace(f"{prefix}_", ""): w for k, w in passed_class_obj.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, **load_kwargs.copy(), **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 <prefix>_<component_name>, e.g. "prior_text_encoder"
|
||||
init_kwargs.update(
|
||||
{"_".join([prefix, name]): component for name, component in connected_pipe.components.items()}
|
||||
)
|
||||
|
||||
# 8. Potentially add passed objects if expected
|
||||
# 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
|
||||
@@ -1417,7 +1361,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
|
||||
use_safetensors
|
||||
and not allow_pickle
|
||||
and not is_safetensors_compatible(
|
||||
model_filenames, variant=variant, passed_components=passed_components
|
||||
model_filenames, passed_components=passed_components, folder_names=model_folder_names
|
||||
)
|
||||
):
|
||||
raise EnvironmentError(
|
||||
@@ -1426,7 +1370,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
|
||||
if from_flax:
|
||||
ignore_patterns = ["*.bin", "*.safetensors", "*.onnx", "*.pb"]
|
||||
elif use_safetensors and is_safetensors_compatible(
|
||||
model_filenames, variant=variant, passed_components=passed_components
|
||||
model_filenames, passed_components=passed_components, folder_names=model_folder_names
|
||||
):
|
||||
ignore_patterns = ["*.bin", "*.msgpack"]
|
||||
|
||||
|
||||
@@ -281,6 +281,16 @@ class StableCascadeDecoderPipeline(DiffusionPipeline):
|
||||
def num_timesteps(self):
|
||||
return self._num_timesteps
|
||||
|
||||
def get_timestep_ratio_conditioning(self, t, alphas_cumprod):
|
||||
s = torch.tensor([0.008])
|
||||
clamp_range = [0, 1]
|
||||
min_var = torch.cos(s / (1 + s) * torch.pi * 0.5) ** 2
|
||||
var = alphas_cumprod[t]
|
||||
var = var.clamp(*clamp_range)
|
||||
s, min_var = s.to(var.device), min_var.to(var.device)
|
||||
ratio = (((var * min_var) ** 0.5).acos() / (torch.pi * 0.5)) * (1 + s) - s
|
||||
return ratio
|
||||
|
||||
@torch.no_grad()
|
||||
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
||||
def __call__(
|
||||
@@ -434,10 +444,30 @@ class StableCascadeDecoderPipeline(DiffusionPipeline):
|
||||
batch_size, image_embeddings, num_images_per_prompt, dtype, device, generator, latents, self.scheduler
|
||||
)
|
||||
|
||||
if isinstance(self.scheduler, DDPMWuerstchenScheduler):
|
||||
timesteps = timesteps[:-1]
|
||||
else:
|
||||
if hasattr(self.scheduler.config, "clip_sample") and self.scheduler.config.clip_sample:
|
||||
self.scheduler.config.clip_sample = False # disample sample clipping
|
||||
logger.warning(" set `clip_sample` to be False")
|
||||
|
||||
# 6. Run denoising loop
|
||||
self._num_timesteps = len(timesteps[:-1])
|
||||
for i, t in enumerate(self.progress_bar(timesteps[:-1])):
|
||||
timestep_ratio = t.expand(latents.size(0)).to(dtype)
|
||||
if hasattr(self.scheduler, "betas"):
|
||||
alphas = 1.0 - self.scheduler.betas
|
||||
alphas_cumprod = torch.cumprod(alphas, dim=0)
|
||||
else:
|
||||
alphas_cumprod = []
|
||||
|
||||
self._num_timesteps = len(timesteps)
|
||||
for i, t in enumerate(self.progress_bar(timesteps)):
|
||||
if not isinstance(self.scheduler, DDPMWuerstchenScheduler):
|
||||
if len(alphas_cumprod) > 0:
|
||||
timestep_ratio = self.get_timestep_ratio_conditioning(t.long().cpu(), alphas_cumprod)
|
||||
timestep_ratio = timestep_ratio.expand(latents.size(0)).to(dtype).to(device)
|
||||
else:
|
||||
timestep_ratio = t.float().div(self.scheduler.timesteps[-1]).expand(latents.size(0)).to(dtype)
|
||||
else:
|
||||
timestep_ratio = t.expand(latents.size(0)).to(dtype)
|
||||
|
||||
# 7. Denoise latents
|
||||
predicted_latents = self.decoder(
|
||||
@@ -454,6 +484,8 @@ class StableCascadeDecoderPipeline(DiffusionPipeline):
|
||||
predicted_latents = torch.lerp(predicted_latents_uncond, predicted_latents_text, self.guidance_scale)
|
||||
|
||||
# 9. Renoise latents to next timestep
|
||||
if not isinstance(self.scheduler, DDPMWuerstchenScheduler):
|
||||
timestep_ratio = t
|
||||
latents = self.scheduler.step(
|
||||
model_output=predicted_latents,
|
||||
timestep=timestep_ratio,
|
||||
|
||||
@@ -353,7 +353,7 @@ class StableCascadePriorPipeline(DiffusionPipeline):
|
||||
return self._num_timesteps
|
||||
|
||||
def get_timestep_ratio_conditioning(self, t, alphas_cumprod):
|
||||
s = torch.tensor([0.003])
|
||||
s = torch.tensor([0.008])
|
||||
clamp_range = [0, 1]
|
||||
min_var = torch.cos(s / (1 + s) * torch.pi * 0.5) ** 2
|
||||
var = alphas_cumprod[t]
|
||||
@@ -557,7 +557,7 @@ class StableCascadePriorPipeline(DiffusionPipeline):
|
||||
if isinstance(self.scheduler, DDPMWuerstchenScheduler):
|
||||
timesteps = timesteps[:-1]
|
||||
else:
|
||||
if self.scheduler.config.clip_sample:
|
||||
if hasattr(self.scheduler.config, "clip_sample") and self.scheduler.config.clip_sample:
|
||||
self.scheduler.config.clip_sample = False # disample sample clipping
|
||||
logger.warning(" set `clip_sample` to be False")
|
||||
# 6. Run denoising loop
|
||||
|
||||
@@ -33,6 +33,20 @@ from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput, StableDiffu
|
||||
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(
|
||||
@@ -105,7 +119,54 @@ class StableDiffusionLatentUpscalePipeline(DiffusionPipeline, StableDiffusionMix
|
||||
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
|
||||
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):
|
||||
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.
|
||||
|
||||
@@ -119,81 +180,100 @@ class StableDiffusionLatentUpscalePipeline(DiffusionPipeline, StableDiffusionMix
|
||||
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.
|
||||
"""
|
||||
batch_size = len(prompt) if isinstance(prompt, list) else 1
|
||||
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]
|
||||
|
||||
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,
|
||||
)
|
||||
text_embeddings = text_encoder_out.hidden_states[-1]
|
||||
text_pooler_out = text_encoder_out.pooler_output
|
||||
|
||||
# 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,
|
||||
if prompt_embeds is None or pooled_prompt_embeds is None:
|
||||
text_inputs = self.tokenizer(
|
||||
prompt,
|
||||
padding="max_length",
|
||||
max_length=max_length,
|
||||
max_length=self.tokenizer.model_max_length,
|
||||
truncation=True,
|
||||
return_length=True,
|
||||
return_tensors="pt",
|
||||
)
|
||||
text_input_ids = text_inputs.input_ids
|
||||
|
||||
uncond_encoder_out = self.text_encoder(
|
||||
uncond_input.input_ids.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_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
|
||||
|
||||
uncond_embeddings = uncond_encoder_out.hidden_states[-1]
|
||||
uncond_pooler_out = uncond_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
|
||||
|
||||
# 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_embeddings = torch.cat([uncond_embeddings, text_embeddings])
|
||||
text_pooler_out = torch.cat([uncond_pooler_out, text_pooler_out])
|
||||
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",
|
||||
)
|
||||
|
||||
return text_embeddings, text_pooler_out
|
||||
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):
|
||||
@@ -207,12 +287,56 @@ class StableDiffusionLatentUpscalePipeline(DiffusionPipeline, StableDiffusionMix
|
||||
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
||||
return image
|
||||
|
||||
def check_inputs(self, prompt, image, callback_steps):
|
||||
if not isinstance(prompt, str) and not isinstance(prompt, list):
|
||||
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)
|
||||
):
|
||||
@@ -222,10 +346,14 @@ class StableDiffusionLatentUpscalePipeline(DiffusionPipeline, StableDiffusionMix
|
||||
|
||||
# verify batch size of prompt and image are same if image is a list or tensor
|
||||
if isinstance(image, (list, torch.Tensor)):
|
||||
if isinstance(prompt, str):
|
||||
batch_size = 1
|
||||
if prompt is not None:
|
||||
if isinstance(prompt, str):
|
||||
batch_size = 1
|
||||
else:
|
||||
batch_size = len(prompt)
|
||||
else:
|
||||
batch_size = len(prompt)
|
||||
batch_size = prompt_embeds.shape[0]
|
||||
|
||||
if isinstance(image, list):
|
||||
image_batch_size = len(image)
|
||||
else:
|
||||
@@ -261,13 +389,17 @@ class StableDiffusionLatentUpscalePipeline(DiffusionPipeline, StableDiffusionMix
|
||||
@torch.no_grad()
|
||||
def __call__(
|
||||
self,
|
||||
prompt: Union[str, List[str]],
|
||||
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,
|
||||
@@ -359,10 +491,22 @@ class StableDiffusionLatentUpscalePipeline(DiffusionPipeline, StableDiffusionMix
|
||||
"""
|
||||
|
||||
# 1. Check inputs
|
||||
self.check_inputs(prompt, image, callback_steps)
|
||||
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
|
||||
batch_size = 1 if isinstance(prompt, str) else len(prompt)
|
||||
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://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
||||
@@ -373,16 +517,32 @@ class StableDiffusionLatentUpscalePipeline(DiffusionPipeline, StableDiffusionMix
|
||||
prompt = [""] * batch_size
|
||||
|
||||
# 3. Encode input prompt
|
||||
text_embeddings, text_pooler_out = self._encode_prompt(
|
||||
prompt, device, do_classifier_free_guidance, negative_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=text_embeddings.dtype, device=device)
|
||||
image = image.to(dtype=prompt_embeds.dtype, device=device)
|
||||
if image.shape[1] == 3:
|
||||
# encode image if not in latent-space yet
|
||||
image = self.vae.encode(image).latent_dist.sample() * self.vae.config.scaling_factor
|
||||
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)
|
||||
@@ -400,17 +560,17 @@ class StableDiffusionLatentUpscalePipeline(DiffusionPipeline, StableDiffusionMix
|
||||
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(text_embeddings.dtype)
|
||||
image_cond = image_cond.to(prompt_embeds.dtype)
|
||||
|
||||
noise_level_embed = torch.cat(
|
||||
[
|
||||
torch.ones(text_pooler_out.shape[0], 64, dtype=text_pooler_out.dtype, device=device),
|
||||
torch.zeros(text_pooler_out.shape[0], 64, dtype=text_pooler_out.dtype, device=device),
|
||||
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, text_pooler_out], dim=1)
|
||||
timestep_condition = torch.cat([noise_level_embed, pooled_prompt_embeds], dim=1)
|
||||
|
||||
# 6. Prepare latent variables
|
||||
height, width = image.shape[2:]
|
||||
@@ -420,7 +580,7 @@ class StableDiffusionLatentUpscalePipeline(DiffusionPipeline, StableDiffusionMix
|
||||
num_channels_latents,
|
||||
height * 2, # 2x upscale
|
||||
width * 2,
|
||||
text_embeddings.dtype,
|
||||
prompt_embeds.dtype,
|
||||
device,
|
||||
generator,
|
||||
latents,
|
||||
@@ -454,7 +614,7 @@ class StableDiffusionLatentUpscalePipeline(DiffusionPipeline, StableDiffusionMix
|
||||
noise_pred = self.unet(
|
||||
scaled_model_input,
|
||||
timestep,
|
||||
encoder_hidden_states=text_embeddings,
|
||||
encoder_hidden_states=prompt_embeds,
|
||||
timestep_cond=timestep_condition,
|
||||
).sample
|
||||
|
||||
|
||||
@@ -1637,6 +1637,21 @@ class StableDiffusionXLControlNetInpaintPipeline(metaclass=DummyObject):
|
||||
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"]
|
||||
|
||||
|
||||
@@ -12,19 +12,26 @@
|
||||
# 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 sys
|
||||
import tempfile
|
||||
import unittest
|
||||
|
||||
import numpy as np
|
||||
import safetensors.torch
|
||||
import torch
|
||||
from transformers import AutoTokenizer, CLIPTextModel, CLIPTokenizer, T5EncoderModel
|
||||
|
||||
from diffusers import FlowMatchEulerDiscreteScheduler, FluxPipeline, FluxTransformer2DModel
|
||||
from diffusers.utils.testing_utils import floats_tensor, require_peft_backend
|
||||
from diffusers.utils.testing_utils import floats_tensor, is_peft_available, require_peft_backend, torch_device
|
||||
|
||||
|
||||
if is_peft_available():
|
||||
from peft.utils import get_peft_model_state_dict
|
||||
|
||||
sys.path.append(".")
|
||||
|
||||
from utils import PeftLoraLoaderMixinTests # noqa: E402
|
||||
from utils import PeftLoraLoaderMixinTests, check_if_lora_correctly_set # noqa: E402
|
||||
|
||||
|
||||
@require_peft_backend
|
||||
@@ -90,3 +97,51 @@ class FluxLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
|
||||
pipeline_inputs.update({"generator": generator})
|
||||
|
||||
return noise, input_ids, pipeline_inputs
|
||||
|
||||
def test_with_alpha_in_state_dict(self):
|
||||
components, _, denoiser_lora_config = self.get_dummy_components(FlowMatchEulerDiscreteScheduler)
|
||||
pipe = self.pipeline_class(**components)
|
||||
pipe = pipe.to(torch_device)
|
||||
pipe.set_progress_bar_config(disable=None)
|
||||
_, _, inputs = self.get_dummy_inputs(with_generator=False)
|
||||
|
||||
output_no_lora = pipe(**inputs, generator=torch.manual_seed(0)).images
|
||||
self.assertTrue(output_no_lora.shape == self.output_shape)
|
||||
|
||||
pipe.transformer.add_adapter(denoiser_lora_config)
|
||||
self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in transformer")
|
||||
|
||||
images_lora = pipe(**inputs, generator=torch.manual_seed(0)).images
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
denoiser_state_dict = get_peft_model_state_dict(pipe.transformer)
|
||||
self.pipeline_class.save_lora_weights(tmpdirname, transformer_lora_layers=denoiser_state_dict)
|
||||
|
||||
self.assertTrue(os.path.isfile(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors")))
|
||||
pipe.unload_lora_weights()
|
||||
pipe.load_lora_weights(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors"))
|
||||
|
||||
# modify the state dict to have alpha values following
|
||||
# https://huggingface.co/TheLastBen/Jon_Snow_Flux_LoRA/blob/main/jon_snow.safetensors
|
||||
state_dict_with_alpha = safetensors.torch.load_file(
|
||||
os.path.join(tmpdirname, "pytorch_lora_weights.safetensors")
|
||||
)
|
||||
alpha_dict = {}
|
||||
for k, v in state_dict_with_alpha.items():
|
||||
# only do for `transformer` and for the k projections -- should be enough to test.
|
||||
if "transformer" in k and "to_k" in k and "lora_A" in k:
|
||||
alpha_dict[f"{k}.alpha"] = float(torch.randint(10, 100, size=()))
|
||||
state_dict_with_alpha.update(alpha_dict)
|
||||
|
||||
images_lora_from_pretrained = pipe(**inputs, generator=torch.manual_seed(0)).images
|
||||
self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in denoiser")
|
||||
|
||||
pipe.unload_lora_weights()
|
||||
pipe.load_lora_weights(state_dict_with_alpha)
|
||||
images_lora_with_alpha = pipe(**inputs, generator=torch.manual_seed(0)).images
|
||||
|
||||
self.assertTrue(
|
||||
np.allclose(images_lora, images_lora_from_pretrained, atol=1e-3, rtol=1e-3),
|
||||
"Loading from saved checkpoints should give same results.",
|
||||
)
|
||||
self.assertFalse(np.allclose(images_lora_with_alpha, images_lora, atol=1e-3, rtol=1e-3))
|
||||
|
||||
@@ -32,7 +32,7 @@ from utils import PeftLoraLoaderMixinTests # noqa: E402
|
||||
@require_peft_backend
|
||||
class SD3LoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
|
||||
pipeline_class = StableDiffusion3Pipeline
|
||||
scheduler_cls = FlowMatchEulerDiscreteScheduler()
|
||||
scheduler_cls = FlowMatchEulerDiscreteScheduler
|
||||
scheduler_kwargs = {}
|
||||
uses_flow_matching = True
|
||||
transformer_kwargs = {
|
||||
@@ -80,8 +80,7 @@ class SD3LoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
|
||||
Related PR: https://github.com/huggingface/diffusers/pull/8584
|
||||
"""
|
||||
components = self.get_dummy_components()
|
||||
|
||||
pipe = self.pipeline_class(**components)
|
||||
pipe = self.pipeline_class(**components[0])
|
||||
pipe = pipe.to(torch_device)
|
||||
pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
|
||||
@@ -124,71 +124,6 @@ class LoraSDXLIntegrationTests(unittest.TestCase):
|
||||
gc.collect()
|
||||
torch.cuda.empty_cache()
|
||||
|
||||
def test_sdxl_0_9_lora_one(self):
|
||||
generator = torch.Generator().manual_seed(0)
|
||||
|
||||
pipe = StableDiffusionXLPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-0.9")
|
||||
lora_model_id = "hf-internal-testing/sdxl-0.9-daiton-lora"
|
||||
lora_filename = "daiton-xl-lora-test.safetensors"
|
||||
pipe.load_lora_weights(lora_model_id, weight_name=lora_filename)
|
||||
pipe.enable_model_cpu_offload()
|
||||
|
||||
images = pipe(
|
||||
"masterpiece, best quality, mountain", output_type="np", generator=generator, num_inference_steps=2
|
||||
).images
|
||||
|
||||
images = images[0, -3:, -3:, -1].flatten()
|
||||
expected = np.array([0.3838, 0.3482, 0.3588, 0.3162, 0.319, 0.3369, 0.338, 0.3366, 0.3213])
|
||||
|
||||
max_diff = numpy_cosine_similarity_distance(expected, images)
|
||||
assert max_diff < 1e-3
|
||||
pipe.unload_lora_weights()
|
||||
release_memory(pipe)
|
||||
|
||||
def test_sdxl_0_9_lora_two(self):
|
||||
generator = torch.Generator().manual_seed(0)
|
||||
|
||||
pipe = StableDiffusionXLPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-0.9")
|
||||
lora_model_id = "hf-internal-testing/sdxl-0.9-costumes-lora"
|
||||
lora_filename = "saijo.safetensors"
|
||||
pipe.load_lora_weights(lora_model_id, weight_name=lora_filename)
|
||||
pipe.enable_model_cpu_offload()
|
||||
|
||||
images = pipe(
|
||||
"masterpiece, best quality, mountain", output_type="np", generator=generator, num_inference_steps=2
|
||||
).images
|
||||
|
||||
images = images[0, -3:, -3:, -1].flatten()
|
||||
expected = np.array([0.3137, 0.3269, 0.3355, 0.255, 0.2577, 0.2563, 0.2679, 0.2758, 0.2626])
|
||||
|
||||
max_diff = numpy_cosine_similarity_distance(expected, images)
|
||||
assert max_diff < 1e-3
|
||||
|
||||
pipe.unload_lora_weights()
|
||||
release_memory(pipe)
|
||||
|
||||
def test_sdxl_0_9_lora_three(self):
|
||||
generator = torch.Generator().manual_seed(0)
|
||||
|
||||
pipe = StableDiffusionXLPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-0.9")
|
||||
lora_model_id = "hf-internal-testing/sdxl-0.9-kamepan-lora"
|
||||
lora_filename = "kame_sdxl_v2-000020-16rank.safetensors"
|
||||
pipe.load_lora_weights(lora_model_id, weight_name=lora_filename)
|
||||
pipe.enable_model_cpu_offload()
|
||||
|
||||
images = pipe(
|
||||
"masterpiece, best quality, mountain", output_type="np", generator=generator, num_inference_steps=2
|
||||
).images
|
||||
|
||||
images = images[0, -3:, -3:, -1].flatten()
|
||||
expected = np.array([0.4015, 0.3761, 0.3616, 0.3745, 0.3462, 0.3337, 0.3564, 0.3649, 0.3468])
|
||||
|
||||
max_diff = numpy_cosine_similarity_distance(expected, images)
|
||||
assert max_diff < 5e-3
|
||||
|
||||
pipe.unload_lora_weights()
|
||||
release_memory(pipe)
|
||||
|
||||
def test_sdxl_1_0_lora(self):
|
||||
generator = torch.Generator("cpu").manual_seed(0)
|
||||
|
||||
|
||||
@@ -976,7 +976,6 @@ class ModelTesterMixin:
|
||||
self.assertTrue(actual_num_shards == expected_num_shards)
|
||||
|
||||
new_model = self.model_class.from_pretrained(tmp_dir, device_map="auto")
|
||||
new_model = new_model.to(torch_device)
|
||||
|
||||
torch.manual_seed(0)
|
||||
if "generator" in inputs_dict:
|
||||
|
||||
@@ -26,7 +26,7 @@ from ..test_modeling_common import ModelTesterMixin
|
||||
enable_full_determinism()
|
||||
|
||||
|
||||
class SD3TransformerTests(ModelTesterMixin, unittest.TestCase):
|
||||
class AuraFlowTransformerTests(ModelTesterMixin, unittest.TestCase):
|
||||
model_class = AuraFlowTransformer2DModel
|
||||
main_input_name = "hidden_states"
|
||||
# We override the items here because the transformer under consideration is small.
|
||||
@@ -73,3 +73,7 @@ class SD3TransformerTests(ModelTesterMixin, unittest.TestCase):
|
||||
}
|
||||
inputs_dict = self.dummy_input
|
||||
return init_dict, inputs_dict
|
||||
|
||||
@unittest.skip("AuraFlowTransformer2DModel uses its own dedicated attention processor. This test does not apply")
|
||||
def test_set_attn_processor_for_determinism(self):
|
||||
pass
|
||||
|
||||
@@ -44,8 +44,8 @@ class FluxTransformerTests(ModelTesterMixin, unittest.TestCase):
|
||||
hidden_states = torch.randn((batch_size, height * width, num_latent_channels)).to(torch_device)
|
||||
encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
|
||||
pooled_prompt_embeds = torch.randn((batch_size, embedding_dim)).to(torch_device)
|
||||
text_ids = torch.randn((batch_size, sequence_length, num_image_channels)).to(torch_device)
|
||||
image_ids = torch.randn((batch_size, height * width, num_image_channels)).to(torch_device)
|
||||
text_ids = torch.randn((sequence_length, num_image_channels)).to(torch_device)
|
||||
image_ids = torch.randn((height * width, num_image_channels)).to(torch_device)
|
||||
timestep = torch.tensor([1.0]).to(torch_device).expand(batch_size)
|
||||
|
||||
return {
|
||||
@@ -80,3 +80,31 @@ class FluxTransformerTests(ModelTesterMixin, unittest.TestCase):
|
||||
|
||||
inputs_dict = self.dummy_input
|
||||
return init_dict, inputs_dict
|
||||
|
||||
def test_deprecated_inputs_img_txt_ids_3d(self):
|
||||
init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
|
||||
model = self.model_class(**init_dict)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
|
||||
with torch.no_grad():
|
||||
output_1 = model(**inputs_dict).to_tuple()[0]
|
||||
|
||||
# update inputs_dict with txt_ids and img_ids as 3d tensors (deprecated)
|
||||
text_ids_3d = inputs_dict["txt_ids"].unsqueeze(0)
|
||||
image_ids_3d = inputs_dict["img_ids"].unsqueeze(0)
|
||||
|
||||
assert text_ids_3d.ndim == 3, "text_ids_3d should be a 3d tensor"
|
||||
assert image_ids_3d.ndim == 3, "img_ids_3d should be a 3d tensor"
|
||||
|
||||
inputs_dict["txt_ids"] = text_ids_3d
|
||||
inputs_dict["img_ids"] = image_ids_3d
|
||||
|
||||
with torch.no_grad():
|
||||
output_2 = model(**inputs_dict).to_tuple()[0]
|
||||
|
||||
self.assertEqual(output_1.shape, output_2.shape)
|
||||
self.assertTrue(
|
||||
torch.allclose(output_1, output_2, atol=1e-5),
|
||||
msg="output with deprecated inputs (img_ids and txt_ids as 3d torch tensors) are not equal as them as 2d inputs",
|
||||
)
|
||||
|
||||
@@ -76,3 +76,7 @@ class SD3TransformerTests(ModelTesterMixin, unittest.TestCase):
|
||||
}
|
||||
inputs_dict = self.dummy_input
|
||||
return init_dict, inputs_dict
|
||||
|
||||
@unittest.skip("SD3Transformer2DModel uses a dedicated attention processor. This test doesn't apply")
|
||||
def test_set_attn_processor_for_determinism(self):
|
||||
pass
|
||||
|
||||
@@ -20,6 +20,7 @@ from diffusers import (
|
||||
)
|
||||
from diffusers.models.attention import FreeNoiseTransformerBlock
|
||||
from diffusers.utils import logging
|
||||
from diffusers.utils.import_utils import is_xformers_available
|
||||
from diffusers.utils.testing_utils import torch_device
|
||||
|
||||
from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_PARAMS
|
||||
@@ -329,6 +330,13 @@ class AnimateDiffControlNetPipelineFastTests(
|
||||
inputs["prompt_embeds"] = torch.randn((1, 4, pipe.text_encoder.config.hidden_size), device=torch_device)
|
||||
pipe(**inputs)
|
||||
|
||||
@unittest.skipIf(
|
||||
torch_device != "cuda" or not is_xformers_available(),
|
||||
reason="XFormers attention is only available with CUDA and `xformers` installed",
|
||||
)
|
||||
def test_xformers_attention_forwardGenerator_pass(self):
|
||||
super()._test_xformers_attention_forwardGenerator_pass(test_mean_pixel_difference=False)
|
||||
|
||||
def test_free_init(self):
|
||||
components = self.get_dummy_components()
|
||||
pipe: AnimateDiffControlNetPipeline = self.pipeline_class(**components)
|
||||
|
||||
@@ -19,6 +19,7 @@ from diffusers import (
|
||||
UNetMotionModel,
|
||||
)
|
||||
from diffusers.utils import logging
|
||||
from diffusers.utils.import_utils import is_xformers_available
|
||||
from diffusers.utils.testing_utils import torch_device
|
||||
|
||||
from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_PARAMS
|
||||
@@ -393,6 +394,13 @@ class AnimateDiffSparseControlNetPipelineFastTests(
|
||||
inputs["prompt_embeds"] = torch.randn((1, 4, pipe.text_encoder.config.hidden_size), device=torch_device)
|
||||
pipe(**inputs)
|
||||
|
||||
@unittest.skipIf(
|
||||
torch_device != "cuda" or not is_xformers_available(),
|
||||
reason="XFormers attention is only available with CUDA and `xformers` installed",
|
||||
)
|
||||
def test_xformers_attention_forwardGenerator_pass(self):
|
||||
super()._test_xformers_attention_forwardGenerator_pass(test_mean_pixel_difference=False)
|
||||
|
||||
def test_free_init(self):
|
||||
components = self.get_dummy_components()
|
||||
pipe: AnimateDiffSparseControlNetPipeline = self.pipeline_class(**components)
|
||||
|
||||
@@ -163,3 +163,7 @@ class AuraFlowPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
|
||||
assert np.allclose(
|
||||
original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
|
||||
), "Original outputs should match when fused QKV projections are disabled."
|
||||
|
||||
@unittest.skip("xformers attention processor does not exist for AuraFlow")
|
||||
def test_xformers_attention_forwardGenerator_pass(self):
|
||||
pass
|
||||
|
||||
@@ -30,7 +30,12 @@ from diffusers.utils.testing_utils import (
|
||||
)
|
||||
|
||||
from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
|
||||
from ..test_pipelines_common import PipelineTesterMixin, to_np
|
||||
from ..test_pipelines_common import (
|
||||
PipelineTesterMixin,
|
||||
check_qkv_fusion_matches_attn_procs_length,
|
||||
check_qkv_fusion_processors_exist,
|
||||
to_np,
|
||||
)
|
||||
|
||||
|
||||
enable_full_determinism()
|
||||
@@ -275,6 +280,48 @@ class CogVideoXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
|
||||
"VAE tiling should not affect the inference results",
|
||||
)
|
||||
|
||||
@unittest.skip("xformers attention processor does not exist for CogVideoX")
|
||||
def test_xformers_attention_forwardGenerator_pass(self):
|
||||
pass
|
||||
|
||||
def test_fused_qkv_projections(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
pipe = self.pipeline_class(**components)
|
||||
pipe = pipe.to(device)
|
||||
pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
frames = pipe(**inputs).frames # [B, F, C, H, W]
|
||||
original_image_slice = frames[0, -2:, -1, -3:, -3:]
|
||||
|
||||
pipe.fuse_qkv_projections()
|
||||
assert check_qkv_fusion_processors_exist(
|
||||
pipe.transformer
|
||||
), "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
|
||||
assert check_qkv_fusion_matches_attn_procs_length(
|
||||
pipe.transformer, pipe.transformer.original_attn_processors
|
||||
), "Something wrong with the attention processors concerning the fused QKV projections."
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
frames = pipe(**inputs).frames
|
||||
image_slice_fused = frames[0, -2:, -1, -3:, -3:]
|
||||
|
||||
pipe.transformer.unfuse_qkv_projections()
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
frames = pipe(**inputs).frames
|
||||
image_slice_disabled = frames[0, -2:, -1, -3:, -3:]
|
||||
|
||||
assert np.allclose(
|
||||
original_image_slice, image_slice_fused, atol=1e-3, rtol=1e-3
|
||||
), "Fusion of QKV projections shouldn't affect the outputs."
|
||||
assert np.allclose(
|
||||
image_slice_fused, image_slice_disabled, atol=1e-3, rtol=1e-3
|
||||
), "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled."
|
||||
assert np.allclose(
|
||||
original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
|
||||
), "Original outputs should match when fused QKV projections are disabled."
|
||||
|
||||
|
||||
@slow
|
||||
@require_torch_gpu
|
||||
|
||||
@@ -13,10 +13,13 @@ from diffusers.utils.testing_utils import (
|
||||
torch_device,
|
||||
)
|
||||
|
||||
from ..test_pipelines_common import PipelineTesterMixin
|
||||
from ..test_pipelines_common import (
|
||||
PipelineTesterMixin,
|
||||
check_qkv_fusion_matches_attn_procs_length,
|
||||
check_qkv_fusion_processors_exist,
|
||||
)
|
||||
|
||||
|
||||
@unittest.skipIf(torch_device == "mps", "Flux has a float64 operation which is not supported in MPS.")
|
||||
class FluxPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
|
||||
pipeline_class = FluxPipeline
|
||||
params = frozenset(["prompt", "height", "width", "guidance_scale", "prompt_embeds", "pooled_prompt_embeds"])
|
||||
@@ -143,6 +146,46 @@ class FluxPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
|
||||
max_diff = np.abs(output_with_prompt - output_with_embeds).max()
|
||||
assert max_diff < 1e-4
|
||||
|
||||
def test_fused_qkv_projections(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
pipe = self.pipeline_class(**components)
|
||||
pipe = pipe.to(device)
|
||||
pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
image = pipe(**inputs).images
|
||||
original_image_slice = image[0, -3:, -3:, -1]
|
||||
|
||||
# TODO (sayakpaul): will refactor this once `fuse_qkv_projections()` has been added
|
||||
# to the pipeline level.
|
||||
pipe.transformer.fuse_qkv_projections()
|
||||
assert check_qkv_fusion_processors_exist(
|
||||
pipe.transformer
|
||||
), "Something wrong with the fused attention processors. Expected all the attention processors to be fused."
|
||||
assert check_qkv_fusion_matches_attn_procs_length(
|
||||
pipe.transformer, pipe.transformer.original_attn_processors
|
||||
), "Something wrong with the attention processors concerning the fused QKV projections."
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
image = pipe(**inputs).images
|
||||
image_slice_fused = image[0, -3:, -3:, -1]
|
||||
|
||||
pipe.transformer.unfuse_qkv_projections()
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
image = pipe(**inputs).images
|
||||
image_slice_disabled = image[0, -3:, -3:, -1]
|
||||
|
||||
assert np.allclose(
|
||||
original_image_slice, image_slice_fused, atol=1e-3, rtol=1e-3
|
||||
), "Fusion of QKV projections shouldn't affect the outputs."
|
||||
assert np.allclose(
|
||||
image_slice_fused, image_slice_disabled, atol=1e-3, rtol=1e-3
|
||||
), "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled."
|
||||
assert np.allclose(
|
||||
original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
|
||||
), "Original outputs should match when fused QKV projections are disabled."
|
||||
|
||||
|
||||
@slow
|
||||
@require_torch_gpu
|
||||
|
||||
@@ -28,6 +28,7 @@ from diffusers import (
|
||||
LattePipeline,
|
||||
LatteTransformer3DModel,
|
||||
)
|
||||
from diffusers.utils.import_utils import is_xformers_available
|
||||
from diffusers.utils.testing_utils import (
|
||||
enable_full_determinism,
|
||||
numpy_cosine_similarity_distance,
|
||||
@@ -256,6 +257,13 @@ class LattePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
|
||||
max_diff = np.abs(to_np(output) - to_np(output_loaded)).max()
|
||||
self.assertLess(max_diff, 1.0)
|
||||
|
||||
@unittest.skipIf(
|
||||
torch_device != "cuda" or not is_xformers_available(),
|
||||
reason="XFormers attention is only available with CUDA and `xformers` installed",
|
||||
)
|
||||
def test_xformers_attention_forwardGenerator_pass(self):
|
||||
super()._test_xformers_attention_forwardGenerator_pass(test_mean_pixel_difference=False)
|
||||
|
||||
|
||||
@slow
|
||||
@require_torch_gpu
|
||||
|
||||
@@ -119,6 +119,10 @@ class LuminaText2ImgPipelinePipelineFastTests(unittest.TestCase, PipelineTesterM
|
||||
max_diff = np.abs(output_with_prompt - output_with_embeds).max()
|
||||
assert max_diff < 1e-4
|
||||
|
||||
@unittest.skip("xformers attention processor does not exist for Lumina")
|
||||
def test_xformers_attention_forwardGenerator_pass(self):
|
||||
pass
|
||||
|
||||
|
||||
@slow
|
||||
@require_torch_gpu
|
||||
|
||||
271
tests/pipelines/pag/test_pag_controlnet_sdxl_img2img.py
Normal file
271
tests/pipelines/pag/test_pag_controlnet_sdxl_img2img.py
Normal file
@@ -0,0 +1,271 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2024 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 inspect
|
||||
import random
|
||||
import unittest
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
|
||||
|
||||
from diffusers import (
|
||||
AutoencoderKL,
|
||||
ControlNetModel,
|
||||
EulerDiscreteScheduler,
|
||||
StableDiffusionXLControlNetImg2ImgPipeline,
|
||||
StableDiffusionXLControlNetPAGImg2ImgPipeline,
|
||||
UNet2DConditionModel,
|
||||
)
|
||||
from diffusers.utils.testing_utils import enable_full_determinism, floats_tensor
|
||||
|
||||
from ..pipeline_params import (
|
||||
IMAGE_TO_IMAGE_IMAGE_PARAMS,
|
||||
TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS,
|
||||
TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS,
|
||||
)
|
||||
from ..test_pipelines_common import (
|
||||
IPAdapterTesterMixin,
|
||||
PipelineFromPipeTesterMixin,
|
||||
PipelineLatentTesterMixin,
|
||||
PipelineTesterMixin,
|
||||
SDXLOptionalComponentsTesterMixin,
|
||||
)
|
||||
|
||||
|
||||
enable_full_determinism()
|
||||
|
||||
|
||||
class StableDiffusionXLControlNetPAGImg2ImgPipelineFastTests(
|
||||
IPAdapterTesterMixin,
|
||||
PipelineLatentTesterMixin,
|
||||
PipelineTesterMixin,
|
||||
PipelineFromPipeTesterMixin,
|
||||
SDXLOptionalComponentsTesterMixin,
|
||||
unittest.TestCase,
|
||||
):
|
||||
pipeline_class = StableDiffusionXLControlNetPAGImg2ImgPipeline
|
||||
params = TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS.union({"pag_scale", "pag_adaptive_scale"})
|
||||
batch_params = TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS
|
||||
image_params = IMAGE_TO_IMAGE_IMAGE_PARAMS
|
||||
image_latents_params = IMAGE_TO_IMAGE_IMAGE_PARAMS
|
||||
callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS.union(
|
||||
{"add_text_embeds", "add_time_ids", "add_neg_time_ids"}
|
||||
)
|
||||
|
||||
# Copied from tests.pipelines.controlnet.test_controlnet_sdxl_img2img.ControlNetPipelineSDXLImg2ImgFastTests.get_dummy_components
|
||||
def get_dummy_components(self, skip_first_text_encoder=False):
|
||||
torch.manual_seed(0)
|
||||
unet = UNet2DConditionModel(
|
||||
block_out_channels=(32, 64),
|
||||
layers_per_block=2,
|
||||
sample_size=32,
|
||||
in_channels=4,
|
||||
out_channels=4,
|
||||
down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
|
||||
up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
|
||||
# SD2-specific config below
|
||||
attention_head_dim=(2, 4),
|
||||
use_linear_projection=True,
|
||||
addition_embed_type="text_time",
|
||||
addition_time_embed_dim=8,
|
||||
transformer_layers_per_block=(1, 2),
|
||||
projection_class_embeddings_input_dim=80, # 6 * 8 + 32
|
||||
cross_attention_dim=64 if not skip_first_text_encoder else 32,
|
||||
)
|
||||
torch.manual_seed(0)
|
||||
controlnet = ControlNetModel(
|
||||
block_out_channels=(32, 64),
|
||||
layers_per_block=2,
|
||||
in_channels=4,
|
||||
down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
|
||||
conditioning_embedding_out_channels=(16, 32),
|
||||
# SD2-specific config below
|
||||
attention_head_dim=(2, 4),
|
||||
use_linear_projection=True,
|
||||
addition_embed_type="text_time",
|
||||
addition_time_embed_dim=8,
|
||||
transformer_layers_per_block=(1, 2),
|
||||
projection_class_embeddings_input_dim=80, # 6 * 8 + 32
|
||||
cross_attention_dim=64,
|
||||
)
|
||||
torch.manual_seed(0)
|
||||
scheduler = EulerDiscreteScheduler(
|
||||
beta_start=0.00085,
|
||||
beta_end=0.012,
|
||||
steps_offset=1,
|
||||
beta_schedule="scaled_linear",
|
||||
timestep_spacing="leading",
|
||||
)
|
||||
torch.manual_seed(0)
|
||||
vae = AutoencoderKL(
|
||||
block_out_channels=[32, 64],
|
||||
in_channels=3,
|
||||
out_channels=3,
|
||||
down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
|
||||
up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
|
||||
latent_channels=4,
|
||||
)
|
||||
torch.manual_seed(0)
|
||||
text_encoder_config = CLIPTextConfig(
|
||||
bos_token_id=0,
|
||||
eos_token_id=2,
|
||||
hidden_size=32,
|
||||
intermediate_size=37,
|
||||
layer_norm_eps=1e-05,
|
||||
num_attention_heads=4,
|
||||
num_hidden_layers=5,
|
||||
pad_token_id=1,
|
||||
vocab_size=1000,
|
||||
# SD2-specific config below
|
||||
hidden_act="gelu",
|
||||
projection_dim=32,
|
||||
)
|
||||
text_encoder = CLIPTextModel(text_encoder_config)
|
||||
tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
|
||||
|
||||
text_encoder_2 = CLIPTextModelWithProjection(text_encoder_config)
|
||||
tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
|
||||
|
||||
components = {
|
||||
"unet": unet,
|
||||
"controlnet": controlnet,
|
||||
"scheduler": scheduler,
|
||||
"vae": vae,
|
||||
"text_encoder": text_encoder if not skip_first_text_encoder else None,
|
||||
"tokenizer": tokenizer if not skip_first_text_encoder else None,
|
||||
"text_encoder_2": text_encoder_2,
|
||||
"tokenizer_2": tokenizer_2,
|
||||
"image_encoder": None,
|
||||
"feature_extractor": None,
|
||||
}
|
||||
return components
|
||||
|
||||
# based on tests.pipelines.controlnet.test_controlnet_sdxl_img2img.ControlNetPipelineSDXLImg2ImgFastTests.get_dummy_inputs
|
||||
# add `pag_scale` to the inputs
|
||||
def get_dummy_inputs(self, device, seed=0):
|
||||
controlnet_embedder_scale_factor = 2
|
||||
image = floats_tensor(
|
||||
(1, 3, 32 * controlnet_embedder_scale_factor, 32 * controlnet_embedder_scale_factor),
|
||||
rng=random.Random(seed),
|
||||
).to(device)
|
||||
|
||||
if str(device).startswith("mps"):
|
||||
generator = torch.manual_seed(seed)
|
||||
else:
|
||||
generator = torch.Generator(device=device).manual_seed(seed)
|
||||
|
||||
inputs = {
|
||||
"prompt": "A painting of a squirrel eating a burger",
|
||||
"generator": generator,
|
||||
"num_inference_steps": 2,
|
||||
"guidance_scale": 6.0,
|
||||
"pag_scale": 3.0,
|
||||
"output_type": "np",
|
||||
"image": image,
|
||||
"control_image": image,
|
||||
}
|
||||
|
||||
return inputs
|
||||
|
||||
def test_pag_disable_enable(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
|
||||
# base pipeline
|
||||
pipe_sd = StableDiffusionXLControlNetImg2ImgPipeline(**components)
|
||||
pipe_sd = pipe_sd.to(device)
|
||||
pipe_sd.set_progress_bar_config(disable=None)
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
del inputs["pag_scale"]
|
||||
assert (
|
||||
"pag_scale" not in inspect.signature(pipe_sd.__call__).parameters
|
||||
), f"`pag_scale` should not be a call parameter of the base pipeline {pipe_sd.__class__.__name__}."
|
||||
out = pipe_sd(**inputs).images[0, -3:, -3:, -1]
|
||||
|
||||
# pag disabled with pag_scale=0.0
|
||||
pipe_pag = self.pipeline_class(**components)
|
||||
pipe_pag = pipe_pag.to(device)
|
||||
pipe_pag.set_progress_bar_config(disable=None)
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
inputs["pag_scale"] = 0.0
|
||||
out_pag_disabled = pipe_pag(**inputs).images[0, -3:, -3:, -1]
|
||||
|
||||
# pag enable
|
||||
pipe_pag = self.pipeline_class(**components, pag_applied_layers=["mid", "up", "down"])
|
||||
pipe_pag = pipe_pag.to(device)
|
||||
pipe_pag.set_progress_bar_config(disable=None)
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
out_pag_enabled = pipe_pag(**inputs).images[0, -3:, -3:, -1]
|
||||
|
||||
assert np.abs(out.flatten() - out_pag_disabled.flatten()).max() < 1e-3
|
||||
assert np.abs(out.flatten() - out_pag_enabled.flatten()).max() > 1e-3
|
||||
|
||||
def test_save_load_optional_components(self):
|
||||
pass
|
||||
|
||||
def test_pag_cfg(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
|
||||
pipe_pag = self.pipeline_class(**components, pag_applied_layers=["mid", "up", "down"])
|
||||
pipe_pag = pipe_pag.to(device)
|
||||
pipe_pag.set_progress_bar_config(disable=None)
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
image = pipe_pag(**inputs).images
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
|
||||
assert image.shape == (
|
||||
1,
|
||||
64,
|
||||
64,
|
||||
3,
|
||||
), f"the shape of the output image should be (1, 64, 64, 3) but got {image.shape}"
|
||||
expected_slice = np.array(
|
||||
[0.5562928, 0.44882968, 0.4588066, 0.63200223, 0.5694165, 0.4955688, 0.6126959, 0.57588536, 0.43827885]
|
||||
)
|
||||
|
||||
max_diff = np.abs(image_slice.flatten() - expected_slice).max()
|
||||
assert max_diff < 1e-3, f"output is different from expected, {image_slice.flatten()}"
|
||||
|
||||
def test_pag_uncond(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
|
||||
pipe_pag = self.pipeline_class(**components, pag_applied_layers=["mid", "up", "down"])
|
||||
pipe_pag = pipe_pag.to(device)
|
||||
pipe_pag.set_progress_bar_config(disable=None)
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
inputs["guidance_scale"] = 0.0
|
||||
image = pipe_pag(**inputs).images
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
|
||||
assert image.shape == (
|
||||
1,
|
||||
64,
|
||||
64,
|
||||
3,
|
||||
), f"the shape of the output image should be (1, 64, 64, 3) but got {image.shape}"
|
||||
expected_slice = np.array(
|
||||
[0.5543988, 0.45614323, 0.4665692, 0.6202247, 0.5598917, 0.49621183, 0.6084159, 0.5722314, 0.43945464]
|
||||
)
|
||||
|
||||
max_diff = np.abs(image_slice.flatten() - expected_slice).max()
|
||||
assert max_diff < 1e-3, f"output is different from expected, {image_slice.flatten()}"
|
||||
@@ -178,6 +178,46 @@ class StableDiffusionLatentUpscalePipelineFastTests(
|
||||
max_diff = np.abs(image_slice.flatten() - expected_slice).max()
|
||||
self.assertLessEqual(max_diff, 1e-3)
|
||||
|
||||
def test_stable_diffusion_latent_upscaler_negative_prompt(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
sd_pipe = StableDiffusionLatentUpscalePipeline(**components)
|
||||
sd_pipe = sd_pipe.to(device)
|
||||
sd_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
negative_prompt = "french fries"
|
||||
output = sd_pipe(**inputs, negative_prompt=negative_prompt)
|
||||
image = output.images
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
|
||||
assert image.shape == (1, 256, 256, 3)
|
||||
expected_slice = np.array(
|
||||
[0.43865365, 0.404124, 0.42618454, 0.44333526, 0.40564927, 0.43818694, 0.4411913, 0.43404633, 0.46392226]
|
||||
)
|
||||
|
||||
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
|
||||
|
||||
def test_stable_diffusion_latent_upscaler_multiple_init_images(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
sd_pipe = StableDiffusionLatentUpscalePipeline(**components)
|
||||
sd_pipe = sd_pipe.to(device)
|
||||
sd_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
inputs["prompt"] = [inputs["prompt"]] * 2
|
||||
inputs["image"] = inputs["image"].repeat(2, 1, 1, 1)
|
||||
image = sd_pipe(**inputs).images
|
||||
image_slice = image[-1, -3:, -3:, -1]
|
||||
|
||||
assert image.shape == (2, 256, 256, 3)
|
||||
expected_slice = np.array(
|
||||
[0.38730142, 0.35695046, 0.40646142, 0.40967226, 0.3981609, 0.4195988, 0.4248805, 0.430259, 0.45694894]
|
||||
)
|
||||
|
||||
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
|
||||
|
||||
def test_attention_slicing_forward_pass(self):
|
||||
super().test_attention_slicing_forward_pass(expected_max_diff=7e-3)
|
||||
|
||||
|
||||
@@ -68,25 +68,21 @@ class IsSafetensorsCompatibleTests(unittest.TestCase):
|
||||
"unet/diffusion_pytorch_model.fp16.bin",
|
||||
"unet/diffusion_pytorch_model.fp16.safetensors",
|
||||
]
|
||||
variant = "fp16"
|
||||
self.assertTrue(is_safetensors_compatible(filenames, variant=variant))
|
||||
self.assertTrue(is_safetensors_compatible(filenames))
|
||||
|
||||
def test_diffusers_model_is_compatible_variant(self):
|
||||
filenames = [
|
||||
"unet/diffusion_pytorch_model.fp16.bin",
|
||||
"unet/diffusion_pytorch_model.fp16.safetensors",
|
||||
]
|
||||
variant = "fp16"
|
||||
self.assertTrue(is_safetensors_compatible(filenames, variant=variant))
|
||||
self.assertTrue(is_safetensors_compatible(filenames))
|
||||
|
||||
def test_diffusers_model_is_compatible_variant_partial(self):
|
||||
# pass variant but use the non-variant filenames
|
||||
def test_diffusers_model_is_compatible_variant_mixed(self):
|
||||
filenames = [
|
||||
"unet/diffusion_pytorch_model.bin",
|
||||
"unet/diffusion_pytorch_model.safetensors",
|
||||
"unet/diffusion_pytorch_model.fp16.safetensors",
|
||||
]
|
||||
variant = "fp16"
|
||||
self.assertTrue(is_safetensors_compatible(filenames, variant=variant))
|
||||
self.assertTrue(is_safetensors_compatible(filenames))
|
||||
|
||||
def test_diffusers_model_is_not_compatible_variant(self):
|
||||
filenames = [
|
||||
@@ -99,25 +95,14 @@ class IsSafetensorsCompatibleTests(unittest.TestCase):
|
||||
"unet/diffusion_pytorch_model.fp16.bin",
|
||||
# Removed: 'unet/diffusion_pytorch_model.fp16.safetensors',
|
||||
]
|
||||
variant = "fp16"
|
||||
self.assertFalse(is_safetensors_compatible(filenames, variant=variant))
|
||||
self.assertFalse(is_safetensors_compatible(filenames))
|
||||
|
||||
def test_transformer_model_is_compatible_variant(self):
|
||||
filenames = [
|
||||
"text_encoder/pytorch_model.fp16.bin",
|
||||
"text_encoder/model.fp16.safetensors",
|
||||
]
|
||||
variant = "fp16"
|
||||
self.assertTrue(is_safetensors_compatible(filenames, variant=variant))
|
||||
|
||||
def test_transformer_model_is_compatible_variant_partial(self):
|
||||
# pass variant but use the non-variant filenames
|
||||
filenames = [
|
||||
"text_encoder/pytorch_model.bin",
|
||||
"text_encoder/model.safetensors",
|
||||
]
|
||||
variant = "fp16"
|
||||
self.assertTrue(is_safetensors_compatible(filenames, variant=variant))
|
||||
self.assertTrue(is_safetensors_compatible(filenames))
|
||||
|
||||
def test_transformer_model_is_not_compatible_variant(self):
|
||||
filenames = [
|
||||
@@ -126,9 +111,69 @@ class IsSafetensorsCompatibleTests(unittest.TestCase):
|
||||
"vae/diffusion_pytorch_model.fp16.bin",
|
||||
"vae/diffusion_pytorch_model.fp16.safetensors",
|
||||
"text_encoder/pytorch_model.fp16.bin",
|
||||
# 'text_encoder/model.fp16.safetensors',
|
||||
"unet/diffusion_pytorch_model.fp16.bin",
|
||||
"unet/diffusion_pytorch_model.fp16.safetensors",
|
||||
]
|
||||
variant = "fp16"
|
||||
self.assertFalse(is_safetensors_compatible(filenames, variant=variant))
|
||||
self.assertFalse(is_safetensors_compatible(filenames))
|
||||
|
||||
def test_transformer_model_is_compatible_variant_extra_folder(self):
|
||||
filenames = [
|
||||
"safety_checker/pytorch_model.fp16.bin",
|
||||
"safety_checker/model.fp16.safetensors",
|
||||
"vae/diffusion_pytorch_model.fp16.bin",
|
||||
"vae/diffusion_pytorch_model.fp16.safetensors",
|
||||
"text_encoder/pytorch_model.fp16.bin",
|
||||
"unet/diffusion_pytorch_model.fp16.bin",
|
||||
"unet/diffusion_pytorch_model.fp16.safetensors",
|
||||
]
|
||||
self.assertTrue(is_safetensors_compatible(filenames, folder_names={"vae", "unet"}))
|
||||
|
||||
def test_transformer_model_is_not_compatible_variant_extra_folder(self):
|
||||
filenames = [
|
||||
"safety_checker/pytorch_model.fp16.bin",
|
||||
"safety_checker/model.fp16.safetensors",
|
||||
"vae/diffusion_pytorch_model.fp16.bin",
|
||||
"vae/diffusion_pytorch_model.fp16.safetensors",
|
||||
"text_encoder/pytorch_model.fp16.bin",
|
||||
"unet/diffusion_pytorch_model.fp16.bin",
|
||||
"unet/diffusion_pytorch_model.fp16.safetensors",
|
||||
]
|
||||
self.assertFalse(is_safetensors_compatible(filenames, folder_names={"text_encoder"}))
|
||||
|
||||
def test_transformers_is_compatible_sharded(self):
|
||||
filenames = [
|
||||
"text_encoder/pytorch_model.bin",
|
||||
"text_encoder/model-00001-of-00002.safetensors",
|
||||
"text_encoder/model-00002-of-00002.safetensors",
|
||||
]
|
||||
self.assertTrue(is_safetensors_compatible(filenames))
|
||||
|
||||
def test_transformers_is_compatible_variant_sharded(self):
|
||||
filenames = [
|
||||
"text_encoder/pytorch_model.bin",
|
||||
"text_encoder/model.fp16-00001-of-00002.safetensors",
|
||||
"text_encoder/model.fp16-00001-of-00002.safetensors",
|
||||
]
|
||||
self.assertTrue(is_safetensors_compatible(filenames))
|
||||
|
||||
def test_diffusers_is_compatible_sharded(self):
|
||||
filenames = [
|
||||
"unet/diffusion_pytorch_model.bin",
|
||||
"unet/diffusion_pytorch_model-00001-of-00002.safetensors",
|
||||
"unet/diffusion_pytorch_model-00002-of-00002.safetensors",
|
||||
]
|
||||
self.assertTrue(is_safetensors_compatible(filenames))
|
||||
|
||||
def test_diffusers_is_compatible_variant_sharded(self):
|
||||
filenames = [
|
||||
"unet/diffusion_pytorch_model.bin",
|
||||
"unet/diffusion_pytorch_model.fp16-00001-of-00002.safetensors",
|
||||
"unet/diffusion_pytorch_model.fp16-00001-of-00002.safetensors",
|
||||
]
|
||||
self.assertTrue(is_safetensors_compatible(filenames))
|
||||
|
||||
def test_diffusers_is_compatible_only_variants(self):
|
||||
filenames = [
|
||||
"unet/diffusion_pytorch_model.fp16.safetensors",
|
||||
]
|
||||
self.assertTrue(is_safetensors_compatible(filenames))
|
||||
|
||||
@@ -551,37 +551,94 @@ class DownloadTests(unittest.TestCase):
|
||||
assert sum(f.endswith(this_format) and not f.endswith(f"{variant}{this_format}") for f in files) == 3
|
||||
assert not any(f.endswith(other_format) for f in files)
|
||||
|
||||
def test_download_broken_variant(self):
|
||||
for use_safetensors in [False, True]:
|
||||
# text encoder is missing no variant and "no_ema" variant weights, so the following can't work
|
||||
for variant in [None, "no_ema"]:
|
||||
with self.assertRaises(OSError) as error_context:
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
tmpdirname = StableDiffusionPipeline.from_pretrained(
|
||||
"hf-internal-testing/stable-diffusion-broken-variants",
|
||||
cache_dir=tmpdirname,
|
||||
variant=variant,
|
||||
use_safetensors=use_safetensors,
|
||||
)
|
||||
def test_download_safetensors_only_variant_exists_for_model(self):
|
||||
variant = None
|
||||
use_safetensors = True
|
||||
|
||||
assert "Error no file name" in str(error_context.exception)
|
||||
|
||||
# text encoder has fp16 variants so we can load it
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
tmpdirname = StableDiffusionPipeline.download(
|
||||
# text encoder is missing no variant weights, so the following can't work
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
with self.assertRaises(OSError) as error_context:
|
||||
tmpdirname = StableDiffusionPipeline.from_pretrained(
|
||||
"hf-internal-testing/stable-diffusion-broken-variants",
|
||||
use_safetensors=use_safetensors,
|
||||
cache_dir=tmpdirname,
|
||||
variant="fp16",
|
||||
variant=variant,
|
||||
use_safetensors=use_safetensors,
|
||||
)
|
||||
assert "Error no file name" in str(error_context.exception)
|
||||
|
||||
# text encoder has fp16 variants so we can load it
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
tmpdirname = StableDiffusionPipeline.download(
|
||||
"hf-internal-testing/stable-diffusion-broken-variants",
|
||||
use_safetensors=use_safetensors,
|
||||
cache_dir=tmpdirname,
|
||||
variant="fp16",
|
||||
)
|
||||
all_root_files = [t[-1] for t in os.walk(tmpdirname)]
|
||||
files = [item for sublist in all_root_files for item in sublist]
|
||||
# None of the downloaded files should be a non-variant file even if we have some here:
|
||||
# https://huggingface.co/hf-internal-testing/stable-diffusion-broken-variants/tree/main/unet
|
||||
assert len(files) == 15, f"We should only download 15 files, not {len(files)}"
|
||||
|
||||
def test_download_bin_only_variant_exists_for_model(self):
|
||||
variant = None
|
||||
use_safetensors = False
|
||||
|
||||
# text encoder is missing Non-variant weights, so the following can't work
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
with self.assertRaises(OSError) as error_context:
|
||||
tmpdirname = StableDiffusionPipeline.from_pretrained(
|
||||
"hf-internal-testing/stable-diffusion-broken-variants",
|
||||
cache_dir=tmpdirname,
|
||||
variant=variant,
|
||||
use_safetensors=use_safetensors,
|
||||
)
|
||||
assert "Error no file name" in str(error_context.exception)
|
||||
|
||||
# text encoder has fp16 variants so we can load it
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
tmpdirname = StableDiffusionPipeline.download(
|
||||
"hf-internal-testing/stable-diffusion-broken-variants",
|
||||
use_safetensors=use_safetensors,
|
||||
cache_dir=tmpdirname,
|
||||
variant="fp16",
|
||||
)
|
||||
all_root_files = [t[-1] for t in os.walk(tmpdirname)]
|
||||
files = [item for sublist in all_root_files for item in sublist]
|
||||
# None of the downloaded files should be a non-variant file even if we have some here:
|
||||
# https://huggingface.co/hf-internal-testing/stable-diffusion-broken-variants/tree/main/unet
|
||||
assert len(files) == 15, f"We should only download 15 files, not {len(files)}"
|
||||
|
||||
def test_download_safetensors_variant_does_not_exist_for_model(self):
|
||||
variant = "no_ema"
|
||||
use_safetensors = True
|
||||
|
||||
# text encoder is missing no_ema variant weights, so the following can't work
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
with self.assertRaises(OSError) as error_context:
|
||||
tmpdirname = StableDiffusionPipeline.from_pretrained(
|
||||
"hf-internal-testing/stable-diffusion-broken-variants",
|
||||
cache_dir=tmpdirname,
|
||||
variant=variant,
|
||||
use_safetensors=use_safetensors,
|
||||
)
|
||||
|
||||
all_root_files = [t[-1] for t in os.walk(tmpdirname)]
|
||||
files = [item for sublist in all_root_files for item in sublist]
|
||||
assert "Error no file name" in str(error_context.exception)
|
||||
|
||||
# None of the downloaded files should be a non-variant file even if we have some here:
|
||||
# https://huggingface.co/hf-internal-testing/stable-diffusion-broken-variants/tree/main/unet
|
||||
assert len(files) == 15, f"We should only download 15 files, not {len(files)}"
|
||||
# only unet has "no_ema" variant
|
||||
def test_download_bin_variant_does_not_exist_for_model(self):
|
||||
variant = "no_ema"
|
||||
use_safetensors = False
|
||||
|
||||
# text encoder is missing no_ema variant weights, so the following can't work
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
with self.assertRaises(OSError) as error_context:
|
||||
tmpdirname = StableDiffusionPipeline.from_pretrained(
|
||||
"hf-internal-testing/stable-diffusion-broken-variants",
|
||||
cache_dir=tmpdirname,
|
||||
variant=variant,
|
||||
use_safetensors=use_safetensors,
|
||||
)
|
||||
assert "Error no file name" in str(error_context.exception)
|
||||
|
||||
def test_local_save_load_index(self):
|
||||
prompt = "hello"
|
||||
|
||||
@@ -20,12 +20,7 @@ import numpy as np
|
||||
import torch
|
||||
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
|
||||
|
||||
from diffusers import (
|
||||
AutoencoderKL,
|
||||
DDIMScheduler,
|
||||
TextToVideoSDPipeline,
|
||||
UNet3DConditionModel,
|
||||
)
|
||||
from diffusers import AutoencoderKL, DDIMScheduler, TextToVideoSDPipeline, UNet3DConditionModel
|
||||
from diffusers.utils import is_xformers_available
|
||||
from diffusers.utils.testing_utils import (
|
||||
enable_full_determinism,
|
||||
@@ -64,7 +59,7 @@ class TextToVideoSDPipelineFastTests(PipelineTesterMixin, SDFunctionTesterMixin,
|
||||
def get_dummy_components(self):
|
||||
torch.manual_seed(0)
|
||||
unet = UNet3DConditionModel(
|
||||
block_out_channels=(4, 8),
|
||||
block_out_channels=(8, 8),
|
||||
layers_per_block=1,
|
||||
sample_size=32,
|
||||
in_channels=4,
|
||||
@@ -134,10 +129,7 @@ class TextToVideoSDPipelineFastTests(PipelineTesterMixin, SDFunctionTesterMixin,
|
||||
return inputs
|
||||
|
||||
def test_dict_tuple_outputs_equivalent(self):
|
||||
expected_slice = None
|
||||
if torch_device == "cpu":
|
||||
expected_slice = np.array([0.4903, 0.5649, 0.5504, 0.5179, 0.4821, 0.5466, 0.4131, 0.5052, 0.5077])
|
||||
return super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice)
|
||||
return super().test_dict_tuple_outputs_equivalent()
|
||||
|
||||
def test_text_to_video_default_case(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
@@ -151,9 +143,8 @@ class TextToVideoSDPipelineFastTests(PipelineTesterMixin, SDFunctionTesterMixin,
|
||||
frames = sd_pipe(**inputs).frames
|
||||
|
||||
image_slice = frames[0][0][-3:, -3:, -1]
|
||||
|
||||
assert frames[0][0].shape == (32, 32, 3)
|
||||
expected_slice = np.array([0.7537, 0.1752, 0.6157, 0.5508, 0.4240, 0.4110, 0.4838, 0.5648, 0.5094])
|
||||
expected_slice = np.array([0.8093, 0.2751, 0.6976, 0.5927, 0.4616, 0.4336, 0.5094, 0.5683, 0.4796])
|
||||
|
||||
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
|
||||
|
||||
|
||||
Reference in New Issue
Block a user